효율적인 베어링 고장 진단을 위한 다중 센서 융합 딥러닝 모델

다국어 초록 (Multilingual Abstract)

Bearings are key components in modern mechanical systems, used in various fields such as wind turbines and machine tools. During long-term operation, faults such as wear or cracks may occur, and failing to detect them early can lead to equipment downtime or accidents. Therefore, developing fault diagnosis methods with high accuracy and efficiency is an important engineering task. Recently, deep learning–based approaches have been actively studied, and in particular, multi-sensor fusion can complement the sensitivities of individual sensors to improve diagnostic accuracy and stability.
However, existing methods are difficult to apply to resource-constrained edge devices due to complex signal preprocessing and large-scale model architectures, and they lack sufficient feature extraction and fusion capabilities under complex operating conditions and noisy environments. In this study, we propose a multi-sensor fusion deep learning model for efficient bearing fault diagnosis. The proposed method achieves high-efficiency diagnostics while reducing computational burden and avoiding reliance on manual feature extraction. First, multi-sensor signals are transformed into frequency-domain features using the Fourier transform. These features are then fed into a 1D convolutional neural network–based branching module, which extracts features progressively and enhances important features through channel attention. Subsequently, multi-scale attention is applied to deeply fuse complementary information across different scales, and the final diagnosis is performed via a classification module. Experiments on bearing datasets under various conditions achieved approximately 99% accuracy, and the model maintained over 96% performance even in a Gaussian noise environment with SNR of −6 dB, demonstrating robustness. With only about 1.2 million parameters, the model has low computational resource requirements, showing its potential for deployment on edge devices.

목차 (Table of Contents)

• 제 1 장 서 론 ....................................................................................... 1
• 제 2 장 연구에 관한 이론적 배경 ......................................................... 6
• 제 1 절 고속 푸리에 변환(FFT : Fast Fourier Transform) ......... 6
• 제 2 절 1 차원 합성곱 신경망 (One-Dimensional Convolutional Neural Network, 1D CNN) ............................................................. 7
• 제 3 장 제안 방법 .............................................................................. 11
• 제 1 절 Feature Extraction Branch 모듈 .................................. 11
• 제 2 절 Multi-scale Fusion 모듈 및 Classification 모듈 .......... 15
• 제 4 장 실험 결과 .............................................................................. 21
• 제 1 절 실험 기본 설정 .............................................................. 21
• 제 2 절 Case 1 Ottawa 회전 베어링 데이터셋............................ 21
• 1. 데이터셋 설명 ...................................................................... 21
• 2. 제안 방법 성능 ..................................................................... 25
• 3. 다른 베어링 진단 방법과 비교 .............................................. 27
• 4. 노이즈 환경에서의 성능 비교 ............................................... 30
• 제 3 절 Case 2 JUST Slewing 베어링 데이터셋 ........................ 33
• 1. 데이터셋 설명 ...................................................................... 33
• 2. 제안 방법 성능 ..................................................................... 36
• 3. 다른 베어링 진단 방법과 비교 .............................................. 36
• 4. 노이즈 환경에서의 성능 비교 ............................................... 39
• 제 4 절 실험 결과 분석 및 논의 ................................................. 41
• 제 5 장 결론 ...................................................................................... 43
• 참 고 문 헌 ....................................................................................... 45
• ABSTRACT ..................................................................................... 55