유압 실린더 고장 진단을 위한 최적 딥러닝 모델 개발 = Developing an Optimal Deep Learning Model for Hydraulic Cylinder Fault Diagnosis

다국어 초록 (Multilingual Abstract)

This study aims to develop and compare optimal deep learning models for diagnosing faults in hydraulic cylinders, with a particular focus on internal leakage caused by seal wear. Hydraulic cylinders, which serve as key actuators in industrial systems such as construction equipment and press machines, are frequently subjected to high pressure, vibration, and temperature variations. These harsh operating conditions accelerate seal degradation, leading to performance deterioration and energy loss. Therefore, reliable and real-time fault diagnosis techniques are essential to ensure system stability and predictive maintenance.
To address this, a comprehensive comparison was conducted among three neural network models—Multi-Layer Perceptron (MLP), One-Dimensional Convolutional Neural Network (1D-CNN), and Two-Dimensional Convolutional Neural Network (2D-CNN)—under identical experimental conditions. The dataset was obtained from a hydraulic test bench that recorded multi-channel time-series signals, including pressure, velocity, and displacement, during reciprocating cylinder motion. The data represent five fault levels (G/A/B/C/N), ranging from normal (G) to severe leakage (N), and were split into training and validation sets with an 80:20 ratio. Each model was trained to perform both regression (leakage estimation in LPM) and classification (binary and five-class fault categorization).
Experimental results demonstrate that the MLP model achieved the highest regression performance with an R² of 0.9992 and an MSE of 0.0155, indicating that statistical features extracted from load segments (Segment 2 and 5) effectively capture leakage progression trends. The 1D-CNN (sequence length = 1024) achieved the best binary classification performance with 95.8% accuracy and an F1-score of 0.972, showing strong sensitivity to minor pressure–velocity fluctuations near the fault threshold. In the five-class classification task (G/A/B/C/N), all models achieved accuracy above 0.997, with the 1D-CNN (sequence length = 512) exhibiting the highest accuracy of 0.9996. Regarding computational efficiency, training time increased in the order of MLP < 1D-CNN < 2D-CNN, while the 2D-CNN under global normalization exhibited a lower recall (0.7248), indicating its sensitivity to domain differences.
These findings suggest that the MLP model is the most suitable for lightweight, real-time monitoring, while the 1D-CNN offers a balance between accuracy and computational cost for on-site fault detection. Although the 2D-CNN model is computationally intensive, it provides high interpretability and precision, making it appropriate for post-diagnosis analysis. Overall, this study establishes a fair and reproducible framework for comparing deep learning architectures in hydraulic fault diagnosis and proposes a hierarchical fault detection strategy that integrates model precision with real-time applicability.

목차 (Table of Contents)

• 목 차 i
• Abstract iii
• List of Figures and Tables ⅵ
• 제 1 장 서론 1
• 1.1 연구 배경 1
• 1.2 선행 연구 현황 3
• 1.3 연구 내용 및 범위 9
• 제 2 장 데이터 수집 및 전처리·입력 구성 12
• 2.1 실험 장치 및 데이터 취득 개요 13
• 2.2 데이터 정리 및 전처리 과정 16
• 2.3 사이클 검출 및 6구간 구분 18
• 2.4 입력 데이터 구성 20
• 2.5 데이터 분할 및 라벨 구성 22
• 제 3 장 모델 설계 및 학습 구성 24
• 3.1 모델 설계 개요 25
• 3.2 MLP 모델 구조 27
• 3.3 1D-CNN 모델 구조 29
• 3.4 2D-CNN 모델 구조 32
• 3.5 학습 구성 및 손실 함수 35
• 3.6 모델 비교 및 검증 전략 36
• 제 4 장 모델 학습 결과 및 성능 분석 38
• 4.1 학습 환경 및 실험 조건 39
• 4.2 회귀 성능 분석(Regression Performance) 40
• 4.3 이진 분류 성능 분석(Binary Classification) 44
• 4.4 5-클래스 분류 성능 분석(G/A/B/C/N) 51
• 4.4.1 비균등 경계 설정에 따른 5-클래스 분류 성능 비교(Weighted Interval) 55
• 4.5 모델 효율성 비교 분석(연산 성능) 58
• 4.6 모델별 오분류 패턴 및 특성 분석 59
• 4.7 본 장 요약 60
• 제 5 장 결론 및 향후 연구 62
• 5.1 연구 요약 62
• 5.2 주요 결론 64
• 5.3 연구의 의의 66
• 5.4 연구의 한계 68
• 5.5 향후 연구 방향 69
• 참고문헌 71
• Acknowledgment 73
• 부록 A. MLP 기반 누유량 회귀 모델 코드 74
• 부록 B. 1D-CNN 기반 시계열 회귀 모델 코드 79
• 부록 C. 2D-CNN 기반 이미지 회귀 및 분류 평가 코드 82