Transfer-Learning-Aided Fault Detection for Traction Drive Systems of High-Speed Trains
Chao Cheng, Xuedong Li, Pu Xie, Xiaoyue Yang
Abstract
Long-term operation may lead to performance degradation of the traction drive systems. It will naturally increase the difficulty of fault detection (FD). To ensure the safe and stable operation of the traction drive system, data-driven FD has received considerable attention, especially deep learning methods. By exploiting the idea of transfer learning, this article proposes a new FD method for traction converter faults in the traction drive systems of high-speed trains. Its structure consists of a federal neural network based on a variational autoencoder. The significant advantages of the proposed FD method based on transfer learning are summarized as follows: 1) FD is still valid for the systems with performance degradation; 2) it can also realize the FD function even if the physical model and related parameters are not provided; and 3) the proposed framework can adaptively adjust the model parameters by storing and reusing the prior knowledge in the neural network. Finally, the effectiveness of the proposed method is demonstrated through the platform of the traction drive control system.