MDTCNet: A Novel Multiscale Denoising Transformer Convolutional Network for Fault Diagnosis of Partial Discharge

Abstract

The types of partial discharge (PD) defects are closely related to the severity of insulation faults in electrical equipment, and the accurate recognition of these defects is essential to guarantee the stability of the power supply system. Current methods are hindered by a lack of adaptive denoising capabilities and an inability to learn multiscale fault features, which limits their effectiveness in processing complex and noisy PD signals. Furthermore, these methods are primarily based on convolutional neural networks (CNNs), which also fail to capture global features of PD. To overcome these challenges, we propose a novel multiscale denoising transformer convolutional network (MDTCNet), integrating a multiscale residual attention denoising (MRAD) module and a fault diagnosis transformer (FDT) module. The MRAD module employs dilated convolutions with varying dilation rates to extract multiscale features, while the advanced convolutional block attention module (CBAM) and a soft thresholding function work in concert to adaptively adjust the denoising threshold based on the characteristics of the PD, effectively suppressing noise. Additionally, the FDT module is developed to enhance the model’s ability to extract global PD features, leveraging the transformer model’s exceptional long-range dependency modeling capabilities. Experimental results on both our on-site PD dataset and a public PD dataset demonstrate that the proposed MDTCNet outperforms other excellent methods in terms of classification performance, generalization capabilities, and robustness, achieving accuracy rates of 98.46% and 100%, respectively.

Topics & Concepts