Electromechanical Actuator Fault Diagnosis under Variable Working Conditions Using Self-distillation and Contrastive Learning

The electromechanical actuator (EMA), as a critical driving component, plays a vital role in ensuring the reliable operation of aerospace systems. Therefore, fault diagnosis of EMAs is essential. With the advancement of artificial intelligence, data-driven algorithms, particularly those based on deep learning, have achieved significant success in fault diagnosis. However, these algorithms typically assume that the training and diagnostic data share the same distribution. Given the complex working environment and variable operating conditions of EMAs, it is challenging to ensure that these two types of data come from the same domain. In this paper, we propose a domain adaptation algorithm based on self-distillation and contrastive learning. By leveraging pseudo-labeling techniques, we filter out target domain data with high confidence and fine-tune the feature extraction network for the target domain. Through the incorporation of contrastive learning, we measure the similarity of fault features from a statistical perspective, thereby enhancing the model’s ability to distinguish between features of different fault states. The discriminator module is designed to utilize adversarial learning to reduce the discrepancy in feature distributions between the source and target domains. We validate our approach on a real-world EMA dataset from NASA, where the diagnostic data and training data come from different working conditions. Experimental results demonstrate the superiority of our method, and ablation studies further confirm the effectiveness of the proposed approach.