Vol 76 No 3 76303

Abstract

Dual fuel engines are crucial for ensuring the safe navigation of ships. Predicting the working status of these engines can provide advanced knowledge of their condition and thereby guarantee safe navigation. In this study, a novel deep learning model, the CNN-BiLSTM-KAN, was designed to forecast exhaust gas temperature (EGT) in dual fuel engines operating in gas mode. The model integrated convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM) networks, and Kolmogorov-Arnold networks (KAN) to perform feature extraction from multi-dimensional time series data, autonomously identify temporal patterns within the data, and directly learn parameterized nonlinear activation functions, respectively. The results reveal that the model obtained a mean square error (MSE) of 0.000051, a root mean square error (RMSE) of 0.007135, a mean absolute error (MAE) of 0.003185, and a mean absolute percentage error (MAPE) of 0.000386. The proposed model demonstrated higher accuracy compared to other forecasting models. Additionally, residual value distribution curves and statistical process control methods were employed to set alarm thresholds for residuals. A sliding window approach was used to establish the alarm threshold for residual standard deviation, with an upper boundary of the residual threshold set at 0.15 and a lower boundary at -0.1. The upper boundary of the residual standard deviation was set at 0.343. Furthermore, the model was validated through a fault dataset. The findings suggest that this approach effectively achieved fault warnings for marine dual-fuel engines. This research provides new references for studies on fault prediction and health management of dual-fuel engines for ships.