Condition Monitoring of Flow-Based Industrial and Mechanical Equipment using Advanced Signal Processing and Deep Learning : Condition Monitoring of mechanical equipment fault diagnosis using signal processing and deep learning

다국어 초록 (Multilingual Abstract)

This dissertation presents a unified and scalable framework for intelligent fault diagnosis across critical industrial systems, including pipelines, centrifugal pumps, bearings, and milling cutting tools. The research addresses the challenges of early fault detection under noisy, non-stationary, and data-scarce conditions by integrating advanced signal processing, enriched time–frequency representations, optimization-driven modeling, and data-efficient deep learning strategies. For pipeline leak detection, multiple scalogram-enhancement techniques are introduced to strengthen fault signatures in acoustic emission signals. Leak-augmented scalograms generated using Gaussian and Laplacian filtering are analyzed through parallel CNN and convolutional autoencoder streams, with fused features classified by a shallow ANN. An Enhanced Leak-Induced Scalogram (ELIS) framework is also developed, employing adaptive histogram equalization and denoising, followed by feature learning via a genetic algorithm- optimized deep belief network and classification with a least squares SVM. These approaches significantly reduce false alarms while improving sensitivity to small leaks and are useful for kind of fluids. For centrifugal pump fault diagnosis, a hybrid framework is proposed that integrates wavelet coherence analysis and Stockwell transform scalograms, refined using Sobel edge detection and non-local means filtering. A CNN extracts discriminative features, and final classification is performed by a Kolmogorov–Arnold Network (KAN), enabling accurate identification of impeller and mechanical seal faults under varying pressures. For bearing fault diagnosis, two complementary deep learning frameworks are presented. The first employs Mel-scalograms from adaptively segmented acoustic emission signals, classified with a FOX-optimized ANN to ensure robust learning under limited data. The second introduces a hybrid model combining Continuous Wavelet Transform scalograms with a Multi-Head Self-Attention, Bidirectional LSTM, and 1D Convolutional ResNet network, capturing spatiotemporal dependencies and improving robustness against noise. Both frameworks are validated on laboratory and benchmark datasets, achieving high accuracy and generalization across domains. For milling cutting tool fault diagnosis, three advanced frameworks are introduced. The first a semi-supervised ViT framework that integrates pseudo-labeling, uncertainty-aware refinement, and a teacher–student knowledge distillation loop, enabling effective learning from minimal labeled data and extending generalization across cutting tools, gears, and bearings. The second adopts combines log-CWT scalograms with Canny Edge Filtering, extracting features through a dual-branch ResNet-18 and Vision Transformer network, followed by classification with a stacking ensemble of machine learning models. The third proposes a Multi-Stage Transfer Learning (MSTL) strategy, which systematically transfers ImageNet-pretrained features across intermediate RPM domains before fine-tuning on the target spindle speed. This progressive adaptation reduces source–target distribution gaps, alleviates issues of data scarcity and noise sensitivity, and ensures strong generalization under variable operating conditions. Overall, this dissertation delivers a comprehensive and interpretable framework that unifies signal enhancement, multi-scale feature extraction, optimization-driven learning, semi- supervised training, and domain adaptation. The proposed methods achieve high-fidelity fault detection across pipelines, pumps, bearings, and machining tools, contributing to predictive maintenance and advancing the goals of Industry 4.0.

목차 (Table of Contents)

• DECLARATION OF AUTHORSHIP I
• TERMS AND CONDITIONS RELATED TO COPYRIGHT II
• ACKNOWLEDGEMENTS VI
• DEDICATION VII
• ABSTRACT IX
• TABLE OF CONTENTS A
• LIST OF FIGURES E
• NOMENCLATURE I
• I. INTRODUCTION 2
• 1.1 Motivation 2
• 1.2 Thesis Objectives and Contributions 3
• 1.3 Thesis Organization 5
• 1.4 Performance Metrics 5
• II. EXPERIMENTAL PROCEDURE AND DATA COLLECTION 8
• 2.1 Pipeline Test Rig Setup Concerning Fluid Flow Equipment 8
• 2.1.1 System development for AE data acquisition 9
• 2.1.2 Dataset recording and its description 10
• 2.2 Centrifugal Pump Test Rig Setup Concerning Mechanical Equipment Defects 11
• 2.3 Bearing Dataset Acquisition and Experimental Setup 14
• 2.4 Milling Cutting Tools Dataset Acquisition and Experimental Setup 16
• III. PIPELINE LEAK DETECTION: SCALOGRAM ENHANCED DEEP LEARNING WITH LEAK AUGMENTED AND OPTIMIZED HYBRID MODELSCENTRIFUGAL PUMP FAULT DIAGNOSIS 22
• 3.1 Introduction 22
• 3.2 Proposed Methods 26
• 3.2.1 Framework I: Leak-Augmented Scalograms with CNN-CAE-ANN 26
• 3.2.2 Framework II: Enhanced Leak-Induced Scalograms with DBN–GA–LSSVM 27
• 3.3 Technical Background 28
• 3.3.1 Gaussian and Laplace Filter 28
• 3.3.2 Autoencoder for Feature Extraction 29
• 3.3.3 Introduction to Convolutional Autoencoders 29
• 3.3.4 Susceptible Feature Extraction 30
• 3.3.5 Non-Local Means and Adaptive Histogram Equalization 31
• 3.3.6 Feature Extraction Using a Deep Belief Network Model 32
• 3.3.7 LSSVM 34
• 3.3.8 Genetic Algorithm 34
• 3.4 Results and Discussions 35
• 3.4.1 Framework I: Leak-Augmented Scalograms with CNN-CAE-ANN 35
• 3.4.2 Framework II: ELIS with Optimized DBN–LSSVM Model 38
• 3.5 Conclusions 40
• IV. A DEEP LEARNING APPROACH FOR FAULT DIAGNOSIS IN CENTRIFUGAL PUMPS THROUGH WAVELET COHERENT ANALYSIS AND S-TRANSFORM SCALOGRAMS WITH CNN-KAN 43
• 4.1 Introduction 43
• 4.2 Proposed Fault Diagnosis Method 47
• 4.2.1 Wavelet Coherent Analysis 48
• 4.2.2 Stockwell Transform 49
• 4.2.3 Sobel and Wavelet Denoising Filters 50
• 4.2.4 Convolutional Neural Networks 51
• 4.2.5 Kolmogorov–Arnold Networks 52
• 4.3 Results and Performance Evaluation 53
• 4.3.1 Performance Comparison of the Proposed and Reference Methods 54
• 4.4 Conclusion 58
• V. ADVANCED DEEP LEARNING TECHNIQUES FOR BEARING FAULT DIAGNOSIS USING MEL AND WAVELET-BASED TIME-FREQUENCY REPRESENTATIONS 61
• 5.1 Introduction 62
• 5.1.1 Related Research Works 62
• 5.2 Technical Background 64
• 5.2.1 1D Residual Network 64
• 5.2.2 Multi-Head Self-Attention Mechanism 65
• 5.2.3 Bidirectional Long Short-Term Memory 67
• 5.2.4 Mel Transformation 70
• 5.2.5 Fox Optimizer 72
• 5.2.6 FOX ANN 72
• 5.3 Proposed Method 73
• 5.4 Results and Discussion 76
• 5.4.1 Framework I: Mel-Scalograms with CAE and FOX-ANN 76
• 5.4.2 Framework II: CWT Scalograms with MHSA–BiLSTM–ConvResNet 78
• 5.5 Conclusions 81
• VI. ADVANCED DEEP LEARNING FRAMEWORKS FOR MILLING CUTTING TOOL FAULT DIAGNOSIS UNDER DATA SCARCITY, DOMAIN SHIFTS, AND NOISY CONDITIONS 83
• 6.1 Introduction 84
• 6.2 Related Research Works 84
• 6.2.1 Research Gaps 86
• 6.3 Technical Background 87
• 6.3.1 Acoustic Emission and Vibration Signals in Tool Condition Monitoring 87
• 6.3.2 Time–Frequency Representations: Continuous Wavelet Transform Scalograms 87
• 6.3.3 Semi-Supervised Learning Framework 88
• 6.3.4 Vision Transformer 89
• 6.3.5 Knowledge Distillation 91
• 6.3.6 Transfer Learning and Domain Adaptation 93
• 6.4 Proposed Method 94
• 6.4.1 Framework I: Semi-supervised vision transformer framework with pseudo-label refinement and knowledge distillation 95
• 6.4.2 Framework II: Hybrid CNN–ViT architecture for MCT fault diagnosis 95
• 6.4.3 Framework III: Multi-Stage Transfer Learning framework for milling cutting tool fault diagnosis 96
• 6.5 Results and Discussion 97
• 6.5.1 Framework I: Semi-supervised vision transformer framework with pseudo-label refinement and knowledge distillation 97
• 6.5.2 Framework II: Hybrid CNN–ViT architecture for MCT fault diagnosis 99
• 6.5.3 Framework III: Multi-Stage Transfer Learning framework for milling cutting tool fault diagnosis 101
• 6.6 Conclusions 102
• VII. CONTRIBUTIONS SUMMARY AND FUTURE WORKCONTRIBUTIONS SUMMARY AND FUTURE WORK 106
• 7.1 Summary 106
• 7.2 Future Work 107
• 7.2.1 Pipelines 107
• 7.2.2 Centrifugal Pumps 108
• 7.2.3 Bearings 108
• 7.2.4 Milling Cutting Tools 108