Wind Turbine Planetary Gearbox Multi-Domain Feature Fusion File Level Diagnosis = 풍력터빈 유성기어박스의 다중 영역 특징 융합 기반 파일 단위 고장 진단

다국어 초록 (Multilingual Abstract)

The vibration signals of wind turbine planetary gearboxes often exhibit different characteristics under varying rotational speed conditions. Even for the same gear tooth fault, impact components may be clearly observed in some signal intervals but masked by noise or speed fluctuations in others. To address this problem, this study integrates features from multiple domains, including time-domain and frequency-domain features, wavelet energy, envelope components, Teager energy, and Hjorth parameters. The purpose of this feature construction is not merely to increase the number of features, but to capture the fault-related signatures of tooth breakage, wear, and root cracks from complementary perspectives.

For the classification stage, XGBoost is employed as the base diagnostic model. ANOVA scores are not used as a direct criterion for feature elimination; instead, they are used to assign lower weights to less informative features during model training. Particle swarm optimization is further applied to simultaneously optimize the key parameters of XGBoost and the feature weight intensities. In the final diagnostic stage, the prediction result is determined at the file level by averaging the prediction probabilities of all signal windows extracted from the same file, rather than relying on individual window-level predictions.

Experiments conducted on the WT-Planetary Gearbox Dataset demonstrate that the proposed method achieves more stable file-level diagnostic performance than BP, LSTM, and standard XGBoost models. In particular, the proposed method is effective in reducing misclassification among fault types with similar decision boundaries, such as tooth damage and wear.

목차 (Table of Contents)

• Ⅰ. INTRODUCTION 1
• 1. RESEARCH BACKGROUND AND SIGNIFICANCE 1
• 2. LITERATURE REVIEW 5
• 2.1 Research Status of Wind Turbine Planetary Gearbox Reliability 5
• 2.2 Research Status of Reliability Analysis Methods 11
• 2.3 Research Status of Fault Diagnosis for Wind Turbine Planetary Gearboxes 13
• 2.4 Research Status of Fault Diagnosis Based on Multi-Domain Features 17
• 2.5 Research Status of Ensemble Learning and Heuristic Optimization 18
• 3. MAIN RESEARCH CONTENT AND RESEARCH ROUTE 20
• 4. INNOVATIONS OF THIS STUDY 23
• 5. ORGANIZATION OF THE THESIS 25
• Ⅱ. WIND TURBINE PLANETARY GEARBOX 29
• 1. STRUCTURE OF THE WIND TURBINE DRIVETRAIN 29
• 2. STRUCTURE AND TYPICAL FAULT MECHANISMS OF PLANETARY GEARBOXES 32
• 3. DATA ACQUISITION PLATFORM AND EXPERIMENTAL OPERATING CONDITIONS 35
• 4. FUNDAMENTALS OF VIBRATION SIGNAL ANALYSIS AND MULTI-SCALE FEATURE EXTRACTION 41
• 5. BASIC PRINCIPLES OF XGBOOST, ANOVA, AND PSO 44
• 5.1 Basic Principle of XGBoost 44
• 5.2 Basic Principle of ANOVA 46
• 5.3 Basic Principle of PSO 47
• 6. SUMMARY 49
• Ⅲ. MULTI-DOMAIN FEATURE CONSTRUCTION AND DATA ORGANIZATION 50
• 1. OVERALL FEATURE EXTRACTION PROCEDURE FOR FAULT DIAGNOSIS 50
• 2. SIGNAL WINDOWING AND PREPROCESSING METHODS 53
• 3. WAVELET DECOMPOSITION AND OPTIMAL SELECTION OF THE DECOMPOSITION LEVEL 56
• 4. CONSTRUCTION OF TIME-DOMAIN FEATURES 59
• 5. CONSTRUCTION OF WAVELET ENERGY AND RATIO FEATURES 62
• 6. CONSTRUCTION OF FREQUENCY-DOMAIN FEATURES 65
• 7. CONSTRUCTION OF ENVELOPE-DOMAIN MODULATION FEATURES 68
• 8. CONSTRUCTION OF DYNAMIC FEATURES, TEAGER ENERGY AND HJORTH PARAMETERS 70
• 9. FEATURE STANDARDIZATION AND DATASET ORGANIZATION 73
• 10. VISUALIZATION ANALYSIS AND PRELIMINARY DISCUSSION OF SEPARABILITY 78
• 11. CHAPTER SUMMARY 84
• Ⅳ. HYBRID FAULT DIAGNOSIS MODEL BASED ON PSO-ANOVA-XGB 85
• 1. TASK DEFINITION OF FAULT DIAGNOSIS AND OVERALL MODELING STRATEGY 85
• 2. PRINCIPLE OF THE XGBOOST CLASSIFICATION MODEL 88
• 3. ANOVA-BASED FEATURE EVALUATION AND WEIGHTING MECHANISM 92
• 4. PSO-BASED HYPERPARAMETER OPTIMIZATION STRATEGY 96
• 5. CONSTRUCTION OF THE PSO-ANOVA-XGB HYBRID MODEL 102
• 6. FILE-LEVEL DECISION-MAKING AND EVALUATION METRIC DESIGN 105
• 7. MODEL TRAINING AND TESTING PROCEDURE 111
• 8. CHAPTER SUMMARY 114
• Ⅴ. EXPERIMENTAL DESIGN AND RESULTS ANALYSIS 116
• 1. EXPERIMENTAL ENVIRONMENT AND PARAMETER SETTINGS 116
• 2. PERFORMANCE ANALYSIS OF MAIN AND BENCHMARK MODELS 119
• 3. DESIGN AND RESULTS ANALYSIS OF THE ABLATION EXPERIMENTS 125
• 4. RECOGNITION ANALYSIS BY FAULT CATEGORY AND BROKEN/WEAR BOUNDARY 131
• 5. ANALYSIS OF MODEL GENERALIZATION ABILITY AND STABILITY 136
• 6. CHAPTER SUMMARY 139
• Ⅵ. CONCLUSIONS AND FUTURE PROSPECTS 141
• 1. RESEARCH CONCLUSIONS 141
• 2. SUMMARY OF INNOVATIONS 144
• 3. LIMITATIONS AND FUTURE RESEARCH DIRECTIONS 146
• 4. FUTURE PROSPECTS 148
• REFERENCES 150
• APPENDIX 156
• 국문초록 162