A Semantic Alignment Guided Transfer-Learning Method for Electro-Mechanical Actuator Fault Diagnosis Under Variable Working Conditions

The electromechanical actuators (EMAs) are safety-critical components in the new generation of aircraft and spacecraft, which makes the fault diagnosis of EMA an important task. However, EMA works under different environments and loads. The data distribution differences between training domains and target testing domains decrease the diagnosis accuracy of traditional data-driven methods, limiting their practical applications. To tackle this problem, this article outlines a semantic alignment-guided transfer-learning method. The model takes advantage of both labeled data from source working conditions and unlabeled data from target working conditions that are to be diagnosed. Unlike previous transfer learning methods, our model not only reduces domain discrepancy by learning domain-invariant features but also considers semantic-discriminative features by defining the novel mixed-center clustering and separability (MCCS) strategy. Benefiting from MCCS, the distributions of the source and the target domain can be actively aligned with each other, and samples are better semantically clustered or separated according to their fault status. Consequently, a powerful domain adaptation metric is generated, which can simultaneously reduce the distribution margin at both the global and semantic levels. As a result, the efficiency of knowledge transfer and diagnosis accuracy can be enhanced. Extensive experiments are conducted on 26 actuator’s domain adaptation scenarios, considering different working loads, velocities, and driving waveforms. Results show that our model exceeds both the traditional supervised learning approaches and the recent state-of-the-art transfer-learning-based models, demonstrating the effectiveness of our method. This study can support the application needs of EMA health monitoring, especially under variable working conditions.