MFL 기법 기반 체인 국부 파단 진단을 위한 자속 신호 처리 및 인공지능 적용 연구

다국어 초록 (Multilingual Abstract)

Chain is widely used across industrial systems as a critical component for power transmission and load support. However, cracks caused by repeated loading and corrosion can lead to unexpected fracture and safety accidents. Therefore, a reliable non-contact diagnostic technology for real-time chain condition monitoring and early crack detection is essential.
This study develops a magnetic flux leakage (MFL) sensor system for the non-destructive inspection of an RS-80 double roller chain, which has a discontinuous magnetic path. A permanent-magnet carbon-steel yoke circuit magnetizes the chain into magnetic saturation, and a top-and-bottom 8-channel Hall-sensor array
measures the spatial leakage-flux signal from the cracked region. The experiment uses a 3-meter-long chain specimen with artificially machined crack depths from 1 to 4 mm, where 400 measurement sequences per crack level are collected using a multi-channel DAQ board.
The acquired leakage-flux sequence contains noise and baseline drift. To address this, moving-average-based baseline detrending removes unnecessary low-frequency distortion while preserving the local crack peaks. FFT the spatial leakage-flux pattern into the frequency domain, where dominant spectral peaks show clear amplitude growth proportional to crack severity. Band-pass filtering retains only the core frequency components, and IFFT reconstructs the filtered signal back to the spatial distance axis. This process enables visual confirmation of crack position and severity-dependent amplitude variation along the chain.
Finally, an LSTM-based regression model learns the leakage-flux sequence for crack-depth estimation. The model captures both the repeating chain-link magnetic pattern and the short-duration leakage peak caused by each crack. After learning, the model predicts crack depth with stable performance, reaching RMSE = 0.4270 mm, MAE = 0.3054 mm, and R² = 0.9014, as experimentally validated.
Through the development and experimental verification of this compact, low-power MFL system, this study confirms that crack signatures in a discontinuous chain can be reliably preserved and learned in a sequence-based regression framework. The result indicates that MFL-LSTM integration can become a practical solution for early-stage chain crack diagnosis. This work shows the potential for applying real-time chain severity monitoring and supports the feasibility of future MFL-based AI diagnostic systems for discontinuous power-transmission components.

국문 초록 (Abstract)

체인은 동력 전달 및 하중 지지의 핵심 구성 요소로 산업 설비 전반에 널리 사용되며, 반복 하중·마모·부식 등에 의해 발생하는 국부 파단은 예기치 못한 파손과 안전사고로 이어질 수 있다. 따라서 체인의 상태를 실시간으로 모니터링하고 초기 파단을 조기에 검출할 수 있는 비접촉식 진단 기술이 필수적이다.
본 연구에서는 불연속 구조물인 RS-80 더블 체인의 국부 파단을 비파괴 방식으로 검출하기 위해 MFL 센서를 설계하였다. 네오디뮴 영구자석-탄소강 요크 자화부를 적용하여 체인을 포화시키고, 상·하부 8채널 홀 센서 배열을 통해 파단 부위에서 발생하는 누설 자속 신호를 계측하였다. 실험은 길이 3m의 체인에 파단 깊이 1-4mm를 가공하여 파단 깊이 단계별 400회 반복 측정으로 수행되었으며, DAQ 보드를 통해 신호를 수집하였다.
수집된 신호는 노이즈가 포함되어 있어 이동평균 기반 드리프트 제거 기법을 적용하여 신호의 불필요한 저주파 성분을 제거하고 유효 신호만을 분리하였다.
유효 신호에 FFT를 이용해 공간 영역에서 주파수 영역으로 변환해 특정 주파수 대역에서 결함에 따라 뚜렷하게 증가하는 피크가 확인되었으며, 해당 피크를 핵심 주파수 성분으로 확인하였다. 이를 대역통과 필터링을 통해 핵심 주파수 성분만 남겨, IFFT를 통해 다시 공간 영역 신호로 변환하여 결함의 위치 및 진폭 변화를 시각적으로 확인하였다.
이후 파단 깊이 단계별로 반복 계측된 400개의 실험 데이터를 기반으로 신호의 시간-공간 패턴을 학습시키기 위해 LSTM 기반 모델을 구축하여 파단 깊이를 예측하였다. 위 과정들을 통해 체인 국부 파단을 직관적으로 모니터링할 수 있으며 높은 정확도로 예측할 수 있음을 검증하였다.

목차 (Table of Contents)

• 제1장 서 론 ·········································································································· 1
• 제1절 연구의 배경 및 필요성 ····································································· 1
• 제2절 연구의 동향 분석 ··············································································· 4
• 1. 자속 누설 기술의 발전 동향 ······························································ 4
• 2. 자속 누설 신호 처리 및 분석 기술 동향 ······································· 7
• 3. 불연속 구조물에 대한 연구 공백과 기술적 필요성 ···················· 9
• 제3절 연구의 목적 및 절차 ······································································· 10
• 제2장 연구에 관한 이론적 배경 ·································································· 11
• 제1절 체인 파단 시 누설 자속 발생 원리 ············································ 11
• 제2절 요크 및 네오디뮴 자석 자화부 설계 ·········································· 13
• 제3절 자속 측정을 위한 홀 센서 ···························································· 17
• 제4절 누설 자속 검출을 위한 신호 보정 ·············································· 20
• 제5절 체인 국부 파단 깊이 예측을 위한 딥러닝 모델 ····················· 23
• 제3장 누설 자속 계측을 위한 MFL 시스템 ············································ 27
• 제1절 체인용 MFL 센서 설계 및 제작 ················································· 27
• 제2절 다채널 신호 수집 장치 ··································································· 30
• 제4장 누설 자속 계측을 위한 실험 연구 ·················································· 33
• 제1절 누설 자속 계측을 위한 장치 구성 ·············································· 33
• 제2절 누설 자속 계측을 위한 실험 과정 ·············································· 34
• 제3절 체인 플레이트 파단 정보 ······························································ 35
• 제5장 연구 결과 ································································································ 37
• 제1절 원신호 및 드리프트 제거 신호 ···················································· 37
• 제2절 주파수 해석을 위한 신호 처리 과정 ·········································· 39
• 제3절 정량적 평가를 위한 파단 지점 피크 비교 ······························· 42
• 제6장 체인 국부 파단 깊이 예측 LSTM 모델 ······································· 44
• 제1절 LSTM 모델 구성 ········································································· 44
• 제2절 LSTM 모델 학습 결과 ······························································ 46
• 제7장 결론 ·········································································································· 48
• 참 고 문 헌 ·································································································· 50
• ABSTRACT ······································································································ 70