음향방출 신호를 이용한 저속회전 베어링 결함 진단을 위한 부분 양자화 기반 경량 딥러닝 모델

다국어 초록 (Multilingual Abstract)

Fault diagnosis of rotating machinery is essential for ensuring operational stability and preventing unexpected system downtime. In particular, bearing defects generate distinct high-frequency patterns that can be effectively captured through Acoustic Emission (AE) signals. However, deploying deep learning models for on-device diagnosis remains challenging due to the limited memory capacity and computational constraints of embedded Edge devices. This study proposes a lightweight convolutional neural network (CNN) model optimized through partial quantization for bearing fault diagnosis. The proposed model converts the final classifier layers into INT8 precision while preserving FP32 or FP16 precision in the feature extraction layers, thereby reducing quantization-induced accuracy degradation.
AE spectrograms were used as input features, and the model was evaluated on multiple edge platforms including the Jetson Orin Nano Super, Raspberry Pi 5, Google Coral Dev Board, and a PC environment. Experimental results demonstrate that partial quantization maintains accuracy equivalent to the full-precision FP32 model, while achieving up to a 40–60 percent reduction in model size and noticeable improvements in inference latency. In particular, the Jetson Orin Nano Super achieved the fastest inference performance, and the Raspberry Pi 5 showed additional speed gains due to optimized XNNPACK execution. These findings confirm that the proposed approach enables efficient on-device fault diagnosis and provides a promising direction for scalable predictive maintenance of rotating machinery.

목차 (Table of Contents)

• 제1장 서 론 ······························································································· 1
• 제2장 이론적 배경 ··················································································· 5
• 2.1 스펙트로그램(시간-주파수 도메인 특징) ····························· 5
• 2.1.1 푸리에 변환 (FT : Fourier Transform) ····················· 5
• 2.1.2 단시간 푸리에 변환 (STFT : Short Time Fourier Transform) ············································································ 6
• 2.1.3 스펙트로그램 (Spectrogram) ··········································· 8
• 2.2 딥러닝 모델 ···················································································· 9
• 2.2.1 합성곱 신경망 (CNN : Convolutional Neural Network) ·············································································· 10
• 2.3 양자화(Quantization) ······························································· 12
• 2.3.1 Post-Training Quantization (PTQ) ·························· 12
• 제3장 베어링 결함진단 딥러닝 모델 ··············································· 14
• 3.1 Jiangsu University of Science and Technology (JUST) Slew Bearing Fault Dataset ······························· 14
• 3.2 DongGuk University Bearing Dataset (DGU-Dataset) ········································································· 17
• 3.3 베어링 결함 진단 CNN 기반 경량화 딥러닝 모델 ······· 20
• 3.4 양자화 및 경량화 과정 ··························································· 21
• 제4장 실험 결과 및 분석 ·································································· 22
• 4.1 실험 환경 ······················································································ 22
• 4.2 데이터셋에 따른 실험 결과 ···················································· 28
• 4.2.1 JUST-Slew Bearing Dataset 실험결과 ···················· 28
• 4.2.2 DGU-Dataset 실험결과 ··················································· 31
• 제5장 결 론 ····························································································· 42
• 참 고 문 헌 ······················································································· 45
• ABSTRACT ··························································································· 52