The National Institute of Fisheries Science is developing a vibration-based anomaly detection system for small fishing vessel engine rooms. Since the fault types that can occur in such engines are highly diverse and fault-labeled data are difficult to collect for every type, the system is designed to learn the range of normal operating signals and flag deviations from that range as anomalies rather than learning each fault type individually. This normal-data-based approach requires a reliable way to define and measure the normal operating state, and engine speed (RPM) is the most basic such variable. As part of this effort, this study proposes a deep-learning method that estimates RPM directly from engine-room vibration and simultaneously classifies the accelerometer mounting position, using only magnet-mounted accelerometers that require no engine modification, and quantifies which mounting position most clearly captures this normal signal. Four-channel vibration signals were measured on a 9.77-ton coastal purse seiner with an inline-six marine diesel engine whose speed ranged 447-1,444 RPM during the sea trial off Tongyeong, Korea. Hilbert-envelope spectrograms (0-100 Hz) of the four channels were fed into an attention-branch convolutional network with approximately 0.39 M parameters, trained in two stages for RPM regression and position classification. The proposed method achieved a mean absolute error of 42.9 ± 2.1 RPM (about 5% relative error; RMSE 66.3 RPM,  = 0.935 for the representative model) and a position classification accuracy of 99.6% on the time-sequential test set. Ablation experiments showed that the envelope demodulation and the 100-Hz bandwidth, which covers the firing frequency and up to the 13th rotational harmonic, were the dominant factors. The learned attention weights were highest for the channel near the power-transmission end of the engine and varied with the operating range, indicating which position most reliably captures the normal signal under each condition and demonstrating the benefit of adaptive multi-channel fusion.

서 론
재료 및 방법
    실험 선박 및 엔진
    기준 RPM 계측
    데이터세트 구성
    신호 전처리
    Hilbert 변환 기반 포락선 추출
    데시메이션
    포락선 스펙트로그램 생성
    특징 표현에 따른 추론 성능 비교
    딥러닝 모델 및 학습
    채널별 인코더(branch encoder)
    어텐션 융합(Attention Fusion)
    채널 분류 브랜치
    2단계 학습 전략
    학습 설정
결과 및 고찰
    특징 표현 비교
    시간-주파수 표현의 필요성
    Hilbert 포락선 추출의 필요성
    최적 대역폭 결정
    RPM 추정 성능
    채널 분류 성능
    어텐션 가중치 분석
    포락선 스펙트로그램 표현의 유효성
    최적 대역폭 100 Hz의 물리적 근거
    적응형 대역 선택 대비 전대역 포락선의 강건성
    연구의 한계 및 향후 과제
결 론
사 사
References

• 박민석(국립수산과학원 수산공학과 연구사) | Minseuk PARK (Researcher, Division of Fisheries Engineering, National Institute of Fisheries Science, Busan 46083, Korea) Corresponding author
• 최규석(국립수산과학원 수산공학과 연구사) | Kyusuck CHOI (Researcher, Division of Fisheries Engineering, National Institute of Fisheries Science, Busan 46083, Korea)
• 김현영(국립수산과학원 수산공학과 연구관) | Hyunyoung KIM (Senior researcher, Division of Fisheries Engineering, National Institute of Fisheries Science, Busan 46083, Korea)
• 강동하(국립수산과학원 수산공학과 인턴연구원) | Dongha KANG (Intern researcher, Division of Fisheries Engineering, National Institute of Fisheries Science, Busan 46083, Korea)