Digital twin enhanced multitask learning framework for fault diagnosis of electromechanical coupling system

The electromechanical system is highly coupled that involves multiple physics, components, and operating conditions, which leads to the various and complicated fault characteristics. In this regard, effective fault diagnosis is critical for ensuring its reliable and safe operation. However, the coupling characteristics makes it difficult and expensive to collect the sufficient failure data of physical entities, and it puts forward the higher requirement for the capability to identify the system faults. In such circumstance, we propose a digital twin enhanced multi task learning fault diagnosis framework. No longer relying on the fault injection of physical entities, we utilized the digital twin (DT) technology to model the virtual space of the electromechanical system. The DT model simulates the system's operating dynamics and generate the high-quality failure data without any physical damage. Furthermore, inspired by the multitasking learning (MTL), we decompose the complicated fault diagnosis of the coupled system into three collaborative subtasks: fault location identification, fault type identification, and operating condition identification. The training dynamics among these subtasks is adaptively controlled through a gradient-based task weight assignment mechanism. Based on the virtual failure data, the knowledge reflecting the coupled system's fault characteristics is well mined and shared among multiple subtasks, which greatly improves the fault diagnosis performance under the situation of few physical fault samples. Through an electromechanical fault test bench, we demonstrated and validated the effectiveness of the above framework.