Soft Modeling in Industrial Manufacturing PDF

This book discusses the problems of complexity in industrial data, including the problems of data sources, causes and types of data uncertainty, and methods of data preparation for further reasoning in engineering practice. Each data source has its own specificity, and a characteristic property of industrial data is its high degree of uncertainty. The book also explores a wide spectrum of soft modeling methods with illustrations pertaining to specific cases from diverse industrial processes. In soft modeling the physical nature of phenomena may not be known and may not be taken into consideration. Soft models usually employ simplified mathematical equations derived directly from the data obtained as observations or measurements of the given system. Although soft models may not explain the nature of the phenomenon or system under study, they usually point to its significant features or properties.

107 downloads 4K Views 8MB Size

Report

Download pdf

Recommend Stories

Empty story

Idea Transcript

Studies in Systems, Decision and Control 183

Przemyslaw Grzegorzewski Andrzej Kochanski Janusz Kacprzyk Editors

Soft Modeling in Industrial Manufacturing

Studies in Systems, Decision and Control Volume 183

Series editor Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland e-mail: [email protected]

The series “Studies in Systems, Decision and Control” (SSDC) covers both new developments and advances, as well as the state of the art, in the various areas of broadly perceived systems, decision making and control–quickly, up to date and with a high quality. The intent is to cover the theory, applications, and perspectives on the state of the art and future developments relevant to systems, decision making, control, complex processes and related areas, as embedded in the ﬁelds of engineering, computer science, physics, economics, social and life sciences, as well as the paradigms and methodologies behind them. The series contains monographs, textbooks, lecture notes and edited volumes in systems, decision making and control spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor Networks, Control Systems, Energy Systems, Automotive Systems, Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems, Robotics, Social Systems, Economic Systems and other. Of particular value to both the contributors and the readership are the short publication timeframe and the world-wide distribution and exposure which enable both a wide and rapid dissemination of research output.

More information about this series at http://www.springer.com/series/13304

Przemyslaw Grzegorzewski Andrzej Kochanski Janusz Kacprzyk •

Editors

Soft Modeling in Industrial Manufacturing

123

Editors Przemyslaw Grzegorzewski Faculty of Mathematics and Information Science Warsaw University of Technology Warsaw, Poland and Systems Research Institute Polish Academy of Sciences Warsaw, Poland

Andrzej Kochanski Faculty of Production Engineering, Institute of Manufacturing Technologies Warsaw University of Technology Warsaw, Poland Janusz Kacprzyk Systems Research Institute Polish Academy of Sciences Warsaw, Poland

ISSN 2198-4182 ISSN 2198-4190 (electronic) Studies in Systems, Decision and Control ISBN 978-3-030-03200-5 ISBN 978-3-030-03201-2 (eBook) https://doi.org/10.1007/978-3-030-03201-2 Library of Congress Control Number: 2018959257 © Springer Nature Switzerland AG 2019 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microﬁlms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional afﬁliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Contents

Part I

Theory

Data and Modeling in Industrial Manufacturing . . . . . . . . . . . . . . . . . . Przemyslaw Grzegorzewski and Andrzej Kochanski

3

From Data to Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Przemyslaw Grzegorzewski and Andrzej Kochanski

15

Data Preprocessing in Industrial Manufacturing . . . . . . . . . . . . . . . . . . Przemyslaw Grzegorzewski and Andrzej Kochanski

27

Part II

Applications

Tool Condition Monitoring in Metal Cutting . . . . . . . . . . . . . . . . . . . . . Krzysztof Jemielniak

45

Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marcin Perzyk and Artur Soroczynski

75

Application of Data Mining Tools in Shrink Sleeve Labels Converting Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Krzysztof Krystosiak

89

Study of Thickness Variability of the Floorboard Surface Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Agnieszka Kujawińska, Michał Rogalewicz, Magdalena Diering, Krzysztof Żywicki and Adam Hamrol Applying Statistical Methods with Imprecise Data to Quality Control in Cheese Manufacturing . . . . . . . . . . . . . . . . . . . . . 127 Ana Belén Ramos-Guajardo, Ángela Blanco-Fernández and Gil González-Rodríguez

v

vi

Contents

Monitoring Series of Dependent Observations Using the sXWAM Control Chart for Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Olgierd Hryniewicz and Katarzyna Kaczmarek-Majer Diagnosis of Out-of-Control Signals in Complex Manufacturing Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Marcin Perzyk, Jacek Kozlowski and Agnieszka Rodziewicz

Introduction

Mathematical modeling seems to be a fundamental device for doing science and engineering. Indeed, mathematics has provided suitable concepts and tools, an appropriate language and a way of thinking that enables an awesome progress in various ﬁelds, including physics, chemistry, biology, ecology, economics, sociology, as well as many disciplines of engineering. Nowadays, in the era of overarching computer technology the need and importance of mathematical modeling have become indisputable. Mathematical models of processes, objects, or systems are frequently divided into hard and soft models. In the so-called hard modeling, the physical nature of the system (process or object) is taken into consideration through known physical laws and characterized in terms of principles of mechanics, thermodynamics, acoustics, electromagnetism, material constants, as well as the geometry of the system under analysis, supplemented with selected boundary conditions. As a result, a hard model, as long as we succeed in developing one, makes possible a deep understanding of the nature of the phenomenon or system under analysis and, in particular, of the cause–effect mechanisms which are at work in the considered situation. Although hard models typically treat variability in a simpliﬁed manner, they exhibit stability, especially over long periods of time. In turn, in the so-called soft modeling, the physical nature of phenomena may not be known and may not be taken into consideration. Soft models usually employ simpliﬁed mathematical equations derived directly from the data obtained as observations or measurements of the given system. Hence, the system (process, or object) which is being modeled is frequently treated as the so-called black box whose internal functioning we know nothing about. And although a soft model may not explain the nature of the phenomenon or system under analysis, it will, nevertheless, point to its signiﬁcant features or properties. Moreover, models of this kind are usually capable of explaining the observed variability. Of course, as long as it is possible to construct an adequate hard model which satisfactorily explains a phenomenon, object or system, such a model will be more desirable than a soft one. However, when a little is known about the observed reality, a soft model may be treated as an introductory step toward the development vii

viii

Introduction

of a hard model. It may also happen that obtaining an acceptable hard model turns out impossible and a soft model is the only sensible model of the phenomenon under study. Finally, sometimes the optimal solution is a compromise which combines hard and soft modeling leading thereby to a model which is satisfactory both from the perspective of understanding the process under analysis and the utilitarian gains connected with the simplicity of solutions, implementations, predictions, etc. In the case of industrial data, widely used in hard modeling, we have to face a pervasive uncertainty of different origin. Consequently, we should examine to what extent various factors, like material parameters, which are established with a high degree of uncertainty, influence the ﬁnal results and analyzes. We should also assess sensitivity of the considered models for diverse simpliﬁcations used in modeling. And therefore, keeping all this in mind, it seems well motivated to talk about different degrees of softness in mathematical modeling. The starting point for any reasoning in empirical sciences is the data. Its quality has essential influence on the quality of the reasoning itself. Moreover, even the most sophisticated reasoning methods may turn out useless when the input data will be of doubtful quality. It is the source of the data which exerts a decisive influence on data quality. In particular, industrial data are different from social, medical, or economic data and, in fact, each of the above kinds of data has its own speciﬁcity. A characteristic property of industrial data is its high degree of uncertainty. This uncertainty may result, among others, from the lack of conﬁdence in the correctness of the observations made, means and methods of taking measurements, required diligence in analysis, and the quality of control and measuring instruments. No matter what production process we are dealing with, the data gathered in databases are frequently incorrect, due to carelessness or inaccuracy of a measurement, where carelessness relates to actions of the person taking measurements, while inaccuracy pertains to the measuring or control instruments of a low quality. In the case of data coming from production processes, we may also deal with inaccuracy resulting from a weakness of the adopted measuring method (for instance, temporal restrictions on the use of more advanced methods) or, directly, from the very nature of the object or process under analysis. It may also happen that uncertainty an additional variability produced by some undesired causes may superpose on the relevant natural variability, typical of the process. The book Soft Modeling in Industrial Manufacturing discusses the problems of industrial data in their overall complexity, that is, the problems of data sources, causes and types of data uncertainty, as well as methods of dealing with preparing this type of data for reasoning in the engineering practice. The reader will also have a chance to learn about the wide spectrum of soft modeling methods, with illustrations pertaining to speciﬁc cases, coming from diverse kinds of industrial processes. The book consists of two main parts: Part I: Theory—containing three chapters and Applications with seven chapters. In Part I, Chapter “Data and Modeling in Industrial Manufacturing”, can be perceived as the starting point for any modeling and for further scientiﬁc analysis. Here we discuss the speciﬁcity of industrial data

Introduction

ix

and its impact on scientiﬁc modeling. Following some general remarks on mathematical modeling, the idea of a hard modeling and soft modeling in engineering is developed. The main goal of Chapter, “From Data to Reasoning”, is to discuss and clarify the meaning of such basic notions like data, information, knowledge to indicate their interrelations and to set them in the broad framework of the cognition oriented activity. Finally, three basic types of reasoning used both in science and in applications are briefly characterized. In Chapter “Data Preprocessing in Industrial Manufacturing”, the importance of data preparation in the whole process of data analysis is discussed. Firstly, the structure of the analytical process is discussed and the data preparation stage is located within this structure. Part II contains a critical discussion of the differences between datum quality and data quality. Finally, the methodology of data preparation dedicated speciﬁcally to industrial data is considered. Part II: Applications consists of seven independent chapters. These chapters represent both various areas of applications and diverse methodologies which have to do with a wide range of soft modeling methods. In addition to a detailed discussion of the proposed modeling method, each chapter contains also an analysis of a speciﬁc practical problem. In accordance with a classiﬁcation offered in Chapter “Data and Modeling in Industrial Manufacturing”, which distinguishes two groups of soft modeling methods, both groups are illustrated in the book. And thus, artiﬁcial neural networks, decision and regression trees, boosted trees, supported vector machines, and rough set theory, which are instances of computational intelligence methods, are discussed and employed in Chapters “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing”, “Application of Data Mining Tools in Shrink Sleeve Labels Converting Process” and “Diagnosis of Out-ofControl Signals in Complex Manufacturing Processes”. In turn, modiﬁed Shewhart charts and statistical hypothesis testing tools, which exemplify statistics and expert systems, are considered in Chapters “Tool Condition Monitoring in Metal Cutting”, “Study of Thickness Variability of the Floorboard Surface Layer”, “Applying Statistical Methods with Imprecise Data to Quality Control in Cheese Manufacturing” and “Monitoring Series of Dependent Observations Using the sXWAM Control Chart for Residuals”. The diversity of topics and themes in Part II is also manifested in the range of practical applications discussed in the respective chapters. They include applications of soft modeling in such areas of industrial production as food processing, machining, printing, metallurgy, casting, floorboard production, or economic analyzes. The very aims for which the selected tools were employed are widely diverse as well. Modeling was used to predict process parameters (Chapter “Application of Data Mining Tools in Shrink Sleeve Labels Converting Process”), to create rules (Chapter “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing”), to monitor a process (Chapters “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing”, “Monitoring

x

Introduction

Series of Dependent Observations Using the sXWAM Control Chart for Residuals” and “Diagnosis of Out-of-Control Signals in Complex Manufacturing Processes”), but also for fault diagnosis (Chapter “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing” and “Diagnosis of Out-ofControl Signals in Complex Manufacturing Processes”), for technological process optimization (Chapter “Study of Thickness Variability of the Floorboard Surface Layer”) or for quality control (Chapter “Applying Statistical Methods with Imprecise Data to Quality Control in Cheese Manufacturing”). The cases discussed in Part II differ also considerably with respect to the size of the input data set and methods applied for gathering data. We show models based on relatively small data sets (a few attributes and about 2000 observations), as well as models designed using large data sets, characterized by hundreds of attributes and hundreds of thousands of observations. The data used for modeling come from a production documentation with manually collected records (Chapters “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing”, “Application of Data Mining Tools in Shrink Sleeve Labels Converting Process” and “Diagnosis of Out-of-Control Signals in Complex Manufacturing Processes”) or from technological line sensors (Chapter “Tool Condition Monitoring in Metal Cutting”), from industrial laboratories (Chapters “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing” and “Diagnosis of Out-of-Control Signals in Complex Manufacturing Processes”), questionnaires (Chapter “Applying Statistical Methods with Imprecise Data to Quality Control in Cheese Manufacturing”) or they are artiﬁcially generated (Chapters “Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing” and “Monitoring Series of Dependent Observations Using the sXWAM Control Chart for Residuals”). However, what all data sets analyzed in the chapters included in Part II have in common is uncertainty. It has a diverse origin, and it differs in type, signiﬁcance, and impact on the problem under study. Nevertheless, its presence is a serious challenge both to the theoreticians and practitioners who face mathematical modeling and decision making. Some illustrations, how to cope with uncertain data, are given in this very volume. Warsaw, Poland July 2018

Przemyslaw Grzegorzewski Andrzej Kochanski

Part I

Theory

Data and Modeling in Industrial Manufacturing Przemyslaw Grzegorzewski and Andrzej Kochanski

Abstract Data can be perceived as a staring point for any modeling and further scientific analysis. Here we discuss the specifity of industrial data and its impact on scientific modeling. Following some general remarks on mathematical modeling the idea of a hard modeling and soft modeling in engineering is developed. Keywords Hard modeling · Industrial data · Manufacture data · Mathematical modeling · Model · Soft modeling

1 Reality and Modeling The main goal of a science, in general, is to understand and explain the world. However, reality is too complicated to be described precisely and completely, without any simplifications and approximations. Therefore, to describe any phenomenon we usually take into account only such aspects of reality which seem to have a significant impact on this phenomenon. All other less significant factors are neglected or at least simplified. Perhaps this very act of the abstraction of the significant factors from all possible ones turns out to be the most important phase of cognition. Indeed, all great and impressive achievements in physics, chemistry, astronomy and engineering were P. Grzegorzewski (B) Systems Research Institute, Polish Academy of Sciences, Newelska 6, 01-447 Warsaw, Poland e-mail: [email protected] Faculty of Mathematics and Information Science, Warsaw University of Technology, Koszykowa 75, 00-662 Warsaw, Poland A. Kochanski Faculty of Production Engineering, Warsaw University of Technology, Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_1

3

4

P. Grzegorzewski and A. Kochanski

possible as soon as people have learned how to recognize which factors are crucial and which are less important for describing a phenomenon under study. To find an explanation one may start from simple objects, believing that this is the right way to discover some laws, rules or mechanisms that are also valid in more complicated situations. Following such reasoning we come to the crucial concept—a model. What does it mean a model? Dictionaries deliver various meanings of this concept. Here we focus our attention on those related to science and cognition. In this context the word “model” is perceived, for instance, as (see [20]): • • • • •

a representation of a system, something which accurately resembles or represents something else, a set of entities that satisfies all the formulae of a given formal or axiomatic system, a conceptual or mental representation, a simplified or idealized description or conception of a particular system, process or situation, etc.

We may consider physical or formal models. A mockup is an example of a physical model. Formal models use mathematical concepts, rules and equations to describe a phenomenon or a system. Thus formal models are often called mathematical models. A relation between a given phenomenon and its mathematical model is sketched in Fig. 1. There, following [5, 6], we have identified a surrounding “real world”, where we observe various phenomena and systems of interest, and a “conceptual world”, i.e. a world of our mind, when we try to understand what is going on in that real, external world. By modeling we call the process of creating an appropriate representation of a certain phenomenon (system). The aforementioned appropriateness is related to the purpose of the model and its quality. It may happen that the model is built simply as

The real world

The conceptual world

Observations

Phenomena

Models

Predictions

Fig. 1 The real world and a mathematical model—a sketch of relation (Source [6])

Data and Modeling in Industrial Manufacturing

5

a demonstration tool to show how the final system works and looks like. Then the only need is to capture the main idea of the real system. Otherwise, one designs a model to predict system’s behavior in the presence of a certain stimulus. Clearly, in this such a desired model should behave as close as possible to the real system. When discussing about the quality of a model we can distinguish the following two aspects: its accuracy and explanability. The first term denotes a capability to reproduce faithfully the stimulus/response relation of the modeled system. The second notion stands for the capability to describe and explain clearly the behavior of the real system, i.e. a mechanism producing, or the knowledge justifying, those inputoutput relations (see, e.g. [17]). These two aspects are not necessarily equally present in each model. Sometimes the only goal in designing a model is to guarantee its high accuracy in replicating the input-output relations of the system, with no particular interest in explaining how it really works. Such models are known as black-box models. A completely different situation occurs if we are interested not only in obtaining output, but also in finding an explanation of the input-output relation, i.e. why we obtain just such response for a given stimulus. Here we can distinguish two types of models: white-box models, i.e. delivering an explanation understandable for “all”, and grey-box models which provide an explanation understandable for the domain experts. Please note, that the explanability of a model, although sometimes identified, is not equivalent with its interpretability (see, e.g. [18, 19]). Turning back to Fig. 1 let us emphasize two links between the real and conceptual worlds, i.e. observations and predictions. Observations give us some insight to the phenomena showing at least partially what is happening in the real world. Therefore, available data, even if not complete or uncertain, play a fundamental role for any further cognitive activity connected with modeling (for a broader discussion on data, information and related notions we refer the reader to Chap. 2). On the other hand, predictions obtained as a model output are not only a desired product of the model but, confronted with observations, serve to validate the model or to suggest why the model is inadequate. One can list various advices and principles that might be helpful in designing a satisfying mathematical model. Clearly, it depends on a domain of the study: one may expect that different aspects may appear crucial in models applied in medicine, social studies, engineering, etc. However, there are some ideas of more philosophical nature that seem to be useful in any situation. Such set of methodological meta-principles was expressed by Dym and Ivey [6] in the following list of questions and answers: • • • • • •

What are we looking for? Identify the need for the model. What do we want to know? List the data we are seeking. What do we know? Identify the available relevant data. What can we assume? Identify the circumstances that apply. How should we look at this model? Identify the governing physical principles. What will our model predict? Identify the equations that will be used, the calculations that will be made, and the answers that will result. • Are the predictions valid? Identify tests that can be made to validate the model, i.e. is it consistent with its principles and assumptions?

6

P. Grzegorzewski and A. Kochanski

• Are the predictions good? Identify tests that can be made to verify the model, i.e. is it useful in terms of the initial reason it was done? • Can we improve the model? Identify parameter values that are not adequately known, variables that should have been included, and/or assumptions/restrictions that could be lifted. Implement the iterative loop that we can call “model-validateverify-improve-predict”. • How will we exercise the model? What will we do with the model? Dym and Ivey [6] underlined that this list of questions and instructions is not an algorithm for building a good mathematical model but rather indicate ways of thinking about mathematical modeling. Moreover, a researcher should expect that some other questions closely related to the particular situation may occur in modeling. And this is also the case of modeling in industry which is the main topic of the present volume. Anyway, whatever is the domain of the phenomena being modeled and whatever methodology is applied in modeling one should never confuse the model with the reality we are trying to model. Model only depicts reality. Sometimes better, otherwise worse, but it is always only a simplified description of the true phenomenon or system. Its complexity, having a fundamental impact on its computability and interpretability, involves a trade-off between simplicity and accuracy of the model. Keeping all these in mind one should remember the aphorism attributed to the famous statistician Box [1] who said: All models are wrong but some are useful.

2 Specificity of Industrial Data Numerous data sets gathered in various areas of human activity and research form smaller or bigger databases which could be then applied for further analysis. Data may be divided by the criterion of their origin: for instance, we may consider social, economic, or natural science data, data coming from physics or chemistry, industrial or medical data, and many other data types. Thus a natural question arises whether industrial data have their own specificity. Can they be differentiated from the data that appear in other areas of scientific research and applications? The experience of the authors makes it possible to state that this is exactly the case. Industrial data have a number of features that do not make such an impression, when each of them is analyzed on its own. So, when considered independently, each given feature may appear in any data set. However, the way in which such features manifest, as well as their simultaneous co-appearance apparently makes industrial data distinct from other data types. Before we demonstrate the distinctive properties of the industrial data, two terms often used interchangeably (not always with a sufficient justification) must be clarified. These two notions in question are industrial data and manufacture data denote data that come from a manufacturing process. The term “manufacture data” is wider since manufacturing can be executed in a range of formats, for instance, as

Data and Modeling in Industrial Manufacturing

7

handicraft. On the other hand “industrial data” stand for the data collected from the process of manufacturing executed in a factory, manufacturing plant or coming from any technological line. The aforementioned statement indicates the first significant feature of industrial data, i.e. a great volume characterized both by the number of recorded observations and by the number of parameters characterizing a single observation. At the time when technological machinery is massively equipped with different sensors and when extended SCADA-type systems of collecting and analyzing data are used, industrial data sets containing millions of records are a norm rather than an exception [2, 4, 8, 15, 25]. Manufacturing a particular product on a technological line requires a procedure or operation that most frequently takes from several to ten-odd seconds. This means that, e.g. in an automated manufacturing process, in a 3-shift system and a product for which a single operation takes 10 s, within a single year (estimated as comprising 200 working days) we obtain 6 × 60 × 24 × 200 = 1,728,000 observations. This figure can be much higher since some operations, like cutting, are performed several hundred times per minute (e.g. 200 times per min). A technological process of production of an advanced product often involves several tens of operations, each of which may provide several measurable parameters. Consequently, databases containing data from an industrial process include hundreds of attributes. Thus the number of attributes combined with the number of collected observations results in a huge data set (big data). One can distinguish two types of industrial data. The first one is related to a direct man-imposed impact on the manufactured item or raw materials used to produce this item (e.g. a particular temperature set as required for given manufacturing process). The second one refers to data coming from inspection or measuring activities performed in the course of the production process. In both cases data may reveal either stiltedly restricted or unnaturally wide range of values of parameters. It happens as a result of the intensification or slowing down of natural processes (e.g. with wood drying or milk curdling processes) in order to increase efficiency, improve effectiveness, optimize costs, etc. Another feature of industrial data is a frequent co-occurrence of observations that differ by one, two or even more orders of magnitude. For example, in the case of alloy melts of a cast iron the same element may not be present in various alloys (0.01%— trace content) or may be present at a range of levels (0.1%—admixture content; 1.0%—alloy contents; 10.0%—prevailing alloy contents), whereas graphite releases in a metal structure may reach the level of tens, hundreds, or thousands per area unit (20–200–2,000 releases of graphite spheroids/sq.mm) [14]. Similarly, the elongation of plastic foil may vary from 2 to 2,000% [9]. Another characteristic of industrial data is connected with level of measurement. Actually, all scales of measurements: nominal, ordinal, interval, and ratio (also including the absolute scale) occur in industrial data [3, 21, 24]. A nominal scale may involve, for instance, names of input material suppliers or codes of teams working on a given shift. An ordinal scale may refer to measurements of such values as, e.g. a descriptive specification of the output material quality (inadequate, low, good, perfect) or the order of samples taken for tests (the first, the second, the third).

8

P. Grzegorzewski and A. Kochanski

Most measurements in industrial data is done on the ratio scale, e.g. weight, pressure, force, or displacement. This kind of scale may also pertain to temperature measured with the Kelvin scale (please note that the Celsius temperature scale is the interval scale). Since the interval scale seems to be a little more to learn, let us consider the following example. Example 1 Material hardness is a parameter which is continuously checked and which is frequently used as a criterion of a product’s (quality) acceptance. Hardness may be measured in many ways and there are a number of measuring methods of general applicability, such as Rockwell’s method, Brinell’s method, Vicker’s method, Leeb’s method, or Shore’s method. Additionally, the most frequently employed methods come in a number of variants involving the use of different shapes penetrators and different loads (e.g. HRA, HRB, HRC, HBS, HBW10/3000, HV5, HV30). Measurements are recorded with the accuracy up to a unit (for instance, 176HB, 45HRC, 107HV). The results obtained within a particular system may be expressed in another system with the help of appropriate tables and calculators (converters). Thus, for example, for a material with the hardness of 70HRA in the other systems we obtain the hardness of 38.8HRC, 360HB, and 380HV, respectively. In turn, for a material of the measured hardness of 60HRA we obtain the hardness of 20HRC, 228HB, and 240HV, respectively. What follows is that the former material is harder by 10HRA (18.8HRC, 132HB, and 140HV) than the latter. However, it is not 70/60 = 1.1(6) times harder, since the respective proportions for the other systems are different and take the following values: HRA: 38.8/20 = 1.94; HB: 360/228 = 1.579; and HV: 380/240 = 1.58(3). Another criterion that distinguishes industrial data is their subjectivity. Obviously, as compared, for instance, to social data, the number of parameters assessed subjectively in industrial data is relatively small. Anyway, in some cases subjectivity is inevitable and the researcher should distinguish between the objective and subjective measurements. The following example illustrates nicely the problem. Example 2 In industrial practice, we frequently deal with data which could potentially be measured in the interval scale. However, this is not what happens in practice where the relevant data are often measured in the ordinal scale. This is a consequence of the objective difficulties which follow from the impossibility of developing the standards for such data which would be reliable and practical in use. As an example let us consider the quality of a scrap metal added to the furnace in the process of metal melting (for instance, cast steel melting). In foundries the scrap metal quality is assessed on the basis of the degree of its corrosion [12]. Thus, for the scrap metal of high quality (non-corroded or with a low degree of corrosion) the label “0” is assigned. In turn, the assignment of the label “1”–medium quality—or “2”–low quality—which is an arbitrary decision of the furnace personnel that depends on various changing factors. For instance, the lighting of the site where the quality

Data and Modeling in Industrial Manufacturing

9

assessment is performed. The same worker may make different quality assessments with respect to the same metal scrap depending on whether the assessment is made in artificial or in natural lighting conditions. In turn, in natural lighting conditions assessments made at different day times may give different results (the red lighting of the setting sun intensifies the impression of corrosion). Finally, the last feature typical for technological processes are absent data, i.e. both missing values and empty values [13]. Moreover, we may have to do with all kinds of missing values, i.e. may appear as Missing at random (MAR), Missing completely at random (MCAR) or Missing not at random (MNAR) [16]. It is worth noting that the high proportion of absent data, which often happens in industrial data sets, requires non-standard analytic methods (including data imputation, etc.). A separate issue is the selection of criteria characterizing the quality of the industrial data collected in databases. In Chap. 3 (Sect. 2) Data quality: datum quality or database quality? we discuss this problem and indicate still another aspect that distinguishes industrial data from other data types.

3 Hard and Soft Modeling in Engineering 3.1 General Remarks Many a time have we come across either an openly asked question or an opinion, phrased in a veiled fashion, on the superiority of soft modeling over hard modeling or, vice versa, of hard modeling over soft modeling. Each of these two types of modeling has its advantages and disadvantages. How they relate to one another as well as to phenomena which are modeled using hard or soft approach is nicely illustrated in Fig. 2, taken from [22]. Once we accept the taxonomy offered in Fig. 2, we should acknowledge that the main difference between these two approaches is the degree of understanding of the phenomenon which is to be modeled. The hypothesis is that, roughly speaking, a hard modeling is dedicated to situations when we know what makes the object or process under modeling behave in a particular way. Conversely, a soft modeling is usually used when we still do not have this kind of knowledge.

3.2 Hard (Constitutive) Modeling What is crucial for hard models is a profound understanding of the problem under analysis. This follows from the fact that in a hard modeling relationships between attributes follows directly from fundamental dependencies between characteristics of the phenomenon under study and they refer directly to basic equations of physical laws. A natural consequence is the use of mean values of the discussed parameters,

P. Grzegorzewski and A. Kochanski

Phenomena understanding high low middle

10

Hard models (CAD, CAE)

Soft models Expert systems Machine learning

low

medium Complexity of a problem

high

Fig. 2 The scope of application of various types of process models (Source [22])

e.g. with respect to the properties of a material. Consider, for instance, the problem of strength modeling of a given object. In crystallization the process of nucleation is random. Whether or not a temporary cluster becomes a nucleus/grain depends on a number of factors, including the degree of local undercooling, appearance of a heterogeneous cluster, element concentration, or the convection mixing of the cooling alloy. As a result within the volume of the cast nuclei of different sizes and crystallographic orientations are randomly distributed. The size of the nuclei, their orientation, and the appearance of phases on nuclear boundaries all have considerable influence on the material strength. Of course, there are attempts to model the size of nuclei and their orientation, but they are very time-consuming. An analogous situation may be observed in the case of a paper production. The main component of a paper are organic fibers, to which some organic or non-organic fillers are added. The fibers used in the process are not of identical length or diameter and an orientation that largely depends on their accidental composition which results in the process of pouring the mass onto a sieve. Despite the precise dosing of the fillers their distribution is not uniform. Moreover, the ultimate strength of paper is influenced by various contaminations present in the paper mass. Finally, all those randomly varying factors bring about considerable differences in the strength exhibited within a paper web. Of course, one may develop a hard model based on the mean values of the underlying parameters but its prediction cannot be fortified by the estimation of the prediction error. Clearly, one may try to make predictions using the minimal and the maximal nucleus/grain size and then to estimate the possible range of the predicted variable/characteristic. However, this approach cannot indicate which value within this range is the most probable one.

Data and Modeling in Industrial Manufacturing

11

As indicated in Fig. 2, hard modeling requires an in-depth knowledge of the relevant problem and thus the applicability of this approach is restricted to a relatively limited number of areas. Applying hard modeling to new areas might be difficult, if ever impossible. New modeling tasks are based on the knowledge pertaining to similar phenomena and, as a result, they are vitiated, from the very start, by the sin of approximation. It should be noted at this point that modeling errors, which result from the assumed simplifications, may either add up or cancel each other, which results in a considerable ultimate error in the first case or to accidentally correct results in the second. An unquestionable advantage of hard models, whose importance cannot be exaggerated, is their stability in the long run. However, the most important reason behind the hard modeling is the possibility of a deep understanding of the considered object or system, as well as of identifying parameters or factors crucial for its behavior.

3.3 Soft (Empirical) Modeling In contrast to the constitutive (hard) modeling, the empirical (soft) modeling does not require a deep knowledge of reality under study. The available information about the analyzed process (object, system) may be limited or fragmentary because a soft model is built out of simplified dependencies or rules which are derived directly from the data. Observations of industrial processes are usually uncertain. The degree of uncertainty depends on the process under analysis. In the case of data coming from production processes it is often quite high [13]. Consequently, the resulting rules or dependencies between process parameters may be marked by a lower reliability, support or coverage [14]. Empirical models usually offer an adequate characterization of the processes under analysis and enable initial approximations to their understanding. By their very nature they do not provide the in-depth explanation of the principles which govern the processes in question but they make possible to identify some general properties of the process sufficiently well. Thanks to advanced methods [23] it is often possible to establish the influence of the considered parameters on the process. Moreover, we are often able to determine the significance of each parameter. Although empirical models are often used for modeling complex phenomena (cf. Fig. 2) they are frequently relatively simple regardless of the data set size. Indeed, a small data set usually results in a rough approximation of the object/system under study in the form of simple and uncertain rules. On the other hand, the simplicity of a model based on a large data set may be a consequence of the underlying algorithm. For instance, [10] discusses a rule-based model for controlling the production process of ductile iron casts involving heat treatment, based on the discretized data. It was shown that using only four attributes from among dozens of the registered process parameters appeared sufficient for the stable process control. Another example involve models based on decision trees which operate on numerous variables but make divisions in decision nodes with a limited number of attributes.

12

P. Grzegorzewski and A. Kochanski

An advantage of soft models is their simplicity combined with relatively good capability for describing the behavior of the relevant process (system or object) and identification of its most important parameters. The limited number of signals makes possible a simple and quick analysis of its behavior. However, since soft models are based directly on observed data sets obtained in a specific time moment, any process changes, changes in supply sources, etc., impose the model actualization, calibration or its entire change.

3.4 Summary Already at the beginning of this section we said that there is no unambiguous answer which modeling approach—hard or soft—is better. Both are used in industrial process modeling. However, we can try to indicate three common situations for which we can suggest a better solution: • if a physical model is tractable a hard modeling seems to be the preferred one, • if nothing (or a little) is known at the outset, empirical modeling always indicates a direction of where to put the effort if a physical model is later to be derived, • in all cases when deriving a hard model is not possible, try to design a soft model (since it would be much better than nothing). Recognizing advantages and disadvantages of the constitutive and empirical modeling researchers sometimes attempt to combine both soft and hard approaches in one general model (see, e.g. [7, 11] ). As a result we may facilitate modeling, simplify measurement (instrumentation), provide a structure (constraints) to the soft model based on physical properties of the system, etc.

References 1. Box, G.E.P.: Robustness in the strategy of scientific model building. In: Launer R.L., Wilkinson G.N. (eds.) Robustness in Statistics, pp. 201–236. Academic Press (1979) 2. Czyzewski, P., Ernt, M.: Modernization of the work centre in accordance to the Industry 4.0 concept on the example of position for the execution of blanking process. Welding. Technol. Rev. 90, 21–24 (2018) 3. Czyzewski, P., Kochanski, A., Moszczynski, L.: Modeling of blanking process parameters for different punch wear stage. Prz. Mech. 5, 23–26 (2016) 4. Dao, P.B., Staszewski, W.J., Barszcz, T., Uhl, T.: Condition monitoring and fault detection in wind turbines based on cointegration analysis of SCADA data. Renew. Energy. 116, 107–122 (2018) 5. Dym, C.L.: Principles of Mathematical Modeling, 2nd edn. Elsevier, Academic Press (2004) 6. Dym, C.L., Ivey E.S.: Principles of Mathematical Modeling. 1st edn. Academic Press (1980) 7. Hawryluk, M., Jakubik, J.: Analysis of forging defects for selected industrial die forging processes. Eng. Fail. Anal. 59, 396–409 (2016) 8. Ignaszak, Z., Sika, R., Perzyk, M., Kochanski, A., Kozlowski, J.: Effectiveness of SCADA systems in control of green sands properties. Arch. Found. Eng. 16, 5–12 (2016)

Data and Modeling in Industrial Manufacturing

13

9. Kochanski, A., Assman, K., Kubera, H., Czaja-Jagielska, N.: Data preparation and the preliminary assumptions of the artificial neural network structure for the evaluation of biodegradable packaging materials. Sci. Noteb. Pozn. Univ. Econ. B. 217, 36–44 (2011) 10. Kochanski, A., Grzegorzewski, P., Soroczynski, A., Olwert, A.: Modeling of austempered ductile iron using discrete signals. Comput. Methods. Mater. Sci. 14, 190–196 (2014) 11. Kochanski, A., Perzyk, M.: Ductile cast iron classification by combined modelling. Acta. Metall. Sl. 7, 50–55 (2001) 12. Kochanski, A., Perzyk, M.: Identification by artificial neural networks of the porosity defect causes in steel castings. (in Polish), Arch. Found. 2, pp. 87–92 (2002) 13. Kochanski, A., Perzyk, M., Kł¸ebczyk, M.: Knowledge in imperfect data. In: Ramiraz, C. (Ed.), Advances in Knowledge Representation, pp. 181–210. InTech (2012) 14. Kochanski, A., Soroczynski, A., Kozlowski, J.: Applying rough set theory for the modeling of austempered ductile iron properties. Arch. Found. Eng. 13, 70–73 (2013) 15. Kozowski, J., Sika, R., Gorski, F., Ciszak, O.: Modeling of foundry processes in the era of industry 4.0. In: Ivanov, V. et al. (Eds.), Advances in Design, Simulation and Manufacturing (DSMIE-2018), pp. 62–71. Springer (2019) 16. Little, R., Rubin, D.: Statistical Analysis with Missing Data. Wiley (2002) 17. Magdalena, L.: Do hierarchical fuzzy systems really improve interpretability? In: Medina, J. et al. (Eds.), Information Processing and Management of Uncertainty in Knowledge-Based Systems (IPMU 2018), CCIS 853, 1626, (2018) 18. Miller, T.: Explanation in Artificial Intelligence: Insights from the Social Sciences. arXiv:1706.07269v2 (2018) 19. Molnar, C.: Interpretable Machine Learning. https://christophm.github.io/interpretable-mlbook/index.html (2018) 20. Oxford English Dictionary 21. Perzyk, M., Kochanski, A.: Detection of causes of casting defects assisted by artificial neural networks. Proc. Inst. Mech. Eng. Part B: J. Eng. Manuf. 217, 1279–1284 (2003) 22. Perzyk, M., Kochanski, A., Biernacki, R., Kozłowski, J., Soroczynski, A.: Modelowanie procesów produkcyjnych w odlewni, Post¸epy teorii i praktyki odlewniczej, 325–344 (2009) 23. Perzyk, M., Kochanski, A., Kozłowski, J., Soroczynski, A., Biernacki, R.: Comparison of data mining tools for significance analysis of process parameters in applications to process fault diagnosis. Inf. Sci. 259, 380–392 (2015) 24. Sadlowska, H., Kocanda, A.: On the problem of material properties in numerical simulation of tube hydroforming. Arch. Civ. Mech. Eng. 10, 77–83 (2010) 25. Wodecki, J., Stefaniak, P., Polak, M., Zimroz, R.: Unsupervised anomaly detection for conveyor temperature SCADA data. In: Timofiejczuk, A. et al. (Eds.), Advances in Condition Monitoring of Machinery in Non-Stationary Operations (CMMNO2016), pp. 361–369. Springer (2018)

From Data to Reasoning Przemyslaw Grzegorzewski and Andrzej Kochanski

Abstract Data appear at the beginning and at the end of any reasonable modeling. Indeed, data deliver a motivation and a starting point for a model construction. But data are also necessary to validate a resulting model. Data bring information on a considered phenomenon. Gathering information enables to widen our knowledge. But, on the other hand, without some knowledge one would not be able to extract information from data and interpret the received information adequately. Such terms like data and information are widely used in the context of scientific modeling and applications. Although sometimes treated interchangeably they are not synonyms. Another closely related concepts are knowledge, uncertainty, reasoning, etc. The main goal of the present chapter is to discuss and clarify the meaning of the aforementioned notions, to indicate their interrelations and set them in the broad framework of the cognition oriented activity. Finally, three basic types of reasoning used both in science as well as in practice is briefly characterized. Keywords Abduction · Data · Data quality · Deduction · Induction Information · Knowledge · Reasoning · Uncertainty · Wisdom

P. Grzegorzewski (B) Systems Research Institute, Polish Academy of Sciences, Newelska 6, 01-447 Warsaw, Poland e-mail: [email protected] Faculty of Mathematics and Information Science, Warsaw University of Technology, Koszykowa 75, 00-662 Warsaw, Poland A. Kochanski Faculty of Production Engineering, Warsaw University of Technology, Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_2

15

16

P. Grzegorzewski and A. Kochanski

1 DIKW Hierarchy Data make a starting point of any modeling. However, data alone are not sufficient to reach a satisfying model. It requires also some information and knowledge. Since these terms: data, information and knowledge, are closely related and therefore they are often mixed up, it seems desirable to clarify their meaning before we make a next step towards modeling. The word data appears very often in different contexts, both alone and in connection with other terms, e.g. raw data, data mining, data set, data analysis, data acquisition, database, data ware-house, data stream, packets of data, data structure, big data, metadata, missing data, etc. One may easily notice that the multitude of expressions connected with data has appeared due to outpouring of data produced both by human beings and different sensors as well as by the immense development of information science. Even this last remark indicates a close relation between data and information. Data (from latin datum) means literally something actually given (see [5]). Common dictionaries usually explain this term as a fact known from the direct observation or as a proposition assumed or given from which conclusions may be drawn. When striving for the essence of this notion we can find data as facts or the products of observation which are unorganized and unprocessed, and do not convey any specific meaning. Data denote everything what is/can be processed/transformed in computational and mental processes (see [13]). Thus data can assume various forms, like symbols, signals, events and so on. However, it has no significance beyond its existence—it simply exists (see [1]). Moreover, data can exist in any form, usable or not, but is of no use until it is in a relevant form (see [30]). In engineering data are usually collected by mapping object under study to symbolic representation by means of a measurement procedure, which associates given property of the object with a value of some variable. The measurement output is often represented by numbers. However, since other representation are also possible, one should be aware of a measurement scale (nominal, ordinal, interval or ratio scale) being used and its limitations. A collection of measurements, i.e. the data items, form a data set. Data are often denoted by symbols. But although data themselves have no meaning, it may happen that the choice of symbols may impose or suggest their interpretation. Information is inferred from data. Someone said that information is data with meaning. It is contained in descriptions and answers to questions that begin with such words as Who?, What?, When? and How many? (see [1]). Thus information may be perceived as data that have been given meaning and endowed with relevance. This meaning can be useful, but does not have to be. Information is data that have been organized, interpreted and understood by a recipient (see [30]). One who is more pragmatic may also claim that information is an aggregation of data that makes decision making easier (see [3]).

From Data to Reasoning

17

One should remember that different data may deliver the same information. And conversely, given data set may contain various information. Moreover, some words or numbers may be considered as information but if we do not understand their meaning they are just data. Knowledge applies information to answer How? questions. It is know-how, and is what makes possible the transformation of information into instructions (see [1]). Knowledge is the interpretation of information and its use in a problem solving context. But knowledge may also refer to data, indicating the way how to extract information from data. Knowledge can lead to new insights. It is also defined as every abstract property of human/artificial agent which has ability to process/transform information into other information or in another knowledge (see [13]). Finally, knowledge might be also viewed as a mix of information, understanding, capability, experience, skills and values (see [30]). We also usually claim that the knowledge comes from that information which is verified in practice. Moreover, only the information essential for the user forms the knowledge. The other, i.e. not significant, remains just as information. These three words: data, information and knowledge, are evidently the key concepts of Information Science and Knowledge Management (the extensive review of conceptual approaches to define these three basic concepts: data, information and knowledge, can be found in [40]). Together with a wisdom they form the well-known cognitive model known as the DIKW (data-information-knowledge-wisdom) hierarchy or DIKW pyramid. It is often illustrated as diagram given in Fig. 1 or a linearly increasing chain shown in Fig. 2 with respect to the context and understanding. Both figures show that the higher elements in the hierarchy can be explained in terms of the lower elements by identifying an appropriate transformation process. For the origin of the hierarchy and its further development we refer the reader to [1, 7, 8, 39]. It is worth noting that Ackoff [1] suggested also an expanded DIKW hierarchy containing one more level, called understanding, located between knowledge and wisdom. In his approach understanding is the ability to synthetize new knowledge from previously stored information and knowledge, while wisdom is the evaluated understanding. However, most of researches claim that understanding is

Fig. 1 DIKW hierarchy

Wisdom Knowledge Information Data

18

P. Grzegorzewski and A. Kochanski

not a separate level but it appears in each step of the transition from data to wisdom. We simply deal with understanding which refers to different aspects of cognition: in transition from data to information we find understanding of relations, when moving from information to knowledge we have understanding of patterns and finally, the transformation from knowledge to wisdom is related with understanding principles. The aforementioned distinction of understanding, suggested by Bellinger et al. [6], is depicted in Fig. 3.

Context

Wisdom

Knowledge

Information

Data

Understanding

Fig. 2 DIKW hierarchy with respect to understanding and context Connectedness

Wisdom Understand principles

Knowledge Understand patterns

Information Understand relations

Data Fig. 3 Understanding in DIKW hierarchy (Source [6])

Understanding

From Data to Reasoning

19

On the other hand, Zeleny [39] proposed a model with the fifth level at the top of the hierarchy. He called it enlightenment and interpreted as “attaining the sense of truth, the sense of right and wrong, and having it socially accepted, respected and sanctioned”. Actually, none of the extended DIKW models met any significant interest.

2 Data Versus Information A user is necessary so the data might become an information. And a user is the one who interprets data and declares whether they are comprehensible. He also specifies the extent of their understanding. Therefore, since information depends on the recipient and his personal knowledge and skills, it is inevitable subjective (at least to some extent). Let us consider the following examples. Example 1 Suppose that given data set contains several symbols, e.g. 20 and 10. Can we say that the intensity of a feature represented by these two symbols and corresponding to the first object is twice bigger than that corresponding to the second object? The answer is, of course, negative. Firstly, it may happen that both symbols, although numerical, are just labels and hence mathematical operations carried with them have no sense. Next, even if both symbols are used there as numbers, still not always all mathematical operations are allowed or have a reasonable interpretation. For example, if both numbers represent the length of the element expressed in the same units, say centimeters, then actually, the element described by 20 cm is twice longer than that characterized by 10 cm. However, if both numbers describe a temperature of an object expressed in the Celsius scale, we may conclude that the temperature of the first object is 10◦ bigger than of the second one, but a sentence the first object is warm twice more than the second one has no sense at all. Example 2 Consider some interval-valued observations expressed by the closed intervals, like [20.8, 21.4], [20.6, 21.1], etc. What can we say about the information delivered by such intervals? Does it have a straightforward and univocal interpretation, clear in any case for any user? Actually, a closed interval may be used to model two different types of information: the imprecise description of a point-valued quantity or the precise description of a set-valued entity. Quite often the results of an experiment are imprecisely observed or are so uncertain that they are recorded as intervals containing the precise outcomes. Sometimes the exact value of a variable are hidden deliberately for some confidentiality reasons (see [21]). In all such cases intervals are considered as disjunctive sets representing incomplete information. Such perspective we perceive intervals are, according to Couso and Dubois [9], known as the epistemic view. In other words, an epistemic set A contains an ill-known actual value of a point-valued quantity x, so we can write x ∈ A. Since it represents the epistemic state of an agent, hence it does not exist per se.

20

P. Grzegorzewski and A. Kochanski

There are also situations when the experimental data appear as essentially intervalvalued data describing a precise information (e.g. ranges of fluctuations of some physical measurements, time interval spanned by an activity, etc.). Such intervals are called conjunctive and correspond to the ontic view (see [9]). Thus an ontic set is the precise representation of an objective entity, i.e. A is a value of a set-valued variable X , so we can write X = A. Please note, that it is the user who decides what kind of information is delivered by a given interval—whether we treat it as an epistemic or ontic set. Of course, both ontic and epistemic view yield different approaches to data analysis and statistical inference. Some persons conduct never-ending discussions whether mathematicians discover or create mathematics. Although there are groups that identify themselves with one of these opposite and disjoint standpoints it seems that there is no obvious reason to treat these two options as the exclusive ones. Many mathematicians claim that they sometimes feel as creators or as discoverers another time. Slightly similar situation happens when we face data: we generally try to extract information hidden in data but simultaneously we interpret and assign some meaning to data. Thus, the user is the one who attributes the meaning to data and similarly who tries to discover its deepest sense.

3 Uncertainty The main problem in transition from data to information and later to knowledge is omnipresent uncertainty. Indeed, as nothing is perfect, the same applies to the DIKW hierarchy. Although some persons claim that uncertainty results of some information deficiency and a new information reduce recipient’s ignorance, the problem does not appear so straightforward. Information uncertainty may be caused by uncertain data. But even reliable data do not guarantee high reliability of information. Actually, inappropriate or wrong method used for extracting information from data, may yield in misleading information or increase information uncertainty. Both data and information may be incomplete, imprecise, fragmentary, not fully reliable, vague, contradictory, or deficient in some other way. In general, these various data/information deficiencies may result in different types of uncertainty. What does it mean uncertainty? Is it subjective or objective? Some persons claim that: “...uncertainty is a personal matter; it is not the uncertainty but your uncertainty” (see [23]). Other try to develop objective theories fitting at least to some practical cases. Without settling the matter let us briefly review some well established approaches. Hartley [16] was probably the first who tried to quantify somehow the uncertaintybased information. But the term information theory usually brings to mind a theory established by Shannon [32] in 1948 and based upon probability theory. Although

From Data to Reasoning

21

mathematically elegant and effective, the classical information theory so strongly connected with classical set theory and probability cannot grasp all aspects of uncertainty since there uncertainty is manifested only in one form. Then several mathematical theories, distinct from probability theory, began to emerge in the 1960s. And soon it became clear that uncertainty can be manifested in different forms. The most famous are the theory of fuzzy sets proposed by Zadeh [37], evidence theory due to Dempster [11, 12] and Shafer [31], possibility theory by Zadeh [38], the theory of fuzzy measures by Sugeno [33, 34] and the theory of rough sets introduced by Pawlak [24]. Basing on these newborn ideas and theories the so-called generalized information theory began in the early 1980s. In the framework of this generalized information theory based on five mathematical theories (classical set theory, fuzzy set theory, probability theory, possibility theory, and the theory of evidence) three types of uncertainty are now recognized in which measurement of uncertainty is currently well established. These three uncertainty types are (see [18, 19]): • nonspecificity (or imprecision)—connected with sizes (cardinalities) of relevant sets of alternatives, • fuzziness (or vagueness)—resulting from imprecise boundaries of fuzzy sets, • strife (or discord)—expressing conflicts among the various sets of alternatives. It is worth noting that two of the uncertainty types: nonspecificity and strife, might be viewed as species of a higher uncertainty type, called ambiguity and associated with any situation in which it remains unclear which of several alternatives should be accepted as the genuine one. In general, ambiguity results from the lack of certain distinctions characterizing an object (nonspecificity) or from conflicting distinctions or from both of these. Fuzziness is different from ambiguity—it results from the lack of sharpness of relevant distinctions (see [18, 19]).

4 Three Types of Reasoning Any transition in DIKW hierarchy, discussed in Sect. 1, requires an appropriate mental process. This is also the case if we try to cope with any kind of uncertainty mentioned in Sect. 3. Two notions: reasoning and inference, are naturally associated with this context. Both terms denote a mental process of moving from premises to conclusions. But there is a subtle difference between these two concepts. Whereas in inference we accept a conclusion that results from available premises, reasoning means also potential situations (i.e. if we admit some premises being true then the appropriate conclusion inferred would be valid too) [2]. Moreover, reasoning can be based on using one or more logical frameworks, while inference is typically based on one logical framework. Thus inference is just a step of reasoning.

22

P. Grzegorzewski and A. Kochanski

Following Charles Sanders Peirce (see [25]) we can distinguish three types of reasoning: deduction, induction and abduction. Deduction (deductive reasoning) is generally defined as inference in which the conclusion about particulars follows necessarily from general or universal premises. In other words, deduction means the process of making conclusions which follow necessarily from generally accepted statements or facts. Suppose, for example, that a given foundry got an order for 20,000 particular casts. A technologist knows that the production line in his plant may deliver 400 casts per h. Therefore, from those two facts he deduce that he needs at least 50 h to execute the order. Whereas in deduction the truth of the conclusion is guaranteed by the truth of the statements or facts considered, induction (inductive reasoning) is a method of reasoning involving possibility. It refers to inference of a generalized conclusion that follow from particular instances. In other words, induction denotes such type of reasoning that forms generalizations based on what is known or observed. To illustrate this type of reasoning let us consider a typical situation of acceptance sampling used in statistical quality control. A general idea of such plan, in brief, is that instead of inspecting the whole lot of size N we select at random a sample of size n < N . Then, if the observed number d of defective items in the sample does not exceed the acceptance number a we accept the entire lot or reject otherwise. Suppose, for example, we have a lot of N = 8000 items. An appropriate sampling plan designed for the accepted quality level AQL = 1% is given by n = 200 and a = 5 (see International Standard , ISO 2859-1 Sampling procedures for inspecting by attributes). In this case if d 5 we induce that the quality all other items in the lot is good enough so we the accept the entire lot. Otherwise, if d > 5 we induce that the quality of other items in the lot is poor so we reject this lot. Of course, contrary to the previous example of the deductive reasoning, here we cannot be sure that are conclusion is right. We simply believe it is so since it can be shown that the probability of a poor quality lot if d 5 is very small (and, similarly, if d > 5 the the probability of a lot with a satisfying quality is very small too). There is also the third method of reasoning, called abduction (abductive reasoning), also known as retroduction, introduced into the modern logic by Peirce [25]. Sometimes it is defined as a syllogism in which the major premise is evident but the minor premise and therefore the conclusion are only probable. One can understand this kind of reasoning as the inference which starts with an observation and then seeks to find the simplest and most likely explanation. This type of reasoning can be used to develop a hypothesis which, in turn, can be tested with other reasoning or data. Let us consider the following example concerning SPC (statistical process control). Suppose we have noticed an alarm signal that appeared on a X -chart used for measuring a diameter of steel cylinders turned with a metalworking lathe. We know that an alarm signal may result when a tool bit breaks. Thus abduction leads us to hypothesis that in this very situation the alarm was caused by the tool bit breaking. It is clear that our conclusion is just a possible, or maybe even very probable, explanation however not necessarily sure.

From Data to Reasoning

23

Keeping in mind the example shown above one can formulate the abductive reasoning accomplishes the following scheme: 1. A fact A is observed. 2. If explanation (hypothesis) H were true it would imply A. 3. Hence there is a reason to suspect that H holds. However, abduction might be also perceived as the so-called inference to the best explanation [15]. Thus, following [17], we can modify the aforementioned scheme into: 1. 2. 3. 4.

A fact A is observed. If explanation (hypothesis) H were true it would imply A. There is no other hypothesis which explains A better than H . Hence H probably holds.

Please note, that the last scheme is typical for the probabilistic or statistical inference. All kinds of reasoning discussed above can be met in industrial manufacturing modeling. Induction and abduction may lead to some new knowledge which is uncertain. Deduction leads to certain conclusion but, from the philosophical standpoint one may notice that by this type of reasoning no new knowledge is created beyond the premises. Indeed, by the definition of deduction, all the derived propositions are implicit in the axioms. However, the deductive reasoning has also another logical flaw. The reliability of its conclusion depends on the preassumed axioms. But, by the famous Gödel theorem no system can demonstrate its own consistency (a consistent theory is one that does not contain a contradiction). All in all one should accept the truth about inevitable uncertainty. The last remark, however, does not need to sound pessimistic. The breakthrough came in the beginning of the twentieth century, when it was realized that although the knowledge created by any rule of generalizing from the particular is uncertain, it becomes certain knowledge, although of a different kind, once we can quantify the amount of uncertainty in it [28]. Famous statistician Rao [28] formulated a new paradigm, depicted in Fig. 4, and explains the whole idea as follows: • If we have to take a decision under uncertainty, mistakes cannot be avoided. • If mistakes cannot be avoided, we better know how often we make mistakes (knowledge of amount of uncertainty) by following a particular rule of decision making (creation of new but uncertain knowledge).

Uncertain knowledge

+

Knowledge of the extent of uncertainty in it

=

Useable knowledge

Fig. 4 New way of thinking—from uncertain to useable knowledge (see [28])

24

P. Grzegorzewski and A. Kochanski

• Such a knowledge could be put to use in finding a rule of decision making which does not betray us too often or which minimizes the frequency of wrong decisions, or which minimizes the loss due to wrong decisions.

References 1. Ackoff, R.L.: From data to wisdom. J. Appl. Syst. Anal. 16, 3–9 (1989) 2. Ajdukiewicz, K.: Pragmatic Logic (in Polish). PWN, Warsaw (1974) 3. Awad, E.M., Ghaziri, H.M.: Knowledge Management. Pearson Education International, Upper Saddle River, NJ (2004) 4. Ballou, D.P., Pazer, H.L.: Modeling data and process quality in multi-input, multi-output information systems. Manage. Sci. 31 (1985) 5. Bandemer, H.: Mathematics of Uncertainty. Springer (2006) 6. Bellinger, G., Castro, D., Mills, A.: Data, information, knowledge, and wisdom. http://www. systems-thinking.org/dikw/dikw.htm (2004) 7. Cleveland, H.: Information as a resource. The Futurist 34–39 (1982) 8. Cooley, M.: Architecture or Bee?. Hogarth Press, London (1987) 9. Couso, I., Dubois, D.: Statistical reasoning with set-valued information: Ontic vs. epistemic views. Int. J. Approx. Reason. 55, 1502–1518 (2014) 10. Data management association: the six primary dimensions for data quality assessment. Defining Data Quality Dimensions Report (2016) 11. Dempster, A.P.: Upper and lower probabilities induced by a multivalued mapping. Ann. Math. Stat. 38, 325–339 (1967) 12. Dempster, A.P.: Upper and lower probability inferences based on a sample from a finite univariate population. Biometrika 54, 515–528 (1967) 13. Gadomski, A.M.: Meta-ontological assumptions: information, preferences and knowledge universal interrelations (cognitive IPK architecture). ENEA’S paper (1999) 14. Hand, D., Manilla, H. Smith, P.: Principles of Data Mining. MIT Press (2001) 15. Harman, G.: Inference to the best explanation. Philos. Rev. 74, 88–95 (1965) 16. Hartley, R.V.L.: Transmission of information. Bell Syst. Tech. J. 7, 535–563 (1928) 17. Josephson, J.R., Josephson, S.G. (eds.): Abductive Inference: Computation, Philosophy, Technology. Cambridge University Press, Cambridge (1994) 18. Klir, G.J., Wierman, M.J.: Uncertainty-Based Information. Springer (1998) 19. Klir, G.J., Yuan, B.: Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice Hall (1995) 20. Kochanski, A.: Data preparation. Comput. Method Mater. Sci. 10, 25–29 (2010) 21. Kreinovich, V., Servin, C.: How to test hypotheses when exact values are replaced by intervals to protect privacy: case of t-tests. Departmental Technical Reports (CS), Paper 892, University of Texas at El Paso (2015) 22. Laudon, K.C.: Data quality and due process in large interorganizational record systems. Commun. ACM 29, 4–11 (1986) 23. Lindley, D.V.: Understanding Uncertainty. Wiley (2006) 24. Pawlak, Z.: Rough sets. Int. J. Comput. Inf. Sci. 11, 341–356 (1982) 25. Peirce, C.S.: Collected Works. Harvard University Press, Cambridge (1958) 26. Pipino, L.L., Lee, Y.W., Yang, R.Y.: Data quality assessment. Commun. ACM 45 (2002) 27. Pyle, D.: Data collection, preparation, quality and visualization. In: Nong, Y. (ed.), The Handbook of Data Mining. LEA Inc. (2003) 28. Rao, C.R.: Statistics and Truth. World Scientific Publishing (1997) 29. Redman, T.C.: Data Driven: Profiting from Your Most Important Business Asset. Harvard Business Press (2008)

From Data to Reasoning

25

30. Rowley, J.: The wisdom hierarchy: representations of the DIKW hierarchy. J. Inf. Sci. 33, 163–180 (2007) 31. Shafer, G.: A Mathematical Theory of Evidence. Princeton University Press (1976) 32. Shannon, C.E.: The mathematical theory of communication. Bell Syst. Tech. J. 27(379–423), 623–656 (1948) 33. Sugeno, M.: Theory of fuzzy integrals and its applications. Ph.D. Dissertation. Tokyo Institute of Technology, Tokyo (1974) 34. Sugeno, M.: Fuzzy measures and fuzzy integrals: a survey. In: Gupta, M.M., Saridis, G.N., Gaines, B.R. (Eds.), Fuzzy Automata and Decision Processes, North-Holland, 89–102 (1977) 35. Wand, Y., Wang, R.Y.: Anchoring data quality dimensions in ontological foundations. Commun. ACM 39, 86–95 (1996) 36. Wang, Y.W., Strong, D.M.: Beyond accuracy: what data quality means to data consumers. J. Manage. Inf. Syst. 12, 5–33 (1996) 37. Zadeh, L.A.: Fuzzy sets. Inf. Control 8, 338–353 (1965) 38. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst. 1, 3–28 (1978) 39. Zeleny, M.: Management support systems: towards integrated knowledge management. Human Syst. Manage. 7, 59–70 (1987) 40. Zins, C.: Conceptual approaches for defining data, information and knowledge. J. Am. Soc. Inf. Sci. Technol. 58, 479–493 (2007)

Data Preprocessing in Industrial Manufacturing Przemyslaw Grzegorzewski and Andrzej Kochanski

Abstract Each scientific modeling starts from data. However, even most sophisticated mathematical methods cannot produce a satisfying model if the data is of low quality. Before concluding about the quality of available data it is worth realizing the difference between datum quality or database quality. Moreover, most of data mining algorithms deal with the data in the form of an appropriately prepared single matrix. Unfortunately, the raw data is rarely stored in such form but is scattered over several databases, may contain observations which differ in formats or units, may abound with “garbage”, etc. Thus an adequate data preparation is an inevitable stage that should precede any modeling and further analysis. Both problems of data quality and data preparation are discussed in this chapter. Keywords Data · Database · Data cleaning · Data integration · Data mining Data preparation · Data reduction · Data quality · Datum quality · Data transformation · Empty value · Knowledge discovery from data (KDD) · Missing data · Missing value

P. Grzegorzewski (B) Systems Research Institute, Polish Academy of Sciences, Newelska 6, 01-447 Warsaw, Poland e-mail: [email protected] Faculty of Mathematics and Information Science, Warsaw University of Technology, Koszykowa 75, 00-662 Warsaw, Poland A. Kochanski Faculty of Production Engineering, Warsaw University of Technology, Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_3

27

28

P. Grzegorzewski and A. Kochanski

1 Data Analysis Structure Thousands of books and scientific papers contain a phrase data analysis. However, this term not always means the same. Sometimes it denotes the whole process of reasoning from data to conclusions. Otherwise, the authors reduce its meaning to some of its aspect, like data description or particular type of inference. To clarify the situation we have depicted the whole process of data analysis with its main steps in Fig. 1. Actually, any data analysis starts with some rankling questions or problems to be solved. Without them no progress is possible, as it was shown in the famous chat between Alice and the Cheshire Cat (see [6]): Alice: Would you tell me, please, which way I ought to go from here? The Cheshire Cat: That depends a good deal on where you want to get to. Alice: I don’t much care where. The Cheshire Cat: Then it doesn’t much matter which way you go.

Fig. 1 Structure of data analysis

Questions/problems formulation

Data collection

Data preprocessing

Exploratory data analysis

Modeling

Inferential data analysis

Conclusions

Data Preprocessing in Industrial Manufacturing

29

Next step after setting a goal is to gather necessary data. Usually we obtain data by conducting a sample survey or performing an experiment. Sometimes this step has to be preceded by designing of experiments. In some cases the reasoning is based on historical records, e.g. on some published materials. Note that collecting data may not be reduced to measurements recording but involves, for example, expert’s opinions or other prior information as well. Obtained raw data should not be included into essential inference without their cross-examination. This step, called data preprocessing, is necessary to examine whether data are genuine or faked, to detect possible measurement errors, recording errors and outliers, to test validity of prior information, to get rid of irrelevant or redundant information and so on. So prepared data can be used in exploratory data analysis (EDA). It is done to understand the nature of the data, to describe, visualize and summarize their main characteristics. Basically, EDA is intended to make the data speak so we would be able to prepare a reasonable and efficient model. However, sometimes as a result of EDA one also formulate new hypotheses which could lead to new data collection and experiments. Model is roughly speaking a mathematical construction applied in science to describe observed objects or phenomena. Model helps to understand a object/ phenomenon under study, to explain its relation with other objects/phenomena and effects of its different components, and to make predictions about its behavior. Thus the next step of reasoning is the specification of a suitable model. Then one has to test the validity of a specified model or select a more appropriate model (or a class of models) for further data analysis. Here the desired tasks are not limited to checking assumptions required for model fitting only but include also such activities like making transformations of data/variables as needed, handling missing values, etc. The problem of specification is not a simple one but is of a extreme importance since incorrect specification may lead to wrong inference and false final conclusions. Therefore, in the case of fears of misspecification one should choose between examining data under different possible models or applying a robust procedure which reduces the sensitivity against possible alternative models. As soon as a model is accepted one may start the next step, i.e. inferential data analysis. Classical statistical inference comprises of estimation (point or interval), hypothesis testing and prediction. The results of the inference lead to decision making. Usually, data analysis provides not only answers/solutions to questions/problems that initialize the whole process but also helps with improving various stages of the data analysis, raise some new questions and bring motivations for future investigations.

2 Data Quality: Datum Quality Or Database Quality? The quality of the mathematical modeling can be perceived as its adequacy, i.e. sufficient correspondence with reality. If so then the data quality seems to be its

30

P. Grzegorzewski and A. Kochanski

preliminary condition. There are many definitions of data quality. Generally, data is considered to be of high quality if “they are fit for their intended uses in operations, decision making and planning” (see [21]). According to Wang [26] we could find a general definition of data quality: “Following this general quality literature, we define data quality as data that are fit for use by data consumers. In addition, we define a data quality dimension as a set of data quality attributes that represent a single aspect or construct of data quality”. Since the overall data quality may be influenced by many factors, we can consider, at least theoretically, an infinite number of data quality dimensions. Therefore, it seems crucial to establish a particular closed set of generally accepted and precisely defined parameters which make possible such an evaluation and to have a selection method allowing us to reduce the number of factors only to those which are relevant in a particular situation. Wang [26] demonstrated that establishing a generally accepted set of criteria is very difficult. This aspect will be discussed in more detail later on in this section. The very selection of the criteria is not accessible without controversies as well. For instance, Wand and Wang [25] listed four considerably different approaches that could be applied here: • • • •

intuitive understanding, industrial experience, empirical studies, literature review.

Ten years later Bandemer [4] stated that “The quality of the data is assessed by how reliably they characterize the given problem. So the question is, whether they are the right data: Are they suitable to reflect the given problem in the model or are they only minor matters of neglectable importance? Do they belong really to the given practical problem or do they describe a quite different situation? Are they carefully and responsibly obtained and specified, suitably for the problem, as numbers or as other mathematical objects?”. Research concerning the influence of data quality on models supporting decision making systems is not new—it has been conducted for at least the last 30 years (see [3]). From the very beginning [15] data quality has been defined as a multidimensional problem/notion, involving the following quality parameters: record completeness, record inaccuracy and record ambiguity. The list of criteria for data quality assessment offered in [17], which is frequently used in the case of data bases such as data bases of financial organizations, consumer organizations, or health care organizations, contains 16 criteria. These are criteria such as objectivity, understood as data impartiality, believability, perceived as truthfulness, but also timeliness, demonstrating that the data does not pertain to the past, or the quality of being free-of-error, demonstrating that the data are correct and reliable. Wang and Strong [26] aimed not only at preparing a list of parameters characterizing data quality, but also at postulating a hierarchical structure resulting from ordering these parameters with respect to their significance. To this end, a questionnaire was prepared which was sent out to current data base users, as well as to other

Data Preprocessing in Industrial Manufacturing

31

people with considerable experience working with data bases. A significant percentage of respondents were people connected with the production industry. The authors took the trouble and prepared, on the basis of the available literature and their own experience, a preliminary list of aspects which should be captured in the criteria for the data quality assessment. Thus the data should be: • available, i.e. the user should know how to obtain the necessary data, • comprehensible for the user, i.e., for instance, recorded in a language that he fluently uses, • adequate, i.e. selected appropriately with respect to the aims for which they are to be employed, • exact. Although these four conditions apply to the general notion of data quality, the last criterion pertains only to a datum, the remaining three have to do with database quality. These distinction reflects two different perspectives that should be taken into consideration. Three criteria related to database quality versus the only one applicable to a single datum may suggest that the quality of the database is more important. A questionnaire containing 179 criteria of quality assessment was filled in by the respondents. In the vast majority these were criteria for the assessment of database quality such as cost of collecting data, actualization ease, division ease. In the second stage, 118 selected parameters were ordered with respect to their importance. Even though the respondents used the whole range of grades in their assessment (i.e. from “non-significant” to “most significant”), as much as 85% of the criteria was considered as significant, while only two criteria: accuracy and correctness, were recognized as the most important ones. What has to be emphasized is that both criteria pertain to datum quality, rather than to database quality, as was the case with the majority of other criteria assessed in the questionnaire. It should also be noted that a significant part of the parameters under analysis cannot be considered in general terms, that is, in isolation from a particular database. A good example is the non-redundancy criterion. The respondents assessed this criterion as non-significant (the assessment value 6.279, in the scale from 1 to 9, where 9 means insignificant). This is where the unambiguity of the criterion comes into play. It remains uncertain whatever redundancy the respondents have in mind: attribute redundancy or observation redundancy. Attribute redundancy, for instance, double appearance of the same attribute under different names can be detected, and the detection methods are widely discussed in the literature. Let us consider the other kind of redundancy then. In this context two questions arise: What does the observation redundancy really mean? and Isn’t redundancy informative? Dwelling on this a bit more we should answer the question of whether similar observations are redundant and—if so—how similar they need to be to be considered redundant. In databases containing data from industrial processes the vast majority of observations are similar or identical, as a matter of the very nature of a production process in which repeated actions and operations result in a replicable product. The lack of replicability is usually a sign of the process instability.

32

P. Grzegorzewski and A. Kochanski

Table 1 Two synthetic databases containing the same number of missing data No. A1 A2 A3 A4 A5 A6

No. A1 A2 A3 A4 A5 A6

1 2 3 4 5 6 7 8 9 10 11 12

1 2 3 4 5 6 7 8 9 10 11 12

X X X X X X X X X X X

X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X

X X X X X X X X X

X X X X X

X X X X X X X X X X X X

X X X X X X X X X X X X X X X X X X X

X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X

A similar mixture of criteria pertaining to datum quality and database quality is considered in one of the more recent studies. The Data Management Association in its 2016 report [7] recommends 6 criteria of data quality assessment: completeness, uniqueness, timeliness, validity, accuracy, and consistency. For each of the proposed criteria the report suggests measures which make possible to compare different sets of collected data. The first of the proposed parameters has to do with the database completeness. This is measured via the percentage of the database cells containing absent data. Employing this single criterion without additional information is not a reliable quality assessment parameter. Table 1 shows two synthetic databases containing the same number of absent data but exhibiting significant differences with respect to their quality and, what is even more important, to chances of their quality enhancement. Just like the quality assessment criteria discussed so far, in this case the proposed criteria pertain also to both database quality and datum quality. And the majority of the criteria, except for accuracy, have to do with database quality. It is visible from the discussion above, no generally accepted and recognized criteria and methods of data quality assessment have been worked out. This results, among other reasons, from the fact that data quality has not been properly defined so far. Pyle [20] states that the notion of quality cannot be defined and analyzed without reference to the aim of the data value and the usefulness assessment. In data mining we can distinguish two stages: data preparation and model development. Pyle regards the notion of quality both to database quality and datum quality, depending on the data mining stage. In data preparation he uses the label “quality” with respect to datum quality. He also indicates three main reasons why one should concern for a data quality assessment. Firstly, because of the very nature of measurements making and recording we always deal with noisy data. Such noise results mainly from

Data Preprocessing in Industrial Manufacturing

33

the lack of calibration of the measuring equipment or from some carelessness of a person responsible for measurements. Secondly, we always deal with some natural variability of the observed process or phenomenon. The third possible reason is a “bias”, i.e. a stable measurement disturbance which make observations either mostly bigger or mostly smaller than they are in reality. In the second stage, i.e. a model development, the quality issue focuses, in Pyle’s opinion, on data consistency and their adequacy with respect to the model being developed. And—what is ultimately important—the quality of the data used for modeling should correspond to the quality of the data used in the latter model application. A clear distinction between data quality for individual measurement and for collections of data was proposed in [11]. Hand et al. [11] identified two requirements pertaining to the quality of an individual measurement: • measurement precision, which is characterized by a small dispersion frequently measured by the variance, • measurement accuracy, which takes into consideration not only a small measurement variation but assumes also that the results of measurements are close to the value considered as the true one. However, the proposed requirements of data quality remain reasonable if and only if the underlying sample is representative of the population. The second criterion for the data set quality assessment is the appearance of missing data, outliers and dubious data. Some authors suggest that such problematic data should be removed from the data set before the model is being developed. In the present paper the authors recommend the following two notions defining data quality: • datum quality—pertaining to the quality of individual data, • database quality—pertaining to the quality of a data set. These two notions constitute two distinct aspects of data quality. Consequently, keeping in mind these two aspects we should define adequate quality criteria and design actions oriented on the quality improvement. The methodology of data preparation, put forward in [14] and discussed in more detail in Sect. 3, focuses on four tasks of particular importance for industrial data: • • • •

data cleaning, data integration, data transformation, data reduction.

Data cleaning involves three actions performed with respect to the collected data: replacement of missing and empty values, accuracy improvement, and inconsistency removal. These actions relate to the accuracy and completeness considered and calculated independently for observations and for the attribute consistency. The first two parameters define the datum quality, while the third one the database quality. Data integration also comprises three actions: identification, redundancy removal and unification. The last two actions are connected with the quality parameters. The

34

P. Grzegorzewski and A. Kochanski

removal of unnecessary redundancies enhance the database quality characterized via the criterion of uniqueness. However, the unification is unequivocal: it may change nothing with respect to data or database quality, but it may bear in negatively upon the quality (e.g., it may decrease the comprehensibility of data, in particular, by introducing some rarely used units). It is also worth noting that some unit transformations may result in a loss of information contained in the data. For instance, it seems that by the transformation of the temperature measurements recorded in the Kelvin scale into the corresponding values in the Celsius (or Fahrenheit) scale nothing changes except of numerical values. Moreover, such transformation increases the readability and comprehensibility of the data for the vast majority of users. Unfortunately, here we have not only a transformation in units but in measurement scales as well, since a measurement originally taken in the absolute ratio scale is transformed into a measurement in the interval scale which results in decreasing the number of admissible mathematical operations. Data transformation and data reduction comprise mainly actions undertaken for the sake of simplifying or speeding up the modeling. Unluckily, such actions may lower the data quality. Is happens, for instance, as a result of transforming or reducing the volume of the data set which is unavoidably connected with a partial loss of information stored in the data. Some transformation tasks may improve the quality of data. Consider, for instance, data containing some noise generated by the measuring equipment of low quality which results in the lack of measurement precision or by employing inappropriate measuring methods. Here, by data smoothing we remove noise which enhances the quality of an individual datum. Another example is the attribute selection, which is a part of the reduction task. It may cause a decline in quality if one cut off a significant variable. But in may also improve the quality through the elimination of parameters which are not connected with a phenomenon under study (here it is equivalent to the removal of an accidental noise) or by the elimination of some redundant correlated variables (here we improve the data quality by intensifying the explaining power of the remaining attributes). To sum up, it seems that the use of a single notion to describe data quality is not sufficient for selecting appropriate tools for the quality improvement. Two notions: datum quality and database quality, should be considered independently. As has been demonstrated in a number of contributions, the selection of criteria for assessing database quality should depend on various factors, while the choice of assessment parameters is dictated by numerous aspects and the specificity of a particular situation under study. This stands in contrast with the datum quality characterized by more or less accepted parameters which depend on understanding of the very nature of measurement.

Data Preprocessing in Industrial Manufacturing

35

Fig. 2 Time to complete during Data Mining (Source [19])

Fig. 3 What data scientists spend most time on (Source [18])

3 A Taxonomy and Overview of Data Preparation 3.1 General Remarks The two key tasks in data exploration are data preparation and model construction. This is reflected, among others, in two well-known and widely discussed methodologies of process modeling: CRISP-DM and SEMMA [2, 23, 24]. It is widely recognized that data preparation consumes about 80% of the time devoted to data exploration [16]. A more detailed division of labor which confirms this relation is illustrated in Fig. 2 (cited from [19]). Recent research, whose results were published in [18], confirmed these percentages, as shown in Fig. 3. At the same time the study demonstrated that data preparation is treated as the stage of data exploration which is considered as particularly tedious. It should be stressed that the time which required for data preparation is directly related to the size of the data set. Thus in the big data era it becomes indispensable

36

P. Grzegorzewski and A. Kochanski

to develop powerful tools of data preparation. One of the distinguishing big data characteristics is their architecture which is distributed across multiple machines and clusters [5]. Hence data preparation and data consolidation tasks gain in importance. Below we sketch the idea of data preparation and its basic methods. For more detail we refer the readers, e.g., to [10, 19, 22, 27].

3.2 An Overview of Data Preparation For a long time we have been observing a perpetual development of data exploration algorithms. New methods are emerging and those already in use are constantly being refined. However, no such progress can be observed in the domain of data preparation. Studies on the relationship between the data quality and the model quality are not numerous (see [12, 13]). Moreover, they are usually focused on individual data sets which preclude further generalized conclusions. In some fields, like missing data or discretization one can observe a constant development (see, e.g., [9] and [8], respectively). Nevertheless, there are simple situations where the suggested algorithms cannot manage the task. Data preparation becomes substantial nowadays in the presence of more and more large and diversified data sets. Huge velocity, variety and volume, typical for big data (see [1]), require new tools dedicated for data preparation. Anyway, data preparation is highly recommended because of many reasons including the following: • an improvement of datum quality or of database quality which enhances the quality of the resulting model, • an acceleration of the process of modeling and further analysis, • the possibility of using algorithms which otherwise could not be applied (e.g. because of noisy or absent data), • the extension of the model applicability (by broadening the range of admissible data), • an increase of the model reliability (via the use of a more complete data set in modeling). High quality models requires high quality data. Thus data preparation should be conducted in accordance with a carefully designed procedure which guarantees relatively good effects measured in terms of the data quality at possibly small time expenditure. To achieve this goal an appropriate methodology of data preparation is desirable. On the other hand, inappropriate data preparation may decrease the model quality and result in the loss of the capability for generalization.

Data Preprocessing in Industrial Manufacturing

37

Fig. 4 Tasks realized in the process of data preparation with the distinction into the initial and the main stage (Source [14])

3.3 A Taxonomy of Data Preparation The methodology of data preparation dedicated specifically to industrial data was firstly proposed in [14]. As was demonstrated in Chap. 1 (Sect. 2), industrial data have their own specificity and therefore, a methodology of data preparation for such data may differ in certain aspects from a methodology relevant for other types of data. A general scheme of data preparation is shown in Fig. 4. One can distinguished two stages there: the initial and the main stage. The area encompassed by the dashed lines (squared area) represents the initial stage, while the grey area correspond to the main stage. It seems that the initial stage will decline for big data, because their volume precludes non-automatized methods of data preparation. Each stage (initial or main) may consist of different tasks. A task is a separate selfcontained part of data preparation which may be and which in practice is performed

38

P. Grzegorzewski and A. Kochanski

in the data preparation stage. We can distinguish four tasks: data cleaning, data transformation, data integration, and data reduction. Depending on the process under study as well as on the data preparation stage a task may consist of a different number of operations.

3.4 Data Preparation Tasks In the case of industrial data preparation we can distinguish the following tasks: • data cleaning—used to eliminate any inconsistency or incoherence in the collected data, • data integration—which makes possible an integration of databases coming from various sources into a single two-dimensional table (and hence enables an application of well-known data mining algorithmized tools), • data transformation—which includes various operations making possible a model building, accelerating calculations and improving accuracy, • data reduction—used to decrease the dimensionality of a database (and hence to cut down the time necessary for data mining). Initial data cleaning involves a one-time elimination of all kinds of errors resulting from a human negligence in data collection. Initial data cleaning is a laborious process of assessing the source data (often in the form of paper notes, laboratory forms or printouts from the lab and measuring equipment) to supplying the lacking data manually. Obviously, for big data such activity becomes unrealistic. The initial data integration denotes the elimination of obvious repetitions in a data set. Initial data transformation usually includes individual operations forced by data integration through aggregation, generalization, or attribute construction. Finally, initial data reduction is usually restricted to attribute selection and is based on the data analyst knowledge and experience. Initial data preparation may start with any task. However, no matter what tasks are realized and regardless of the operation sequence, data cleaning is the indispensable task of the initial stage. Data cleaning in the main stage involves the following three operations: • missing and empty values imputation—towards enriching the available data set with the missing values representing the absent data (in technological processes databases the proportion of the missing data in selected attributes may reach tens of percents; moreover, the distribution of the missings may be totally accidental but may also reveal some correlations), • accuracy improvement—through the elimination or replacement of the erroneous data (e.g. outliers) with appropriate or corrected values (which is always problematic since a decision whether given outlier is a correct or erroneous observation is often disputable),

Data Preprocessing in Industrial Manufacturing

39

• inconsistency removal—realized by some special procedures (e.g. control codes) for indicating unrealistic values or tools designed to discover incompatibilities or correlated attributes (like a temperature and the amount of emitted heat or the rotation speed of an engine and a gear which is linked with it) which may weaken the power of the model. Data mining typically requires integrating data sets coming from many separate bases. A typical situation in a production plant is such that we deal with at least three different data sets: data from the production line, data from the supplier or plant laboratory of materials and raw materials and data from the quality control section. Integration data coming from different sources requires the following operations: • identification—involves methods of identifying properties or parameters in separate data sets, • unification—adjustment of formats and meanings in the diversity if data registration methods (which may differ as a result of utilizing measurement instruments obtained from different producers), • redundancy removal—elimination of all repetitions and redundancies (e.g. multiple appearing of the same attributes hidden under different labels). Data transformation involves the following operations on the data which make possible their exploration: • smoothing—the elimination of a noise which appear in data as a result of carelessness in taking measurements (human factor) or non-calibrated instruments; smoothing includes techniques such as binning, clustering or regression, • aggregation—a summary of some data or counting up some values (e.g. the number of faulty items or defects registered in a given period), • generalization—the replacement of some measurements by the higher-order value (e.g. via their discretization), • normalization—the adjustment (rescaling) of the values into the specified range, • attribute (feature) construction—a transformation of some attributes (features) which results in a new attribute which seems to operate better than the original ones, • accommodation—an adjustment of the data format for a specific algorithm or a tool. Data reduction aims at arriving at a significantly reduced data representation which at the same time preserves basic features of the initial data set. Data reduction includes the following five operations: • attribute selection—the elimination of the less significant attributes, • dimensionality reduction—the process of reducing the number of variables under consideration by obtaining a set of some new attributes without loosing information contained in the original data set, • discretization—the replacement of the original data with some newly created higher-order values (which are easier for the the interpretation and/or implementation),

40

P. Grzegorzewski and A. Kochanski

• numerosity reduction—the process of reducing the volume of data through the elimination of the identical or very similar cases, • aggregation—understood as in data transformation. It is worth noting that the same operation is sometimes realized in different tasks. Moreover, each operation may be executed using different techniques or methods.

References 1. Assunção, M.D., Calheiros, R.N., Bianchi, S., Netto, M.A.S., Buyya, R.: Big data computing and clouds: trends and future directions. J. Parallel Distrib. Comput. 79–80, 3–15 (2015) 2. Azevedo, A., Santos, M.F.: KDD, SEMMA and CRISP-DM: a parallel overview. In: Proceedings of the IADIS European Conference on Data Mining, pp. 182–185 (2008) 3. Ballou, D.P., Pazer, H.L.: Modeling data and process quality in multi-input, multi-output information systems. Manag. Sci. 31 (1985) 4. Bandemer, H.: Mathematics of Uncertainty. Springer (2006) 5. Begoli, E., Horey, J.L.: Design principles for effective knowledge discovery from big data. In: Proceedings of the Joint Working IEEE/IFIP Conference on Software Architecture (WICSA) and European Conference on Software Architecture (ECSA), pp. 215–218 (2012) 6. Caroll, L.: Alice’s Adventures in Wonderland 7. Data Management Association: The six primary dimensions for data quality assessment. Defining Data Quality Dimensions, Report (2016) 8. García, S., Luengo, J., Herrera, F.: Discretization. In: Data Preprocessing in Data Mining. Intelligent Systems Reference Library, Springer (2015) 9. Grzymala-Busse, J.W., Hu, M.: A Comparison of several approaches to missing attribute values in data mining. In: Ziarko, W., Yao, Y. (eds.) Rough Sets and Current Trends in Computing, RSCTC 2000. Lecture Notes in Computer Science, pp. 378–385 (2005) 10. Han, J., Kamber, M., Pei, J.: Data Mining Concepts and Techniques. Morgan Kaufmann Publisher (2012) 11. Hand, D., Manilla, H. Smith, P.: Principles of Data Mining. MIT Press (2001) 12. Kochanski, A.: Prediction of ductile cast iron properties by artificial neural networks. Ph.D. Thesis, Warsaw University of Technology (1999) (in Polish) ´ atkowski (ed.) Polish Met13. Kochanski, A.: Aiding the detection of cast defects causes. In: Swi¸ allurgy 2002. Komitet Metalurgii Polskiej Akademii Nauk (2006) (in Polish) 14. Kochanski, A.: Data preparation. Comput. Method Mater. Sci. 10, 25–29 (2010) 15. Laudon, K.C.: Data quality and due process in large interorganizational record systems. Commun. ACM 29, 4–11 (1986) 16. McCue, C.: Data Mining and Predictive Analysis: Intelligence Gathering and Crime Analysis. Butterworth-Heinemann (2007) 17. Pipino, L.L., Lee, Y.W., Yang, R.Y.: Data quality assessment. Commun. ACM 45 (2002) 18. Press, G.: Cleaning Big Data: Most Time-Consuming, Least Enjoyable Data Science Task, Survey Says. Forbes, 23 Mar 2016 19. Pyle, D.: Data Preparation for Data Mining. Morgan Kaufmann Publisher (1999) 20. Pyle, D.: Data collection, preparation, quality and visualization. In: Nong, Y. (ed.) The Handbook of Data Mining. LEA Inc. (2003) 21. Redman, T.C.: Data Driven: Profiting from Your Most Important Business Asset. Harvard Business Press (2008) 22. Refaat, M.: Data Preparation for Data Mining Using SAS. Morgan Kaufmann Publisher (2007) 23. SAS Enterprise Miner—SEMMA. SAS Institute 24. Shearer, C.: The CRISP-DM model: the new blueprint for data mining. J. Data Warehous. 5, 13–22 (2000)

Data Preprocessing in Industrial Manufacturing

41

25. Wand, Y., Wang, R.Y.: Anchoring data quality dimensions in ontological foundations. Commun. ACM CACM 39, 86–95 (1996) 26. Wang, Y.W., Strong, D.M.: Beyond accuracy: what data quality means to data consumers. J. Manag. Inf. Syst. 12, 5–33 (1996) 27. Witten, I.H., Frank, E., Hall, M.A., Pal, C.J.: Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann Publisher (2016)

Part II

Applications

Tool Condition Monitoring in Metal Cutting Krzysztof Jemielniak

Abstract Automatic tool condition monitoring is based on the measurements of physical phenomena which are correlated with the tool wear, and thus can be exploited as the tool wear symptoms. However measured quantities depend not only on the tool wear but also on a variety of other process parameters of random nature, making the relationship between tool wear and measured value very complex which has a statistical rather than strict, predictable nature. Therefore, the development of a robust and reliable tool condition monitoring system requires a combination of different, meaningful signal features, which best describe the tool wear. There are numerous signal features (SFs) that can be extracted from time domain, frequency domain or time-frequency domain signal. As it is really not possible to predict which signal features will be useful in a particular case thus these informative, correlated with tool wear, should be automatically selected. The information extracted from one or several sensors’ signals has to be combined into one tool condition estimation. This can be achieved by various artificial intelligence methods.

1 Introduction The demand for continuous improvements in product quality, dependability, and manufacturing efficiency has imposed strict demands on automated product measurement and evaluation. Therefore, automated process monitoring is crucial in successfully maintaining high quality production at low cost. Reliability of manufacturing processes like turning, milling or drilling can be noticeably improved by robust tool wear monitoring systems. Such system should allow for the exchange of worn tools in time, thus avoiding the production waste. It also allows for application of higher, more effective cutting parameters, as the risk of costly catastrophic tool failures is reduced, or—if such failure occurs—its consequences are minimized K. Jemielniak (B) Faculty of Production Engineering, Warsaw University of Technology, ul. Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_4

45

46

K. Jemielniak

Fig. 1 Structure of tool condition monitoring system

due to immediate detection and termination of machining process. Hence, extensive research work is taking place world-wide in the area of tool and process condition monitoring, being one of the most important focuses of research efforts from more than forty years. Numerous papers have been published since (see e.g. [1]) presenting many ideas, and many approaches have been proposed to accomplish tool condition monitoring. Nonetheless the problem is still far from solved, as the existing systems, both laboratory and commercially available are based on vague, incomplete or randomly-distorted sensory information about the condition of a cutting tool. The structure of tool condition monitoring (TCM) systems is presented in Fig. 1. They are based on measurement of process variables that are dependent on tool wear and can be exploited as tool wear symptoms. The most effective, consequently most often used are cutting forces (or force dependent measures like power, torque), acoustic emission, and vibration [1–4]. These quantities are measured online, during a cutting process. However, they depend not only on the tool wear but also on a variety of other process parameters of random nature, such as inhomogeneity of work material or tool geometry, machine tool temperature etc. The cutting process itself is nonlinear, time-variant system. Therefore the relationship between tool wear and the measured value is very complex and it has a statistical rather than strict, predictable nature. Selected process variables of the monitoring process, which indirectly correlate with tool condition, are measured by appropriate sensors producing analogue signals. The industrial sensors used in TCM systems are different from laboratory ones. They are more robust, immune from the harsh machining zone environment, with disturbances like flying chips, coolant, or overload, especially dangerous for dynamometers. However, they are less accurate, affected by crosstalk between sensor channels, electric interference during the signal transmission through cables and instrumentation [5, 6]. The dependence of the output signal on measured quantity is not as strict and repeatable like in the case of laboratory sensors. Therefore the sensors and measuring chain are an important source of uncertainty. Signals acquired from the sensors are then subject of signal processing, the aim of which is to generate useful signal features, correlated (at least potentially) with tool or process condition. Tool condition monitoring is always based on some form of a model of dependence of signal feature (SF) on tool condition. The model is built

Tool Condition Monitoring in Metal Cutting

47

Fig. 2 Signal processing scheme

during system training. Then the model is reversed, and in monitoring mode tool condition can be evaluated from SF values using a decision making algorithms. The following stages can be distinguished in signal processing (Fig. 2). First—preprocessing (filtering, amplification, A/D conversion, segmentation). Sometimes the signal is transformed into frequency or time-frequency domain (FFT, STFT, WT etc.). The next stage of signal processing is the extraction of signal or signal transform features changing in time, with tool or process condition. There can be many different descriptors from different sensor signals, most of them hardly related to monitoring process, therefore, a feature selection procedure is necessary. Relevant features, are then integrated into a diagnosis of the tool condition. This can be achieved by various means such as statistical methods, auto-regressive modeling, pattern recognition, expert systems, and others. The tool condition estimation can just be announced to the operator or sent to the NC controller executing appropriate action (see Fig. 1). The objective of this chapter is to discuss various aspects of the signal processing and feature integration techniques applied in tool condition monitoring, aiming at removing or reduction information uncertainty in sensory signals and producing reliable tool condition estimation.

2 Signal Preprocessing The analog signal from the sensor usually cannot be connected directly to A/D converter, but has to be prepared by a conditioner specific to the sensor (positron coupler, charge amplifier etc.). For example, the typical procedure of analog AE signal preprocessing follows the pattern schematically shown in Fig. 3.

Fig. 3 A typical measuring chain in an AE measurement system in metal cutting

48

K. Jemielniak

The piezoelectric AE sensor is usually placed as close as possible to the cutting zone, e.g. to the tool shank, the tool post, the head stock or to the spindle. Because of the high impedance of the sensor it must be directly connected to a buffer amplifier which converts charge signal from quartz sensor into the proportional voltage signal. This is typical also to other piezoelectric sensors like dynamometers or accelerometers. The analog sensor signal should be filtered to keep it within the range of the frequency response of the sensor, suppress high frequency noise or continuous biases. E.g. low frequency noise components, which are inevitably present in AE signal, are not correlated with process condition and hence useless. Besides, they can be of high amplitude forcing the usage of lower signal amplification. This results in lower amplification of useful band of the signal. Sometimes, the AE signal is then fed through a low-pass filter to get rid of the high frequency noise components due to electric sparks, etc. or to avoid aliasing. The filtered AE signal is a subject to further processing and/or recording. The frequency range of the raw AE signal reaches 1 MHz (typically 80–700 kHz) so dealing with it requires high sampling frequency (more than 2 MS/s), a lot of computer memory and computation cost. Therefore the most often AE signal is demodulated to into effective value (RMS), to obtain a low frequency variable, which can be further processed with the less expensive signal processing equipment. The integration time constant of the RMS converter should be carefully selected, depending on following signal feature extraction. Integration time constant should be ten times shorter than typical burst duration, which is some 2 ms [7]. The AE signals originating from the cutting zone can be considerably strong. Because of the characteristics of pre-processing units, such high amplitude signals sometimes cause overloading of the buffer amplifier and saturation of the signal. High-pass filtering of saturated signal results in temporal vanishing of the signal value (Fig. 4). This can often result in a misleading evaluation of the data. It should be pointed out that such signal distortion cannot be detected in AE RMS signal, so such signals must be considered as completely distorted, thus useless. To avoid such problems the gain of the buffer amplifier should be as small as possible, and any further necessary amplification should be done after signal filtering. In other words, the AE signals should be high-pass filtered at the earliest possible stage of processing, just after the unavoidable buffering. It is particularly important while using AE RMS signals only instead of AE raw [8]. After conversion of analogue signals into digital time series, they are often subject of further preprocessing. Digital filtering reduces frequency bands not correlated with monitored process or extracts information necessary at a particular stage of pattern recognition. For example, while using a spindle-integrated force sensor system on a machining center, the cutting force signals are distorted when the spindle speed harmonics coincide with the natural modes of the spindle structure. Kalman filter eliminates the influence of structural modes on the force measurements, and significantly increases the frequency bandwidth of the force measurement system [9]. Scheffer and Heyns [10] investigated possible signal features that might correlate with tool wear in interrupted cutting. They applied digital filters to separate two frequency ranges of cutting force signals into low-frequency content that was an

Tool Condition Monitoring in Metal Cutting

49

Fig. 4 Examples of acoustic emission signals saturated in the buffer amplifier than further distorted by high pass filtering, finally demodulated into ARRMS [8]

indication of the static cutting forces and the higher frequency range, the natural frequencies of the toolholder, that resulted from the excitation from the cutting operation. Jemielniak [11] used low-pass filtering of cutting force signal for catastrophic tool failure detection in turning, based on detection of sudden changes of the force value. The filtering allows to apply much lower tolerance band of the force value. Further filtering was applied to obtain the tolerance limits, following slow changes of the cutting force, due to change of uncut chip thickness. In many applications, a digital signal is filtered in order to prevent high frequency noises and signal oscillations due to transient mechanical events e.g. [12, 13]. The signal acquired during operation consists of sequences of positioning movements and working feed—see example in Fig. 5. The working feed consists of air cutting (idle) and actual cutting (removing metal). Duration of air cutting can vary between workpieces, so precise determination of actual cutting reduces the uncertainty of further signal feature extraction results. Thus the next stage of the signal preprocessing is automatic detection of actual cutting during the working feed, identified on the base of digital signals from CNC controller. The simplest, most often used method of cutting recognition is detection of the signal value crossing of the preset threshold [14, 15] in a time window selected by the user. The threshold value is calculated as part of maximum signal value, which makes the method not applicable automatically, online, as the max value is not known before the cutting starts. Another disadvantage of this method is possible disturbances of the signals occurring in many industrial applications. Signal from piezoelectric transducer (e.g. cutting force) may fall down or rise during an air cut or even become negative during cutting due to complex cross coupling between sensor sensitive directions. The latter is characteristic for industrial cutting force sensors mounted

50

K. Jemielniak

Fig. 5 The feed force signal registered during drilling of subsequent holes [5]

under the turret of the lathe, when the force signals are not just proportional to the cutting, feed or passive force but are related in complex ways to all of them and the turret position. Then the measured sensor signal can become negative during cutting, in spite all cutting force components are positive. Some signals are heavily distorted, which might efface cutting process. Therefore, cutting detection should be based on more than one signal, and more than one signal feature. Bombi´nski et al. [5] presented the algorithm which allows for detection of cutting based on all available signals using their low pass filtered values and standard deviation as signal features. The method is presented in Fig. 6, where two signals—cutting force F c and AE RMS acquired simultaneously are analyzed as an example. The first operation is offset removal performed 40 ms after receiving signal “working feed on” from the CNC controller. Standard deviation σ0 and average value S av of the sensor signal S are calculated from the 120 ms segment of the signal. The latter is subtracted from the signal as an offset, thus during air cutting the signal should oscillate around zero. The standard deviation calculated during the offset removal—σ0 (here σ0 (F c ) and σ0 (AE RMS )) is a measure of signal disturbances characterizing the sensor installation, which might be dependant on the spindle rotational speed, feed, position of the current etc. Therefore it can be used for determination of the threshold values for cutting detection. After the offset removal, the actual cutting detection starts. Every 2 ms two signal features are calculated. The first one is S f —signal filtered with low pass, 1 Hz Butterworth II order filter. This feature represents moving average value of the signal and it is the most effective in the absence of the signal drift or change of sign of the signal value due to cross coupling mentioned above. The second feature is σc —standard deviation of a 400 ms fragment of this signal, which is independent of the signal drift or sign changes. If there are more available sensor signals in the system, all of them are used for the cutting detection. The system recognizes the beginning of cutting if S f > 5σ0 or σc > 3σ0 for any of the signals more than 200 ms. In the example presented in Fig. 4 the earliest threshold crossing appeared at 8.35 s for the standard deviation of the cutting force thus the cutting was recognized at 8.55 s. Interruption

Tool Condition Monitoring in Metal Cutting

51

Fig. 6 The cutting detection method [5]

of the cutting is recognized after all filtered signals and standard deviations which were above their thresholds, falls below the thresholds. The multipliers: 5 for filtered signal S f and 3 for the standard deviation of the signal σc were established on the base of experience from several installations and many experiments. If some signals have a strong drift tendency, cutting detection based on filtered signal might be switched off during the system installation. The same applies to detection based on standard deviation for very disturbed signals. None of these are done by the machine tool operator, and cutting detection is performed automatically without any user tuning or even knowledge. Another preprocessing technique which has to be applied to the sensor signal is segmentation, necessary to extract the signal information when the cutting-tool is actually removing metal in a steady state, since only this part of the signal contains information about the tool wear condition—see Fig. 5. This allows for avoiding random changes of the signals during changes of cutting conditions. Dong et al. [16] calculated signal features from the force samples in one spindle rotation instead of one tooth period, in order to reduce the influence of runout. Similarly, in [17], where tool failure detection in interrupted turning was analyzed, a number of data points taken into consideration contained the measured AE data from at least one full revolution of the workpiece. Reñones et al. [3] developed multitooth tool breakage detection system based on spindle motor electrical power consumption. Every group of inserts was configured to work under different cutting conditions (feed and cutting speed) and to perform roughing or finishing. Thus, to make an accurate diagnosis of

52

K. Jemielniak

the whole tool, it was necessary to analyze the different segments of the measured electrical power consumption by decomposition of a signal into stationary or transient pieces with a length adapted to the local properties of the signal. Jemielniak et al. [18] observed, that despite constant cutting conditions during single micro-milling cut, AE signals were not constant, thus separate signal features were calculated for all cut and for the first and second 1/3 of the cut. Bombi´nski et al. [5] developed the algorithms for automatic selection of short, steady state, representative signal segments. The signals acquired during cutting in the first, training operation were divided into 1 s segments. Having three subsequent signal segments A, B and C local fluctuation of the signal corresponding to the B segment was calculated using effective values of these subsequent segments: R M S[C] R M S[A] (1) − 1 + − 1 Fl B R M S[B] R M S[B] The fluctuation Fl is a measure of the segment usability for tool condition monitoring—the lower the better. For strictly a steady state signal it would be zero. The best segments should be selected from the signal registered during cutting uniformly distributed through the entire operation. Therefore segments are collected in clusters, six segments each and the best of every cluster is selected as its representative. Then numerous signal features must be calculated from every selected signal. If several sensor and advanced signal processing methods like e.g. wavelet transform are applied, number of SFs can be very large (582 in [19]). For long operation, there would be many segments, overloading computer memory and increasing computing time without any added value. Therefore, if operation contains more than 2 min (120 segments in 20 clusters) not more than 20 best segments are selected for further processing. When the number of segments exceeds 128, they are clustered in pairs and the better of the two is selected. This segmentation algorithm allows for selection of the all sensor signal fragments from all operations corresponding to the same moment of the operation duration respectively. This selection is carried out only during the first tool life, while the system training. During this and all following tool lives, all available signal features are calculated from all selected signal segments and only the SFs are kept in computer memory. The original signals are erased, which reduces memory consumption.

3 Signal Feature Extraction Discussed in the introduction uncertain, incomplete or randomly-distorted sensory information about the tool condition offered by any signal feature causes that reliable tool wear evaluation based on one signal feature (SF) is not possible. Therefore, the key issue in a TCM system is calculating a sufficient number of SFs related to tool and/or process conditions. The sensor signal has to be transformed into features that

Tool Condition Monitoring in Metal Cutting

53

could describe the signal adequately, while simultaneously maintaining the relevant information about tool conditions in the extracted features.

3.1 Time Domain Signal Features There are several signal features (SFs) that can be extracted from any time domain signal. The most often used are: • • • • • • • • •

arithmetic mean, average value, effective value (root mean square—RMS) variance (or standard deviation) skewness (the third statistical moment of a distribution) kurtosis (the fourth statistical moment of a distribution) signal power peak-to-peak, range or peak-to-valley amplitude, crest factor (the ratio of the peak level to the rms level) the ratios of the signals, signal increments, especially cutting force components, used to eliminate or to reduce the influence of variable cutting conditions • logarithmic energy. Some features are applied only to vibration and AE signals: • ring down count or pulse rate: number of times AE raw signal crosses the threshold level, • pulse width: the percentage of time during which AE raw remains above the threshold level, • burst rate Number of times AE RMS signal exceeds preset thresholds per second, • burst width—percentage of the time AE RMS signal remains above each threshold, Kannatey-Asibu and Dornfeld [20] assumed that AE RMS signal has β distribution. They showed that the skew and kurtosis are sensitive to both the stick-slip transition to chip contact along the tool rake face and progressive tool wear on the flank of the cutting tool. Jemielniak and Otman applied these parameters to catastrophic tool failure detection [17, 21]. Three main time series modeling techniques are frequently utilized within tool monitoring: auto regressions (AR), moving average (MA) and auto regressive moving average (ARMA) models (see e.g. [4, 16, 22]). The ARMA model usually expresses the dependence of one variable on its own past values, or the effect of some disturbances w(n) on the behavior of subsequent values of the variable. The early research work developed AR models which were of a high order, up to 28th order [4]. These were considered as being of little practical use because of the very high computational load which was inappropriate for online tool monitoring. Thus, the first or first and the second AR, MA or ARMA coefficients were chosen as features [2, 16, 23] sometimes higher (AR coefficients of order 3–5 [22, 24]). Recently Suprock et al. [25] applied 100th order AR model for failure prediction of end-milling. They

54

K. Jemielniak

noticed that while lower-order models may have achieved “adequacy” as defined by statistical terms, higher-order models produce more stable trends. Nevertheless they also remarked that because of computational requirements associated, the 100thorder models exceeded the current real-time computational capabilities of a typical commercially available computer. Principal Component Analysis (PCA), also known as the Karhunen–Loeve transformation, has been widely used in system identification and dimensionality reduction in dynamic systems. Shi and Gindy [26] investigated the PCA technique to extract features from multiple sensory signals. Multiple sensory signals were treated as a high-dimensional multivariate random matrix composed of several vectors formed by the signals. By implementation of PCA, the signals can be reduced to a new, reduced feature vectors. Shi and Gindy used two perpendicular cutting force signals for monitoring of broaching tool wear. The pattern of cutting forces in two-dimensional space of orbit diagram formed as scatter ellipse and had a close relation to the condition of tool wear. This relation was quantitatively evaluated by PCA in the form of signal features: the length (a/b) of the major/minor axes and inclination angle (β) of the ellipse. Moreover, the origin (F y , F z ) of scatter ellipse was related to the average value of cutting force in two perpendicular directions and could be included in the feature set as well. Finally, the normalized elements of feature set were specified as {F y , F z , a, b, β} and was used as the input to the tool wear prediction model. Abellan-Nebot and Subirón [27] who extracted several conventional SFs from cutting force signals (RMS, mean, standard deviation, skew, kurtosis etc.) applied Principal Component Analysis for further reduction of signal feature number. They found a new set of features that were a combination of the original SFs. The Singular Spectrum Analysis (SSA) is a new non-parametric technique of time series analysis incorporating the elements of classical time series analysis, multivariate statistics, multivariate geometry, dynamical systems and signal processing [28]. It decomposes a given time series into the sum of a three independent components: a slowly varying trend representing the signal’s local mean, difference between the original signal and the trend, called detrended signal or oscillatory component, and a structureless noise, presenting no latent structure. Fundamentally the method projects the original time series onto a vector basis obtained from the series itself, following the procedure of principal component analysis (PCA). Salgado and Alonso applied SSA to the analysis of the vibration signals acquired in a turning process in order to extract information correlated with the state of the tool [28]. They decomposed two vibration signals (longitudinal and transverse) into the trend and the detrended signal, and extracted from them several statistical features (mean variance, RMS, range, median, skewness and kurtosis). It appeared that only the RMS and variance of the detrended signals showed a monotonic behavior with the tool wear, which meant that, the information in the vibration signals about flank wear was contained mostly in the high-frequency components. Another relatively new parameter of the complexity measure of time series, applied in tool and process condition monitoring is a permutation entropy [13]. For time series {x t , t 1, …, T } consisting n different signal values, there can be

Tool Condition Monitoring in Metal Cutting

55

Fig. 7 A long feed-motor current signal and their permutation entropy H p during normal cutting and tool breakage [13]

n! permutations π of order patterns. The permutation entropy for the time series is defined as: H p (n) −

n!

p(πi ) ln p(πi )

(2)

i1

where p(πi )—relative frequency of permutation πi . The normalized permutation entropy is then: Hp

H p (n) ln(n!)

(3)

The smaller the value of H p , the more regular is the time series. Li et al. [13] applied permutation entropy as a feature of feed-motor current signals in end milling, to detect tool breakage. Figure 7 shows long-term feed-motor current signals and corresponding normalized permutation entropy. The feed-motor current during normal cutting condition is similar to a regular periodic signal, therefore the H p values were small—from 0.75 to 0.8 between 8th and 38th second of the experiment. That meant that the feature values were insensitive to the noise influences like the different effects of friction coefficients at the different positions. When a flute of the cutter was broken, the regular periodic characteristics of motor current were disturbed and H p value increased.

56

K. Jemielniak

3.2 Frequency and Time-Frequency Domain Signal Features Fast Fourier Transform The determination of signal features in frequency domain is usually based on a discrete, windowed Fourier Transform. Discrete Fourier transform (DFT) maps discretetime sequence of N samples x[n] (n 0, …, N − 1) into discrete-frequency representations X[m] (m 0, …, N − 1) and is given by: X [m]

N −1

x[n]e− j2πnm / N

(4)

n0

Usually a Hanning window is applied to the raw, digital signal before DFT to reduce leakage. Practically one of several commonly known Fast Fourier Transform (FFT) algorithms is used. If the tool wear influenced the frequency contents of the measured signal, FFT gives an inside view of this process. Spectrum X[m] has the same number of elements (half of them symmetrical) as time domain signal x[n]. The use of single Fourier coefficients X[m] is rather not practical due to leakage effects. Therefore, further signal feature extraction is usually applied. Typical features are (see e.g. [1, 10, 29]): • • • •

amplitudes of dominant spectral peaks, power of the signal in specific frequency ranges, sometimes overlapping, energy in frequency bands, statistic features of band power spectrum like mean frequency (the first moment of the power spectrum), variance, skewness, kurtosis of the power spectrum distribution, • frequency of the highest peak of the spectrum, Frequency domain signal features are often used together with time domain features, e.g. [10]. Fourier transform has some important weaknesses. Despite sensor signals registered during machining are essentially non-stationary, FFT averages frequency composition over the duration of the signal. The second weakness is fixed resolution of the entire frequency spectrum which is the inverse of sampling time. To address these issues, the time-frequency analysis, as short-time Fourier transform (STFT) can be applied. The short-time Fourier transform (STFT) uses a window w[n] sliding successively along the time axis to characterize the change of frequency components at different time intervals. Spectral coefficients are calculated for this short length of data, the window is then moved to a new position k and the calculation repeated: X [k, m]

Nw n0

x[k − n]w[n]e− j2πnm / Nw

(5)

Tool Condition Monitoring in Metal Cutting

57

where k is a time shift parameter, w [n] is the analysis window, which is assumed to be non-zero only in the interval [0, N w − 1], and N w is the width of the window. Thus STFT provides the time information by computing a different FTs for consecutive time intervals, and then putting them together. Wavelet Transform Despite Short Time Fourier Transform is a form of joint time-frequency analysis, it has an important drawback: the window width which decides both on time and frequency resolution. Both time and frequency resolutions cannot be arbitrarily high (Heisenberg’s uncertainty principle). Wide analysis window means poor time resolution and good frequency resolution. Once the window is chosen, the resolution is set for both time and frequency. To overcome the preset resolution problem of the STFT the wavelet theory was developed in the late 1980s by Mallat [30] and Daubechies [31]. Wavelet transform (WT) uses windows of different lengths for different frequencies: high frequencies are analyzed with narrower windows for better time resolution, while at low frequencies wider windows are used for better frequency resolution. Therefore WT can extract more information in the time domain at different frequency bands. The wavelet transform decomposes a signal through the wavelet scale function, and scaled and shifted versions of the mother wavelet. Practically it can be reduced to filtering the signal by high-pass and low-pass filters derived from the wavelet and the scaling function. The Discreet Wavelet Transform (DWT) decomposes the signal into the scaling coefficients (approximations A) and the wavelet coefficients (details D) by convolution of the signal and the impulse response of the low-pass and high-pass filters. The filter’s outputs are subsampled by 2. At the first level the original signal is decomposed into A1 and D1 , then approximation A1 can be decomposed again into A2 and D2 . A three level DWT decomposition tree is presented in Fig. 8. Generally approximations (Aj+1 ) and details (Dj+1 ) at level j + 1 can be expressed by convolutions: A j+1 [n] D j+1 [n]

∞ k−∞ ∞

h[2n − k]A j [k]

(6)

g[2n − k]A j [k]

(7)

k−∞

where h and g are impulse responses of low-pass and high-pass filters respectively, which are discrete equivalents to scaling function and wavelet. From a mathematical point of view, the structure of computations in a DWT is exactly an octave-band filter band, thus approximations and details are band pass signals. Another type of wavelet transform is Wavelet Packet Transform (WPT) where both—approximations and details are decomposed, providing much more frequency bands (see Fig. 8b). This provides more opportunities to find useful signal features. On the other hand, for n levels of decomposition, the DWT produces 2n sets of coefficients as opposed to (2n+1 − 2) sets for the WPT. Thus the computational cost of the DWT is much less than the WPT.

58

K. Jemielniak

Fig. 8 Three levels wavelet decomposition tree; a DWT, b WPT; blackened fields indicates the frequency band of the original signal

The wavelet transform has been used in process condition monitoring for more than a decade (e.g. [1, 32, 33]). Kamarthi and Pittner [34] who used force and vibration signals for flank wear estimation in turning, compared performances of fast Fourier transforms (FFTs) and wavelet transforms in this application. They noticed, that unlike in the FFT, short time delays of the signal can cause large changes in wavelet coefficients especially at fine scales. Therefore, they recommended WT for force signals, while the FFT appeared to be better matched to vibration signals. The usual type of the wavelet is selected arbitrary, without any explanation. Occasionally a sentence like: “The coiflet 3 wavelet was chosen for analysis because it yielded the best results after experimentation with a number of different wavelets” [23] can be found. Sometimes, especially in earlier works, the wavelet coefficients were applied directly, e.g. Tansel et al. [32] used them as inputs of neural networks for tool failure detection in milling based on cutting force measurement. As the output of the wavelet transforms has relatively large size, informative features must be extracted from the coefficients. Wavelet coefficients are usually treated as separate signals, and each is characterized by signal features used for time domain signals: average value, RMS, standard deviation or variance, crest factor, peak-to-valley or peak-to-peak value, kurtosis [1]. Wu and Du [35] introduced an automatic feature extraction and assessment procedure using a wavelet packet transform for monitoring of machining processes. They selected the wavelet packets accordingly to their energy, as such packets contained large amounts of information. To identify the effectiveness of the selected features, four criteria such as cross-correlation and cross-coherence of signal and reconstructed signal, correlation of the residue, and power spectrum of the residue, were proposed. Scheffer and Heyns [23] used a similar method, but applied Shannon

Tool Condition Monitoring in Metal Cutting

59

entropy to choose the optimal packets. On the other side, energy of wavelet packets itself can be applied as signal feature. de Jesus et al. [36] used asymmetry of compressed cutting force signals in milling, for catastrophic tool failure detection. During normal milling, with both inserts in good conditions, cutting force signals are alike and the asymmetry is close to zero. When breakage takes place, cutting force signals are different and the waveform is asymmetric. The asymmetry was calculated as the point-to-point variance between detail level-5 DWT coefficients of cutting force signals for each insert in a full revolution of the tool-head: A

4

(Bi+4 − Bi )2

(8)

i1

The wavelet transform was extensively used in tool condition monitoring systems developed in Warsaw University of Technology, e.g. [19, 37]. Hilbert-Huang Transform Another, new method of time-frequency analysis techniques, recently applied to extract the crucial characteristics of the sensor signals for process condition monitoring is Hilbert-Huang Transform (HHT) [38]. It is especially designed for nonstationary and nonlinear signals. Unlike spectrograms, wavelet analysis, or the WignerVille Distribution, the HHT is more like an algorithm (an empirical approach) that can be applied to a data set, rather than a theoretical tool. The HHT consists of two steps: the empirical mode decomposition (EMD) for decomposition a sensor signal into a set of components, called intrinsic mode functions (IMFs), and application of the Hilbert transform to the IMFs. Using the EMD method, any complicated data set can be adaptively decomposed into a set of monocomponent signals (IMFs). This decomposition method operating in the time domain is adaptive and highly efficient. Since the decomposition is based on the local characteristic time scale of the data, it can be applied to nonlinear and nonstationary processes. The Hilbert spectral analysis (HSA) examines the IMF’s instantaneous frequency and generates effective time-frequency distributions called Hilbert spectra. Peng [39] used this method for tool breakage detection based on cutting force signal of the milling process. The tool breakage could be detected directly in the Hilbert spectrum or by means of the energies of the characteristic IMFs associated with characteristic frequencies of the milling process. When tool breakage occurs the energies of the associated characteristic IMFs change in opposite directions, which is different from the effect of changes of the cutting conditions e.g. the depth of cut and spindle speed. Thus, they were not only able to effectively capture the significant information reflecting the tool condition, but also reduced the sensitivity to the effect of various uncertainties.

60

K. Jemielniak

4 Signal Feature Selection As it is really not possible to predict which signal features will be useful in a particular case, as many as possible should be extracted from the available signals. Then those that are informative, correlated with tool wear, should be selected for tool condition monitoring. There can be many different descriptors from different sensor signals, most of which are hardly related to the monitoring process. Therefore, effective feature selection procedure is necessary. Relevant features are then used for tool or process condition diagnosis. A number of possible features originating from one or more signals, can be very large, however, most of them are very distorted or hardy sensitive to the process or tool condition. Decision making algorithm of the P/TCM system should not be overburdened with those features and avoid unnecessary computation or hinder the subsequent classification process. The selected features should be relevant, sensitive to the tool or process condition. However, every SF, even well correlated, e.g. with the tool wear, can be sometimes disturbed randomly. Therefore the number of features should be high enough to cover these possible random disturbances of any SF—the robustness of the monitoring system can be improved by a certain amount of redundant input information. On the other hand, especially in neural networks based systems, the more features, the more training samples are needed. If the system is supposed to work (monitor the tool wear) already after the first, training tool life, amount of training samples may be not big enough to properly train a big network necessary for a large number of inputs (signal features) [1]. Thus the second objective of signal processing is to preserve as much of the relevant information as possible by removing redundant or irrelevant signal features. In final, industrial application, the feature selection process should be performed automatically, without intervention of an operator. Sick [29] presented an interesting classification of feature selection processes. In 38% out of 138 analyzed publications, features were selected without any reason (or based on literature review), in 26% signal features were defined after analysis of measured signals, in 21% the most appropriate of these features were selected without considering the behavior of the subsequent wear model. Only in 15% of publications the optimal set of features was found after the analysis of the influence of different features on the estimation of tool wear. Quan et al. [40] applied Pearson correlation coefficient r to find those features that can best characterize tool-wear conditions. The correlation coefficient between a selected feature x and a tool-wear value y is can be expressed as follows: 2 ¯ i − y¯ ) (xi − x)(y r ¯ 2 i (yi − y¯ )2 i (x i − x)

2

i

(9)

where x¯ and y¯ are the means of x and y, respectively. The correlation coefficient r is a measure of the strength of linear dependence between x and y. Also Scheffer and Heyns [10] used this coefficient for signal feature selection, assuming that lower the value of r, the lesser the chance for the selected feature to show any trend towards

Tool Condition Monitoring in Metal Cutting

61

tool wear. They ignored the fact, that even if SF is perfectly correlated with the tool wear, but the correlation is not linear, the correlation coefficient is lower than 1. Scheffer and Heyns [10] noticed that automated feature selection methods often select features that are too similar or dependent on one another, and therefore do not achieve the goal of proper sensor fusion. In such cases, they recommended few rules based on “engineering judgment” which means resignation from automatic feature selection and manual intervention of the scientist. Such procedure is hardly acceptable in factory floor condition, making a TCM system purely laboratory. Nevertheless, they pointed important issue. Although every machine tool operator knows what the cutting force or the motor power is, in practice, he/she does not use these quantities and has no intuitive sense of their values and changeability. As far as the tool condition is concerned, the natural categories would rather include “sharp”, “partially worn” or “worn out” (failed). Such tool wear measures as VB or KT are seldom used in factory floor conditions. Therefore, in Warsaw University of Technology (WUT) the concept of the used up portion of tool life (T ), defined as the ratio of the cutting time as performed so far to the overall tool life span: T t/T was introduced [19, 41, 37, 42, 43]. Each signal feature can be correlated with T , using e.g. two-degree polynomial approximation and the root mean square error of such correlation can be a measure of the feature applicability to tool wear monitoring. Comprehensive methodology of signal feature selection was developed in WUT [19, 37, 43]. It consists of three stages: • selection of meaningful SFs • elimination of redundant SFs, similar to another, • selection of repeatable SFs. First, the signal feature usability for tool condition monitoring is evaluated using the coefficient of determination R2s . It is a statistical measure of how well any SF-tool wear model approximates the real data points or—in other words—how much this model is better than just average value of SF. Low-pass filtered signal feature SF f was accepted as SF(T ) model, which allowed avoiding any uncertain suppositions about the mathematical formula of this model. Rs2

i

2 (S Fi − S Fav )2 − i S Fi − S F f i 2 i (S Fi − S Fav )

(10)

where: (S Fi − S Fav )2 total square sum, 2 i residual square sum, i S Fi − S F f i is a single value of SF and SF f respectively, normalized in time SF i and SF fi (0–100% of T , i 0.0.100), SF av is the average value of SF T . These SFs, for which average value of R2s > 0.4 can be assumed as satisfactory correlated with the tool condition, thus usable for TCM—see examples in Fig. 9 [37].

62

K. Jemielniak

Fig. 9 Examples of signal features usability evaluation, a useful SF, b useless SF [37]

On the other hand, selected SFs should not be strongly correlated one with each other to avoid multiplication of the same information. Therefore, these SFs which meet the criterion, are sorted into descending order, according to the R2s values. Then the first (best) is selected and correlation coefficients r 2 between this SF and every other are calculated. SFs with r 2 > 0.8 are rejected as too much correlated with the best one. From among the remaining signal features, again the best one is selected, and the SFs correlated with it are rejected. The procedure is repeated until no signal feature meeting the R2s > 0.4 criterion remains. After completion of the third tool life, tool feature selection is repeated, using all available data, thus R2s coefficients are calculated for three tool lives and averaged. Now application of second, even more important SF usability criterion can be applied: repeatability. It is evaluated using another determination coefficient R2r : Rr2

j

i

2 2 S F f ji − S F2 f av − j i S F f ji − S F f avi 2 j i S F f ji − S F3 f av

(11)

where SF fji is the value of SF f in i-th point (i 0.0.100) and j-th tool life (j 1.0.2), is average of SF f in i-th point S F f avi 13 j S Fi j 1 S F3 f av 303 S F is average of all SF f values in three tool lives. i j j i These SFs, for which R2r > 0.6 are assumed as repeatable enough. All SFs meeting both criteria are sorted according to the R2r values. Elimination of SFs correlated one to each other is based on three tool lives data. Examples of repeatable and not repeatable SFs are presented in Fig. 10. The selected features are subject of the decision making algorithm.

5 Decision Making Algorithms The monitoring strategies used in TCM systems make use of monotonous increment of some signal feature accompanying the tool wear progress. The most often used

Tool Condition Monitoring in Metal Cutting

63

Fig. 10 Signal feature repeatability evaluation, a accepted SF, b rejected SF [37]

signal feature is the average value of the signal, but also maximum signal value and other signal features also can be used. Figure 11 presents a typical strategy of tool wear monitoring. Learning of the system consist in machining the first workpiece with a new, sharp tool. The obtained value of the signal feature SF0 (here it is the area under the signal vs. time curve) is automatically normalized, that is regarded as 100%. Then the threshold level of the signal feature (SF T ), or the so called wear limit, is calculated as an admissible increase in the signal feature value in percentage terms: d S FT (12) S FT S F0 1 + 100% where dSF T —limit factor, admissible relative increase in the signal feature value: d S FT

S FT − S F0 · 100% S F0

(13)

While monitoring, after each cycle, the system displays the value of the selected signal feature in a digital or graphic form (Fig. 12). When the feature reaches the threshold level, the tool life is assumed to have come to its end (tool failure). Some-

Fig. 11 Tool wear monitoring strategy used in most of TCM systems [5]. 1: area signal learned, 2: larger area signal (e.g. through worn tools), 3: learnt area (bar diagram), 4: pre-alarm limit, wear limit, 6: pre-alarm, 7: wear alarm

64

K. Jemielniak

Fig. 12 Examples of information on signal feature values presented to the user [1]

times, two limits can be set—a warning limit (e.g. 130%) and the proper one (e.g. 150%). The admissible increase in the dSF T measure in percentage terms can be preset by the manufacturer, but its final and conclusive value must be determined by the operator. System tuning consists in correcting the dSF T value. The operator has to make some additional computations according to formula (12) and (13), which is a rather inconvenient and unclear procedure requiring some expertise. There are no reasons why the operator should not be relieved of dealing with signal values by making the system tuning easier, simplifying the communication between the system and the operator. Necessary condition for signal feature applicability for tool condition monitoring is a correlation of the feature with the tool wear. Commercial systems make use of only those measures which are positive, monotonous and increasing. This eliminates a lot of signal features which do not fulfill these conditions, although they are quite well correlated with the tool wear. Strategy of TCM based on not necessarily positive, rising and monotonic signal features was presented in [42] and verified in many applications and installations [19, 41, 37, 43]. Tool condition is estimated there on the base on signal feature—used-up portion of tool live model in the form of function SF f (T ), which is an array of 101 elements (0–100% of T ) − SF f [T ]. After subsequent signal measurements the system calculates signal feature value SFn], where n is number of measurements (data acquisitions). Then SF f [T ] array created after preceding tool lives are searched for the values closest to SF[n]—Fig. 13a. It may happen that after the next measurement the SF value corresponds to a value of the used-up portion of the tool life lower than that reached in the previous operation (Fig. 13b). Such a system indication might be disorienting for the operator. Therefore the search starts from the T value obtained after previous measurement, which means that the used-up portion of the tool life presented to the operator cannot decrease. Sometimes it happens that the SF value affected by some disturbances corresponds to a very large increase in the tool wear. To remedy such a mistake the search for the SF value is limited to 30 elements of the SF f [T ] array created after preceding tool lives, i.e. to 30% of the tool life (see Fig. 13c). This means that, in the case of accelerated tool wear, the system allows three operations to be performed before it signals tool failure. This procedure also

Tool Condition Monitoring in Metal Cutting

65

Fig. 13 Used up portion of tool life evaluation based on single signal feature. a Search out in array SF f [T ] for the value closest to SF[n] obtained after last signal measurement; b the search starts from previous result, thus T value cannot diminish; c the search limited to 30 elements of the SF f [T ] array reduces influence of accidental high values; d and enables to utilize non-monotonic the signal features [19, 42]

has another purpose, namely it enables signal features which are not monotonic with respect to the used-up portion of the tool life to be utilized, at least to some extent, as presented in Fig. 13d. In the example shown here, the signal feature value corresponds to T 63% and T 95%. Restriction of the array search to 30% of T results indicates that T 63%. The combination of different signal features (SFs) today is ever increasing to overcome drawbacks of a single sensor approach. The information extracted from one or several sensors’ signals has to be combined into one tool condition estimation. The signal feature integration minimizes the diagnosis uncertainty, reducing the randomness in one SF and providing more reliable tool condition estimation. It can be achieved by various means, such as statistical methods, auto-regressive modeling, pattern recognition, expert systems and others [1]. Artificial intelligence (AI)

66

K. Jemielniak

plays a key role in the development of modern tool condition monitoring (TCM) systems [1, 6]. Generally, training of artificial intelligence system requires test cuts to be made, and thus is only viable for series production situations [44]. The most frequently chosen methods are neural network (NN) [3, 45], Mamdani fuzzy logic (FL) [46], takagi-sugeno-kang (TSK) FL [18] or a combination of FL and an automatic generating method, i.e., genetic algorithm (GA) [47]. All these methods have a similar objective—matching the estimate of average cutting tool wear with the directly measured wear value. In [6] the use of three artificial intelligence methods for tool condition monitoring: FF-BP neural network (FFBPNN), a fuzzy decision support system (FDSS) and an artificial neural network based fuzzy inference system (ANNBFIS) were compared. The input data were feed f , feed force F f and cutting force F c . Not only the accuracy of the tool wear prediction, but also on practical usability of the presented methods was discussed there. An important aspect of this usability is the dependence of obtained results on AI system parameters set by the operator. All these three artificial intelligence methods gave similar, acceptable results. Important differences in the “internal” structure of these systems would be unknown and thus irrelevant for the operator. A major difference in their usage, however, is a critical factor. Construction of a knowledge base for the fuzzy logic system necessitates skill and expert knowledge. The operator has to analyze the dependence of cutting forces on tool wear, which means that the results of preliminary experiments have to be presented to the operator in the convenient and transparent form. This makes fuzzy logic rather difficult for practical implementation. The number of neurons in the hidden layer of the neural network and the number of iterations can be selected arbitrarily as they have very little influence on system performance. Results of preliminary (learning) tests do not have to be presented to the operator in this case, instead they are simply fed to the system input. The disadvantage of the method is a considerably long training time, which makes it inconvenient in practice. Similarly for the neuro-fuzzy system, the structure (number of rules) and the number of iterations do not have an important influence on system performance and the operator does not have to know the results of preliminary tests. The most important difference between these last two methods is learning time; it is so short for the neuro-fuzzy system that it can be easily optimized and implemented on the factory floor. Most of the papers (including mentioned above) describing application of AI method for signal feature integration in TCM monitoring use scanty number of SFs, usually less than ten. The signal features are selected by scientist on the base of previous experience or visual analysis of the SF dependence on the tool condition. However, as was shown above, it is impossible to decide in advance which SFs might be useful in a particular case, so a large number of them should be calculated and useful ones automatically selected for the integration. On the other hand, the use of many SFs in a single AI structure requires extensive experimental data. They are not available if the TCM system is supposed to be trained during the first tool life and be ready to monitor the tool during the next ones. A different approach was presented by Kuo and Cohen [48], who proposed a TCM employing several measures of the cutting force, acoustic emission and vibration. Proposed system consists of two modules where the first one estimates the tool wear from all SFs taken from

Tool Condition Monitoring in Metal Cutting

67

Fig. 14 Hierarchical tool condition estimation [19, 41]

one sensor and the cutting parameters. A single radial basis function artificial neural network was used here. The results obtained in the first module were integrated into the final system’s response in the second module, in which a fuzzy neural network was used. Jemielniak and Bombi´nski [41] presented a comparison of efficiency of tool wear monitoring strategies based on one signal feature, on a single neural network with several input signals, and on a hierarchical algorithm and a large number of signal features. In the first stage of the hierarchical algorithms, the tool wear was estimated separately for each signal feature [42]—see Fig. 13. In the second stage, the results obtained in the first one, were integrated into the final tool wear evaluation (Fig. 14). All T estimations are averaged and displayed as the final tool condition evaluation. This value is used as the initial value T B in the next iteration of the algorithm (after next measurement). The integration was carried out by the use of either a neural network or averaging. They showed, that decomposition of the multi signal feature tool wear estimation into hierarchical algorithms had an important advantage over the single step approach, meaning conventional industrial solutions (single signal feature) and typical labo-

68

K. Jemielniak

ratory solutions (single, large neural network) as many more SFs could be used. Determination of single feature dependence on tool condition is simple and the dependence is easy to reverse [42]. On the other hand the direct determination of tool condition dependence on the signal feature values needs numerous training data and long learning time. This means, that the TCM system necessitates many tool lives to be trained. Unfortunately, the accuracy of laboratory TCM systems is usually tested based on experimental data collected at several wear levels and at cutting conditions different from those used in training, but during the same tool life. This makes obtaining good prediction results relatively easy, but it is far from factory-floor practice. Generally, the reliability and user friendliness are the most important concerns of those who actually are using some form of TCM [1, 44]. Most laboratory systems presented in the literature are “manually” tuned and cannot work without the author. Thus, it is obviously vital to minimize the complexity of any future TCM system so that it can be employed on many different machines for many different applications and can be used by a machine tool operator without any knowledge of the complex strategy involved. Any threshold values determination, signal feature selection and as well as their integration, should be performed by the system without any operator intervention, who should only point the end of the first, training tool life.

6 Case Study In [19] a tool wear monitoring strategy developed at Warsaw University of Technology, based on a large number of signal features in rough turning of Inconel 625 was presented. The workpieces were impeller cases made of Inconel 625 (Fig. 15a), and machined with subsequent perpendicular cuts from diameter 406 to 268, with the depth of the cut ap 2.5 mm, feed f 0.2 mm/rev, and cutting speed vc 220 m/min. The tool was a CRSNL with whisker-reinforced round ceramic inserts, RNGN CC670. Tool life was limited by three phenomena: tool notch wear (Fig. 15b), burr formation (Fig. 15c) and drastic decrease in the surface finish (Fig. 15c). All three phenomena appeared autonomously, making the determination of the tool life end difficult, subjective, and dependant on the machine tool operator’s experience. Three workpieces were machined, during which seven tools were worn out. Higher number of used tools (tool lives) then machined workpieces is characteristic for machining of big, aerospace parts. The experiments were performed on a turning center TKX 50 N equipped with an industrial AE sensor (Kistler 8152B121) and accelerometer (PCB PIEZOTRONICS 356A16) mounted on the turret and cutting force sensor (Kistler 9017B) mounted under the turret. A raw AE (AE raw ) signal was acquired with a sampling frequency of 2 MHz using a DAQ card, NI PCI 6111. As this sampling frequency produces an enormously large amount of data, only 0.05 s (100,000 samples) out of every 10 s period was recorded and analyzed. The demodulated amplitude of AE signal (AE RMS ), two cutting force signals (F x and F z ) and two vibration signals (V y and V z )

Tool Condition Monitoring in Metal Cutting

69

Fig. 15 Workpiece and tool (a), and tool life criteria: tool wear (b), burrs (c), surface roughness (d) [19]

were acquired simultaneously with a sampling frequency 30 kHz using NI-PCI 6221 DAQ card at the same points of time during 1.66 s (5000 samples each). Each cut lasted 96 s, during which eight such recordings of the signals were taken and treated as separate, subsequent measurements, used for tool wear monitoring. Then, from each of 6 signals the five time domain SFs features were extracted: effective value (e.g. F x,RMS ), standard deviation (e.g. F x,StDev ), skewness (e.g. F x,Skew ), kurtosis (e.g. F x,Kurt ), crest factor (the ratio of the peak level to the rms level, e.g. F x,Crest ). Fast Fourier Transform was applied to obtain eight frequency domain features: dominant frequency (e.g. F x,DF ), power in dominant band (e.g. F x,PDB ), power in 6 selected bands (e.g. F x,P250-500 ). Finally, three level Wavelet Packet Transform (WPT) decomposition was used to obtain fourteen coefficients, called approximations A and details D, which are band bass signals (see Fig. 8). From each of these coefficients six timefrequency domain features were calculated: logarithmic energy (e.g. F x,ADA,E is the energy of the wavelet coefficient ADA of signal F x ), skewness (e.g. F x,ADA,Skew ), kurtosis (e.g. F x,ADA,Kurt ), effective value (e.g. F x,ADA,RMS ), threshold crossing rate (number of times the signal crosses the threshold level, e.g. F x,ADA,Count ) and pulse width (the percentage of time during which the signal remains above this threshold e.g. F x,ADA,Pulse ), so there were 84 wavelet based SFs calculated from each signal. Altogether there were 582 signal features calculated automatically, (97 from each of six available signals, 194 SFs from each sensor). Signals originating from each sensor were treated separately. While the number of extracted signal features is very large, some of them are very distorted, hardly dependant on tool wear (e.g. Fig. 16a), others are dependant mainly on the tool position on the workpiece (e.g. Fig. 16b). There are, however SFs which are dependant on the tool condition (e.g. Fig. 16c) even if some depend also on the

70

K. Jemielniak

Fig. 16 Examples of signal features calculated from available signals during all 7 tool lives; skew of AEraw signal (a), kurtosis of Fz signal (b), energy of WPT coefficient ADD of Vy signal (c), and standard deviation of Fx signal (d) [19]

Fig. 17 Examples of evaluation of signal features’ repeatability based on data from three tool lives; SF which met the criterion (a), SF which was rejected as not repeatable, despite being correlated with tool wear (b) [19]

tool position (e.g. Fig. 16d). Only those SFs should be selected, which are relevant and sensitive to tool conditions, using a selection of meaningful SFs procedure based on the coefficient of determination R2s > 0.4 (Eq. 10). To avoid multiplication of the same information, these SFs which meet the first usability criterion, and were correlated with the best SF (correlation coefficients r 2 > 0.8) were rejected. The procedure was repeated until no signal feature meeting the R2s > 0.4 criterion remains. After the end of the second and the third tool life, again feature selection is repeated, using all available data, thus R2s coefficients were calculated for tool lives available and averaged. Then repeatability criterion was applied using another determination coefficient R2r (see Eq. 11). Elimination of SFs correlated one to each other was carried out again, basing on the available tool lives data. Figure 17 presents examples of signal features accepted by both criteria and recognized as correlated with the tool wear but not repeatable thus rejected. Signal features selected this way were used for tool wear monitoring in all subsequent tool lives, beginning from the fourth one. Thus, signal features selected after each of three tool lives were different. E.g. the signal measures selected from

Tool Condition Monitoring in Metal Cutting

71

Fig. 18 Number of useful SFs selected from each sensor signal, before and after elimination of similar SFs [19]

accelerometer, sorted accordingly to R2s after 1st tool life and accordingly to R2r after the 2nd and 3rd tool life were: • after 1st tool life: V z,ADA,E , V y,RMS , V y,Crest , V z,D,Skew , V z,DDA,Kurt , V z,AD,Pulse , V y,P250-500 , V z,AAD,Kurt , V z,Kurt , V y,A,RMS , V y,AD,Kurt , V z,AD,Kurt , V z,DAA,Kurt , V z,DDA,Count , • after two tool lives: V y,Crest , V z,StDev , V y,D,Pulse , V z,AAD,RMS , V y,P4000-8000 , V y,AAD,Kurt , V y,DA,E , V y,AAA,Kurt , V y,AD,RMS , V y,A,RMS , V y,AD,Count , V z,P250-500 , V z,P500-1000 , V z,AAD,Kurt , • after three tool lives: V y,D,Pulse , V y,AAD,RMS , V y,P4000-8000 , V y,AAA,Kurt , V z,StDev , V y,A,RMS . Figure 18 presents the numbers of SFs selected after the first tool life and the three tool lives, for each sensor. The latter are also shown for each signal. As can be seen, the force sensor produced the highest number of useful, relevant SFs, meaning those that were well-correlated with the tool condition, repeatable, and not similar to each other. The data presented in Fig. 18 show that the cutting force component, F x (perpendicular to the cutting speed vector), is a source of more potentially useful SFs than F z (parallel to the cutting speed vector), and the vibration in the y direction is also more informative than that parallel to the cutting speed. The AE sensor produced the smallest number of useful signal features from both the low and high frequency signals (AE RMS and AE raw , respectively). Tool condition was first estimated on the base on each useful signal feature separately (see Fig. 13) and then integrated in the next step of the algorithm as average value (Fig. 14). Results of tool condition monitoring obtained for each sensor used separately and for all sensors used together are presented in Fig. 19 as used-up portion of tool lives evaluated by the system T ev versus actual values of T . As the first tool life was used only for system training, results of six following tool lives are presented there. The second tool (dashed line with circles) was monitored only on data gathered during the first tool life, the third tool (dashed line with triangles) was monitored on data from two first tool lives, while tools 4–7 (solid lines) were monitored using

72

K. Jemielniak

Fig. 19 Used-up portion of tool life evaluated by the TCM system (T ev ) versus actual portion (T ) after training on selected signals [19]

data from the tool lives 1–3. Accuracy of tool condition monitoring evaluation can be assessed by using root mean square error (RMSE): 1 RMSE (14) (Tev − T )2 n i The T values are expressed in percent, thus RMSE can be interpreted as average percentage errors (also presented in Fig. 19). Results achieved by vibration sensor are relatively good (Fig. 19b, RMSE 13.4), but much worse than obtained using the force sensor, which could be expected from lower number of useful signal features (see Fig. 18). Not much worse were results based on AE signals (Fig. 19c, RMSE 14.4) which produced similar number of good SFs. All signals used together (Fig. 19d) produced results little worse than the force sensor alone, which means, that poorly repeatable AE signal features had negative influence on this result. The presented case study proved the effectiveness of the methodology even under very difficult cutting conditions, where the number of tool lives is less than the number of machined parts. It was also implemented in several other applications like [5, 41], including micromachining [43]. Recently it was applied in the aerospace industry in factory floor conditions [49].

Tool Condition Monitoring in Metal Cutting

73

References 1. Teti, R., Jemielniak, K., O’Donnell, G., Dornfeld, D.: Advanced monitoring of machining operations. CIRP Ann.—Manuf. Technol. 59, 717–739 (2010) 2. Li, X.: A brief review—acoustic emission method for tool wear monitoring during turning. Int. J. Mach. Tools Manufact. 42, 157–165 (2002) 3. Jemielniak, K., Kwiatkowski, L., Wrzosek, P.: Diagnosis of tool wear based on cutting forces and acoustic emission measurements as inputs to a neural network. J. Intell. Manuf. 9, 447–455 (1998) 4. Prickett, P.W., Johns, C.: An overview of approaches to end milling tool monitoring Int. J. Mach. Tools Manuf. 39, 105–122 (1999) 5. Bombi´nski, S., Bła˙zejak, K., Nejman, M., Jemielniak, K.: Sensor signal segmentation for tool condition monitoring. Procedia CIRP 46, 155–160 (2016) 6. Balazinski, M., Czogala, E., Jemielniak, K., Leski, J.: Tool condition monitoring using artificial intelligence methods. Eng. Appl. Artif. Int. 15, 73–80 (2002) 7. Jemielniak, K.: Some aspects of acoustic emission signal pre-processing. J. Mater. Process. Technol. 109, 242–247 (2001) 8. Jemielniak, K.: Some aspects of AE application in tool condition monitoring. Ultrasonics 38, 604–608 (2000) 9. Altintas, Y., Park, S.S.: Dynamic compensation of spindle-integrated force sensors. Ann. CIRP 53(1), 305–308 (2004) 10. Scheffer, C., Heyns, P.C.: An industrial tool wear monitoring system for interrupted turning. Mech. Syst. Signal Process. 18, 1219–1242 (2004) 11. Jemielniak, K.: Detection of cutting edge breakage in turning. Ann. CIRP 41(1), 97–100 (1992) 12. Ghosh, N., et al.: Estimation of tool wear during CNC milling using neural network-based sensor fusion. Mech. Syst. Signal Process. 21, 466–479 (2007) 13. Li, X., et al.: Complexity measure of motor current signals for tool flute breakage detection in end milling Int. J. Mach. Tools Manufact. 48, 371–379 (2008) 14. Nordmann International GmbH, NORDMANN Tool Monitoring. http://www.toolmonitoring. com/pdf/Nordmann-Praesentation.pdf. Accessed 28 Nov 2015 15. ARTIS GmbH. Artis Marposs. http://www.artis.de/en/. Accessed 28 Nov 2015 16. Dong, J., et al.: Bayesian-inference-based neural networks for tool wear estimation. Int. J. Adv. Manuf. Technol. 30, 797–807 (2006) 17. Jemielniak, K., Otman, O.: Catastrophic tool failure detection based on AE signal analysis. Ann. CIRP 47(1), 31–34 (1998) 18. Ren, Q., Baron, L., Balazinski, M., Jemielniak, K.: TSK fuzzy modeling for tool wear condition in turning processes: an experimental study. Eng. Appl. Artif. Int. 24(2), 260–265 (2011) 19. Jemielniak, K., Urba´nski, T., Kossakowska, J., Bombi´nski, S.: Tool condition monitoring based on numerous signal features. Int. J. Adv. Manuf. Technol. 59(1–4), 73–81 (2012) 20. Kannatey-Asibu, E., Dornfeld, D.: A study of tool wear using statistical analysis of metal cutting acoustic emission. Wear 76, 247–261 (1982) 21. Jemielniak, K., Otman, O.: Tool failure detection based on analysis of acoustic emission signals. J. Mater. Process. Technol. 76, 192–197 (1998) 22. Ravindra, H., Srinivasa, Y., Krishnamurthy, R.: Acoustic emission for tool condition monitoring in metal cutting. Wear 212(1), 78–84 (1997) 23. Scheffer, C.: Heyns, P.C Wear monitoring in turning operations using vibration and strain measurements. Mech. Syst. Signal Process. 15(6), 1185–1202 (2001) 24. Song, D.Y.: A new approach to cutting state monitoring in end-mill machining. Int. J. Mach. Tools Manufact., 45, 909–921 (2005) 25. Suprock, C.A., Piazza, J.J., Roth, J.T.: Directionally independent failure prediction of endmilling tools during pocketing maneuvers. J. Manuf. Sci. Eng. 129(4), 770–779 (2007) 26. Shi, D., Gindy, N.N.: Tool wear predictive model based on least squares support vector machines. Mech. Syst. Signal. Pr. 21, 1799–1814 (2007)

74

K. Jemielniak

27. Abellan-Nebot, J.V., & Subirón, F.R.: A review of machining monitoring systems based on artificial intelligence process models. Int. J. Adv. Manuf. Technol. 47 (1–4), 237–257 (2010) 28. Salgado, D.R., Alonso, F.J.: Tool wear detection in turning operations using singular spectrum analysis. J. Mater. Process. Technol. 171, 451–458 (2006) 29. Sick, B.: On-line and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research. Mech. Syst. Signal Process. 16(4), 487–546 (2002) 30. Mallat, S.G.: A theory of multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Mach. Intell. 11(7), 674–693 (1989) 31. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Trans. Inf. Theory 36, 961–1005 (1990) 32. Tansel, I.N., et al.: Detection of tool failure in end milling with wavelet transformations and neural networks. Int. J. Mach. Tools Manuf. 35(8), 1137–1147 (1995) 33. Hong, G.S., Rahman, M., Zhou, Q.: Using neural network for tool condition monitoring based on wavelet decomposition. Int. J. Mach. Tools Manuf. 36, 551–566 (1996) 34. Kamarthi, S.V., Pittner, S.: Fourier and wavelet transform for flank wear estimation—a comparison. Mech. Syst. Signal Process. 11(6), 791–809 (1997) 35. Wu, Y., Du, R.: Feature extraction and assessment using wavelet packets for monitoring of machining processes. Mech. Syst. Signal Process. 10(1), 29–53 (1996) 36. de Jesus, R.-T.R., et al.: FPGA based on-line tool breakage detection system for CNC milling machines. Mechatronics 14, 439–454 (2004) 37. Jemielniak, K., Kossakowska, J., Urba´nski, T.: Application of wavelet transform of acoustic emission and cutting force signals for tool condition, monitoring in rough turning of Inconel 625. Proc. IMechE Part B: J. Eng. Manuf. 225, 1 (2011) 38. Huang, N., Samuel, S. (eds.): Hilbert-Huang Transform and its Application. World Scientific Publishing (2005) 39. Peng, Y.: Empirical model decomposition based time-frequency analysis for the effective detection of tool breakage. J. Manuf. Sci. Eng. 128(1), 154–166 (2006) 40. Quan, Y., et al.: On-line robust identification of tool-wear via multi-sensor neural-network fusion. Eng. Appl. Artif. Intell. 11, 717–722 (1998) 41. Jemielniak, K., Bombi´nski, S.: Hierarchical strategies in tool wear monitoring. J. Eng. Manuf. 223(B3), 375–382 (2006) 42. Jemielniak, K.: Tool wear monitoring based on a non-monotonic signal feature. J. Eng. Manuf. 220(B2), 163–170 (2006) 43. Jemielniak, K., Bombinski, S., Aristimuno, P.X.: Tool condition monitoring in micromilling based on hierarchical integration of signal measures. CIRP Ann.—Manuf. Technol. 57, 121–124 (2008) 44. Rehorn, A.G., Jiang, J., Orban, P.E.: State-of-the-art methods and results in tool condition monitoring: a review. J. Adv. Manuf. Technol. 26, 693–710 (2005) 45. Das, S., Bandyopadhyay, P.P., Chattopadhyay, A.B.: Neural-networks-based tool wear monitoring in turning medium carbon steel using a coated carbide tool. J. Mater. Process. Technol. 63, 187–192 (1997) 46. Baron, L., Archiche, S., Balazinski, M.: Fuzzy decisions system knowledge base generation using a genetic algorithm. Int. J. Appl. Reason. 28(2–3), 125–148 (2001) 47. Achiche, S., Balazinski, M., Baron, L., Jemielniak, K.: Tool wear monitoring using geneticallygenerated fuzzy knowledge bases. Eng. Appl. Artif. Int. 15(3–4), 303–314 (2002) 48. Kuo, R.J., Cohen, P.H.: Multi-sensor integration for on-line tool wear estimation through radial basis function networks and fuzzy neural network. Neural Netw. 12(2), 355–370 (1999) 49. Project “Advanced techniques of manufacturing of aircraft transmission—INNOGEAR” No. INNOLOT/I/10/NCBR/2014 (2013–2016), co-financed by the European Regional Development Fund under the Operational Programme Innovative Economy

Assessment of Selected Tools Used for Knowledge Extraction in Industrial Manufacturing Marcin Perzyk and Artur Soroczynski

Abstract General requirements for knowledge representation in the form of logic rules, applicable to design and control of industrial processes, are formulated. Characteristic behavior of Decision Trees and Rough Sets Theory in rules extraction from recorded data is discussed and illustrated. The significance of the models’ drawbacks was evaluated, using simulated and industrial data sets. It is concluded that performance of Decision Trees may be considerably poorer in several important aspects, compared to Rough Sets Theory, particularly when not only a characterization of a problem is required, but also detailed and precise rules are needed, according to actual, specific problems to be solved.

1 Introduction Large amounts of data related to designs, manufacturing processes, materials and equipment, collected in most of manufacturing enterprises, can be potentially used for improvement of the quality and economics of production. Since the year 2000 a vast growth of various applications, aimed at supporting and improvement of manufacturing processes, has been observed. A comprehensive and insightful characterization of the problems in manufacturing enterprises as well as the potential benefits from application of Data Mining (DM) in this area is presented in [1] whereas numerous examples and general characteristics of problems related to the usage of DM techniques and systems in manufacturing environment can be found in [2–14]. DM techniques can provide various types of information. Most frequently, methods of automated knowledge extraction from the recorded past data in the form of logic rules of the type: ‘IF (conditions) THEN (decision class)’ are utilized. Also another types of information may be important for industrial applications, such as prediction of continuous-type output (usually process results) based on process inputs M. Perzyk (B) · A. Soroczynski Faculty of Production Engineering, Warsaw University of Technology, Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_5

75

76

M. Perzyk and A. Soroczynski

as well as relative significances of input variables (usually process parameters) [15, 16]. Generally, for extraction of logic rules from data any classification system or model can be used. Typical learning algorithms include direct rule induction, Decision Trees (DT), naïve Bayesian classifier and algorithms based on the Rough Sets Theory (RST). Detailed information on these methods can be found in [17] and the literature quoted there. Artificial neural networks have also been successfully utilized for logic rules extraction [18–21], often involving fuzzy numbers. This approach facilitates processing continuous-valued variables, handling uncertainties appearing in data and usage of linguistic variables. For manufacturing problems DTs are probably the most frequently used tools for rules extraction from data (e.g. [1, 4, 6, 22–24]), whereas the RST-based methods seem to be their newer alternative (e.g. [10, 13, 17, 25]). Both algorithms are relative simple, especially compared to neural or fuzzy-neural systems, and easy to interpret by users. Both of them treat the data in a natural way however, they are based on completely different principles and algorithms. Practical aspects of application of those tools are also different. The computation times of DT are generally short and the interpretation of rules obtained from DT can be facilitated by the graphical representation of the trees. The RST may require long computational times and may lead to much larger number of rules, compared to DT, if one seeks a detailed information from the knowledge system. It should be noticed, that whereas DTs are widely spread both in handbooks and in commercially available DM software, the RST can be rather seldom found, except for scientific literature. Making a right choice of the rules extraction algorithm is important, particularly in construction of DM systems. However, there are very little comparative studies available, which could show the advantages and weakness of individual tools [13, 17]. In this chapter important differences in performances of the two algorithms mentioned above, i.e. DT-based and RST-based, chiefly from the standpoint of supporting design and control of manufacturing processes, are demonstrated and discussed. The work presented in this chapter comprises three main parts. Firstly, the expectations and requirements for the knowledge rules systems which could be successively used in design and control of industrial manufacturing processes are formulated and characterized. Then the characteristic behavior of DT-type models is presented and compared with RST, using simple demonstration data sets. Finally, the significance of drawbacks of the DT models as rules extraction system is evaluated, using several simulated and real data sets.

Assessment of Selected Tools Used for Knowledge Extraction …

77

2 Requirements for Knowledge Rules Applicable for Industrial Processes General requirements for knowledge rules which could be useful in manufacturing industry are rather obvious and similar to those for other areas of applications. Firstly, the rules should be reliable, which means that there is a real chance that application of a rule will bring the predicted result. This can be expressed by rules quality parameters, of which the most important are confidence (also known as consistency or precision) and support (also known as strength). A rule can be expressed as: if C then D

(1)

where C is the conditional part (input, i.e. premise) of the rule and D is its decision part (output, i.e. conclusion). The rule confidence is defined as a ratio: confidence

nCD nC

(2)

where nCD is the number of instances (records) records with given combination of input values and the given output value and nC is the number of records which have that combination of input values. Confidence estimates the probability, that application of input values appearing in the rule will give the result expressed by the rule’s output (decision class), i.e. it indicates how often the rule has been found to be true. The rule support is defined as the ratio: support

nCD n

(3)

where n is the total number of records and nCD is defined as above. Support reflects the breadth of observation basis of the rule. The second general requirement is that the rules should not be unnecessarily demanding, i.e. they do not comprise conditions which are not important, particularly redundant. The tools used for knowledge extraction are above all oriented at generation of a set of rules which would best characterize the problem, i.e. most reliable ones. However, in many industrial processes, particularly in manufacturing, some more specific requirements are relevant, related to design and development of new processes or control of currently running ones. The typical questions to be answered by using the rules can be formulated as follows: • What are the most effective and reliable ways (i.e. values of process parameters—input variables) to achieve the assumed result (class variable)? It

78

M. Perzyk and A. Soroczynski

is important that this does not necessary mean finding the most reliable rule set since the requested result may be included in the rules of minor quality only. • What will happen if we are not able to apply certain input values, i.e. what will we get using process parameters different from those found as most appropriate? Do we still have a chance (and how big) to get the required result? • What will be the predictions (and how reliable) when some input variables cannot be specified, e.g. they may be out of control? • What are all alternative ways to achieve our goal? How reliable they are? It should be noticed that answering some of the above questions may result in necessity of predictions for combination of parameters (input variables values) which have never appeared in the past (i.e. are not present in the data). The users may be interested not only in obtaining a one-time prediction for such input values but also in having a general rule or rules with certain quality parameters. The requirements for rules system and the knowledge extraction tools, suitable for manufacturing industry applications, are not only a consequence of the issues described above, but also the specificity of available data. Typically, the number of independent variables (i.e. problem dimensionality) is not large, it seldom exceeds 10. Number of available records can vary within broad ranges, from only a few to many thousands, especially when the automatic data acquisition system is utilized. The variables can be continuous or categorical. In the former case the discretization, necessary for some classification systems also for input variables, can be often done by the user, based on his/her experience and feeling. Typical industrial data are noisy which results in their inconsistency, i.e. an occurrences of different output variable values (decision classes) for identical combinations of input values (conditions) in the rule. The consistent data, i.e. where a given combination of input values always points at the same output value, is seldom. Finally, strongly unbalanced representativeness of classes can be often observed, both for input and output variables. From the characteristics of industrial processes problems presented above the following requirements for rules systems seem to be essential or at least important: • The rules should make use of all relevant information in data. This means, for example, that all output values (classes) must be represented. Even single cases can be valuable and therefore should be reflected in the rules system. • The rules should not contain redundant conditions as they can be misleading for the user. • It should be possible to find a rule ‘tailored’ to the user’s specifications, including combinations of input variables values which are not represented in the data. • Reliability of all rules should be evaluated, using the confidence and support as the primary parameters.

Assessment of Selected Tools Used for Knowledge Extraction …

79

3 Characteristic Behavior of DTs and RST in Rules Extraction A decision tree model structure of is uniquely defined by a set of the logic expressions, corresponding to the knowledge rules. The nature of DT models, based on recursive partitioning of the data records, results in a set of conditions, which may be different from the combinations of input variables in the training data records. Some of the combinations appearing in the data set may be absent in the tree and vice versa, also some sequences of conditions in the data may be abbreviated. The lack of some combinations of input values in DTs which are present in training data, may result in the rule system without some important rules. Another consequence is that DTs can give wrong predictions for training data. In case of consistent data, this may be a result of improper tree structure in which the given combination of input values (attributes) is connected with a class of the output variable being different from that appearing in the data. Partly incorrect predictions may be a consequence of the fact, that DTs are able to give only one prediction for a given combination of input variables values. For noisy, inconsistent data it must always lead to a fraction of false predictions. Considering a DT as a knowledge rules system means that for this type of data DT must omit some rules, potentially also important for a user. In particular, those omitted rules can be the only ones which give a certain output. An illustrative example is given in Table 1 (details concerning computation algorithms are given in the methodology description in the next section). Due to ambiguous output classes D appearing in the last four records, the DT has ignored the output class D u, although its occurrence in the data was more frequent that D r. In contrast with DTs, RST is able to offer all possible rules resulting from the data, with specification of their confidence and other quality parameters. Rules obtained from DTs may include redundant conditions as the splitting variable used in the root node must appear in all rules (generally, the splitting variable in a node must appear in all rules resulting from subsequent splits). In contrast, RST provides ‘fitted’ rules, i.e. without unnecessary conditions. That type of behavior of the both algorithms was commented in detail in [17]. The rules extracted by RST were described as ‘more individualized’ and made the authors to choose RST for their application. In principle, both DTs and RST can offer predictions for combinations of input values absent in the data, however, they treat such cases in entirely different ways. Interrogation of DTs for such values may lead to one of the following results: (a) prediction which is a good reflection of the general dependencies in the training data (b) prediction which is far from the expectations (c) impossibility of the prediction, when the requested path does not exist in DT (some DT induction algorithms provide mechanisms which may help in such situations).

80

M. Perzyk and A. Soroczynski

Cases (a) and (b) mean that DT includes paths (leading to the tree leaves) which, in principle, correspond to rules not supported by data. However, these logic expressions cannot be treated as valuable rules not only because they may not meet the actual user’s demands but also because the quality parameters of such rules cannot be specified. The confidence of such rule would be indeterminate (zero divided by zero) whereas the support would be equal to zero. For combinations of input values absent in data RST-based algorithm will find a rule with reduced number of conditions, so that they include only those combinations of input values which appear in data. This ‘substitute’, shorter rule will have its confidence and support defined. However, it may result in obtaining predictions (rules) which are not expected. An illustrative example is given in Table 2. The training data contains ordinal-type variables and were obtained by computing output values from the simple formula: D A1 + A2, for random values of inputs, followed by the normalization and categorization of the output. Thus, the expected class for the new input combination can be known (calculated). It can be seen that the RST-based algorithm has led to the result which is far from the expected one.

Table 1 Example of decision tree used for rules extraction; A1 and A2 are the names of input variables (attributes), D is the output (class) variable name, lower case letters are the variable values A1 A2 D Comments Record No. Training data set 1

b

f

r

Record represented in rules

2

b

g

s

3 4

c c

f g

s t

Record represented in rules

5

c

g

t

Record represented in rules

6

c

g

u

Record missing in rules (1 of 2 identical records)

7

c

g

u

Record missing in rules (1 of 2 identical records)

Rule No.

Rules obtained from DT

1 2

b b

f g

r s

3 4

c c

f g

s t

The only rule obtained for A1 c and A2 g

Assessment of Selected Tools Used for Knowledge Extraction …

81

4 Significance of DTs’ Drawbacks Used as Rules Extraction Systems 4.1 Methodology A potential practical significance of problems which may appear in application of DT models for rules extraction described in the previous section, was assessed with a use of simulated and industrial data sets. The artificial data sets were obtained by assuming formulae of the type Y f(X1 , X2 , …), from which, for random values of continuous input variables X1 , X2 , …, the dependent continuous-type variable Y was calculated. Then a Gaussian noise with maximum deviation 20% was imposed on the input variables, and finally all the continuous values variables were converted to categorical ones assuming the equal intervals method. Each of the simulated data sets had 1000 records. Three numbers of the discretization intervals were assumed: 3, 5 and 7, and the following two basic formulas were used: Y X1 + 2 · X2 + 3 · X3 + 4 · X4 + 5 · X5 (data sets denoted as Sim1 3cl, Sim1 5cl, Sim1 7cl) and Y X1 · X2 + X3 + X4 + X5 (data sets denoted as Sim2 3cl, Sim2 5cl, Sim2 7cl). The first group of data sets reflects the situation, where the input variables have highly differentiated effect on output whereas the second data sets is an example of interaction between two variables with overall significance equal to significances of the remaining variables. Similar situations often appear in practice. All the real (industrial) data sets concern foundry production. The Ind1 data set correlates chemical composition of ductile cast iron, defined by 5 main elements (Mn, Si, Cr, Ni and Cu) with its four grades, obtained as a result of the melting process, as the output class variable. The number of classes (categories) for all 5 input variables

Table 2 Example of RST rule for new data; A1 and A2 are the names of input variables (attributes), D is the output (class) variable name, variables values explained in the text

A1

A2

D

1 1 2 4 2 5 2 3 4

1 2 2 3 3 4 4 5 4

3

3

Training data set 1 2 1 1 3 1 4 5 2 New data 2

Matched (shorter) rule found from RST (Any)

3

5

82

M. Perzyk and A. Soroczynski

was assumed equal 5. The second type of industrial data (data sets denoted as Ind2 3cl, Ind2 5cl, Ind2 7cl) correlates chemical composition of the ductile cast iron, defined by 9 elements with its tensile strength (details can be found in [26]). Another type of industrial data were obtained as readouts from a semi-empirical nomograph which permits to calculate solidification shrinkage of grey cast iron as a function of four variables: carbon contents (5 different values—categories), sum of silicon and phosphorus content (4 values), casting modulus (4 values) and pouring temperature (4 values). In the data set denoted as Ind3 the output was the iron shrinkage expressed by 7 different levels (classes) and in the last two data sets the outputs were the decision concerning necessity and size of application of feeders to avoid shrinkage defects: in data set denoted as Ind4 the output variable named ‘Feeder’ had 2 classes (No and Yes) and in data set denoted as Ind5 the same output had 3 classes (No, Small and Large). Each of the last 3 data sets contained 190 records. Further details can be found in [27]. The requirements for rules and knowledge extraction tools formulated in Sect. 2 have brought about utilization of the procedures which ensured possibly the largest choice of rules available from the data, passing over possible overfitting of the models. Binary DTs were obtained using CART algorithm and MineSet commercial software package. Various splitting conditions, stopping criteria and pruning parameters were tried out. The smallest trees which ensured the smallest fraction of false predictions for training sets were chosen. RST procedure, oriented at generation of full set of rules, was written by the present authors with a somewhat similar approach as used in the Explore algorithm [28]. Firstly, all the combinations of single input variables appearing in the data are placed in the rules (i.e. rules with only one conditions were generated) and their confidences are calculated. Then the further conditions are added, providing the confidence of a rule thus obtained is increased, compared to the rule with shorter conditional part.

4.2 Results In Fig. 1 the fractions of wrong predictions obtained from DTs for all consistent data subsets (i.e. all the discernible input values combinations pointing at one output value only) are shown, for all the training data sets. It can be observed that the fraction of wrong predictions for simulated data increases with the number of classes (categories) which can be attributed to a limited accuracy of DT models. The general level of false predictions for real data is much lower, compared to simulated data. An interpretation of this observation would require a deeper analysis of the data sets structures, e.g. representativeness of the classes of input and output variables. In Figs. 2 and 3 some statistical information obtained for inconsistent data subsets is shown. The ratios of the number of inconsistent data subsets to all subsets of the same input values, were quite similar in all presented data sets (20–30%). The fractions of false predictions also result from the distributions of the classes in the inconsistent data subsets. It is interesting to note that in several cases DTs have

Assessment of Selected Tools Used for Knowledge Extraction …

83

20% 16,4% 14,8%

15%

10%

7,9%

7,6% 7,7%

5%

3,0%

2,5%

1,9%

3,8%

0,0%

0% Sim1 Sim1 Sim1 Sim2 Sim2 Sim2 3cl 5cl 7cl 3cl 5cl 7cl

Ind1

Ind2 3cl

Ind2 5cl

Ind2 7cl

Fig. 1 Average fractions of false predictions obtained from DTs for consistent data subsets (including single records)

0

30%

0

67%

Maximum % of wrong predictions for all inconsistent subsets

2

47% 50%

49% 50%

44%

29%

50%

67%

3

Average % of wrong predictions for all inconsistent subsets

44%

75%

5

Number of inconsistent subsets with over 50% of wrong predictions

0

Sim1 3cl Sim1 5cl Sim1 7cl Sim2 3cl Sim2 5cl Sim2 7cl

Fig. 2 Statistics of false predictions obtained from DTs for inconsistent data subsets for simulated data sets

pointed at the decision classes which are not predominant for the given combination of input values. The results presented in Figs. 1, 2 and 3 indicate that the rules systems represented by DTs may be significantly incorrect for inconsistent data as well as for consistent data with variables having large numbers of classes (categories). In Fig. 4 the fractions of rules included in DTs which are not supported by data are shown, exhibiting quite large values in several cases. In principle, this can be a positive feature of DTs as such rules may be desired by a user (see comments in Sect. 2). However, the usefulness of such rules may be questionable. Firstly, because they do not necessarily meet the user’s specific

M. Perzyk and A. Soroczynski Average % of wrong predictions for all inconsistent subsets 50%

67%

3

50%

67%

84

31%

Maximum % of wrong predictions for all inconsistent subsets Number of inconsistent subsets with over 50% of wrong predictions

16%

30%

33%

2

0 Ind1

0

Ind2 3cl

Ind2 5cl

Ind2 7cl

Fig. 3 Statistics of false predictions obtained from DTs for inconsistent data subsets for industrial data sets (for Ind3, Ind4 and Ind5 no inconsistent subsets were found) 80%

60%

40%

20%

Ind5

Ind4

Ind3

Ind2 7cl

Ind2 5cl

Ind2 3cl

Ind1

Sim2 7cl

Sim2 5cl

Sim2 3cl

Sim1 7cl

Sim1 5cl

Sim1 3cl

0%

Fig. 4 Fractions of rules in DTs not supported by data

needs and secondly because their reliability, defined by confidence and support, is not determined, as pointed in Sect. 3. In Fig. 5 the numbers of rules missing in DTs, but extracted by RST, are presented, together with total numbers of rules in DTs and from RST. Note, that if a conditional part corresponding to a RST rule was found in a longer DT rule, then such rule was not qualified as ‘missing in DT’. The average confidence and support values of the missing rules in DTs are shown in Fig. 6, together with those for the rules which are present in DTs. It can be seen that the missing rules may be valuable for a user as their confidences are relatively high and comparable with those for the rules included in DTs. It is worth noticing that for some of the simulated data sets, some of the missing rules had 100% confidence. The support values are generally low for both groups of rules, which is obviously a result

Assessment of Selected Tools Used for Knowledge Extraction …

85

Number of rules in DT Number of absent rules in DT (obtained from RST) Number of rules from RST 6000 5000 4000 3000 2000 1000 0 Sim1 Sim1 Sim2 Sim2 5cl 7cl 5cl 7cl

Ind1 Ind2 3cl

Ind2 5cl

Ind2 7cl

Ind3

Ind5

Fig. 5 Quantities of rules in DTs and obtained from RST—total and missing in DTs

Average confidence of absent rules, %

Average confidence of DT rules, %

Average support of absent rules, %

Average support of DT rules, %

100% 80% 60%

Sim1 5cl

Sim1 7cl

Sim2 5cl

Sim2 7cl

Ind1

Ind2 3cl

Ind2 5cl

Ind2 7cl

Ind3

3,6% 5,7%

3,2% 1,3%

6,9% 0,5%

3,5% 2,0%

1,2%

0,3% 0,2%

0,5% 0,3%

0,4% 0,2%

0%

0,9% 0,3%

20%

5,2% 5,5%

20,7%

40%

Ind5

Fig. 6 Quality parameters of rules obtained from RST and omitted by DTs

of the nature of the data sets. However, the support values for rules missing in DTs are often higher than for the rules present in DTs which is probably a consequence of the fact that the rules from RST do not have redundant conditions. In Fig. 7 fractions of DT rules with redundant conditions are shown. Obviously, the RST rules taken as a reference had this same confidence values. In Fig. 8 some characteristics of DT rules containing redundant conditions (denoted as ‘oversized’) are presented with reference to the corresponding rules obtained from RST (denoted as ‘fitted’), for selected data sets. It can be seen that

86

M. Perzyk and A. Soroczynski 100% 80% 60% 40% 20%

Ind2 5cl

Ind2 3cl

Ind1

Ind5

Ind4

Ind3

Sim2 7cl

Sim2 5cl

Sim2 3cl

Sim1 7cl

Sim1 5cl

Sim1 3cl

0%

Fig. 7 Fractions of rules in DTs with redundant conditions

1.96 2.09

Average length ratio oversized/fitted

2.90

Average support ratio oversized/fitted 0.04 0.04 Sim1 3cl

0.53 Ind1

Ind5

Fig. 8 Characteristics of DT rules with redundant conditions (oversized) compared to the corresponding rules obtained from RST (fitted)

the fraction of redundant input variables in DT rules is high. The conclusion is that the presence of redundant conditions in rules obtained from DTs, being a result of the nature of that type models, may be their significant disadvantage. However, it is worth noticing that some DT induction algorithms, such as C4.5, contain a mechanism of dropping conditions that are irrelevant to the class, which may reduce the redundancies appearing in the rules.

Assessment of Selected Tools Used for Knowledge Extraction …

87

5 Summary and Conclusion The characterization of rule systems based on decision trees as knowledge extraction tools has drawn attention to their potential disadvantages. The presented numerical results revealed that these shortcomings are real and may often be encountered and suggest that the rules obtainable from RST seem to be generally better. Although the present paper is focused on industrial manufacturing processes, it can be expected that the obtained results can be useful also for other application areas where not only a characterization of a problem is required, but also detailed and precise rules are needed, according to specific problems to be solved. Acknowledgements The authors are grateful to prof. Jerzy Stefanowski from Institute of Computing Science of Poznan University of Technology, Poland, for his highly valuable advice in preparation of the earlier version of this work. We would also like to thank International Science Council of World Academy of Science, Engineering and Technology for the permission to use the copyrighted material from our paper “Comparative Study of Decision Trees and Rough Sets Theory as Knowledge Extraction Tools for Design and Control of Industrial Processes” at http://waset.org/publications/7119/.

References 1. Wang, K.: Applying data mining to manufacturing: the nature and implications. J. Intell. Manuf. 18(4), 487–495 (2007) 2. Chen, W.-C., et al.: A data mining projects for solving low-yield situations of semiconductor manufacturing. In: Advanced Semiconductor Manufacturing, 2004. ASMC’04. IEEE Conference and Workshop. IEEE (2004) 3. Harding, J., Shahbaz, M., Kusiak, A.: Data mining in manufacturing: a review. J. Manuf. Sci. Eng. 128(4), 969–976 (2006) 4. Huang, H., Wu, D.: Product quality improvement analysis using data mining: a case study in ultra-precision manufacturing industry. In: Fuzzy Systems and Knowledge Discovery, pp. 485–485 (2005) 5. Kamal, A.M.M.: A data mining approach for improving manufacturing processes quality control. In: The 2nd International Conference on Next Generation Information Technology (ICNIT), 2011. IEEE (2011) 6. Koonce, D.A., Fang, C.-H., Tsai, S.-C.: A data mining tool for learning from manufacturing systems. Comput. Ind. Eng. 33(1), 27–30 (1997) 7. Kusiak, A.: Data mining: manufacturing and service applications. Int. J. Prod. Res. 44(18–19), 4175–4191 (2006) 8. Perzyk, M.: Data mining in foundry production. Research in Polish metallurgy at the beginning of XXI century, Committee of Metallurgy of the Polish Academy of Sciences, Cracow, Poland, pp. 255–275 (2006) 9. Shahbaz, M., et al.: Product design and manufacturing process improvement using association rules. Proc. Inst. Mech. Eng. Part B: J. Eng. Manuf. 220(2), 243–254 (2006) 10. Shen, L., et al.: Fault diagnosis using rough sets theory. Comput. Ind. 43(1), 61–72 (2000) 11. He, S.-G., Li, L., Qi, E.-S.: Study on the continuous quality improvement systems of LED packaging based on data mining. In: International Conference on Wireless Communications, Networking and Mobile Computing, 2007. WiCom 2007. IEEE (2007) 12. Tsang, K., Lau, H., Kwok, S.: Development of a data mining system for continual process quality improvement. Proc. Inst. Mech. Eng. Part B: J. Eng. Manuf. 221(2), 179–193 (2007)

88

M. Perzyk and A. Soroczynski

13. Tseng, T.-L.B., et al.: Applying data mining approaches for defect diagnosis in manufacturing industry. In: IIE Annual Conference. Proceedings. Institute of Industrial and Systems Engineers (IISE) (2004) 14. Tanuska, P., et al.: Data mining model building as a support for decision making in production management. In: Advances in Computer Science, Engineering & Applications, pp. 695–701 (2012) 15. Perzyk, M., Biernacki, R., Kozlowski, J.: Data mining in manufacturing: significance analysis of process parameters. Proc. Inst. Mech. Eng. Part B: J. Eng. Manuf. 222(11), 1503–1516 (2008) 16. Perzyk, M., et al.: Comparison of data mining tools for significance analysis of process parameters in applications to process fault diagnosis. Inf. Sci. 259, 380–392 (2014) 17. Kusiak, A., Kurasek, C.: Data mining of printed-circuit board defects. IEEE Trans. Robot. Autom. 17(2), 191–196 (2001) 18. Etchells, T.A., Lisboa, P.J.: Orthogonal search-based rule extraction (OSRE) for trained neural networks: a practical and efficient approach. IEEE Trans. Neural Networks 17(2), 374–384 (2006) 19. Brouwer, R.K.: Fuzzy rule extraction from a feed forward neural network by training a representative fuzzy neural network using gradient descent. Int. J. Uncertain. Fuzziness Knowl. Based Syst. 13(06), 673–698 (2005) 20. Duch, W., Adamczak, R., Grabczewski, K.: A new methodology of extraction, optimization and application of crisp and fuzzy logical rules. IEEE Trans. Neural Networks 12(2), 277–306 (2001) 21. Huang, S.H., Xing, H.: Extract intelligible and concise fuzzy rules from neural networks. Fuzzy Sets Syst. 132(2), 233–243 (2002) 22. Chen, R.-S., Wu, R.-C., Chang, C.-C.: Using data mining technology to design an intelligent CIM system for IC manufacturing. In: Sixth International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing, 2005 and First ACIS International Workshop on Self-Assembling Wireless Networks. SNPD/SAWN 2005. IEEE (2005) 23. Hur, J., Lee, H., Baek, J.-G.: An intelligent manufacturing process diagnosis system using hybrid data mining. In: Industrial Conference on Data Mining. Springer (2006) 24. Rokach, L., Maimon, O.: Data mining for improving the quality of manufacturing: a feature set decomposition approach. J. Intell. Manuf. 17(3), 285–299 (2006) 25. Sadoyan, H., Zakarian, A., Mohanty, P.: Data mining algorithm for manufacturing process control. Int. J. Adv. Manuf. Technol. 28(3–4), 342–350 (2006) 26. Perzyk, M., Kocha´nski, A.: Prediction of ductile cast iron quality by artificial neural networks. J. Mater. Process. Technol. 109(3), 305–307 (2001) 27. Perzyk, M., Soroczynski, A.: Comparison of selected tools for generation of knowledge for foundry production. Arch. Foundry Eng. 8(4), 263–266 (2008) 28. Stefanowski, J., Vanderpooten, D.: Induction of decision rules in classification and discoveryoriented perspectives. Int. J. Intell. Syst. 16(1), 13–27 (2001)

Application of Data Mining Tools in Shrink Sleeve Labels Converting Process Krzysztof Krystosiak

Abstract In manufacturing practices of most big printing companies there are collected data and records of process parameters. Gathering information from this data in form of developed models and rules such as data mining, which uses statistical methods or Artificial Intelligence, Artificial Neural Networks, Decision Trees, Expert Systems, and others are subjects of inter-disciplinary fields of science called Data Mining. In the shrink sleeve production process, the effects of using data mining tools not only improve the quality of the shrink sleeve and winding process but also reduce manufacturing costs. This paper describes how developed models of data mining tools can be used for prediction of initial tension parameters and winding speed for each every new design of shrink sleeve labels. Every one design of shrink sleeve label has a lot of factors. Some of them are more significant, some of them less. The aim of this paper is to choose significant factors and compute a models in learning process using collected data. Finally when models will be computed, can be used for prediction of key winding parameters of each new shrink sleeve label designs. This will bring for company saved time for experimental selection during converting of winding parameters like tension and speed and also will minimize risk of occurrence of defects with incorrect winding parameters.

1 Shrink Sleeve Label Manufacturing Process In many references and online resources different definitions of shrink sleeve labels are found, as for example in [2, 4–9, 11], but for the purposes of implementation of this article the following definition was established as below: • Shrink sleeve labels are manufactured by printing on the plastic film followed by seaming the printed film into a sleeve, which under the influence of a determined temperature for a given material, shrinks and clings onto the target surface. K. Krystosiak (B) Faculty of Production Engineering, Warsaw University of Technology, Narbutta 85, 02-524 Warsaw, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_6

89

90

K. Krystosiak

Fig. 1 Shrink sleeve labels [5]

Shrink sleeve labels (Fig. 1) are produced using flexographic printing technology, more specifically it is a rotary printing method that uses a flexible printing form to compensate for the surface irregularities of the substrate. Ink is transferred from the inkpot onto the flexible printing form by anilox roller and from there directly onto the substrate. Labels printed with this technique are used to perform unstable mass production. The group includes, for example shrink sleeve labels as mentioned in [4, 7, 8]. Shrink sleeve labels are made from many types of macromolecular materials, as listed below [7]: • • • • • •

PVC; PET; PET-G; OPS; PLA (polylactic acid); And other hybrid materials for specific applications.

Winding quality of wounded rolls on every stage of the manufacturing process and beam quality are essential to achieve optimal quality of shrink sleeve labels and roll to roll (R2R) winding process without any problems. Main steps of the shrink sleeve labels production process are four typical production stages which result in a finished product like wounded roll of sleeve. These stages are: printing, slitting, seaming and inspection, what is described as well in [2, 3, 12–15]. On the printing stage material in the form of foil film is printed on the printing press with using printing techniques like flexographic or rotogravure and sometimes offset. In next stage—which is slitting, rewinding—printed material is cut on small rolls and prepared to seaming stage of the manufacturing process. On the seaming stage printed and cut material is seamed in a form of sleeve. There is also inspection stage after the seaming but it is not necessary. The common feature of mentioned steps of the production process is winding from roll to roll (R2R) and that is typical for flexographic printing process (Fig. 2). The most important parameter characterizing of winding quality is free from significant defects both: in manufacturing process and in final application process, like [13, 15]:

Application of Data Mining Tools in Shrink Sleeve Labels …

91

1 2

1

2

Fig. 2 Schematic diagram of R2R process

• • • • • • •

glued of sleeve layers, roll telescoping, foil film blocking, foil film web deformation (stretching), web wrinkling, damage of sleeve edge (U-fold), or finally, wound damage.

In case of the quality it should be noted that significant factors which have influence for winding quality are initial tension parameters and winding speed which maintain appropriate web tension in the process. Winding quality issues indicate one basic problem which is incorrect tension parameters setup on converting machines. There are two problems related to in-correct tension parameters: too high or too low tension. Both of them are causes of occurrence of different defects. It is therefore expedient to collect, process and analyze information from the production process, and draw conclusions. It can be concluded that manufactured products speak to us through the data that we collect on them. Without these data, the enterprise doesn’t have enough information related to the products which are produced. Identification of variables in the converting process will enable a better understanding of the phenomena occurring in the process and examine the possibility of adjusting such parameters as web tension setup.

2 Winding Quality Issue The winding quality issue is a topic of research at many universities and institutes all over the world. It is worth mentioning, at this point, the research center at Oklahoma State University—the Web Handling Research Center, where issues concerning the processing of winding, packaging materials in the form of wounded rolls, are examined. One major question arises when speaking about the correct winding quality of converted product: what is the right tension? This issue is the subject of researches, and there are many different methods for determining the tension and it can be found in [15, 17, 20–22]. However, this knowledge is based on experience and con-

92

K. Krystosiak

Table 1 Typical problems related to wrong tension setup

Too high tension

Too low tension

Glued sleeve inside

Telescoping

Film blocking

U-fold damages

Stretched film Web breaking

Wrinkles Beam damaged

stant experimentation; and each organization should strive to improve their methods through this knowledge. The winding quality issue points to one fundamental problem of improperly adjusted tension values on the converting machines. The most common types of defects caused by too high or low tension setup on the converting machines are listed in the following table (Table 1). Preliminary researches were conducted within the packaging industry where shrink sleeve labels are core product. The main reason for these researches was to determine which problems occurred during the converting processes. For this purpose Pareto-Lorentz analysis was used which is an excellent tool for indicating the factors responsible for a specific problem, also known as the Pareto principle 80/20 states that, for many events, roughly 80% of the effects come from 20% of the causes, described in [11].

Pareto Analysis: Type of defect 120

100%

100 80% 80 60%

66 60

40% 40 20%

18

20

11

9 3

3

2

2

2

tc

he

lo w

s

te

on

ns

fil

io

m

n

ng oc

Sc

ra

To

bl Fi lm

g in itt Sl

tc re St

ki

fe de

fil he

d

ns te

To

hi

gh

am

ct

m

n io

ct de

ag m da

ld

Se

si in ve U -fo

ee sl ed G

lu

fe

es

de

0

Fig. 3 Pareto-Lorentz analysis of type of defects in converting process

Application of Data Mining Tools in Shrink Sleeve Labels … Table 2 Summary of control variables in converting process Variable Description Unit

93

Data type

x1

Web width

[mm]

Continuous

x2

Web thickness

[µm]

Continuous

x3

Material type

[X, Y, Z, …]

Discrete

x4

Material manufacturer [A, B, C, …]

Discrete

x5

Converting machine

Discrete

[S-1, S-2, S-3]

Table 3 Summary of dependent variables in converting process Variable Description Unit

Data type

x1

Unwind tension

[N]

x2

Rewind tension

[N]

Continuous Continuous

x3

Winding speed

[m/min]

Continuous

Presented data states the quantity of non-conformities from the shrink sleeve label converting process which was registered in the packaging company from 2012 to 2013. The Pareto-Lorentz analysis of the various types of quality defects showed that the most common problem is glued sleeve inside (Fig. 3). This issue may arise as a result of incorrect web tension settings in the converting process. Among the many parameters which determine the quality level of the production process, based on personal experience, the carried out preliminary researches and conducted analyzes found that one of the main reasons for the use of improper tension setting was in fact due to missing or incorrect methods to effectively control the winding quality using web tension settings during the unwinding and rewinding of rolls in the converting process.

3 Input and Output Variables The data records gathered come from October 2014 to March 2015 and include parameters from the seaming step of the manufacturing process. There were 2096 records gathered directly from the manufacturing process. Data was gathered for three seaming machines, all data was recorded from Technological Cards which are the recipe for producing all work orders. Records from work orders where any converting issues occurred were removed from this database. Next step in data mining process is to check correlations between control and dependent variables but first we need to divide all variables into control and dependent (Table 2). These variables are the input data to building prediction models. At the network output there are three dependent variables—the initial settings of web tension on unwinding and rewinding roll in the converting process and winding speed which are listed in the table below (Table 3).

94

K. Krystosiak

Table 4 Correlation matrix of input and output data for ANN model Variable Mean Standard Web Thickness Winding deviation width speed 38.688

1.000

0.066

−0.189

0.638

0.602

4.944 67.728

0.066 −0.189

1.000 −0.084

−0.084 1.000

0.109 0.061

0.209 −0.198

35.458

8.671

0.638

0.109

0.061

1.000

0.502

41.666

12.327

0.602

0.209

−0.198

0.502

1.000

Web 126.876 width Thickness 47.166 Winding 347.947 speed Unwinder tension Rewinder tension

Unwinder Rewinder tension tension

As shown, the input data presented both data types: continuous and discrete. This is not a problem due to the input of prediction models because data mining models can compile different types of data. The mentioned input variables above apply only to certain parameters affecting the roll winding quality of shrink sleeve labels. Therefore, the developed models will not be perfect, as the process may include some interference. Also, the materials used may be of different quality; and certain characteristics can affect the winding quality, although this research needs to demonstrate a desirability of using this method. The target of the prediction models is to find the relationship between the input and output data. A correlation table is presented below (Table 4), and major coefficients have been highlighted which are: web width and also thickness what was proved in [15]. An analysis of the variance ANOVA was also conducted on the collected discrete data type, and it showed that there is a statistically significant variance in material type—material code “Z” was converted with a lower initial tension setup than material code “X” (Fig. 4). It is related to the construction properties of each material types. Some of them are stronger and can be converted faster and with higher tensions setup. But some of them are sensitive—like material “Z”—and setup parameters will be lower. Analysis of the variances between seaming machines (Fig. 5) also found significant variance. All seaming machines are from the same producer, but there are different types and has different construction. It happened even we use machines from one producer—during time every manufacturer modified them. However quality of machines maintenance must be assured in the process, in other way we cannot trust gathered data.

4 Predictors Analysis One of the techniques used to select the most important predictors or reducing less important is multivariate adaptive regression using spline functions known as MAR-

Application of Data Mining Tools in Shrink Sleeve Labels … Fig. 4 ANOVA analysis of material type

95

ANOVA: Material type Lambda Wilksa=,91994, F(4, 4184)=44,566, p=0,0000 48 46 44 42 40 38 36 34 32 30 28 26 X

Y

Z

Material type Unwinder tension [N] Rewinder tension [N]

Fig. 5 ANOVA analysis of seaming machine

ANOVA: Type of the Machine Lambda Wilksa=,71154, F(4, 4184)=194,03, p=0,0000

50 48 46 44 42 40 38 36 34 32 30 S-1

S-2

S-3

Seaming Machine Unwinder tension [N] Rewinder tension [N]

Splines [6, 16, 18]. This technique is used to solve problems of both the classification as well as regression. What is important, MARSplines is a nonparametric procedure does not require assumptions about the functional relationship between the depen-

96

K. Krystosiak

Table 5 MARSplines analysis results Dependent variable Unwind tension

Rewind tension

Winding speed

Independent variables

Web width, Web thickness, Material type, Material manufacturer, Seaming machine

No of records without data No of factors No of basic functions Interaction level Penalty

0

0

0

12 20 3 2.000000

13 20 3 2.000000

13 18 3 2.000000

Threshold Error (GCV)

0.000500 39.578000

0.000500 40.184399

0.000500 1743.476861

Elimination

Yes

Yes

Yes

dent variables and independent. The method described models this relationship by means of a set of coefficients and basis functions work only on the basis of the data. MARSplines characterized by the fact that the input space is divided into areas, which are referred to separate functions regression or classification, proved in [16]. This technique becomes particularly useful when more than two variables at the input, when other techniques compromise the dimensionality begins problem, as in [16, 18]. The table below (Table 5) presents the results of analyzes carried out using techniques MARSplines separately for winding speed, tension for unwinder and rewinder. To perform these evaluations was used Statistica software, which has a built-in tool MARSplines. During preparation for analysis of predictors, dashboard option was changed, interaction level was set from one to three in order to more accurately analyze the relationships between the variables. The error mentioned in the table above is a Generalized Cross Validation (GCV) for comparing the performance of model subsets in order to choose the best subset: lower values of GCV are better. The GCV error is a form of regularization: it trades off goodness-of-fit against model complexity. The GCV error shows how well model will perform new data, not on the trained data, so we can use it to estimate what performance would be on new data [6]. We can see that the MARSpline model of winding speed had the biggest error in this analysis. Below was presented mathematical equations for these three factors like: winding speed, tension for unwinder and rewinder.

Application of Data Mining Tools in Shrink Sleeve Labels …

97

First, mathematical equation for winding speed:

Then mathematical equation for unwinder tension:

And finally an mathematical equation for rewinder tension was computed:

Above graphs (Fig. 6a–c) shows the correlations between observed data of each independent variable and predicted data. As you can see, there is a high correlation.

98

K. Krystosiak

Table 6 Summary of predictors influence on the variables Predictors No of appeals to the base function Web width Web thickness Machine Material producer Type of material

Winding speed

Unwinder tension

Rewinder tension

10 1 4 1

7 1 5 1

6 4 4 1

2

6

5

Having analyzed the correlation graphs above, we can see a set of values substantially deviate, which can be explained in two ways: the data quality was not satisfactory enough or there are some variables or factors not included in the models. Summarizing conducted analysis showed that this technique is useful for determining the essential predictors and the relationship between the variables. Below in the table summarizes the number of appeals to the functions of the base of each predictor developed models. If the number is higher, that means the higher the validity of a predictor. So it is related to its impact on the dependent variable (Table 6). For the winding speed and the tensions: unwinder and rewinder, the most important parameter for the development of models was the web width predictor. Also an important predictor proved to be the independent variable—machine, and type of material, at this point it is worth to rely on previous analyses ANOVA, which showed us that these variables are important that cannot be missed by building mathematical models. In other way, the independent variable—the producer of the material, had the lowest rank, and therefore in the development of models may have a small effect on the dependent variable, however author decided to have this variable in the pocket of the predictors.

5 Prediction Effects with Data Mining Tools 5.1 Artificial Neural Networks Models There were conducted several researches for different network topology. First was used Automated Neural Network named ANN with using Statistica software. A standard sampling method proposed by the software, which is the number of random samples was set: learning—70%, testing and validation both for 15%, while the initial value of the generator set in 1000—is also the standard setting from Statistica software. Number of hidden neurons between 3 and 11 was selected as a default, and the type of network—MLP was set, which is multi-layer perceptron. Was used all the functions of activation and both hidden neurons, as well as the output indicated all possible activation functions. The computed networks was named as a.MLP-

Application of Data Mining Tools in Shrink Sleeve Labels … Scatterplot: Winding speed [m/min] (Observ ed) v s Winding speed [m/min] (Predicted) Y = 129,1745+0,6288*x R 2 = 0,6288

Winding speed [m/min] (Predicted)

(a) 440 420 400 380 360 340 320 300 280 260 240 220 200 180 160 140 120 -50

0

50

100

150

200

250

300

350

400

450

500

Winding speed [m/min] (Observed) Scatterplot: Unwinder tension [N] (Observ ed) v s Unwinder tension [N] (Predicted)

Unwinder tension [N] (Predicted)

(b)

Y = 18,2671+0,4848*x R 2 = 0,48482

75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0 0

5

10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

Unwinder tension [N] (Observed) Scatterplot: Rewinder tension [N] (Observ ed) v s Rewinder tension [N] (Predicted)

(c) Rewinder tension [N] (Predicted)

Fig. 6 Correlation between observed and predicted data for a winding speed, b unwinder and c rewinder tension

99

90 85 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0 -5 -10 -10 -5

Y = 10,7629+0,7417*x R 2 = 0,7417

0

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95

Rewinder tension [N] (Observed)

100

K. Krystosiak

810 800

0.815 0.812

0.812

797

793

790

794

795

0.811 0.809

780

778

775

770

0.805 767

0.806

0.802

0.807

0.807 0.805

0.803

760

0.803

765

0.805

755

739

0.805 0.803 0.801

750 740

0.813

0.799

0.798

740

730

0.797 0.795

a.MLP 17-10-3

5.MLP 17-5-3

4.MLP 17-7-3

2.MLP 2.MLP 2.MLP 17-11-3 17-17-3 17-25-3 Error (validaƟon)

4.MLP 17-5-3

6.MLP 17-7-3

12.MLP 20.MLP 22.MLP 17-11-3 17-17-3 17-25-3

Quality (validaƟon)

Fig. 7 Prediction quality versus error—summary for ANN models

17-10-3 and it is shown on below graph as a first match (Fig. 7). In the next testing trials numbers of hidden neurons was changed and also activation function for hidden neurons was changed from Tanh to logistics. Other parameters wasn’t changed during testing trials, see also [14]. The above described changes in the network topology was applied to the effect of influence of increasing the number of neurons in the hidden layer, and change the activation function of hidden neurons. While increasing the number of neurons in hidden layer does not significantly influence the correlation coefficient prediction of the initial parameters of the winding, so much the quality of prediction expressed network error coefficient slightly worsened—thus proving the theory that the number of neurons in hidden layer greatly exceeding 10 may lead to the loss of the network ability to generalize, which means poorer quality prediction what was proved in [10, 14, 18, 19]. It was noted that the logistic activation function of hidden neurons showed a better prediction than Tanh function, which can be seen in the graph above. Described above testing trials of ANN tool to verify the quality prediction using the interference in the network topology represents only a narrow margin opportunities provided powerful tools of artificial neural networks. The developed ANN model can be used for predicting the initial winding parameters of each new shrink sleeve label design. The variables of new designs will be set as the inputs, and the ANN model will compute the outputs such as the winding parameters—initial tension for the unwinder and rewinder and winding speed.

Application of Data Mining Tools in Shrink Sleeve Labels …

101

5.2 Data Miner Recipes Models In order to develop an effective method to control winding quality, the Data Miner Recipe with Statistica software was used. Some of the data mining tools are usually used to solve problems of classification or regression and some of them only for classification. In this case it is typically a regressive problem. With Data Mining Recipes any classification or regression problem can be handled with such tools as: Classification & Regression Trees, Artificial Neural Networks, Support Vector machines or Boosted Trees. Pilot researches were conducted on the collected historical data of shrink sleeve labels during the seaming step of the converting process, which includes the following control variables mentioned above, see also [15]. As mentioned above, the prediction models was computed with using special Data Miner Recipe with Statistica software. This solution offer for the user predefined path, step by step, the analyst will carry out the entire process of data analysis, from identifying the data file by checking the purification and transformation of data, and further develop a model and apply it to new situations. The user can also define his own rules for data mining studies carried out, and modify them by adjusting own preferences, what was described in [1, 19]. With Data Miner Recipe many variants of networks were tested. First was the choice of data source—it was the same 2096 rows of historical data records as for the variables analysis. After selecting the control and dependent variables was determination of test sample. In this case test sample for the machine learning process was set 20% of data records. The next step was the ability to limit the redundant data by two methods—the correlation coefficient or the Spearman rank correlation coefficient “R” by a coefficient which is normally set to 0.70. In this case, the test showed no redundant variables. In the next stage of the development of automatic models, it is possible to limit the significant variables, namely the rejection of the variables which do not affect test size. Software allows quick or advanced selection of significant variables, but in the present case, there are five variables that have previously been selected at the beginning of data mining process. The final step is to create the models, where the user has the ability to choose which method wants to use. In the present case, selected such data mining methods like above, also it can be found in [15]: • • • •

Classification & Regression Trees (C&RT); Boosted Trees (BT); Artificial Neural Networks (ANN); Support Vector Machines (SVM).

Finally all models was computed. Effects of modelling process was set on the graph below (Fig. 8) where all data mining prediction models are shown on the same graph and for all three dependent variables. Above graph shows that significantly mostly better prediction effect as a correlation between dependent variable from the historical data and computed effect has Boosted Trees and Classification & Regression Trees than other models. In all three

102

K. Krystosiak

1 0.8 0.6 0.4 0.2 0 BT

C&RT

ANN

Winding speed

SVM

BT

C&RT

ANN

SVM

BT

Unwinder tension

CorrelaƟon, learning process

C&RT

ANN

SVM

Rewinder tension

CorrelaƟon, tesƟng process

Fig. 8 Correlation results of computed models for all dependent variables

dependent variables, models of Support Vector Machines reached lowest prediction quality (Fig. 9a–c). Having analyzed the correlation graphs above (Fig. 9a–c), we can see some of values substantially deviate, which can be explained in two ways: the data quality was not satisfactory enough or there are some variables or factors not included in the computed models. The obtained correlation of coefficients is oscillating in the range of 0.65–0.85 and as for the industrial raw data are quite good these models may be useful—model with Boosted Trees founds dependences in the process for all three dependent variables [13]. The developed model of Boosted Trees can be used for predicting the initial winding parameters of each new shrink sleeve label design. The variables of new designs will be set as the inputs, and the BT model will compute the outputs such as the winding parameters—initial tension for the unwinder and rewinder and winding speed. This method can achieve saved time on the experimental way of setting of initial winding parameters during operation on work order, and a reduced risk of defect occurrence related to improper setup of winding parameters.

6 Idea for Application of Data Mining Tools in Manufacturing Process The main idea for application of data mining tools in manufacturing process, where shrink sleeve labels are produced, is to use recorded data from the converting process of each label design. The converting process has a lot of factors. During process analysing there were set five control factors of converting process on seaming machines, which has influence on the initial setup parameters. These initial setup parameters are: unwinder and rewinder tensions and winding speed. Initial setup parameters are set in experimental way during setup of every new design. After

Application of Data Mining Tools in Shrink Sleeve Labels …

103

(a) 420 400 380 360 340 320 300 280 260 240 220 200 180 160 140 120

0

50

100

150

200

250

300

350

400

450

500

Winding speed [m/min]

(b)

SVM Prediction

C&RT Prediction

ANN Prediction

BT Prediction

80 70 60 50 40 30 20 10

0

10

20

30

40

50

60

70

80

90

80

90

Unwinder tension [N]

(c)

C&RT Prediction

SVM Prediction

ANN Prediction

BT Prediction

120

100

80

60

40

20

0

0

10

20

30

40

50

60

70

Rewinder tension [N] SVM Prediction

C&RT Prediction

ANN Prediction

BT Prediction

Fig. 9 Prediction effects with data mining tools for a winding speed, b unwinder and c rewinder tensions

104

K. Krystosiak

Fig. 10 Idea graph for application of data mining tools in manufacturing environment

that, operator or process engineer save the estimated initial winding parameters into a Technology Card. The main idea is to compute effective models based on data recorded in Technology Cards and use them in process. This idea is described above (Fig. 10). This process could look like: every new design of shrink sleeve label is recorded into the data base where are known control variables as web width and thickness, material type and other. But there is no information about the initial setup parameter because it’s unknown until machine operator or process engineer will estimate this in experimental way. Using of data mining tools and techniques like described above in this paper, the initial setup parameters could be predicted. Computed model could be set into any manufacturing information system and based on the known control variables could compute dependent variables which are initial setup parameters, known as unwinder and rewinder tension and winding speed. This information system could be provided by—after computing—directly to the machine station and operators or process engineers which are involved in manufacturing process could use it. This is how initial setup parameters for converting process on seaming machines could be improved. Operators and process engineers could have easy way to get right initial setup parameters for every new design of shrink sleeve label.

Application of Data Mining Tools in Shrink Sleeve Labels …

105

7 Conclusions The application of data mining models in order to optimize the winding quality of the shrink sleeve labels has brought a positive effect. Pilot studies on historical data have shown that this methods may be useful in order to determine the optimal initial winding parameters in the converting process based on several important factors determining the converting process. Research has shown that computed models found dependences occurring in the process. The developed model with using of Boosted Trees tool can be used for predicting the initial winding parameters such as the rewinder and unwinder tension and winding speed. But other models such as Classification & Regression Trees and Artificial Neural Networks also reached very good prediction quality, measured as correlation coefficient and it was approximately 0.70–0.75 for unwinder and rewinder tension. For winding speed ANN model has significantly lower correlation coefficient than C&RT model. However the developed models can be used constantly to predict the winding parameters of each new production work order. Thus we need to bear in mind that in printing industry there are several new work orders every day, with different input parameters like web or sleeve width, type of material, ink amount or type, hot-melt layer etc., predicted model is able to perfectly find even complicated relationships and interactions in the process, and thus its use can be effective and measurable for the quality of the production process and the finished product. Unfortunately only models with using of Support Vector Machines method reached poor prediction quality. The main rationale for the use of data mining tools in order to optimize shrink sleeve labels converting process is their main characteristics and advantages in their ability to reproduce very complicated relationships occurring in the processes, which also was confirmed in this study. Additionally, an important feature of all used data mining tools in this paper is that variables don’t need to represent the normal distribution and may belong to different data types: continuous and discrete. The application of data mining models in order to optimize the winding quality of the shrink sleeve labels has brought a positive effect. Researches on historical data have shown that this method may be useful in order to determine the optimal initial winding parameters in the converting process based on several important factors determining the converting process. Studies has shown also that the Boosted Trees model found dependences occurring in the process, but further researches should focus on collecting data for all relevant factors, which are not included in current prediction models, as well as limiting and less important factors.

References 1. Demski, T.: Creating and using of Data Mining model with STATISTICA Data Miner Recipe on the example of fraud detection, content of the StatSoft website (2009). http://www.statsoft. pl/czytelnia.html

106

K. Krystosiak

2. Ejsmont, K., Krystosiak, K.: Manufacturing of shrink sleeve labels. In: Economics and Organization of Enterprise, pp. 53–61, 12/2014. Publishing House ORGMASZ, Warsaw (2014) 3. Ejsmont, K., Krystosiak, K., Lipiak, J.: Application of selected data mining technique in printing industry. In: Conference: Polish Association for Production Management, vol. II, pp. 75–86. Publishing House of The Polish Society of Production Management, Opole (2015) 4. Flexography: Principles and Practices, 5th edn. Foundation of Flexographic Technical Association, Inc. (1999) 5. Forum for Shrink Sleeve Technology Development in Poland. http://www.forumsleeve.pl/ 6. Friedmann, J.H.: Multivariate adaptive regression splines. Ann. Stat. 1–67 (1991) 7. Heat Shrink Sleeve Label Technical Manual & Test Methods. Technical Publication, AWA Alexander Watson Associates, Amsterdam (2014) 8. Kipphan, H.: Handbook of Print Media. Technologies and Production Methods. Springer, Germany (2001) 9. Kit, L.Y.: The Wiley Encyclopedia of Packaging Technology, 3rd edn. Wiley (2009) 10. Knosala, R.: Applications of Artificial Intelligence in Production Engineering. WNT, Warsaw (2002) 11. Krystosiak, K., Werpachowski, W.: The improvements of the quality level of packaging with overprint. In: Economics and Organization of Enterprise, pp. 55–64, nr 11/2013. Publishing House ORGMASZ, Warsaw (2013) 12. Krystosiak, K., Werpachowski, W.: Advanced data mining methods as a key to improvement of shrink sleeve labels production process. In: Conference: Product & Packaging—Contemporary Challenges 2014, Lodz University of Technology (2014) 13. Krystosiak, K., Werpachowski, W.: Control method of winding quality in shrink sleeve labels converting process. In: Computer Methods in Material Science, vol. 15, 3/2015. Publishing House AGH University of Science and Technology, Krakow (2015) 14. Krystosiak, K.: Impact study of artificial neural network topology on prediction quality of winding parameters. In: Opakowanie, pp. 60–64, 2/2016. Publishing House SIGMA-NOT, Warsaw (2016) 15. Krystosiak, K.: Prediction method for winding parameters in label converting process with data mining tools. In: 7th International Conference on Engineering, Project, and Production Management, Bialystok 2016, Procedia Engineering. Elsevier (2016) 16. Nisbet, R., Elder, J., Miner, G.: Handbook of Statistical Analysis & Data Mining Applications. Academic Press/Elsevier (2009). ISBN: 978-0-12-374765-5 17. Roisum D.R.: How to measure roll quality. Tappi J. 10/1988 (1988) 18. StatSoft Electronic Statistics Textbook, StatSoft. http://www.statsoft.pl/textbook/stathome. html 19. Tadeusiewicz, R.: Neural Networks. Academic Publishing House, Warsaw (1993) 20. Walker, T.J.: Stress & strain. In: Paper, Film & Foil Converter, nr 6/2009. Penton Media Publication (2009) 21. Walker, T.J.: What is the right tension? In: Paper, Film & Foil Converter, nr 12/2009. Penton Media Publication (2009) 22. Web Handling website. http://www.webhandling.com/

Study of Thickness Variability of the Floorboard Surface Layer Agnieszka Kujawinska, ´ Michał Rogalewicz, Magdalena Diering, ˙ Krzysztof Zywicki and Adam Hamrol

Abstract A characteristic feature of the wood machining processes is a high level of data uncertainty from these processes. This uncertainty is due to the influence of material environment, but also of the measurement system. On the one hand, the quality of the data is affected by the inhomogeneity of raw wood material, the uniqueness of its structure, the randomness of its natural defects, wood deformation and because of nature of wood, which depends on the conditions/environmental parameters. Moreover, the uncertainty of the measured data results from the engineering problems associated with the inaccuracy of cutting process and method of measuring the geometrical characteristics of wood. The subject of authors research is production process of the floorboard surface layer (lamella). Due to the material loss in the process of equalization of cut off wood plies, there is a need to redesign (develop) methodology of the rules to select material allowances. One of the stages of research leading to that is to describe statistical models which explain the nature of the observed variability of chosen woodworking operations. One of the critical features in the production process of the floorboard surface layer is its thickness. The paper discusses the engineering problems associated with the measuring method of this feature. Based on the observations and research results, a new methodology for measuring the thickness of the lamella is proposed. The study proposes a method and organization of measurement performance—the number and layout (net) of measuring points (grid points) on the surface of the lamella, the choice of measuring instruments, and the size and frequency of sampling. The proposed approach allowed to increase the usefulness of measurement results. According to the new methodology, date were acquired. Based on this—for chosen woodworking operations: cutting, drying and grinding—the statistical models of lamella thickness variation were built. The article describes these three models. These models will be used to decide whether to redesign a model of tolerance and the methodology of the

˙ A. Kujawi´nska (B) · M. Rogalewicz · M. Diering · K. Zywicki · A. Hamrol Faculty of Mechanical Engineering and Management, Poznan University of Technology, Pl. M. Sklodowskiej-Curie 5, 60-965 Pozna´n, Poland e-mail: [email protected] © Springer Nature Switzerland AG 2019 P. Grzegorzewski et al. (eds.), Soft Modeling in Industrial Manufacturing, Studies in Systems, Decision and Control 183, https://doi.org/10.1007/978-3-030-03201-2_7

107

108

A. Kujawi´nska et al.

rules to select material allowances for each operation. The authors formulate conclusions and recommendations to improve the methodology of the lamella thickness measuring. Keywords Raw wood · Floorboard surface layer (lamella) · Thickness measure Statistical models

1 Introduction The main source of data for the assesment of a manufacturing process are measurement results [1]. Manufacturing process data are characterised by a large degree of uncertainty, caused by inappropriate measurement tools, lack of specific measurement methods, human factors (operators’ errors), wrong sampling, etc. Minimalisation the uncertainty of measurement data facilitates the creation of a reliable process model. The model can be based on empirical statistical distribution, and support the process engineer in decision-making. This is the so-called soft modelling, contrary to hard modelling, which is based on relations taking into consideration the nature of the phenomenon [2–4]. Soft modelling is used especially for complex processes, such as the manufacturing of a three-layer floorboard [5]. The authors studied the manufacturing process of the surface layer (lamella). The results of the study are presented in this paper. Taking into consideration the user needs and industry standards, the critical feature of the floorboard surface layer is its thickness. However, one of the main difficulties in the assessment of thickness is its variability resulting from non-homogeneity of wood and sensitivity to external conditions, such as humidity. Therefore, the measurement system must be characterised by low variability, since the quality of measurement data determines the quality of the lamella thickness variability model. Decisions of operators, made on the basis of an unreliable model, translate into wrong business decisions at further stages of the value added chain [6]. They may lead to reduction of operational efficiency of the production process [5, 7] as well as to the customers’ dissatisfaction and claims, and even damage the brand’s reputation or the corporate image [1, 8]. The article discusses a model of variability of thickness of the surface layer manufactured in the wet cutting technology. The study was conducted at one Poland’s companies operating in the wood industry.

Study of Thickness Variability of the Floorboard Surface Layer

109

2 Study of Variability of the Floorboard Surface Layer Thickness 2.1 Wood Industry Wood industry is one of the major sectors of the Polish economy. In spite of a rapid growth in the use of new materials, wood retains its strong position of a widely used raw material. Its use is growing continuously, driving the increase of wood prices. The value of wood exports amounts to PLN 45bn annually and accounts for 10% of the total Polish exports. The wood industry in Poland creates more than 300 thousand jobs. An estimated 2 million people in Poland benefit from work in forestry and the wood sector, which creates ca. 2% of Poland’s GDP [9]. A common feature of all business entities marketing wood products is the structure of manufacturing costs. The key cost structure item is the cost of wood, which, for example in the sawmill industry, amounts to ca. 70% of total costs [10]. The structure of costs is similar in the manufacturing of multilayer floorboards, e.g. for the operations of cutting wood with frame, belt or disc saws [11]. Seeking to maintain competitiveness, business entities operating in the wood industry cannot transfer the ca. 8% annual growth in prices of the most popular types of timber in Poland (oak, pine) to their customers [9]. Therefore, one of the key challenges faced by the wood industry nowadays is curbing the costs of raw materials through optimizing the manufacturing processes and developing new technologies. In this respect, Poland’s wood industry is supported by the National Centre of Research and Development under the BIOSTRATEG programme. One of the main objectives of the programme is the development of knowledge on sustainable management of natural resources, forestry and wood industry. BIOSTRATEG is aimed to strengthen Poland’s position on the international arena of research and development in these areas. The results of research presented in this article have been developed under the R&D project “Increasing the effectiveness of use of wood in manufacturing processes”, EFFraWOOD in short. The project is carried out by the Poznan University of Technology, the Chair of Management and Production Engineering (in cooperation with a floorboard manufacturer), within the framework of the task entitled Optimization of technological allowances in the process of wet cutting, aimed to reduce the waste of raw materials in the manufacturing of the surface layer [12].

2.2 Timber, Waste of Raw Material Timber is a raw material obtained from cut trees. It is widely used in many industries, including the construction, paper, machine construction, chemical and pharmaceutical industries.

110

A. Kujawi´nska et al.

It is biodegradable. At the end of its life cycle, it decays without harming the environment or is used as fuel [13]. Timber owes its popularity in the manufacturing industry to its structure and the resulting unique benefits. Its advantages include high durability and low specific gravity. The durability of timber relative to its specific gravity is similar to that of steel. Timber also has a low level of thermal conductivity and low sound transmission properties. Owing to these properties, a wall made of wood may be a few times thinner than one made of brick, concrete or stone. The disadvantages of timber include high hygroscopicity and the related shrinking, cracking and warping, as well as difficult preservation. Timber is non-homogenous and anisotropic, it had variable density and features knots, splits, cracks, etc. [14, 15]. One the basic technologies in wood processing is mechanical processing (machining). During the processing, the manufacturer incurs the costs of wasting the material caused by excessive operational allowances. Large material allowances are secured for particular processing operations due to the defects which appear in the processing (e.g. surface irregularities, shape flaws after drying). On the one hand, the allowances are supposed to prevent the production of semi-manufactured items whose dimensions do not meet the specified requirements, on the other hand, however, they create additional costs. Therefore, manufacturing allowances should be economically justified and ensure the required geometric accuracy and quality of the product surface while maintaining minimum waste of material. Excessive waste of material also occurs in the manufacturing of the surface layer of floorboards. The costs of raw material have the largest share (more than 80%) in the total costs of manufacturing of the oak top layer of floorboards, while the costs of tools amount to 6% of the total costs for a frame saw and 5% for a belt saw [4, 11]. Therefore, it is important to identify the sources of losses and ways to minimise them. For this purpose, assessment of the manufacturing process must be performed with the use of soft modelling. The following chapters describe research on the creation of statistical models for the manufacturing of lamellas.

2.3 The Manufacturing Process of Multilayer Floorboards The manufacturing process of a three-layer floorboard was studied (Fig. 1). The top (face) layer is made of selected European or exotic timber of appropriate thickness and hardness (e.g. oak, beech). The middle layer, made of coniferous timber, is positioned transversely to the other two layers for reduced tension and natural deformation (swelling, creaking or gapping). The bottom layer is also made of soft coniferous timber. The three cross-glued layers are varnished for decorative appearance and resistance to mechanical impact. The floorboards are usually installed as floating flooring. The lamella is an especially important component of the floorboard. Its quality determines the aesthetic values of the face layer, and has a significant impact on the

Study of Thickness Variability of the Floorboard Surface Layer

111

TOP LAYER

MIDDLE LAYER

BOTTOM LAYER

Fig. 1 A three-layer floorboard

DRY TECHNOLOGY

WET TECHNOLOGY

1. Drying of timber, ca. 40 days

1. Cutting of wet timber

2. Cutting of dry timber

2. Drying of lamellas, ca. 2-4 days

3. Polishing - top and bottom

3. Polishing – top and bottom

4. Formatting - length and width

4. Formatting - length and width

Fig. 2 Lamella manufacturing technologies (Source own analysis, based on [12])

sound damping properties, maintenance of the room microclimate, durability and strength of the floor. The critical characteristic of lamella production process is thickness of the layer. In compliance with the standard, a cross-glued block may be classified as a floorboard if its surface layer is at least 2.5 mm thick [14]. A cross-glued block with the top layer thinner than 2.5 mm is classified as a panel (a product of lower quality class). The processing of timber layers is considered demanding in the industry sector [14, 15]. In the process under analysis, the lamella is manufactured in two technologies—wet and dry (Fig. 2).

112

A. Kujawi´nska et al.

CUTTING

DRYING

GRINDING

FORMATTINGLENGTH

FORMATTINGWIDTH

thickness length width

Fig. 3 Shaping geometric parameters of lamellas (Source own analysis, based on [12])

Irrespective of the processing technology used, the manufacturing process begins with delivering logs to the timber depot. There, logs are cut with saws into timber. Next, timber is cut into lamellas, or forms. The difference between dry and wet processing is that for dry processing, timber is dried for 40 days, on average, before being cut. In wet processing, on the other hand, fresh timber is cut into lamellas (forms), which are then dried for ca. 48 h (Fig. 2). Before the drying operation, cut lamellas are stacked in baskets which ensure free air flow. Dry lamellas are formed into the desired geometric dimensions (thickness, length and width) through grinding and adjusting the length and width. A ready lamella is sent to the block facility, where it is joined with the other elements (the middle and bottom layers). The three components are glued together, sanded and varnished to form the end product—the floorboard. The process of manufacturing lamellas in the wet technology accounts for only 20% of the company’s manufacturing output. Research aimed to mitigate the process variability and optimise the manufacturing technology in order to increase the share of wet technology in the total output and minimise the waste of material is discussed also in [16], and more detailed presented in further part of the paper. In the wet technology, thickness of the lamella is formatted in three operations—the cutting of timber into lamellas, the drying and the grinding (Fig. 3). A statistical thickness variability model was developed for each of the three operations.

2.4 Methodology 2.4.1

Stages of the Study Methodology

As part of the study aimed to improve the wet technology, a process variability model was developed and further used as a basis for redesigning the operational allowances. The study methodology comprised four stages (Fig. 4). At stage one, the methodology of measurement of lamella thickness was selected, based on the permissible duration of measurement and accuracy—the two criteria determined by the study team. At stage two, the sample size was selected. At stages three and four, data was acquired and analysed, and the process variability model developed. At stage four, conclusions were drawn and a set of recommendations developed.

Study of Thickness Variability of the Floorboard Surface Layer

113

Fig. 4 Study methodology

STAGE 1

SelecƟon of the measurement method; Measurement System Analysis (MSA)

STAGE 2

SelecƟon of the sample size

STAGE 3

Data collecƟon

Variability (quality capability) analysis

STAGE 4

Material waste analysis

Conclusions

2.4.2

Measurement System Analysis, Measurement Methodology

Many factors may have an impact on variability of a measurement system. In the case under review, the study team identified the following key components of the measurement system: operators (individuals performing the measurements), measurement/assessment method (including the number and distribution of measurement positions), measurement tools and control station, external conditions (humidity, lighting, noise), and specification. The measurement system analysis (MSA) was aimed to identify the impact of repeatability and reproducibility on the measurement process, as well as the impact of the measurement process on the picture of the manufacturing process. The main objective at that stage was to develop the measurement methodology and approve it for application in further studies. This was done in four stages: 1. Monitoring of the measurement process (getting to know the process and the external conditions). 2. Diagnosis of the current state—an analysis of the measurement system in place (selection of the procedure and specification of the study design (number of parts, number of operators, number of series of repetitions); preparation of the study stages; collection of data; analysis and conclusions). 3. Recommendations concerning alterations of the measurement system and development of the measurement methodology for the study.

114

A. Kujawi´nska et al.

4. Validation of the methodology—examination and approval of the altered measurement system for study purposes. A statistical analysis of the measurement system in place was performed by the Average Range Method (ARM) [17, 18], in which the %R&R (repeatability and reproducibility) index is determined. A value of the %R&R index below 10% indicates high utility of the measurement system, while a value of more than 30% indicates that the measurement system is useless. The analysis was performed for a set of 3 operators, 2 series of measurements, 10 parts and 6 measurement points on each lamella. The number and positions of measurement points resulted from the measurement method in place. The measurements were performed in real-life conditions, i.e. during daily operation. Results of the analysis showed that the system used for the measurement of lamella thickness was useless—values of the %R&R index for each measurement exceeded 30%. It means that variability was largely affected by the measurement system. As a result, information about the manufacturing process state obtained through measurements was unreliable and might have led to wrong decisions about the product and/or the manufacturing process. Based on observations and results of the analysis, the study team proposed a new methodology of measurement of the floorboard surface layer thickness. The manner and organisation of the measurements (the method), i.e. the number and distribution of measurement positions (point grid) on the surface of lamellas, the selection of measurement tools, and the sample size were determined [12]. In order to develop recommendations concerning the positions and number of measurement points for study purposes, maps of lamella thickness were developed for particular classes of wood and classes of dimensions. The maps were created on the basis of an optical measurement performed with a 3D GOM scanner. Application of an optical scanner facilitated the determination of areas of significantly different variability of the key feature (Fig. 5). A statistical analysis of maps was used for the development of recommendations concerning the number and position of measurement points for particular dimension groups of lamellas. Figure 6 presents an example map and distribution of measurement points. With the number and positions of measurement points determined, the procedure of marking the features (identification) of each piece of timber and lamella measured was developed. Templates were prepared for positioning the measurement points. The number of measurement points was 9 or 17, depending on the lamella dimension group. A digital external measurement gauge was selected for the measurement tool. With the proposed alterations implemented, validation of the new methodology was performed [12]. The measurement system variability components were estimated for each measurement position. The analysis proved that the new measurement methodology could be used for study purposes—each measurement point obtained an acceptable value of the %R&R index (Table 1).

Study of Thickness Variability of the Floorboard Surface Layer

115

Fig. 5 Example distribution of lamella thickness (legend: blue—5 mm, red—5.6 mm) Table 1 Measurement system analysis results MSA for the measurement system in place (diagnosis)

MSA for the altered measurement system

Number of operators

3

3

Number of parts studied

10

10

Number of series (trials)

2

3

Number of measurement position

6

9 (or 17)

Measurement position

Not marked on the lamella

Marked on the lamella (template for measurement point positioning)

Measurement tool

Calliper 300/0.01 mm

External measurement digital gauge 50/0.02 mm

Measurement method

No standards for measurements performed by different operators

Specific measurement method

Results—%R&R index values More than 30% (the measurement system is useless)

Ca. 10–15% (the measurement system is useful)

116

A. Kujawi´nska et al.

125

41,6 41,6

41,6

41,6

250

125

125

41,6

41,6

1225

122,5

367,5

612,5

857,5

1102,5

Fig. 6 Example result of thickness map analysis for lamellas 5.04 × 250 × 2450 mm

Study of Thickness Variability of the Floorboard Surface Layer Table 2 Requirements for thickness for particular operations Material Sample size Reference value [items] [mm] Cutting—input

Timber

Cutting—output

Rough, wet lamella Drying—output Rough, dry lamella Grinding—output Dry, polished lamella

2.4.3

117

Tolerance [mm]

100

30

+1/−2

500

5.04

±0.2

3.0

±0.2

500 500

Sample Size and Data Acquisition

Proper selection of the sample size is one of the key elements of study design. The sample size depends on many technical and organisational factors, such as the budget (costs), duration of the study, availability of studied objects, etc. The statistical requirements must also be taken into consideration for reliability of the study conclusions. A compromising solution was adopted for the selection of sample size for the study. The sample included 100 pieces of timber, 20 pieces of each of the five timber classes (I–V). Each piece of timber was cut into 5 lamellas. Timber classes are groups of particular features which comply with the manufacturer’s technical standards. The features include the type of wood, dimensions of the cross section of the log and the acceptable irregularities. Measurements were made during the real-life manufacturing process, at the beginning and end of each operation in which the lamella thickness is shaped. The requirements concerning the key feature for each operation are presented in Table 2.

2.4.4

Results and Discussion

An analysis of thickness of the input material for the cutting operation (timber) showed that the material supplied by the sawmill is on average 1 mm thicker than the nominal value (Fig. 7a). To find out if the variability of the various classes of timber is equal test for equality of variances was conducted in Minitab software using Levene’s test and multiple comparisons method [19]. They were chosen because do not require the data to follow normal distribution and can be used for any continuous probability distribution. If the p-value for these tests is low (p < 0.05) the hypothesis about the equality of variances should be refused. Comparing confidence intervals for standard deviation makes it possible to check which variability is different (confidence intervals for standard deviations differing from each other do not overlap). The results showed that particular classes of timber have different variances (Fig. 7a). Class II proved to be significantly different in terms of the variability (Fig. 7b)—its variability was much bigger than that in other wood classes. The thickness of input

118

A. Kujawi´nska et al.

(a)

(b)

Fig. 7 Boxplot and test for equal variances for timber thickness output (Source own analysis in Minitab software)

material of this class is a feature of low predictability and it can have an influence on subsequent operations of production process.

Study of Thickness Variability of the Floorboard Surface Layer

119

Fig. 8 Boxplot of lamella thickness after individual operations (Source own analysis in Minitab software)

Figure 8 shows a box-plot of the lamella thickness variability model, by operations and class of wood. It can be observed in Fig. 8 that the operation of cutting has the strongest impact on the variability of thickness. Variability of lamella thickness after the operation of drying is a derivative of that after cutting (in the drying process, a lamella shrinks in thickness by 6–8% [11]). The mean thickness is clearly shifted towards the upper tolerance limit, what may imply that operators on purpose raise the setting of the saw for the nominal value (in the operation of cutting). An analysis of the grinding operation showed that the process has a high potential—its variability relative to the tolerance limits is low, but it requires a shift towards the nominal value (down). What is important, the grinding operation makes up for the irregularities resulting from the previous operations. In conclusion, the company produces lamellas which meet the requirements of customers, paying a high price through substantial waste of material caused by the variability of cutting and drying operations. Therefore, an action plan to mitigate the variability of lamella thickness obtained in the cutting operation would make it possible to redesign (minimise) the operational allowances and reduce the raw material consumption. Results of a study of the cutting operation are presented below. The main goal of this analysis was to compare the variability and mean value of thickness after cutting taking into consideration various classes of wood. The more repeatable and “reaching the target” is this operation the better input for next operations of the production process. As following a normal distribution is a main assumption for

120

A. Kujawi´nska et al.

Fig. 9 Probability plot for thickness after cutting for I-V class (Source own analysis in Minitab software) Table 3 Anderson-Darling normality test results (Source own analysis in Minitab software) Class of wood Mean StDeviation AD statistic p-value I II III IV V

5.207 5.153 5.218 5.200 5.298

0.27 0.34 0.32 0.24 0.25

22.923 25.135 23.108 20.192 21.208

Soft Modeling in Industrial Manufacturing

Recommend Stories

Idea Transcript

Helpful Links

Smile Life

Get in touch