Deep Transfer Fault Diagnosis Using Digital Twin and Generative Adversarial Network

Abundant fault samples are crucial for intelligent fault diagnosis models. However, collecting efficient annotated faulty data are often expensive or infeasible, especially for safety-critical systems. Instead of directly training diagnosis models on hard-to-collected sensory signals, training on simulated data generated from digital twin is feasible since simulated data with annotation is easy to access. However, simulated samples are often not realistic enough and domain gap exists between simulated and real sensory samples. The fault diagnosis model may lack of generalization if the model is trained only on simulated samples. In this paper, we propose a deep transfer fault diagnosis approach which generalizes fault diagnosis knowledge from digital twin to perform fault diagnosis task in real-world. First, a digital twin model with capability of bearing dynamics is composed using numerical simulation method, from which the simulated faulty samples are generated under various faulty depth and operation conditions. Second, a generative adversarial network is utilized to perform domain adaptation and eliminate the domain gap. Specifically, the generator component conditioned on simulated samples as well as random noise and produces target-like samples as realistic as possible to fool the discriminator. The discriminator component is trained to distinguish the samples as real or synthetic. Hence, the classifier component trained on generated samples can generalize well in real fault diagnosis tasks. Extensive experiments on CWRU dataset are performed to validate the effective of proposed method.