Statistical Process Monitoring Using Advanced Data-Driven and Deep Learning Approaches: Theory and Practical Applications

✍ Scribed by Fouzi Harrou, Ying Sun, Amanda S. Hering, Muddu Madakyaru, abdelkader Dairi

• Publisher: Elsevier
• Year: 2020
• Tongue: English
• Leaves: 322
• Edition: 1
• Category: Library

✦ Synopsis

Statistical Process Monitoring Using Advanced Data-Driven and Deep Learning Approaches tackles multivariate challenges in process monitoring by merging the advantages of univariate and traditional multivariate techniques to enhance their performance and widen their practical applicability. The book proceeds with merging the desirable properties of shallow learning approaches – such as a one-class support vector machine and k-nearest neighbours and unsupervised deep learning approaches – to develop more sophisticated and efficient monitoring techniques.

Finally, the developed approaches are applied to monitor many processes, such as waste-water treatment plants, detection of obstacles in driving environments for autonomous robots and vehicles, robot swarm, chemical processes (continuous stirred tank reactor, plug flow rector, and distillation columns), ozone pollution, road traffic congestion, and solar photovoltaic systems.

✦ Table of Contents

Cover
Statistical Process
Monitoring using
Advanced Data-Driven
and Deep Learning
Approaches:
Theory and Practical Applications
Copyright
Contents
Preface
Acknowledgments
1 Introduction
1.1 Introduction
1.1.1 Motivation: why process monitoring
1.1.2 Types of faults
1.1.3 Process monitoring
1.1.4 Physical redundancy vs analytical redundancy
1.2 Process monitoring methods
1.2.1 Model-based methods
1.2.2 Knowledge-based methods
1.2.3 Data-based monitoring methods
1.3 Fault detection metrics
1.4 Conclusion
References
2 Linear latent variable regression (LVR)-based process monitoring
2.1 Introduction
2.2 Development of linear LVR models
2.2.1 Full rank methods
2.2.1.1 Ordinary least squares regression
2.2.1.2 Ridge regression (RR)
2.2.2 Latent variable regression (LVR) models
2.2.2.1 Principal component analysis
Feature extraction with PCA
Criteria for selecting the number of principal components to use
2.2.2.2 Principal component regression
2.2.2.3 Partial least squares
2.3 Dynamic LVR models
2.4 Process monitoring methods
2.4.1 Univariate chart for process monitoring
2.4.1.1 Shewhart-based monitoring scheme
2.4.1.2 Cumulative sum (CUSUM)-based monitoring schemes
2.4.1.3 Exponentially weighted moving average (EWMA) schemes
2.4.1.4 Generalized likelihood ratio (GLR) hypothesis testing approach
2.4.2 Distribution-based process monitoring schemes
2.4.2.1 Kullback-Leibler-based monitoring scheme
2.4.2.2 Hellinger-based monitoring scheme
2.4.2.3 Limitations of univariate monitoring schemes
2.4.3 Multivariate process monitoring schemes with parametric and nonparametric thresholds
2.4.3.1 Multivariate Shewhart schemes
2.4.3.2 Multivariate cumulative sum scheme (MCUSUM)
2.4.3.3 Multivariate exponentially weighted moving average scheme (MEWMA)
2.5 Linear LVR-based process monitoring strategies
2.5.1 Conventional LVR monitoring statistics
2.5.1.1 Hotelling's T2 statistic
2.5.1.2 Q statistic or squared prediction error (SPE)
2.5.2 Fault isolation
2.5.2.1 Fault isolation using modified contribution plots
T2 contribution approach
SPE contribution approach
2.5.2.2 Fault diagnosis using RadViz visualizer
2.6 Cases studies
2.6.1 Simulated example
2.6.2 Monitoring influent measurements at water resource recovery facilities
2.7 Discussion
References
3 Fault isolation
3.1 Introduction
3.1.1 Pitfalls of standardizing data
3.1.2 Shortcomings of contribution plots/scores
3.2 Fault isolation
3.2.1 Variable thinning
3.2.2 Iterative traditional isolation
3.2.2.1 Mason-Young-Tracy method
3.2.2.2 Murphy method
3.2.2.3 Artificial neural network methods
3.2.2.4 Discussion
3.2.3 Variable selection methods
3.2.3.1 Phase I variable selection
3.2.3.2 Phase II variable selection
3.3 Fault classification
3.4 Fault isolation metrics
3.4.1 Fault isolation errors
3.4.2 Precision and recall
3.4.3 Phase I FI metrics
3.4.4 Discussion
3.5 Case studies
3.5.1 Retrospective fault isolation
3.5.2 Real-time fault isolation
3.6 Further reading
References
4 Nonlinear latent variable regression methods
4.1 Introduction
4.2 Limitations of linear LVR methods for process monitoring
4.3 Developing nonlinear LVR methods for process monitoring
4.3.1 Nonlinear partial least squares
4.3.1.1 Polynomial PLS modeling algorithm
4.3.2 ANFIS-PLS modeling framework
4.3.2.1 Nonlinear PLS-based monitoring
4.3.3 Kernel PCA
4.3.4 Kernel principal components analysis (KPCA) model
4.3.5 KPCA-based fault detection procedures
4.4 Cases study: monitoring WWTP
4.4.1 Anomaly detection using KPCA-OCSVM method
4.5 Simulated synthetic data
4.5.1 Application of plug flow reactor
4.5.1.1 Data generation and modeling
4.5.1.2 Detection results
4.5.1.3 Case (A) - abrupt anomaly detection
4.5.1.4 Case (B) - intermittent anomaly detection
4.5.1.5 Case (B) - drift anomaly detection
4.6 Discussion
References
5 Multiscale latent variable regression-based process monitoring methods
5.1 Introduction
5.2 Theoretical background of wavelet-based data representation
5.2.1 Wavelet transform
5.2.2 Multiscale representation of data using wavelets
5.2.3 Advantages of multiscale representation
5.2.3.1 Decorrelating autocorrelated measurements
5.2.3.2 Data are closer to normality at multiple scales
5.3 Multiscale filtering using wavelets
5.3.1 Single scale filter method
5.3.2 Multiscale filtering methods
5.3.3 Advantages of multiscale denoising
5.4 Wavelet-based multiscale univariate monitoring techniques
5.4.1 An illustrative example
5.4.1.1 Impact of autocorrelated data on the conventional Shewhart chart
5.4.1.2 Effect of measurement noise on the conventional Shewhart chart
5.4.1.3 Impact of the violation of normality assumption on the conventional Shewhart chart
5.5 Multiscale LVR modeling
5.5.1 Benefits of multiscale denoising in LVR modeling
5.6 Multiscale LVR modeling
5.7 Results and discussions
5.7.1 Application with synthetic data
5.7.1.1 Simulation results: synthetic data
5.7.1.2 Simulation results: distillation column
5.7.2 Application of monitoring distillation column
5.8 Discussion
References
6 Unsupervised deep learning-based process monitoring methods
6.1 Introduction
6.2 Clustering
6.2.1 Partition-based clustering techniques
6.2.1.1 k-Means clustering
6.2.2 Hierarchy-based clustering techniques
6.2.2.1 BIRCH (hierarchical)
6.2.2.2 Agglomerative clustering
6.2.3 Density-based approach
6.2.3.1 Mean shift clustering
6.2.3.2 k-Nearest neighbor clustering
6.2.4 Expectation maximization
6.3 One-class classification
6.3.1 One-class SVM
6.3.2 Support vector data description (SVDD)
6.4 Deep learning models
6.4.1 Autoencoders
6.4.1.1 Variational autoencoder
6.4.1.2 Denoising autoencoder
6.4.1.3 Contrastive autoencoder
6.4.2 Probabilistic models
6.4.2.1 Boltzmann machine
6.4.2.2 Restricted Boltzmann machine
6.4.3 Deep neural networks
6.4.3.1 Deep belief networks
6.4.4 Deep Boltzmann machine
6.4.4.1 Deep stacked autoencoder
6.5 Deep learning-based clustering schemes for process monitoring
6.6 Discussion
References
7 Unsupervised recurrent deep learning scheme for process monitoring
7.1 Introduction
7.2 Recurrent neural networks approach
7.2.1 Basics of recurrent neural networks
7.2.2 Long short-term memory
7.2.2.1 LSTM implementation steps
7.2.3 Gated recurrent neural networks
7.3 Hybrid deep models
7.3.1 RNN-RBM
7.3.2 RNN-RBM method
7.3.3 LSTM-RBM model
7.3.4 LSTM-DBN
7.4 Recurrent deep learning-based process monitoring
7.4.1 Residuals-based process monitoring approaches
7.4.2 Recurrent deep learning-based clustering schemes for process monitoring
7.4.2.1 RNN-RBM clustering
7.5 Applications: monitoring influent conditions at WWTP
7.6 Discussion
References
8 Case studies
8.1 Introduction
8.2 Stereovision
8.2.1 Deep stacked autoencoder-based KNN approach
8.2.1.1 Preliminary materials: autoencoders
8.2.1.2 The SDA-kNN obstacle detection approach
8.2.2 Data description
8.2.3 Results and discussion
8.2.4 Model trained using data with no obstacles
8.2.5 Evaluation of performance for busy scenes
8.2.6 Obstacle detection using the Bahnhof dataset
8.3 Detecting abnormal ozone measurements using deep learning
8.3.1 Introduction
8.3.2 Data description
8.3.3 Ozone monitoring based on deep learning approaches
8.3.3.1 Results and discussion
8.3.4 Detection results
8.3.4.1 Sensor anomaly detection: false anomalies
8.3.4.1.1 Case A: single abrupt fault
8.3.4.1.2 Case B: multiple abrupt faults
8.3.4.1.3 Case C: intermittent faults
8.3.4.2 Conclusion
8.4 Monitoring of a wastewater treatment plant using deep learning
8.4.1 Introduction
8.4.2 Proposed DBN-based kNN, OCSVM, and k-means algorithms
8.4.3 Real data application: monitoring a decentralized wastewater treatment plant in Golden, CO, USA
8.4.4 Conclusion
References
9 Conclusion and further research
directions
References
Index
Back Cover