Abstract
Diesel engines provide essential power and energy guarantee for vessels. Due to scarce fault samples and complex parameter-fault coupling, traditional methods struggle in marine diesel engine diagnosis, underscoring the need for reliable intelligent approaches based on multi-source thermal parameter fusion. This article develops a reliable intelligent fault diagnosis method based on generalized multi-source information fusion. Key parameters are selected using Pearson correlation and mutual information, while an improved Bayesian optimization algorithm automatically tunes random forest parameters to enhance accuracy. TreeSHAP interprets parameter influence, guiding feature selection for retraining. An improved Dempster-Shafer evidence fusion strategy with Shannon entropy and Jousselme distance strengthens model decision-making. The method achieves 99.45% accuracy, outperforming existing models in F1-score and recall, and identifies critical thermal parameters such as intercooler velocity, maximum pressure during combustion, brake power, and velocity of the exhaust manifold. This approach provides a reliable, interpretable, and robust diagnostic tool for marine diesel engines.