TD-DALN: A Twin Data-Driven Design Anomaly Detection Method for Electrohydraulic Actuators

Abstract

Electrohydraulic actuators (EHAs) are one of the most widely applied high-power-density equipments in mechatronic systems. Anomaly detection is essential for EHA prototypes to avoid potential design errors and ensure the reliability of the final design. However, current data-driven anomaly detection methods rely on extensive previous samples under complete design anomaly modes. This premise is impractical in industry applications, where sufficient intrusive physical anomaly experiment data may not be available or introduce unaffordable design costs. This article develops a twin data-driven design anomaly detection method to address the aforementioned problem. First, digital twins (DTs) of the EHA system are established to broadly simulate dynamic responses to design anomalies. Physics-informed parameter estimation ensures twin model fidelity and data availability. Besides, a twin data-driven domain-adversarial long short-term memory (LSTM) network (TD-DALN) is proposed to facilitate domain-invariant and discriminative feature extraction and cross-domain knowledge transfer for accurate design anomaly classification. Correspondingly, domain reconstruction is designed to bridge initial distribution differences between virtual and physical domains caused by the imbalance of anomaly and dynamic conditions. The experimental results demonstrate the effectiveness of the proposed method and its advantages over the competitors.