Early-Stage Electrical Fault Identification for Traction Transformers Using Vibration Signals Based on Dual-Attention Convolutional Network

Vibration signals typically contain a wealth of information on equipment status, making them widely used in the fault diagnosis for transformers. However, electrical faults, especially early-stage inter-turn short circuit faults are not characterized on vibration signals, which causes difficulties in precisely identifying the signals. To solve this issue, a dual attention-based neural network is proposed in this paper. A novel attention mechanism (AM) called similarity attention (SA) is designed in this network. It is then embedded in a convolutional neural network (CNN) with the conventional channel attention (CA) to form a dual-attention module (DAM). This module uses multi-channel receptive fields to automatically extract fault features from vibration signals. It subsequently expands and adaptively weights each receptive field and channel through the SA and CA blocks. By using the DAM, the unobvious fault-related features can be extracted effectively. Finally, these features are input into an eXtreme Gradient Boosting (XGBoost) classifier to achieve high-accuracy fault detection. The effectiveness of the method is verified using a 50 kVA traction transformer experimental platform. Moreover, the superiority of this method has been compared with other methods. The comparison results indicate that the proposed method can successfully categorize various early-stage faults with an accuracy rate of over 96% and unaffected by load fluctuations.