Multi-system condition monitoring of marine engines: A unified deep learning framework introducing physical prior knowledge
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Marine engines serve as the core components of marine propulsion system. However, in actual operation, it is impossible to accurately capture the internal operating state of the engine. This often results in delayed detection of combustion anomalies and subsystem degradation. To address this challenge, this study proposes a unified deep learning framework that incorporates physics-informed priors for multi-subsystem condition monitoring of marine engines. The goal is to enable accurate and physically consistent predictions of key performance parameters—such as peak firing pressure, peak firing temperature, air–fuel ratio (AFR), compressor work, turbine work, and others—based on limited but commonly available measurements under long-term operating conditions. The proposed Denoising Autoencoder-Physics-informed Multi-Layer Perceptron (DAE-PMLP) model demonstrates superior performance across cylinder, turbocharging, and overall engine power prediction tasks. All predicted tasks achieve R2 values exceeding 0.99, with tightly clustered residual distributions and low prediction volatility.
Abstract
Marine engines serve as the core components of marine propulsion system. However, in actual operation, it is impossible to accurately capture the internal operating state of the engine. This often results in delayed detection of combustion anomalies and subsystem degradation. To address this challenge, this study proposes a unified deep learning framework that incorporates physics-informed priors for multi-subsystem condition monitoring of marine engines. The goal is to enable accurate and physically consistent predictions of key performance parameters—such as peak firing pressure, peak firing temperature, air–fuel ratio (AFR), compressor work, turbine work, and others—based on limited but commonly available measurements under long-term operating conditions. The proposed Denoising Autoencoder-Physics-informed Multi-Layer Perceptron (DAE-PMLP) model demonstrates superior performance across cylinder, turbocharging, and overall engine power prediction tasks. All predicted tasks achieve R2 values exceeding 0.99, with tightly clustered residual distributions and low prediction volatility.