BR-MTFL: A Novel Byzantine Resilience-Enhanced Multitask Federated Learning Framework for High-Speed Train Fault Diagnosis

Abstract

In high-speed train systems deployed across diverse geographical regions, robust fault diagnosis techniques are essential for ensuring operational safety. This article proposes the Byzantine resilience-enhanced multitask federated learning (BR-MTFL) framework, a novel framework tailored for the complexities of fault diagnosis in traction asynchronous motors under varying operational conditions. This framework innovatively introduces multitask federated learning (MTFL) to accommodate regional law restrictions and varying fault diagnosis requirements. In addition, the Byzantine resilience of our proposed framework is specifically enhanced to address the challenges posed by inconsistent and potentially misleading feature distributions across different train networks. BR-MTFL is practically validated through experiments conducted across nine clients, each representing a distinct set of fault types and operational conditions typical of high-speed trains. The experiments demonstrate the ability of BR-MTFL to outperform conventional federated learning frameworks in terms of accuracy and resilience to Byzantine threats. BR-MTFL establishes a new standard for federated learning applications in high-speed train fault diagnosis, particularly where data diversity and privacy dominate.