Transformer fault diagnosis based on hybrid feature selection and an improved differential evolution algorithm for optimizing convolutional neural networks

Abstract

To effectively address the shortcomings of transformer fault diagnosis—namely, the limited variety of characteristic gases and low diagnostic accuracy—this paper proposes a novel method for transformer fault diagnosis that optimizes a convolutional neural network (CNN) using hybrid feature selection (HFS) and an improved differential evolution (IDE) algorithm. First, addressing the issue that fault diagnosis using Dissolved Gases Analysis (DGA) technology relies on a limited number of characteristic gases, making it impossible to comprehensively characterize different fault types, this paper employs the correlation ratio method to expand the 5-dimensional set of characteristic gases to 30 dimensions. To address the problem of feature redundancy caused by excessive dimensionality, the HFS algorithm is proposed for dimensionality reduction and feature selection. A CNN model is employed for transformer fault diagnosis. To address the significant impact of CNN hyperparameters on diagnostic accuracy, this paper proposes an IDE algorithm to optimize and determine the hyperparameters, thereby establishing an IDE-CNN model for transformer fault diagnosis. The final experimental results show that the proposed method achieves a fault diagnosis accuracy of 98.89%. Compared with the Differential Evolution (DE) and Particle Swarm Optimization (PSO) algorithms, the fault diagnosis accuracy is improved by 4.45% and 6.11%, respectively. This validates that the proposed method possesses significant theoretical significance and practical engineering application potential in transformer fault detection.