이미지 기반 딥러닝 기법을 활용한 자동차 부품의 결함 탐지 및 이상 예측에 관한 연구 = A Study on Image-based Deep Learning Techniques for Defect Detection and Abnormality Noise Prediction in Automotive Components

다국어 초록 (Multilingual Abstract)

This thesis investigates image-based deep learning methods for defect detection and abnormality prediction in automotive components, and analyzes their applicability and performance characteristics across different engineering problems. Two representative case studies are considered: turbocharger whine noise prediction and defect detection of a constant velocity (CV) joint boot.
In the first case study, vibration signals were measured from the turbocharger housing during acceleration tests. Due to the high-speed rotational characteristics of turbochargers, direct measurement of rotational speed is difficult; therefore, order-related features were indirectly represented using time–frequency domain analysis. Image-based deep learning classification models were applied to predict the occurrence of whine noise, and the effects of preprocessing methods and model architectures on prediction performance were analyzed. Experimental results show that spectrum maps processed with the “Square+Canny” preprocessing method effectively capture order-related characteristics of whine noise and achieve superior prediction performance. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visually verify that the proposed models focus on frequency regions associated with whine noise.
In the second case study, rotational image data of CV joint boots were analyzed to detect through-hole defects. An image summation method, a U-Net-based object segmentation model, and a YOLO-based object detection model were applied and compared using accuracy, precision, recall, and F1-score metrics. Considering the characteristics of defect detection problems, minimizing false detections is of particular importance; therefore, multiple evaluation metrics were used to analyze the performance of each method. The results demonstrate that the YOLO-based object detection approach provides robust performance under varying environmental conditions and shows superior performance in terms of both detection accuracy and processing speed.
Through these case studies, this thesis confirms that image-based deep learning techniques can be effectively applied to diagnostic problems in automotive components. Furthermore, the results highlight the importance of selecting appropriate preprocessing methods, model architectures, and evaluation metrics according to the target application and diagnostic objectives.

목차 (Table of Contents)

• 제 1장 서론 1
• 1.1 연구 배경 1
• 1.2 연구 목적 9
• 1.3 연구 내용 11
• 1.3.1 이론적 배경 11
• 1.3.2 전처리를 활용한 소음 예측 성능 향상 사례 11
• 1.3.3 모델 선정을 활용한 결함 검출 향상 사례 12
• 제 2장 이론적 배경 13
• 2.1 딥러닝(Deep Learning) 13
• 2.2 Canny Edge Detection 16
• 2.3 차수 18
• 2.4 설명 가능한 AI(XAI) 20
• 2.5 이미지 분류, 객체 감지, 객체 분할 23
• 제 3장 자동차 부품에 대한 딥러닝 적용 사례 연구 26
• 3.1 터보차저 진동 기반 소음 예측 연구 26
• 3.1.1 개요 26
• 3.1.2 실험 환경 28
• 3.1.3 데이터 세트 구성 31
• 3.1.4 머신러닝/딥러닝 모델 학습 40
• 3.1.5 모델 성능 분석 47
• 3.1.6 Grad CAM 시각화 50
• 3.2 자동차 CV 조인트 부트 결함 탐지 연구 54
• 3.2.1 개요 54
• 3.2.2 실험 환경 55
• 3.2.3 데이터 세트 구축 56
• 3.2.4 영상처리 기법 57
• 3.2.5 영상 처리 방법 비교 분석 63
• 제 4장 결론 72
• 참고문헌 74
• Abstract 77