Bidirectional Cross-Attention Domain Adaption Transformer Network for Aircraft EMA Fault Diagnosis Under Varying Working Conditions

In the modern multi-electric aircrafts, electro-mechanical actuator (EMA) plays the critical role in the execution of flight control commands. The undetected EMA faults will induce significant safety hazards and even cause catastrophic accident. However, the working conditions of EMAs are dynamically varying and there are large discrepancies in the monitoring data of different working conditions. Traditional deep learning-based fault diagnosis methods have difficulty in extracting domain-invariant features and cannot generalize to varying working conditions effectively. In this study, a bidirectional cross-attention domain adaption Transformer network (BCA-DATrans) is proposed to implement EMA fault diagnosis under varying working conditions. In order to reduce the difficulty in adapting domains with large distribution difference, the BCA-DATrans is designed as a weight-sharing quadruple-branch structure in which bidirectional cross-attention is designed to generate intermediate domain representations. Subsequently, bidirectional distillation is conducted on the intermediate features and the extracted source/target domain features to narrow the large domain discrepancy of varying working conditions. In this procedure, to prevent the worse sample pairs misleading model training, a center-aware matching strategy is designed to generate high-quality sample pairs to facilitate the model mining domain-invariant features. Furthermore, in order to implement the full alignment of class centers under different working conditions, one class and sample joint weighted maximum mean discrepancy (CSJW-MMD) metric is proposed to correct both the class and sample weight bias in which a discriminative term is further introduced to enhance the feature discriminability. Experimental results on one large aircraft EMA demonstrate the prominent performance of our proposed approach.