Thermodynamic simulation-assisted random forest: Towards explainable fault diagnosis of combustion chamber components of marine diesel engines

Aiming at the challenges that traditional intelligent fault diagnosis methods of marine diesel engines often suffer low generalizability due to the lack of fault training samples, as well as poor explainability due to the insufficient incorporation of domain knowledge on fault mechanism, this paper develops a Thermodynamic Simulation-assisted Random Forest (TSRF), which reveals fault characteristics through thermodynamic simulations and incorporates them as prior knowledge when designing the intelligent fault diagnosis model. Firstly, five thermodynamic fault models are developed by fine-tuning the essential system parameters to correspond with the distinct attributes of different faults. Then, potential thermodynamic indicators of combustion chamber component degradation are identified through numerical simulation results. By calculating SHapley Additive exPlanations (SHAP) values, a parameter selection process is conducted to retain only those variables demonstrating significant correlations with fault states. Finally, the selected parameters are leveraged to assess the condition of the combustion chamber and input into the fault diagnosis model. The proposed TSRF achieved exceptional classification performance, illustrating a mean accuracy of 99.07% on the fault dataset constructed in this paper. The estimation results of the model are interpreted from the local and global perspectives based on SHAP values. As a result, turbocharger exhaust temperature, blow-by heat flow, and cylinder liner heat flow are found to contribute significant to fault diagnosis.