Inhalt

•  
   Abstract
•  
   Zusammenfassung
•  
   Acknowledgments
•  
   Contents
•  
   Notation
•  
   1 Introduction
•  
   1.1 Goal and Focus of Thesis
•  
   1.2 Challenges of Thesis
•  
   1.3 Contributions
•  
   1.4 Summary of Contributions and Thesis Outline
•  
  2 Theoretical Background and Related Work
•  
   2.1 Machine Health Monitoring
•  
   2.1.1 Tool Condition Monitoring
•  
   2.1.2 Imbalance Detection
•  
   2.2 Signal Segmentation
•  
   2.2.1 Piecewise Linear Approximation
•  
   2.2.2 Clustering-based Approaches
•  
   2.2.3 Changepoint Approaches
•  
  2.3 Modeling Non-Stationary Frequency Components
•  
   2.3.1 Parameter Estimation
•  
   2.3.2 Frequency Component Tracking
•  
   2.4 Anomaly Detection
•  
   2.4.1 Shallow Models
•  
   2.4.2 Deep Models
•  
   2.5 Annotations by Human Users
•  
   2.6 Weakly Supervised Learning
•  
   2.7 Summary
•  
   I Task-Specific Machine Monitoring Features and Models
•  
   3 Signal Segmentation
•  
   3.1 Motivation
•  
  3.2 Methods
•  
   3.2.1 Modeling Recurrent Signal Segments with Gaussian Mixture Models
•  
   3.2.2 Bayesian Estimation of Recurrent Signal Segments
•  
  3.3 Experiments on Signal Segmentation
•  
   3.3.1 Data for Signal Segmentation
•  
   3.3.2 Signal Segmentation by Clustering-based Methods
•  
   3.3.3 Quality and Cost of Signal Segmentation
•  
   3.3.4 Signal Segmentation by Bayesian Online Changepoint Detection and Extensions
•  
   3.3.5 Selected Predictive Tasks
•  
   3.4 Conclusions
•  
   3.5 Related Publications
•  
   4 Modeling Non-Stationary Discrete Frequency Components
•  
   4.1 Motivation
•  
   4.2 Methods
•  
   4.2.1 Estimation of Signal Model Parameters
•  
   4.2.2 Discrete Frequency Component (DFC) Tracking
•  
   4.3 Results
•  
   4.3.1 Noise Variations for Artificial Data
•  
   4.3.2 DFC Tracking and Assignment for Measured Sensor Data
•  
   4.4 Conclusions
•  
   II Low-Cost Annotation and Robust Detection of Generic Machine Tool Anomalies
•  
   5 User Study: Quality of Live Annotations and Influencing Factors
•  
   5.1 Motivation
•  
   5.2 Measurement Setup
•  
  5.3 Description of the Visualization and Labeling Prototype
•  
   5.3.1 Design Process of the Labeling Prototype
•  
   5.3.2 Functionality of the Labeling Prototype
•  
  5.4 Assumptions on Evaluation Measures
•  
   5.4.1 Assumptions on Measures for Quality of Label Feedback
•  
   5.4.2 Assumptions on Measures for Annotator Motivation
•  
  5.5 Experiments
•  
   5.5.1 Selection of a Generic Anomaly Detection Algorithm
•  
   5.5.2 Evaluation of Label Feedback
•  
   5.6 Conclusions
•  
   5.7 Related Publications
•  
   6 Neural Anomaly Detection
•  
   6.1 Motivation
•  
  6.2 Methods
•  
   6.2.1 Loss Functions
•  
   6.2.2 Network Layers
•  
   6.2.3 Training and Hyperparameter Optimization
•  
   6.2.4 Label Generation via Probabilistic Graphical Models (PGMs)
•  
   6.3 Results
•  
   6.3.1 Experimental Setup
•  
   6.3.2 Anomaly Detection with Unsupervised Models
•  
   6.3.3 Utilizing Labels for Anomaly Detection Model Extensions
•  
   6.3.4 Anomaly Propositions with Neural Anomaly Detection Models
•  
   6.4 Conclusions
•  
   7 Summary
•  
   7.1 Summary of Contributions
•  
   7.2 Conclusions and Outlook
•  
  A Appendix for User Study (Chapter 5)
•  
   A.1 Original Version of Labeling Prototype Screens
•  
  B Appendix for Neural Anomaly Detection (Chapter 6)
•  
  B.1 Encoder Networks
•  
   B.1.1 Multilayer Perceptron (MLP) Encoder
•  
   B.1.2 Fully Convolutional Network (FCN) Encoder
•  
   B.1.3 Convolutional Encoder
•  
   B.1.4 Temporal Convolutional Network (TCN) Encoder
•  
   B.2 Decoder Networks
•  
   B.2.1 Multilayer Perceptron (MLP) Decoder
•  
   B.2.2 Convolutional Decoder
•  
   B.3 Variational Autoencoder (VAE) Projection Network
•  
   B.4 Training of Neural Anomaly Detection Models
•  
   B.5 Optimization of Hyperparameters
•  
   C List of Figures
•  
   D List of Tables
•  
   Bibliography