A class-center fine-tuning prototypical network for few-shot fault diagnosis of turnout switch machine driven by multi-source signals

Abstract

The turnout switch machine is mainly responsible for controlling the switch converting and locking, which is a crucial component of the rail transit signal system, and its reliability directly impacts the safety of rail transit transportation. However, many existing methods for diagnosing faults in turnout switch machines suffer from low accuracy and poor generalization capabilities. The primary issue is that the fault data of turnout switch machines are scarce, and the single-source data cannot fully reflect the fault characteristics. To address these issues, this paper introduces a fault diagnosis approach for the turnout switch machine based on a class-center fine-tuning prototype network (CFPN). Firstly, current signals of three phases are fused with different weights through an adaptive data fusion strategy, using the fused signal as input data. Secondly, a multiple weighted feature fusion network (MWFFN) based on the activation function Meta-ACON that can learn whether to activate neurons is proposed, which can effectively extract and fuse features at different levels in the fusion signal. In addition, a fine-tuning center balance strategy is proposed to effectively expand the distance between prototypes and enhance intra-class aggregation. Finally, a few-shot sample experiment is conducted to verify the performance of CFPN, using fused current data collected from’ a switch machine. In the comparative experiment, CFPN shows good fault diagnosis performance, the fault diagnosis accuracy reaches up to 99.08%. At the same time, fused vibration data are used to conduct a cross-domain experiment on CFPN, and the results indicate that CFPN still shows good classification effect and has good generalization performance, demonstrating its high practical application value.