Control Chart Pattern Recognition Method Based on Improved One-dimensional Convolutional Neural Network

9th IFAC Conference on Manufacturing Modelling, Management and 9th IFAC Conference on Manufacturing Modelling, Management and Control 9th IFAC Conference on Manufacturing Modelling, Management and Control 9th IFAC Conference on Manufacturing Modelling, Management and online at www.sciencedirect.com Berlin, Germany, August 28-30, 2019 Available Control 9th IFAC IFAC Conference on Manufacturing Manufacturing Modelling, Management and 9th Conference on Modelling, Management and Berlin, Germany, August 28-30, 2019 Control Berlin, Germany, August 28-30, 2019 Control Control Berlin, Germany, August 28-30, 2019 Berlin, Berlin, Germany, Germany, August August 28-30, 28-30, 2019 2019

ScienceDirect

IFAC PapersOnLine 52-13 (2019) 1537–1542

Control Chart Pattern Recognition Method Based on Control Chart Pattern Recognition Method Based on Control Chart Pattern Recognition Method Based on Improved One-dimensional Convolutional Neural Network Control Chart Pattern Recognition Method Based on Control Chart Pattern Recognition Method Based on Improved One-dimensional Convolutional Neural Network ControlOne-dimensional Chart Pattern Recognition Method Based on Improved Convolutional Neural Network Improved One-dimensional Convolutional Neural Network Improved One-dimensional Convolutional Neural Network Jie Xu*. One-dimensional Huichun Lv*. Zilong Zhuang*. Zhiyao Lu*. Dewei Zou*.Network Wei Qin* Improved Convolutional Neural

Jie Jie Xu*. Xu*. Huichun Huichun Lv*. Lv*. Zilong Zilong Zhuang*. Zhuang*. Zhiyao Zhiyao Lu*. Lu*. Dewei Dewei Zou*. Zou*. Wei Wei Qin* Qin* Jie Xu*. Huichun Lv*. Zilong Zhuang*. Zhiyao Lu*. Dewei Zou*. Wei Qin* Qin* Jie Xu*. Huichun Lv*. Zilong Zhuang*. Zhiyao Lu*. Dewei Zou*. Wei * School of Engineering, Shanghai Jiao Zhiyao Tong University, Shanghai, China, CO 200240 JieMechanical Xu*. Huichun Lv*. Zilong Zhuang*. Lu*. Dewei Zou*. Wei Qin* * of Engineering, Shanghai University, Shanghai, China, * School School of Mechanical Mechanical Shanghai Jiao Jiao Tong Tongluzhiyao, University, Shanghai, China, CO CO 200240 200240 (e-mail: {xujie0822,Engineering, huichun, zhuangzilong910126, gritzou, wqin}@sjtu.edu.cn) * School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China, CO 200240 200240 (e-mail: {xujie0822, huichun, zhuangzilong910126, luzhiyao, gritzou, wqin}@sjtu.edu.cn) * School School of Mechanical Mechanical Engineering, Shanghai Jiao Jiao Tong Tongluzhiyao, University, Shanghai, China, CO CO (e-mail: {xujie0822,Engineering, huichun, zhuangzilong910126, gritzou, wqin}@sjtu.edu.cn) * of Shanghai University, Shanghai, China, 200240 (e-mail: {xujie0822, {xujie0822, huichun, huichun, zhuangzilong910126, zhuangzilong910126, luzhiyao, luzhiyao, gritzou, gritzou, wqin}@sjtu.edu.cn) wqin}@sjtu.edu.cn) (e-mail: (e-mail: {xujie0822, huichun, zhuangzilong910126, luzhiyao, gritzou, wqin}@sjtu.edu.cn) Abstract: The application of statistical process control (SPC) has promoted production quality Abstract: The application of process control (SPC) has production quality Abstract: The application of statistical statistical control (SPC)chart hasis promoted promoted production quality improvement of many enterprises. As a coreprocess tool of SPC, control used to reflect the production Abstract: The application of statistical process control (SPC) has promoted production quality improvement of many enterprises. As a core tool of SPC, control chart is used to reflect the production Abstract: The application of statistical process control (SPC) has promoted production improvement of many enterprises. As a coreprocess tool in of the SPC, control chart is promoted used the production state. In addition to normal pattern, abnormalities production process can to be reflect summarized inquality seven Abstract: The application of statistical control (SPC) has production quality improvement of many enterprises. aa core tool of SPC, control chart is used the production state. In addition addition to normal normal pattern,As abnormalities in the production process can to be reflect summarized in seven seven improvement of many enterprises. core tool of SPC, control chart is to reflect the production state. to pattern, abnormalities in the process can be summarized in basic In control chart patterns (CCPs).As The recognition of production CCPs is helpful toused identify quality failures and improvement of many enterprises. As a core tool of SPC, control chart is used to reflect the production state. In addition to normal pattern, abnormalities in the production process can be summarized in seven basic control chart patterns (CCPs). The recognition of CCPs is helpful to identify quality failures and state. In addition to normal abnormalities in the process be summarized in seven basic control chart patterns (CCPs). The recognition of production CCPs is (CNN) helpful to acan identify quality and find root abnormal causes inpattern, time. Convolutional neural network is classical modelfailures in the field state. In addition to normal pattern, abnormalities in the production process can be summarized in seven basic control chart patterns (CCPs). The recognition of CCPs is helpful to identify quality failures and find root abnormal causes in time. Convolutional neural network (CNN) is a classical model in the field basic control chart patterns The of is helpful to quality failures and find root abnormal causes in(CCPs). time. Convolutional neural network is classical model in the field of deep learning. can automatically extract features from the raw data, so the operation of basic control chart CNN patterns (CCPs). The recognition recognition of CCPs CCPs is (CNN) helpful to aaidentify identify quality failures and find root abnormal causes in time. Convolutional neural network (CNN) is classical model in the field of deep learning. CNN can automatically extract features from the raw data, so the operation of find root abnormal causes in time. Convolutional neural network (CNN) is a classical model in the field of deep learning. CNN can automatically extract features from the raw data, so the operation of constructing manual features can be omitted. In this paper, the one-dimensional CNN is applied to the find root learning. abnormal causes in time. Convolutional neural network (CNN) is a classical model in thetofield of deep CNN can automatically extract features from the raw data, so the operation of constructing manual features can be omitted. In this paper, the one-dimensional CNN is applied the of deep can automatically extract features from the raw so operation of constructing features be omitted. In this paper, the one-dimensional is applied to the recognition ofmanual CCPsCNN and achieves 98.96% average recognition accuracy 30 data, tests.CNN What’s more, even if of deep learning. learning. CNN can can automatically extract features from the in raw data, so the the operation of constructing manual features can be omitted. In this paper, the one-dimensional CNN is applied to the recognition of CCPs and achieves 98.96% average recognition accuracy in 30 tests. What’s more, even if constructing manual features can be omitted. In this paper, the one-dimensional CNN is applied to the recognition of CCPs and achieves 98.96% average recognition accuracy in 30 tests. What’s more, even if there is a deviation between the distribution of test data and training data, the model still shows excellent constructing manual features candistribution be omitted. In this paper, theaccuracy one-dimensional CNN isshows applied to the recognition of CCPs and achieves 98.96% average recognition in 30 tests. What’s more, even there is a deviation between the of test data and training data, the model still excellent recognition of performance. CCPsbetween and achieves achieves 98.96% average average recognition accuracy inthe 30 model tests. What’s What’s more, even if if there is a deviation the distribution of test recognition data and training data,in still shows excellent generalization recognition of CCPs and 98.96% accuracy 30 tests. more, even if there is a deviation between the distribution of test data and training data, the model still shows excellent generalization performance. there is a deviation between the distribution of test data and training data, the model still shows excellent generalization performance. there is a deviation between the distribution of test data and training data, the model still shows excellent generalization performance. © 2019, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. generalization performance. generalization performance. © Federation of Control) Hosting by Ltd. reserved. © 2019, 2019, IFAC IFAC(International (International Federation of Automatic Control) Hosting by Elsevier Ltd.rights All rights reserved. © 2019, IFAC (International Federation of Automatic Automatic Control) Hosting by Elsevier Elsevier Ltd. All All rights reserved. © 2019, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: Statistical process control, Control chart, Pattern recognition, Convolutional neural network, © 2019, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. © 2019, IFAC (International Federation of Automatic Control) Hosting by ElsevierConvolutional Ltd. All rights reserved. Keywords: Statistical process control, Control Pattern recognition, neural Keywords: process control, Control chart, chart, Pattern recognition, Convolutional neural network, network, Monte CarloStatistical simulation Keywords: Statistical process control, Control chart, Pattern recognition, Convolutional neural network, Monte Carlo simulation Keywords: Statistical process control, Control chart, Pattern recognition, Convolutional neural Monte Carlo simulation Keywords: Statistical process control, Control chart, Pattern recognition, Convolutional neural network, network, Monte Carlo simulation Monte Monte Carlo Carlo simulation simulation becomes smaller. STR pattern may be caused by 1. INTRODUCTION becomes STR may becomes smaller. smaller. STR pattern pattern may be be caused caused by by computational errors (Zaman et al., 2018). 1. INTRODUCTION 1. INTRODUCTION becomes smaller. STR pattern may be caused by computational errors (Zaman et al., 2018). becomes smaller. STR pattern may be caused by computational errors (Zaman et al., 2018). INTRODUCTION Product quality is a 1. key factor in the market competition of becomes smaller. STR pattern may be caused by 1. INTRODUCTION computational errors (Zaman et al., 2018). (4) Cyclic pattern: Periodic occurrence of peaks and troughs 1. INTRODUCTION Product quality is a key factor in the market competition of computational errors (Zaman et al., 2018). Product quality is a keyprocess factor incontrol the market of computational enterprises. Statistical (SPC)competition has brought errors (Zaman et al., 2018). (4) Cyclic pattern: Periodic occurrence of peaks and troughs (4) Cyclic pattern: Periodic occurrence and troughs can be found in cyclic pattern (Addehofetpeaks al., 2018). CYC Product quality is aa key factor incontrol the market competition of enterprises. Statistical process (SPC) has Product quality is key factor the market competition of enterprises. Statistical process control (SPC) has ofbrought brought significant improvements to thein production quality many (4) Cyclic pattern: Periodic occurrence of peaks and troughs can be found in cyclic pattern (Addeh et al., 2018). CYC Product quality is a key factor in the market competition of (4) Cyclic pattern: Periodic occurrence of peaks and troughs can be found in cyclic pattern (Addeh et al., 2018). CYC pattern may be related to system cycle changes such as enterprises. Statistical process control (SPC) has brought significant improvements to the production quality of many (4) Cyclic pattern: Periodic occurrence of peaks and troughs enterprises. Statistical process control (SPC) has brought significant improvements to the production quality of many enterprises, covering the industries of machining, chemical, can be found in cyclic pattern (Addeh et al., 2018). CYC pattern may be related to system cycle changes such as enterprises. Statistical process control (SPC) has brought can be found in cyclic pattern (Addeh et al., 2018). CYC pattern may be related to system cycle changes such as voltage fluctuations, periodically executed tasks, etc. significant improvements to the production quality of many enterprises, covering the industries of chemical, be fluctuations, found in related cyclic pattern (Addeh ettasks, al., 2018). CYC significant improvements to the the production production quality oftomany many enterprises,improvements covering industries of machining, machining, electronic, etc. The the fundamental idea of quality SPC ischemical, use can pattern may be to system cycle changes such as voltage periodically executed etc. significant to of pattern may be related to system cycle changes such voltage fluctuations, periodically executed tasks, etc. enterprises, covering the industries of machining, chemical, electronic, etc. The fundamental idea of SPC is to use pattern may be related to system cycletasks, changes such as as enterprises, covering industries of machining, electronic, etc. The the fundamental idea of various SPC ischemical, to use mathematical statistics methods to monitor stages of (5) voltage fluctuations, periodically executed etc. Trend pattern: In trend patterns, the data shows a enterprises, covering the industries of machining, chemical, voltage fluctuations, periodically executed tasks, etc. electronic, etc. The fundamental idea of SPC is to use mathematical statistics methods to monitor various stages of voltage fluctuations, periodically executed tasks, etc. (5) Trend pattern: In trend patterns, the data shows electronic, etc. The fundamental idea of SPC is to use mathematical statistics methods to monitor various stages of production process, so the production anomalies can be continuous (5) Trend rise pattern: In named trend patterns, the and data downward shows aa or fall, upward trend electronic, etc. The fundamental idea of SPC is to use mathematical statistics to monitor various stages of production process, somethods the anomalies can (5) Trend pattern: In trend patterns, the data shows a continuous rise or named upward trend and downward mathematical statistics to monitor various stages of production process, the production production can be be detected timely andsomethods measures can beanomalies implemented to trend (5) Trend pattern: In trend the data shows continuous rise or fall, fall, upward trend and respectively. The emergence of trend may beaa mathematical statistics methods to can monitor various stages of (5) Trend pattern: In named trend patterns, patterns, the patterns data downward shows production process, so the production anomalies can be detected timely and measures be implemented to continuous rise or fall, named upward trend and downward trend respectively. The emergence of trend patterns may be production process, so the production anomalies can be detected timely and measures can be implemented to eliminate potential hazards. continuous rise orwear, fall, named upward trend andofdownward downward trend respectively. The operator emergence of trend patterns may be related to rise tool or fatigue, lack effective production process, someasures the production anomalies can be continuous fall, named upward trend and detected timely and can be implemented to eliminate potential hazards. trend respectively. The emergence of trend patterns may be related to wear, operator fatigue, lack of detected timely and measures can be implemented to eliminate potential hazards. trend respectively. The emergence of trend patterns may related to tool tool wear, operator fatigue, lack of effective effective supervision, etc. detected timely and measures can be implemented to trend respectively. The emergence of trend patterns may be be eliminate potential hazards. Control chart is a core tool of SPC. Western Electric related to tool wear, operator fatigue, lack of effective supervision, etc. eliminate potential hazards. related to tool operator fatigue, lack of effective supervision, etc. wear, eliminate potential hazards. Control chart is a core tool of SPC. Western Electric related to tool wear, operator fatigue, lack of effective Control is has a core tool of 15 SPC. Western Electric (6) Companychart (1956) summarized control chart patterns supervision, etc. Shift pattern: The shift patterns appear as a sudden rise or supervision, etc. Control chart is aa core tool of SPC. Western Electric Company (1956) has summarized 15 control chart patterns supervision, etc. The (6) Shift pattern: shift patterns appear as sudden rise Control chart is core tool of SPC. Western Electric Company (1956) has summarized 15 control chart patterns (CCPs), including 8 basic patterns, while others such as (6) Shift pattern: The shiftnamed patterns appearshift as aaand sudden rise or or fall in the mean of data, upward downward Control chart is a core tool of SPC. Western Electric Company (1956) has summarized 15 control chart patterns (CCPs), including 8 basic patterns, while others such as (6) Shift pattern: The shift patterns appear as a sudden rise or fall in the mean of data, named upward shift and downward Company (1956)orhas has summarized 15 while control chart mixture, patterns (CCPs), including 8 basic patterns, others such as shift freaks, grouping bunching, instability, interaction (6) Shift pattern: The shift patterns appear as a sudden rise or fall in the mean of data, named upward shift and downward respectively. Shift patterns may be caused by the Company (1956) summarized 15 control chart patterns (6) Shift pattern: The shiftnamed patterns appear as acaused sudden risethe or (CCPs), including 8 basic patterns, while others such as freaks, grouping or bunching, instability, interaction mixture, fall in the mean of data, upward shift and downward shift respectively. Shift patterns may be by (CCPs), including 8 basic patterns, while others such as freaks, grouping or bunching, instability, interaction mixture, etc. are a including mixture or8special case of these basic patterns. The fall in the mean of data, named upward shift and downward shift respectively. Shift patterns may be caused by the introduction of new operators, materials, methods, minor (CCPs), basic patterns, while others such as fall in the mean of data, named upward shift and downward freaks, grouping or bunching, instability, interaction mixture, etc. are a mixture or special case of these basic patterns. The shift respectively. Shift patterns may be methods, caused by the introduction of operators, materials, minor freaks, or bunching, instability, mixture, etc. are grouping a mixtureare special(NOR), case of cyclic theseinteraction basic patterns. The faults 8 basic patterns normal (CYC), systematic shift respectively. patterns may caused the introduction of new new operators, minor of devices, etc.Shift freaks, grouping oror bunching, instability, interaction mixture, shift respectively. Shift patternsmaterials, may be be methods, caused by by the etc. are aa mixture or special case of these basic patterns. The 8etc. basic patterns are normal (NOR), cyclic (CYC), systematic introduction of new operators, materials, methods, minor faults of devices, etc. are mixture or special case of these basic patterns. The 8etc. basic patterns are normal (NOR), cyclic (CYC), systematic (SYS), stratification (STA), uptrend (UT), downtrend (DT), introduction of new operators, materials, methods, minor faults of devices, etc. are a mixture or special case of these basic patterns. The introduction of new operators, materials, methods, minor 8(SYS), basic patterns are normal (NOR), cyclic (CYC), systematic stratification (STA), uptrend (UT), downtrend (DT), faults of devices, etc. abnormal CCPs the production process are usually 8(SYS), basic patterns are normal (NOR), cyclic (CYC), systematic (STA), uptrend (UT), downtrend (DT), The upward shift (US), and downward shift (DS) patterns (Gauri faults of devices, devices, etc. in 8(SYS), basic stratification patterns are and normal (NOR), cyclic (CYC), systematic faults of etc. The abnormal CCPs in the production process are stratification (STA), uptrend (UT), downtrend (DT), upward shift (US), downward shift (DS) patterns (Gauri The abnormal CCPs in thecauses. production process are usually usually correspond to some specific Hence, the recognition of (SYS), stratification (STA), uptrend (UT), downtrend (DT), upward shift (US), and downward shift (DS) patterns (Gauri et al., 2009). Fig. 1 shows an example of 8 basic CCPs. A (SYS), stratification (STA), uptrend (UT), downtrend (DT), The abnormal CCPs in the production process are usually correspond to some specific causes. Hence, the recognition of upward shift (US), and downward shift (DS) patterns (Gauri et al., 2009). Fig. 1 shows an example of 8 basic CCPs. A The abnormal CCPs in the production process are usually correspond to some specific causes. Hence, the recognition of abnormal patterns can not only help to find the problems in upward shift (US), and downward shift (DS) patterns (Gauri et al., 2009). Fig. 1 shows an example of 8 basic CCPs. A brief introduction of the 8 basic CCPs is stated below. The abnormal CCPs in the production process are usually upward shift (US), and downward shift (DS) patterns (Gauri correspond to some specific causes. Hence, the recognition of abnormal patterns can not only help to find the problems in et al., 2009). Fig. 1 shows an example of 8 basic CCPs. A brief introduction of the 8 basic CCPs is stated below. correspond to some specific causes. Hence, the recognition of abnormal patterns can not only help to find the problems in time, but also narrow the scope of abnormal causes. et al., 2009). Fig. 1 shows an example of 8 basic CCPs. A brief introduction of the 8 basic CCPs is stated below. correspond to some specific causes. Hence, the recognition of et al.,introduction 2009). Fig.of1 the shows an example of 8 basic CCPs. A abnormal patterns can not only help to find the problems in time, but also narrow the scope of abnormal causes. brief 8 basic CCPs is stated below. (1) Normal pattern: The normal pattern indicates the abnormal patterns can not only help to find the problems time, but also narrow the scope of abnormal causes. brief introduction of the the The 8 basic basic CCPs is ispattern stated below. below. abnormal patterns canthe notscope onlyofhelp to findcauses. the problems in in brief introduction of 8 CCPs stated (1) Normal pattern: normal indicates the time, but also narrow abnormal In the early period of control chart applications, human (1) Normalprocess pattern: The normal pattern indicates the time, but also narrow the scope of abnormal causes. production is in-control. time, but also narrow the scope of abnormal causes. In the early period of control chart applications, human (1) Normal pattern: The normal pattern indicates the production process is in-control. In the earlyis period of to control applications, human required judgechart whether the production (1) Normal pattern: The production is in-control. (1) Normalprocess pattern: The normal normal pattern pattern indicates indicates the the experience In the early period of control chart applications, human experience is required to judge whether the production production process is in-control. (2) Systematic pattern: The systematic pattern appears as a process In the early period of control chart applications, human experience is required to judge whether the production is abnormal and find the corresponding cause. With production process is in-control. the early period of find control chart applications, human production process is in-control. (2) Systematic pattern: The systematic pattern appears as aa In experience is required to judge whether the production process is abnormal and the corresponding cause. With (2) Systematic pattern: The systematic pattern appears as high point always follows a low point and vice versa, so the the experience is required to judge whether the production process is abnormal and find the corresponding cause. With development of industrial automation, the rule-based is required to judge whether the production (2) Systematic pattern: The systematic pattern appears as aa experience high point always follows aa low point and vice versa, so the process is abnormal and find the corresponding cause. With the development of industrial automation, the rule-based (2) Systematic pattern: The systematic pattern appears as high point always follows low point and vice versa, so the point-to-point fluctuations can be predicted (Zhou et al., process is abnormal and find the corresponding cause. With the development of industrial automation, the rule-based discriminant system partially replaces the role of manual (2) Systematic pattern: The systematic pattern appears as a process is abnormal and find the corresponding cause. With high point always follows aa can low point and vice versa, so the point-to-point fluctuations be predicted (Zhou et al., the development of industrial automation, the rule-based discriminant system partially replaces the role of manual high point always follows low point and vice versa, so the point-to-point fluctuations can be predicted (Zhou et al., 2018). the development of partially industrial automation, the of rule-based discriminant system replaces the role manual observation. The discriminant rules of control chart are based high point always follows a low point and vice versa, so the the development of industrial automation, the rule-based point-to-point fluctuations can be predicted (Zhou et al., 2018). discriminant system partially replaces the role of manual observation. The discriminant rules of control chart are based point-to-point fluctuations 2018). discriminant system partially replaces the role of manual observation. The discriminant rules control chart based small probability events and canofbe easily implemented. point-to-point fluctuations can can be be predicted predicted (Zhou (Zhou et et al., al., on system partially replaces the role ofare manual 2018). (3) Stratification pattern: The stratification pattern shows that discriminant observation. The discriminant rules of control chart are based on small probability events and can be easily implemented. 2018). observation. The discriminant rules of control chart are on small probability events and can be easily implemented. it The is difficult to cover allofabnormal patterns with 2018). (3) pattern: The pattern shows that observation. discriminant rules control chart are based based (3) Stratification Stratification The stratification stratification shows that However, the data is morepattern: concentrated and the pattern variance of data on small probability events and can be easily implemented. However, it is difficult to cover all abnormal patterns on small probability events and can be easily implemented. However, itthe is complexity difficult to of cover all abnormal patterns with with rules due to production process. (3) Stratification pattern: The stratification pattern shows that the data is more concentrated and the variance of data on small probability events and can be easily implemented. (3) Stratification The stratification shows that the data is morepattern: concentrated and the pattern variance of data However, it is difficult to cover all abnormal patterns with rules due to the complexity of production process. (3) Stratification pattern: The stratification pattern shows that However, it is difficult to cover all abnormal patterns rules due to the complexity of production process. the itthe is complexity difficult to of cover all abnormal patterns with with the data data is is more more concentrated concentrated and and the the variance variance of of data data However, rules due to production process. the data is more concentrated and the variance of data rules the of process. Copyright 2019 IFAC 1557Hosting rules due due to the complexity complexity of production production process. 2405-8963 © 2019, IFAC (International Federation of Automatic Control) by to Elsevier Ltd. All rights reserved. Copyright 2019 responsibility IFAC 1557Control. Peer review© of International Federation of Automatic Copyright ©under 2019 IFAC 1557 Copyright © 2019 IFAC 1557 10.1016/j.ifacol.2019.11.418 Copyright © 2019 IFAC 1557 Copyright © 2019 IFAC 1557

2019 IFAC MIM 1538 Berlin, Germany, August 28-30, 2019

Jie Xu et al. / IFAC PapersOnLine 52-13 (2019) 1537–1542

In response to the problems of rule-based discriminant systems, researchers attempt to mine the specific features of different CCPs. The advent of the era of big data provides a data foundation for the CCPs recognition. At the same time, the development of information technology and artificial intelligence technology provides a technical basis.

Fig. 1. Basic control chart patterns. The rest of this paper is organized as follows. The existing CCPs recognition methods are reviewed in section 2. Section 3 introduces the data simulation method. The model of the improved one-dimensional convolutional neural network (1D-CNN) proposed in this paper is presented in section 4. Section 5 shows the recognition accuracy and generalization ability of the improved 1D-CNN model. The conclusion is showed in section 6. 2. RELATED WORK Many methods have been applied to the recognition of CCPs. Some methods directly feed the raw CCPs data into the recognition model, while others first use statistical knowledge to extract features from raw data, such as mean square amplitude, standard deviation, peak, average, etc. Then these extracted features are fed into the model. Pham et al. (1997) propose that different CCPs have different geometric characteristics, so it is helpful to improve CCPs recognition accuracy by constructing features. However, there are lots of different features that can be constructed, making it a challenge to select the appropriate feature subset so as to reduce the computational complexity and improve the classification accuracy of the model. Gauri et al. (2007) have applied classification and regression tree (CART) to select feature subset. Addeh et al. (2018) apply the association rules (AR) to select the best feature subset. But some of the features selected by these methods are still highly correlated, leading to redundancy. Thus, some dimension reduction methods such as independent component analysis (ICA) (Lu et al., 2011) and principal component analysis (PCA) (Tai-Fu et al., 2012) are used to reduce redundancy between features. After determining the input form of data, whether it is raw data or extracted feature data, a model will be built to realize

accurate recognition of different CCPs. Some classical methods such as fuzzy inference system (FIS), support vector machine (SVM), artificial neural network (ANN) have been applied in CCPs recognition systems. Gulbay et al. (2007) propose a fuzzy method for recognition of unnatural CCPs. Zaman et al. (2018) combine the fuzzy cmean (FCM) with adaptive neuro-fuzzy inference system (ANFIS) to realize the CCPs recognition and get comparable classification accuracy. Ebrahimzadeh et al. (2011) apply SVM method to CCPs recognition for its excellent generalization performance. However, it is difficult to select appropriate parameters for SVM. Therefore, some optimization algorithms such as genetic algorithm (GA) (Zhao et al., 2017) and particle swarm optimization (PSO) (Yongman et al., 2013) are used to automatically optimize the parameters of SVM model. In addition, many studies have introduced ANN into the CCPs recognition. Cheng et al. (1997) construct a modular neural network using a multilayer perceptron (MLP) trained by back-propagation (BP) algorithm and realize the recognition of two kinds of CCPs. The CCPs recognition module designed by Ghomi et al. (2011) combines two types of neural network, one is learning vector quantization (LVQ) and another is MLP. Considering the continuity of control chart data over time, Awadalla et al. (2012) apply the spiking neural network (SNN) to the CCPs recognition. Based on existing researches, it can be found that the CCPs recognition method based on feature extraction usually have better performance. But the construction of features depends on human experience, thus, feature screening methods are needed to select the best feature subset. Deep learning has been extensively studied for its outstanding performance. Deep learning establishes a mapping from inputs to outputs by a network structure (Zhang et al., 2018). As a multi-level representation learning method, the deep learning model is composed by simple but non-linear modules of which can transform a lower level representation to a higher and slightly more abstract level representation (Lecun et al., 2015). The structure of deep learning makes it can automatically extract features from raw data. There are many typical deep learning methods such as deep belief network (DBN), deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN) and etc. CNN is a classical deep learning model and has been applied in many fields, such as image recognition (Krizhevsky et al., 2012), speech recognition (Hinton et al., 2012), fault diagnosis (Zilong et al., 2018) and so on. In recent years, CNN has been used to recognize onedimensional (1D) signals. Kiranyaz et al. (2016) apply the 1D-CNN to the real-time classification of patient-specific electrocardiogram (ECG). Malek et al. (2017) propose a new method to analysis the chemometric data based on 1D-CNN. Wen et al. (2018) convert one-dimensional bearing vibration signal into two-dimensional (2D) image format, and use 2DCNN for bearing fault diagnosis. The literatures above show that CNN can extract features from raw data and has great advantages in the processing of complex classification

1558

2019 IFAC MIM Berlin, Germany, August 28-30, 2019

Jie Xu et al. / IFAC PapersOnLine 52-13 (2019) 1537–1542

problems. It is the reason why this paper attempts to use CNN for recognition of CCPs. 3. DATA SIMULATION Data simulation technology is the most widely used method in the field of recognition of CCPs. With the improvement of automatic production capacity, enterprises have accumulated a lot of data. However, it is difficult to train models with these data for it takes lots of efforts to classify and label the data. Fortunately, once enough data of an in-control production process have been collected, some information such as the distribution, mean estimate, and variance estimate of data can be obtained. Then the data simulation process can be carried out to generate data of various patterns.

1539

In this paper, Monte Carlo method is used for data simulation. Table 1 shows the formulas and parameter settings of the simulation process. Here, 𝜇𝜇 and 𝜎𝜎 are the mean and standard deviation estimate of the controlled production process, respectively, 𝑟𝑟(𝑡𝑡) represents the inevitable accidental fluctuation which subject to gaussian distribution 𝑁𝑁(0,1), 𝑑𝑑 indicates the degree of system state departure, 𝑎𝑎 is the amplitude of cyclic pattern, 𝑇𝑇 is the period of the cycle, 𝑔𝑔 is the gradient of data trend, 𝑃𝑃 represents the time point when the shift anomaly occurs, and 𝑠𝑠 is the amplitude of shift pattern. The setting range of the above parameters is referred to Gauri et al. (2006). The sequence length of simulation data 𝐿𝐿 should not be too long, because longer window width of data means larger lag of anomaly recognition. Generally, the window length is set to be 16-64 sampling points.

Table 1. Parameters and formulas of data simulation Class 0 1 2 3 4 5

Description Normal, NOR Cyclic, CYC Systematic, SYS Stratification, STR Upward Trend, UT Downward Trend, DT

Equations 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 + 𝑎𝑎 sin(2𝜋𝜋𝜋𝜋/𝑇𝑇) 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 + 𝑑𝑑 × (−1)𝑡𝑡 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 ′ 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 + 𝑡𝑡 × 𝑔𝑔 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 − 𝑡𝑡 × 𝑔𝑔 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 + 𝑘𝑘 × 𝑠𝑠 𝑘𝑘 = 1 𝑖𝑖𝑖𝑖 𝑡𝑡 ≥ 𝑃𝑃, 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 𝑘𝑘 = 0

Remarks 𝜇𝜇 = 0,𝜎𝜎 = 1 𝑟𝑟(𝑡𝑡)~𝑁𝑁(0,1) 𝛿𝛿 = 1𝜎𝜎 𝛿𝛿 ′ ∈ (0.2𝜎𝜎, 0.4𝜎𝜎) 𝑑𝑑 ∈ (1𝜎𝜎, 3𝜎𝜎) 𝑎𝑎 ∈ (1.5𝜎𝜎, 2.5𝜎𝜎) 𝑇𝑇 = 16 6 Upward Shift, US 𝑔𝑔 ∈ (0.05𝜎𝜎, 0.25𝜎𝜎) 𝑃𝑃 ∈ (10, 20) 𝑦𝑦𝑡𝑡 = 𝜇𝜇 + 𝑟𝑟(𝑡𝑡) × 𝛿𝛿 − 𝑘𝑘 × 𝑠𝑠 𝑠𝑠 ∈ (1𝜎𝜎, 3𝜎𝜎) 7 Downward Shift, DS 𝑘𝑘 = 1 𝑖𝑖𝑖𝑖 𝑡𝑡 ≥ 𝑃𝑃, 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 𝑘𝑘 = 0 𝑡𝑡 = 1,2, … , 𝐿𝐿 connection and weight sharing greatly reduce the number of 4. METHODS training parameters in the network, leading to lower complexity and better generalization ability. While pooling 4.1 A brief Introduction of CNN operation reduces the number of neurons in the model and The structure of CNN can be divided into two parts. The first enhances the robustness. part is composed of convolutional layers and pooling layers to realize feature extraction, and the second part is composed 4. 2 Architecture of the Improved 1D-CNN The network structure proposed in this paper is shown in Fig. of full connection layers for final classification. 2. The first layer of the network is an input layer. The length Each convolutional layer contains several filters to extract of simulation data is set to 32, which means that the local features from the feature maps of previous layer, dimension of the input layer is 32. respectively, and followed by an activation unit to generate the output feature maps. The weights of filters to calculate Next comes a triple parallel convolutional module, known as each unit of a feature map are same, which is called weight ‘inception’ (Szegedy et al., 2015). The filter sizes of three sharing. The depth of feature maps is determined by the parallel convolutional structures are 1*1, 1*3, 1*5, number of filters. The performance of CNN is affected by the respectively. The stride of the filters in each convolutional size and number of filters in per layer, so the setting of the structure is 1 and the number of filters is 5. Filters of different filters in convolutional layer needs to be well designed to sizes can extract multi-scale information from raw data. It should be noted that if these three parallel convolutional adapt to specific problems. structures are fed with the same input, the size of output The pooling layer is usually set after the convolutional layer, feature maps will be different because of various filters. which is used to reduce the size of extracted feature maps. Therefore, paddings are needed to keep the output in the Max pooling and average pooling are two most widely used same size. For convolutional structure with filters of size 1*3, pooling methods. it is necessary to add one zero before and after the input sequence. Similarly, two zeros need to be padded in the The fully connected layer and the softmax layer are usually convolutional structure with the filters of size 1*5. set at the end of network structure to output the probability that the input sample belongs to each category. Then, all the feature maps obtained in ‘inception’ layer need to be concatenated and be fed into an averaging pooling layer. The typical characteristics of CNN can be summarized as Feature maps obtained after pooling are fed into the next follows, local connection, weight sharing, pooling operation convolutional layer which has 10 filters of size 1*1 and the and multi-layer structure (Lecun et al., 2015). Local 1559

2019 IFAC MIM 1540 Berlin, Germany, August 28-30, 2019

Jie Xu et al. / IFAC PapersOnLine 52-13 (2019) 1537–1542

stride is 1. Then another pair of averaging pooling layer and convolutional layer with the same parameters is followed. The activation function used in convolutional layers is the Rectified Linear Units (RELU) function which has showed great performance in recent years. The action mode of RELU can be expressed as follows, the neuron’s output 𝑓𝑓 can be calculated from the input 𝑥𝑥 by formula 𝑓𝑓(𝑥𝑥) = 𝑚𝑚𝑚𝑚𝑚𝑚 (0, 𝑥𝑥). As a non-saturating nonlinearity function, RELU has been proved to achieve faster convergence rate than other activation function such as sigmoid and tanh (Krizhevsky et al., 2012).

1*1

Conv1D F1

Conv1D 1*3

Average Pooling

F10 Conv1D

Feature Maps 400*8*10 1*2 F1

F10

Feature Maps 400*32*15

Feature Maps 400*2*10

Input Conv1D 1*1

1*2

F1

Feature Maps 400*16*15

F5 Conv1D F1 F5 Inception

1*2

Average Pooling

1*3

1*5

The CNN structure proposed in this paper introduces an ‘inception’ structure to extract information of different scales and the model is named as improved 1D-CNN.

1*2

F5 Conv1D F1

Input 400*32

The convolutional layers showed above mainly realize the function of automatically extracting the features from raw data, and then the fully connected layer is set to realize classification. Before connecting to the fully connected layer, the feature maps need to be flattened by a flatten layer so that the feature maps can be fed into the fully connected layer which is set at the end of the network structure. The softmax function is used as the activation function of the output layer to implement multi-category output.

Conv1D 1*5

Conv1D

Feature Maps 400*4*10

Flatten 400*20 Fully Connected Output 400*8

Concatenate

Inception Structure

Fig. 2. Architecture of the improved 1D-CNN model. 5. 1 Experiment I: Recognition effects of different models The data of 8 kinds of CCPs used in this experiment are Two experiments are carried out to verify the effectiveness of generated by the simulation method mentioned in section 3. improved 1D-CNN model for CCPs recognition. What’s Each sample length is 32, and 500 pieces of samples of each more, several different neural network models, which are pattern are generated, 350 of which are used for training, 50 presented in Table 2, have been used for comparative for validation and 100 for testing. Considering the experiments. The recognition effects of different models are randomness of data simulation, the experiments are repeated compared in experiment I and the generalization performance for 30 times to obtain the average recognition accuracy of each model. The recognition effects of various methods are of improved 1D-CNN is tested in experiment II. summarized in Table 3. Table 2. Structure of the models used for CCPs recognition Model ANN 1L-1D-CNN 2L-1D-CNN 3L-1D-CNN Improved 1D-CNN Input 1*32 1*32 1*32 1*32 1*32 Conv 1D Conv 1D Conv 1D Conv 1D Conv 1D Conv 1D Dense Layer 1 (40) (1*2, 10) (1*2, 10) (1*2, 10) (1*1, 5) (1*3, 5) (1*5, 5) Average Pooling Average Pooling Average Pooling Dense Layer 2 Concatenate (40) (pool size=2) (pool size=2) (pool size=2) Conv 1D Conv1D Dense Conv 1D Layer 3 Flatten (8) (1*2, 10) (1*2, 10) (1*2, 10) Average Pooling Average Pooling Layer 4 Dense (8) Average Pooling (pool size=2) (pool size=2) (pool size=2) Layer 5 Flatten Conv 1D (1*2, 10) Conv1D (1*2, 10) Layer 6 Dense (8) Flatten Average Pooling (pool size=2) Layer 7 Dense (8) Flatten Layer 8 Dense (8) Parameters 3288 678 408 538 748 5. EXPERIMENTS

1560

2019 IFAC MIM Berlin, Germany, August 28-30, 2019

Jie Xu et al. / IFAC PapersOnLine 52-13 (2019) 1537–1542

Table 3. Recognition effects of various models Model ANN 1L-1D-CNN 2L-1D-CNN 3L-1D-CNN Accuracy (%) 96.32 98.64 98.37 98.34 Standard deviation 1.9 0.56 0.52 0.63 Compared with ANN model, CNN has higher recognition accuracy and fewer model parameters. All the models based on CNN show great recognition performance, while the improved 1D-CNN with ‘inception’ structure achieves the highest recognition accuracy, reaching 98.96%. What’s more, the standard deviation of recognition accuracy of improved 1D-CNN is also smaller than other methods in 30 times tests, which shows the good stability of recognition. The result indicates that the 1D-CNN model combined with parallel ‘inception’ structure is more suitable for recognition of time series data for its ability of multi-scale information extraction. The confusion matrix of model recognition accuracy of the improved 1D-CNN is showed in Fig. 3.

1541

Improved 1D-CNN 98.96 0.32

Fig. 4. Recognition accuracy of improved 1D-CNN and ANN when the standard deviation of training data and test data are inconsistent.

Fig. 3. Confusion matrix of recognition accuracy of improved 1D-CNN model. 5. 2 Experiment II: Generalization Performance of improved 1D-CNN Model Even huge amounts of data are obtained, it is still impossible to get the true mean and true standard deviation of the process data. There must be a deviation between the estimate value and the true value. Therefore, experiment is carried out to measure the generalization ability of the model when the mean and standard deviation of the training data and the test data are inconsistent. According to the data simulation method provided in section 3, 500 pieces of data are generated for each CCP. The generating parameters of 400 pieces of data used for model training are set unchanged, that is, the estimate mean value is 0 and the estimate standard deviation is 1. While the mean value of test data is set to -0.5 to 0.5 with an interval of 0.1, the standard deviation value of test data is set to 0.5 to 1.5 with the same interval. Fig. 4 shows the test accuracy of improved 1D-CNN and ANN when the true mean value of test data is fixed as 0 and the true standard deviation value of test data is set between 0.5 and 1.5. Fig. 5 shows the test accuracy of improved 1D-CNN and ANN when the estimate standard deviation value of test data is fixed as 1 and the true mean value of test data is set between -0.5 and 0.5. 30 times repeated tests are implemented for different combination of parameters to get the average accuracy of CCPs recognition.

Fig. 5. Recognition accuracy of improved 1D-CNN and ANN when the mean of training data and test data are inconsistent. The results showed in Fig. 4 and Fig. 5 indicate that when the test data and training data distribution are inconsistent, the prediction accuracy of the model will inevitably decrease. However, the prediction accuracy of the improved 1D-CNN model is still significantly better than that of ANN. It can be found that the model still has high recognition accuracy for different CCPs when the error between the mean or standard deviation estimates value and true value is small, which indicates the excellent generalization performance of the improved 1D-CNN model. 6. CONCLUSIONS The model of improved 1D-CNN proposed in this paper has achieved an average recognition accuracy of 98.96% on the simulation dataset after 30 repeated tests, which is a satisfactory result. The result shows that CNN with an introduction of ‘inception’ structure achieves higher recognition accuracy than purely layer-by-layer CNN. Besides, the improved 1D-CNN model also has better generalization ability when there is an error between the estimate value and true value of mean or standard deviation.

1561

2019 IFAC MIM 1542 Berlin, Germany, August 28-30, 2019

Jie Xu et al. / IFAC PapersOnLine 52-13 (2019) 1537–1542

Though many methods have been applied to the recognition of CCPs, different data simulation methods and parameter settings used in different studies reduce the significance of comparing these results. But in general, the improved 1DCNN model obtains better recognition accuracy when compared with other models which based on raw data and reaches the same level of accuracy when compared with the model which based on artificial designed features. What’s more, thanks to the function of automatic feature extraction of CNN, there is no need to design features by human experience. ACKNOWLEDGMENTS The authors would like to acknowledge financial supports of the National Science Foundation of China (No. 51775348, No. U1637211) and Shanghai Aerospace Science and Technology Innovation Fund (No. SAST2016048). REFERENCES Addeh, A., Khormali, A., and Golilarz, N. A. (2018). Control chart pattern recognition using RBF neural network with new training algorithm and practical features. ISA Transactions, 79, 202-216. Awadalla, M. H., and Sadek, M. A. (2012). Spiking neural network-based control chart pattern recognition. Alexandria Engineering Journal, 51(1), 27-35. Cheng, and C.-S. (1997). A neural network approach for the analysis of control chart patterns. International Journal of Production Research, 35(3), 667-697. Ebrahimzadeh, A., and Ranaee, V. (2011). High efficient method for control chart patterns recognition. Acta technica ČSAV, 56(1), 89-101. Gauri, S. K., and Chakraborty, S. (2006). Feature-based recognition of control chart patterns. Computers and Industrial Engineering, 51(4), 726-742. Gauri, S. K., and Chakraborty, S. (2007). A study on the various features for effective control chart pattern recognition. The International Journal of Advanced Manufacturing Technology, 34(3-4), 385-398. Gauri, S. K., and Chakraborty, S. (2009). Recognition of control chart patterns using improved selection of features. Computers and Industrial Engineering, 56(4), 1577-1588. Ghomi, S. M. T. F., Lesany, S. A., and Koochakzadeh, A. (2011). Recognition of unnatural patterns in process control charts through combining two types of neural networks. Applied Soft Computing, 11(8), 5444-5456. Gulbay, M., and Kahraman, C. (2007). Development of fuzzy process control charts and fuzzy unnatural pattern analyses. Computational Statistics and Data Analysis, 51(1), 434-451. Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A. R., and Jaitly, N., et al. (2012). Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups. IEEE Signal Processing Magazine, 29(6), 82-97. Kiranyaz, S., Ince, T., and Gabbouj, M. (2016). Real-time patient-specific ECG classification by 1-d convolutional

neural networks. IEEE Transactions on Biomedical Engineering, 63(3), 664-675. Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. International Conference on Neural Information Processing Systems, 25, 1090-1098. Lecun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444. Lu, C. J., Shao, Y. E., and Li, P. H. (2011). Mixture control chart patterns recognition using independent component analysis and support vector machine. Neurocomputing, 74(11), 1908-1914. Malek, S., Melgani, F., and Bazi, Y. (2017). One ‐ dimensional convolutional neural networks for spectroscopic signal regression. Journal of Chemometrics, 32(5), e2977. Pham, D. T., and Wani, M. A., (1997). Feature-based control chart pattern recognition. International Journal of Production Research, 35(7), 1875-1890. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., and Anguelov, D., et al. (2015). Going Deeper with Convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1-9. Tai-Fu, L. I., Sheng, H. U., Zheng-Yuan, W., and Ya-Jun, H. (2012). PCA-SVM for control chart recognition of genetic optimization. Application Research of Computers, 29(12), 4538-4541. Wen, L., Gao, L., Li, X., Xie, M., and Li, G. (2018). A new data-driven intelligent fault diagnosis by using convolutional neural network. IEEE International Conference on Industrial Engineering and Engineering Management. 813-817. Western Electric Company. (1956). Statistical quality control handbook. Western Electric Company, US. Yongman, Z., Zhen, H., Shuguang, H., and Min, Z. (2013). Support vector machine based on particle swarm optimization for monitoring mean shift signals in multivariate control charts. Journal of Tianj in University: Science and Technology, 46(5), 469-475. Zaman, M., and Hassan, A., (2018). Improved statistical features-based control chart patterns recognition using anfis with fuzzy clustering. Neural Computing and Applications, (4), 1-15. Zhang, W., Yang, G., Lin, Y., Gupta, M. M., and Ji, C. (2018). On definition of deep learning. World Automation Congress (WAC), 232-236. Zhao, C., Wang, C., Hua, L., Liu, X., Zhang, Y., and Hu, H. (2017). Recognition of control chart pattern using improved supervised locally linear embedding and support vector machine. Procedia Engineering, 174, 281-288. Zhou, X., Jiang, P., and Wang, X. (2018). Recognition of control chart patterns using fuzzy SVM with a hybrid kernel function. Journal of Intelligent Manufacturing, 29(1), 51-67. Zilong, Z., and Wei, Q. (2018). Intelligent fault diagnosis of rolling bearing using one-dimensional multi-scale deep convolutional neural network based health state classification. IEEE 15th International Conference on Networking, Sensing and Control (ICNSC), 1-6.

1562