An Improved Triplet Network for Electromechanical Actuator Fault Diagnosis Based on Similarity Strategy

Deep learning-based methods have been widely used and achieved state-of-art performance in fault diagnosis of electro-mechanical actuator (EMA). Traditional deep learning methods face three major challenges, i.e., when the training dataset is limited, unbalanced, and/or when the model is applied in a different working conditions from the training dataset. In this paper, we propose an improved deep learning-based fault diagnosis framework for EMAs based on triplet network with coupled cluster losses. Unlike the traditional approaches, the proposed framework learns to predict similarity between samples, rather than the fault labels directly. The trained model is used to calculate the distance between testing samples and a set of benchmark samples with known labels. Fault diagnosis is, then, conducted based on K-nearest neighbor algorithm. Experiments on a real-world EMA dataset from NASA show that the developed framework can improve the performance of traditional deep learning-based approaches under the three challenges.