Distribution Transformer Failure Prediction for Predictive Maintenance Using Hybrid One-Class Deep SVDD Classification and Lightning Strike Failures Data

The distribution transformer in field monitoring data failure may be influenced by its maintenance history and risk index in a distribution network associated with keraunic level, average number of lightning strikes, and protection devices employed. Transformer failure is a rare event, and the number of “failed” labels is much smaller than that of “non-failed” labels. Therefore, the transformer failure prediction can be formulated as an anomaly detection or binary classification with an imbalanced dataset, which is challenging to handle. In this paper, we propose a novel distribution transformer failure prediction method through a hybrid one-class deep support vector data description (SVDD) that uses the synthetic minority oversampling technique (SMOTE) to handle the data imbalance between minority and majority class labels. Minimum redundancy maximum relevance (mRMR) is used as a feature selection technique to improve the model's accuracy. The proposed method uses the current condition data of transformers and the distribution network to predict transformer failure for the next year. Real-world field data for 15,066 distribution transformers is used to train and validate the proposed method. It shows superior performance when compared against five benchmark approaches.