검색 상세

One-Class Classification-Based Control Charts for Multivariate Process Monitoring and Diagnosis

  • TuerhongGulanbaier (정보경영공학전문대학원 정보경영공학과 산업시스템전공)

초록/요약

The objective of this dissertation is to develop novel nonparametric control charts that can efficiently and effectively monitor multivariate processes through their integration with one-class classification algorithms. Statistical process control (SPC) provides a set of statistical tools to reduce process variations. A control chart is the most widely used SPC tool because it is not only based on fundamental statistical theory, but also because of its user-friendly graphical display of the behavior of a quality output. Traditional SPC techniques are challenged by the large volume of complex data encountered in many modern industries. To relax the limitations, data mining algorithms can be integrated with SPC approaches, because of their proven capabilities to resolve challenging problems in SPC. In particular, one-class classification algorithms, which are one type of these proven data mining algorithms, can be directly integrated with control charts because both assume only in-control target data are available to measure the degree of abnormality of an in-coming unclassified observation. This dissertation has three main parts. First, we propose a novel nonparametric multivariate control chart, called the hybrid novelty score (HNS)-based chart that integrates novelty score algorithms with control charts. The monitoring statistics of the HNS charts are calculated based on novelty score algorithms. The control limits of the HNS charts are determined by a bootstrap percentile technique. Second, we present a comparative study of various novelty score-based multivariate control charts through vigorous simulation study under various scenarios. The monitoring statistics are obtained from various novelty score algorithms. The control limits are estimated via a percentile bootstrap technique. Third, we propose a novel multivariate control chart based on Gower distance that can efficiently handle the processes characterized by high-dimensional and mixture of continuous and categorical data.

more

목차

1. Introduction 1
1.1 Statistical Process Control 1
1.2 Data Mining Algorithms 7
1.3 One-Class Classification Algorithms 8
1.4 Motivation and Contribution 9
1.5 Outline of This Dissertation 10
2. Multivariate Control Charts Based on Hybrid Novelty Scores 13
2.1 Introduction 13
2.2 HNS-Based Control Chart 20
2.2.1 Monitoring Statistics of an HNS Control Chart 20
2.2.2 Control Limits of an HNS Control Chart 23
2.3 Simulation Study 25
2.3.1 Simulation Setup 25
2.3.2 Performance Comparisons 31
2.4 Case Study 36
2.5 Exponentially Weighted HNS Charts 37
2.6 Conclusions 42
3. Comparison of Novelty Score-Based Multivariate Control
Charts 44
3.1 Introduction 44
3.2 Novelty Score as Monitoring Statistics in Multivariate
Control Charts 48
3.2.1 T2 48
3.2.2 Local T2 49
3.2.3 Dmax Novelty Score 50
3.2.4 Dmean Novelty Score 50
3.2.5 K2 Novelty Score 51
3.2.6 K Nearest Neighbor Data Description 51
3.2.7 Local Distance-Based Outlier Factor 52
3.2.8 Hybrid Novelty Score 53
3.3 Design of Novelty Scores-Based Multivariate Control
Charts- Control Limits 55
3.4 Simulation Study 58
3.4.1 Simulation Setup 58
3.4.2 Performance Comparisons 68
3.5 Conclusions 75
4. Gower Distance-Based Multivariate Control
Charts for HighDimensional and Mixture Data 77
4.1 Introduction 77
4.2 Proposed Gower Distance-Based Multivariate
Control Charts for Mixture Data 83
4.2.1 Gower’s Dissimilarity Coefficient 83
4.2.2 Monitoring Statistics Based on
Gower Distance 84
4.2.3 Control Limits 88
4.3 Simulation Study 90
4.3.1 Simulation Setup 90
4.3.2 Performance Comparisons 93
4.4 Real Case Study 104
4.5 Conclusions 107
5. Summary and Future Study 108
Reference 112
Acknowledgement 122

more
반출 Meta View 목록