A multimodal fusion fault diagnosis method for partial discharge in switchgear

Abstract

Precise diagnosis of partial discharge (PD) faults is of great significance for guaranteeing the secure and stable operation of power systems. However, single-modal monitoring signals are confronted with the issue of inadequate feature representation for fault characterization, and traditional multimodal fusion approaches demonstrate restricted capabilities in cross-modal interaction. To overcome these challenges, a novel mulit-modal fusion fault diagnosis method based on cross-attention transformer is proposed. Transient earth voltage (TEV) and acoustic emission (AE) signals are collected,which were converted into heatmaps by continuous wavelet transform and coordinate matrix transformation, respectively.Then, the heatmap is taken as input by the Transformer encoder, and cross-attention is subsequently used to capture the deep features of multimodal information. A switchgear partial discharge (PD) experiment platform is established to capture both TEV and AE signals, and a PD fault diagnosis dataset is constructed which includes three typical fault types and one normal operating condition. Experimental results demonstrate that our method outperforms other exciting methods on diagnosis accuracy of PD faults.