Interpretability Study on the Fault-Diagnosis Model of the Heat Pipe / Vapor-Compression Composite Air-Conditioning System

Applying data-driven fault-diagnosis models to data center air-conditioning systems can significantly improve operational reliability. However, these models often lack diagnostic interpretability, which limits their application. This study develops a composite fault-diagnosis model based on typical machine-learning algorithms, compares the diagnostic performance of different models, and conducts interpretability research on the diagnostic models using the Shapley additive explanation method. The results demonstrate that the convolutional neural network (CNN)-based fault-diagnosis model achieves optimal performance in both the heat-pipe and vapor-compression modes, with F-1 scores exceeding 0.999 across all classifications. In the heat-pipe mode, the diagnosis of the CNN model primarily relies on the condenser-fan frequency, outdoor temperature, and refrigerant-pump power consumption as key features, whereas in the vapor-compression mode, the dominant features are the outdoor temperature, compressor frequency, and subcooling degree.