Lightweight Fault Diagnosis via Siamese Network for Few-Shot EHA Circuit Analysis

Due to the scarcity of faulty samples of aerospace electro-hydrostatic actuator (EHA) electronic modules, there are important challenges to accurately perform fault diagnosis. To this end, this article proposes a diagnostic framework based on Siamese neural networks, which aims to extend the model training process under small sample size conditions and to achieve fault category identification by evaluating the similarity between different labels. To address the shortcomings of traditional distance metrics in discriminating highly similar data, this article proposes an improved hybrid distance metric, which is optimized based on Euclidean distance. In addition, in order to reduce the possible computational complexity of concatenated neural networks, we design a lightweight SqueezeNet module. Finally, a lightweight Siamese SqueezeNet network based on the hybrid distance function was constructed. In tests on two electronic modules in the EHA, the proposed hybrid distance function significantly improves the diagnostic accuracy by about 10% compared to the traditional distance metric even in small sample situations with only 10 or 20 single-tagged samples, and also reduces the number of required training iterations. These results show that the proposed framework can still achieve good fault diagnosis performance in the small sample condition.