A metro train air conditioning system fault diagnosis method based on explainable artificial intelligence: Considering interpretability and generalization

Abstract

Most of the existing air conditioning system fault diagnosis methods adopt black box models, which lack transparency and interpretability. Given the high-speed, enclosed nature of metro train environments, the requirements for trust and safety in metro train air conditioning fault diagnosis models are even more stringent than those for building. Therefore, this paper presents an interpretable and generalized method for fault diagnosis of metro train air-conditioning system. The importance of features is analyzed a priori, and the XGBoost-Shapely Additional Explanations (XGBoost-SHAP) method is used to explain the single fault diagnosis model. Then the trained single fault model is utilized to predict the simultaneous fault data, obtaining score values for various labels, and a binary classification model is established to differentiate single/simultaneous faults. Additionally, the model's generalization ability is improved by screening generalization features based on the geometric difference across operating conditions. The results show that the features with high contribution to three types of single faults are evaporator outlet enthalpy, condenser outlet air temperature and air flow rate. The scores of various tags for simultaneous faults differ from those for single faults, which is beneficial to the identification of suspicious simultaneous faults. After screening the generalized features, when the number of features is less than 10, the generalization performance of the model across operating conditions is better than other cases. Specifically, the average accuracy increases by 5.84 %, 8.38 %, and the average false alarm rate decreases by 10.22 %, 11.26 %.