Transformer Fault Diagnosis Based on Hybrid-Weighted and Multi-Scale One-Dimensional Convolutional Neural Network

To address the issues of low accuracy and imbalanced datasets in transformer fault diagnosis, this paper proposes a fault diagnosis model based on a multi-scale one-dimensional convolutional neural network (M1DCNN) and a hybrid-weighted sparse categorical cross-entropy (HWSCCE) loss function. Firstly, data preprocessing and feature selection are employed to significantly enhance the discriminative capability of features. Then, M1DCNN is utilized to extract multi-scale information, overcoming the limitations of traditional one-dimensional convolutional neural network (1DCNN) in handling complex features. Meanwhile, the integration of HWSCCE enables dynamic adjustment of sample weights, effectively mitigating the issue of class imbalance. The effectiveness of combining M1DCNN with HWSCCE is validated through ablation and comparative experiments. Ablation experiments demonstrate that the combined model of M1DCNN and HWSCCE achieves an mF1 of 93.07%, representing a 3.21% improvement over the baseline 1DCNN model, indicating a significant performance enhancement. In comparative experiments, the proposed model outperforms other benchmark algorithms on the collected dataset. Furthermore, generalization tests on different datasets constructed based on the IEC TC 10 dataset show that the model maintains consistently high performance, with mP, mR, and mF1 mostly exceeding 86%. The results indicate that the proposed method effectively improves the accuracy and reliability of transformer fault diagnosis, exhibits strong generalization capability, and provides robust support for the safe and stable operation of power systems.