DBTE-CAM: Fault Diagnosis for the Chiller Based on Dual-Branch Transformer Encoder

Aiming at the problem that the characteristic parameters of the chiller fault operation data are coupled and the timing features are difficult to extract, a fault diagnosis model named DBTE-CAM is proposed in this paper. The model combines an ensemble feature selection strategy, a dual-branch Transformer encoder, and a cross-attention mechanism. Firstly, an ensemble feature selection method based on sequence forward search is adopted, combining the minimal-redundancy-maximal-relevance (mRMR) algorithm and the random forest algorithm, to obtain the most indicative and discriminative fault-relevant features. Then, utilizing a dual-branch Transformer encoder structure for feature encoding, taking into account the timing dependence of single variables and the correlation between feature variables. Finally, the feature information from the two branches is fused by the cross-attention mechanism, and the fault diagnosis is performed by the Softmax classifier. The experimental data from ASHRAE RP-1043 is used to verify the effectiveness of the diagnostic framework. The experimental results show that the model can accurately identify faults in chiller units, with a diagnostic accuracy of 98.34% on the dataset of fault level L1 (the lowest severity). Compared with CNN, LSTM and other mainstream fault diagnosis models, it achieves better diagnostic results.