Power transformer fault diagnosis based on a self-strengthening offline pre-training model

Accurate transformer fault diagnosis is crucial for maintaining the power system stability. Due the complex operation condition of the transformer, its faults are with the characteristic of multi-class faults, class-imbalance, and limited diagnosis data of availability. Additionally, some fault samples are only with overheating or discharge labels when collected, it is a challenge that how to how to use these samples. To address these issues, in this paper, a novel transformer fault diagnosis method based on a hybrid model of Res-Variational-Auto-Encoder (ResVAE) and ensemble learning (EL) model is proposed. Through a self-strengthening strategy, fault characteristics are extracted category-by-category by using a residual convolutional neural network, and low dimensional characteristics are mapped into characteristic fusion samples by VAE. Based on this strategy, an offline pre-training model is built based on ResVAE and EL. The hybrid model can obtain more information from offline source domain, enabling the EL to diagnose multiple fault types as well as undetermined faults. Considering 11 categories of imbalanced classification scenarios with limited sample sizes, the comparison is made between eight expansion and six diagnosis algorithms. The results show that the offline pre-training EL model increased the diagnostic accuracy up to 11.224% compared with tradition ratios method. The ResVAE-EL model achieves the highest diagnostic accuracy of 91.011%, which is 10.112% higher than that of the single offline pre-training model.