Детальная информация

The subject of this Masters thesis is the application of deep learning methods to bearing vibration data analysis. The topic is focused on improved bearing fault detection using a Multi-Scale Convolutional Neural Network (MSCNN) and Long Short-Term Memory (LSTM) network. The purpose of the work is to develop and evaluate a hybrid MSCNN-LSTM model for accurate fault diagnosis in wind turbine bearings operating under variable loads, noise conditions, and limited labeled data. The methodology involves designing a hybrid neural architecture where the MSCNN captures features at multiple frequency scales, and the LSTM models temporal dependencies in vibration signals. The model is trained and tested on both a controlled laboratory dataset (CWRU) and a real-world dataset (Paderborn), with additional physics-informed noise augmentation to enhance robustness. The proposed approach achieves high classification accuracy and demonstrates strong generalization across datasets, outperforming baseline models such as SVMs and conventional CNNs. It shows resilience to industrial noise and adaptability to non-stationary conditions. The results are applicable to predictive maintenance systems in wind energy, helping reduce unplanned downtime and maintenance costs. The thesis concludes that the MSCNN-LSTM architecture enables early and reliable fault detection, with future work directed at real-time monitoring and edge-computing deployment.

Предметом данной магистерской диссертации является применение методов глубокого обучения для анализа данных о вибрации подшипников. Тема посвящена усовершенствованному обнаружению неисправностей подшипников с использованием многомасштабной сверточной нейронной сети (MSCNN) и сети долговременной памяти (LSTM). Целью работы является разработка и оценка гибридной модели MSCNN-LSTM для точной диагностики неисправностей подшипников ветряных турбин, работающих в условиях переменных нагрузок, шума и ограниченного объема маркированных данных. Методология включает в себя разработку гибридной нейронной архитектуры, в которой MSCNN фиксирует характеристики в нескольких частотных диапазонах, а LSTM моделирует временные зависимости вибрационных сигналов. Модель обучается и тестируется как на контролируемом лабораторном наборе данных (CWRU), так и на реальном наборе данных (Paderborn) с дополнительным усилением физического шума для повышения надежности. Предлагаемый подход обеспечивает высокую точность классификации и демонстрирует сильное обобщение по наборам данных, превосходя базовые модели, такие как SVM и обычные CNN. Он демонстрирует устойчивость к промышленному шуму и адаптируемость к нестационарным условиям. Полученные результаты применимы к системам прогнозирования технического обслуживания в ветроэнергетике, помогая сократить незапланированные простои и затраты на техническое обслуживание. В диссертации делается вывод о том, что архитектура MSCNN-LSTM обеспечивает раннее и надежное обнаружение неисправностей, а дальнейшая работа направлена на мониторинг в режиме реального времени и внедрение передовых вычислений.

• LIST OF ABBREVIATIONS AND DESIGNATIONS
• SYMBOLS
• ABBREVIATIONS
• INTRODUCTION
• 1 Problem introduction
  • 1.1 Rolling Element Bearings: An Overview
    • 1.2 Failure identification
    • 1.4 Wind Turbine Bearing Fault Diagnosis: Mechanism, Process, and Severity
      • 1.4.1 Fault Diagnosis Mechanism
  • 1.5 The similarities and differences between wind turbines bearings and other machinery:
  • 1.6 Vibration Data Acquisition:
  • 1.7 Signal preprocessing:
  • 1.8 Evolution of Wind Power Bearing Fault Diagnosis: A Literature Review
    • Improving the operational safety and reducing maintenance expenses of wind turbines requires the advancement of fault detection technologies for bearings. Below is an overview of key diagnostic approaches.
    • 1.8.1 Fault Diagnosis of Wind Turbine Bearings Based on Spectrum Analysis
    • Fourier Transform-Based Methods
    • Acoustic-Vibration Fusion Techniques
    • Advanced Feature Extraction and Classification
    • Low-Speed and Indirect Monitoring
    • Challenges and Transition to Wavelet Analysis
    • 1.8.2 Fault Diagnosis of Wind Turbine Bearings Based on Wavelet Analysis
    • Wavelet-Based Signal Denoising
    • Speed-Variant Vibration Analysis
    • Wavelet Transform for Fault Feature Extraction
    • Wavelet Reconstruction and Limitations
    • 1.8.3 Fault Diagnosis of Wind Turbine Bearings Based on Artificial Intelligence
    • 1.8.3.1 Shallow Learning Approaches
    • 1.8.3.2 Deep Learning Architectures
      • Convolutional Neural Networks (CNNs)
      • Multi-Scale and Hybrid Models
      • Temporal and Thermal Models
    • 1.8.4 AI-Driven Fault Diagnosis: Advances and Limitations
      • Acoustic Signal-Based Approaches
      • SCADA-Driven Diagnostic Frameworks
      • Hybrid Architectures for Spatial-Temporal Feature Learning
    • Research Gap and Motivation
    • 1.8.5 Hybrid and Attention-Based Architectures
  • 1.9 Critical Analysis
  • 1.10: Problem Description
    • 1.10.1 Proposed Solution: MSCNN-LSTM
    • 1.10.2 Data structure and validation
  • 2 Proposed Methodology for Bearing Fault Detection
    • 2.1 Early Detection of Failure
    • 2.2 Classification of Failure
    • 2.3 Core Concepts of the Hybrid Model
      • 2.3.1 Convolutional Neural Networks (CNNs)
      • 2.3.2 Long Short-Term Memory Network (LSTM)
    • 2.4 The Proposed Hybrid CNN-LSTM Architecture
    • 2.5 Integrated Bearing Fault Detection Mathematics
      • 2.5.1 Characteristic Frequency Context
      • Bearing fault diagnosis relies on characteristic frequencies derived from bearing geometry and rotational speed :
  • 3 Implementation details
    • 3.1 Overview of the Datasets
      • 3.1.1 Dataset 1: CWRU Bearing Data [87]
      • 3.1.2 Paderborn Data Set [88]
      • 3.1.3 Relevance to Fault Diagnosis
      • 3.1.4 Feature Comparison : CWRU vs. Paderborn
      • The difference between the two datasets are shown in Table 3.2
      • Table 3.2 - Feature Comparison of Bearing Fault Datasets
      • Limitations of Seeded Fault Data
      • 3.1.5 Dataset Size and Problem Requirement Prioritization
    • 3.2 Data Preprocessing & Augmentation
      • 3.2.1 Unified Preprocessing Framework
      • 3.2.2 Physics-Informed Noise Augmentation
      • 3.2.3 Dataset-Specific Adaptations
      • 3.2.4 Noise Robustness Evaluation
      • This study implements a rigorous evaluation framework to validate the effectiveness of the proposed MSCNN-LSTM architecture. It compares the hybrid model against three deep learning baselines and one traditional machine learning model, aiming to isola...
      • 3.3.1 Comparative Baseline Models
      • Each model chosen for this comparison and the rational behind it is shown in Table (3.5)
  • 4 The Results
    • This section presents key findings from the evaluation of the MSCNN-LSTM model across tow benchmark datasets (CWRU and Paderborn) detailed in (Section 3.1). The analysis focuses on classification accuracy, robustness to noise, and comparative performa...
    • 4.1 Raw Data Visualization
    •  CWRU Dataset:
    • Frequency Spectrum and Time-Domain Signal Analysis
    • 4.2 Training results
      • 4.2.1 ROC curves:
      • 4.2.3 Classification report
      •  Model Performance on CWRU
      •  Model Performance on Paderborn
      • Figure 4.7 shows near-perfect training and validation accuracy (~1.0) and low, stable loss curves. This confirms the model’s ability to learn discriminative features effectively without overfitting. The MSCNN-LSTM generalizes well to unseen data, maki...
    • 4.2.4 Noise Robustness Evaluation
    • 4.3 Model comparison results
      • 4.3.1 Validation Accuracy Comparison
      • 4.3.2 Confusion Matrices
      • 4.3.3 Performance Metrics
• The LSTM model shows low accuracy and F1 score, indicating that it is not performing well overall. In contrast, the 1D CNN demonstrates a significant improvement, with both accuracy and F1 score reflecting strong performance and good classification ab...
• CONCLUSION
• REFERENCES
• Assignment for thesis 2025_template.pdf
  • РЕФЕРАТ
• Title page thesis 2025_template.pdf
  • РЕФЕРАТ
• Title page thesis 2025_template en.pdf
  • РЕФЕРАТ
• Title page thesis 2025_template en.pdf
  • РЕФЕРАТ
• Без имени