Fault Diagnosis of Multi-Railway High-Speed Train Bogies by Improved Federated Learning
Na Qin, Jiahao Du, Yiming Zhang, Deqing Huang, Bi Wu
Abstract
Bogie is the unique connection unit between the train body and rails, and degenerations of its key elements could seriously threaten train safety. In previous works that address the fault diagnosis of bogie of high-speed train (HST), only a single railway is adopted for modeling, which holds insufficient data characteristics and thus leads to poor model generalization ability. In this paper, an improved federated learning algorithm is proposed, which reduces the computation costs by wavelet packet decomposition and trains the local diagnosis models by SecureBoost in modeling. By aggregating model parameters of multiple railways in each iteration, the proposed method aims to establish a global model of bogie fault diagnosis on the premise of protecting the security and privacy of bogie data. Experimental results show that the proposed fault diagnosis method is able to identify the faults of bogie for railways that participate in the modeling procedure with an accuracy of more than 88%, and that for those which do not participate in, the fault diagnosis model is also with a remarkable generalization ability. This provides an effective idea for fault diagnosis of HST bogie with less high-quality labeled data, especially under the current “data island” situation of railway industry.