Signal Processing Guided Deep Learning for Robust Fault Diagnosis in Rotating Machines

다국어 초록 (Multilingual Abstract)

Centrifugal pumps (CPs) and milling cutting tool (MCT) machines are the core components of industrial systems. Faults in these assets can trigger costly repairs, unplanned downtime, reduced energy efficiency, and safety risks for operating personnel. Reliable con- dition monitoring is hindered by three major challenges: (i) macrostructural vibration noise and other environmental interference that obscure fault signatures, (ii) the limited diag- nostic capability of single-sensor systems that fail to capture complementary fault informa- tion—necessitating multimodal learning from both vibration and acoustic emission signals, and (iii) in transfer learning, not all features learned from a source domain are beneficial for the target task—some may even hinder adaptation. Identifying which representations contain transferable knowledge and which require re-learning remains a critical challenge, motivating the use of data-driven methods such as Layer-wise Domain Discrepancy (LDD) for selective fine-tuning. Conventional raw statistical features (e.g., amplitude, energy, simple time/frequency descriptors) are often insensitive to weak incipient faults and unstable under severe op- erating variations. Therefore, the thesis begins with a vibration-driven edge-enhanced time–frequency representation for centrifugal pumps. Vibration signals are pre-processed and band-selected to isolate fault-specific content. A Stockwell transform produces time–frequency scalograms, after which a Sobel filter emphasizes energy transitions to form SobelEdge scalograms that suppress interference and highlight discriminative fault cues. These repre- sentations are provided to a convolutional neural network for robust classification of pump faults under noisy conditions. Second, to overcome the limitations of single-sensor systems in machine tools, a multimodal dual-input CNN is proposed for milling machine diagnosis using synchronized acoustic emission and vibration signals. Each modality is converted to wavelet scalograms and processed by parallel ResNet-50 backbones with selective fine-tuning to retain modality- specific structure. A new interleaving fusion strategy integrates the learned features for classification, enabling state-invariant recognition in both active milling and idle conditions. Finally, to address the problem of domain shift between different operating speeds, the dissertation introduces the LDD-guided ScaloNet, a lightweight transfer learning framework for adaptive fault diagnosis. ScaloNet, a compact CNN pretrained on multi-RPM scalograms, extracts domain-rich hierarchical features that generalize across conditions. The LDD score quantitatively measures the alignment between source and target domains at each network layer, allowing only layers with high discrepancy to be fine-tuned while freezing the rest. This selective adaptation minimizes overfitting, reduces computational cost, and ensures robust generalization to unseen operating speeds. All three contributions are validated on datasets collected from industrial testbeds. The proposed methods consistently surpass reference baselines, achieving near-perfect ac- curacy on centrifugal-pump datasets under varying inlet pressures and approximately 99% accuracy on milling-machine data. The LDD-guided ScaloNet further demonstrates over 97% accuracy with strong cross-speed adaptability, confirming its efficiency for real-world transfer scenarios. Overall, this work demonstrates that coupling targeted time–frequency preprocessing (edge enhancement, multimodal fusion) with adaptive deep learning architec- tures (CNNs, domain-guided transfer) yields accurate, noise-robust, and generalizable fault diagnosis suitable for practical industrial deployment.

목차 (Table of Contents)

• 1 Introduction 1
• 1.1 Motivation 1
• 1.2 Thesis Objectives and Contributions 2
• 1.3 Organization of the Thesis 4
• 2 Experimental Setup and Data Collection 6
• 2.1 Centrifugal Pump Experimental Setup and Test Rig Setup 6
• 2.1.1 Mechanical Seal Faults 8
• 2.1.2 Impeller Fault 9
• 2.1.3 Centrifugal Pump Dataset 10
• 2.2 Milling Cutting Tool Experimental Setup and Test Rig Setup 11
• 2.2.1 Milling Cutting Tool Datasets 12
• 3 Centrifugal Pump Fault Diagnosis Based on a Novel SobelEdge Scalogram and CNN 18
• 3.1 Introduction 18
• 3.2 Proposed Method 23
• 3.3 Technical Background 23
• 3.3.1 Stockwell Transform 23
• 3.3.2 SobelEdge Scalograms 25
• 3.3.3 Materials and Methods 27
• 3.3.4 Convolutional Neural Network 28
• 3.3.5 Sobel Filter 29
• 3.4 Results and Discussions 30
• 3.4.1 Performance and Comparison 31
• 3.5 Conclusion 34
• 4 A New Dual-Input CNN for Multimodal Fault Classification Milling Cut-ting Tool Machines Using Acoustic Emission and Vibration Signals 37
• 4.1 Introduction 38
• 4.2 Background 42
• 4.2.1 Acoustic Emission 42
• 4.2.2 Vibration Signals 43
• 4.2.3 Continuous Wavelet Transform 43
• 4.2.4 Feature Extraction Using ResNet-50 45
• 4.2.5 Feature Fusion 45
• 4.2.6 Classification with Artificial Neural Network 46
• 4.3 Proposed Technique for Milling Machine Fault Diagnosis 47
• 4.3.1 Feature Extraction 49
• 4.3.2 Feature Fusion 49
• 4.3.3 Classification 49
• 4.4 Results and Performance Evaluation 49
• 4.4.1 Performance Evaluation Metrics 50
• 4.4.2 Cross-Validated Performance of the Proposed Model 50
• 4.4.3 Ablation Study 51
• 4.4.4 Comparison of the Proposed Model with Existing Methods 53
• 4.5 Conclusion 59
• 5 Layer-wise Domain Discrepancy Guided Transfer Learning with ScaloNet for Fault Diagnosis in Milling Cutting Tools Machines 62
• 5.1 Introduction 63
• 5.2 Related Literature 65
• 5.3 Proposed Transfer Learning Technique 69
• 5.3.1 Data Pre-processing 69
• 5.3.2 Lightweight ScaloNet 70
• 5.3.3 Layer-wise Domain Discrepancy (LDD) Score 72
• 5.3.4 Selective Fine-tuning 73
• 5.4 Results and Performance Evaluation 74
• 5.4.1 Performance Evaluation Metrics 74
• 5.4.2 Performance Evaluation Across RPMs Using LDD-Guided Fine-Tuning 75
• 5.4.3 Analysis of Experimental Results Across Different RPMs 76
• 5.4.4 Confusion Matrix and Class-wise Analysis 77
• 5.5 Comparison with Recent Models 79
• 5.6 Conclusion 83
• 6 Conclusion and Future Work 85
• 6.1 Summary of Key Findings 85
• 6.2 Limitations and Future Research Directions 86
• 6.3 Concluding Remarks 87
• Publications 90
• References 92