Transformer Fault Diagnosis Model with Unbalanced Samples Based on SMOTE Algorithm and Focal Loss

摘要

The transformer fault diagnosis method based on neural network is an important method to evaluate the status of transformers. However, it cannot work well when dealing with datasets with unbalanced samples. The results are biased towards majority samples. For the oil chromatography dataset with a maximum imbalance ratio of 266:19, a BPNN model based on SMOTE algorithm and Focal Loss is proposed for transformer fault diagnosis. First, the minority samples of the dataset are expanded based on the SMOTE algorithm. Then a five-layer neural network is constructed, with focal loss as the loss function of the network, so as to pay more attention to the identification and differentiation of minority samples. Results show that compared with the original BPNN model, the proposed model in this paper has a faster convergence speed, improves the overall accuracy by 6.48% and has a higher F1 score for all types. Compared with the k-nearest neighbor (KNN) and random forest (RF) models, the accuracy rate is improved by 16.53% and 7.98%, respectively. The recommended model has better diagnostic performance.