Interpretable Online Fault Detection and Localization for Space Thermal Control Systems Using Permutation Feature Importance

Thermal control systems for space applications (SA-TCS) pose significant challenges for fault diagnosis, as they must simultaneously cope with concept drift in data streams and real-time processing requirements. Traditional fault diagnosis methods based on static models are often inadequate handling concept drift and real-time diagnostic demands. To overcome these limitations, an interpretable online fault diagnosis framework based on incremental permutation feature importance (iPFI) is proposed. The framework employs an incremental learning model to predict key system parameters and health status of the SA-TCS in real time, while using the iPFI algorithm to dynamically track the global importance of each sensor feature with respect to the model predictions. By monitoring abrupt changes in feature importance, fault alerts are bidirectionally verified, and both the key sensors and their associated system components indicating faults are identified in real time. This study simulates pipeline leakage and component failure scenarios under multiple operating conditions of the SA-TCS. The effectiveness and advantages of the proposed online fault diagnosis framework are validated using the simulated dataset. Experimental results show that the constructed model can accurately capture dynamic changes in feature importance caused by condition shifts and fault events, thereby achieving accurate and real-time fault detection and localization for multi-condition SA-TCS.