ZD6 turnout switch machine anomaly detection framework based on a multimodal lightweight end-to-end approach

Abstract

The ZD6 turnout switch machine is an important part of rail transit signaling equipment. It is one of the most widely used turnout switch machine models to switch train tracks. Any abnormality in its operation could result in serious consequences, such as train derailment. To achieve condition-based maintenance of ZD6 turnout switch machines, this paper proposes a multimodal lightweight end-to-end deep learning framework for anomaly detection in ZD6 turnout switch machines. The input data for this framework are multimodal, comprising two-dimensional time-frequency diagrams obtained by a short-time Fourier transform of the armature current of a ZD6 turnout switch machine series-excited direct current motor, as well as one-dimensional data manually extracted from the armature current characteristics. This framework uses partial convolution and depthwise separable convolution to achieve lightweight processing. SimAM’s parameter-free attention mechanism enhances the network’s feature extraction capabilities. Experimental verification based on a dataset collected from the armature current of two actual ZD6-D and ZD6-G turnout switch machines demonstrates that the proposed anomaly detection framework performs well and has high application value.