Transformer Fault Diagnosis Method Based on TLR-WSMOTE and IAOO-SVM-AdaBoost
TL;DR
An ensemble support vector machine (SVM)-based fault diagnosis model is proposed, which incorporates the Tomek link removal (TLR)–weighted SMOTE (WSMOTE) oversampling technique and the hybrid Improved Animated Oat Optimization (IAOO) algorithm, which validates its effectiveness in enhancing diagnostic accuracy and stability.
Abstract
In view of the issues of sample imbalance and low diagnostic accuracy associated with traditional methods for oil-immersed power transformer fault diagnosis, an ensemble support vector machine (SVM)-based fault diagnosis model is proposed, which incorporates the Tomek link removal (TLR)–weighted SMOTE (WSMOTE) oversampling technique and the hybrid Improved Animated Oat Optimization (IAOO) algorithm. First, the dataset is divided into training and testing sets, and TLR-WSMOTE is applied exclusively to the training set to alleviate class imbalance and reduce the influence of overlapping boundary samples. Second, the AOO algorithm is enhanced by introducing Latin hypercube sampling and Thinking Innovation Strategy to strengthen its global exploration and local exploitation ability. The performance of the IAOO algorithm is verified using six typical test functions, demonstrating competitive optimization performance and stability on the selected benchmark functions; the algorithm is then applied to optimize the penalty factor and kernel parameters of the SVM. Finally, multiple SVM weak classifiers are integrated via the AdaBoost framework to further boost the classification performance and generalization ability of the model. Experimental results indicate that the proposed method achieves an accuracy of 97.17% for transformer fault diagnosis, markedly outperforming traditional and other comparative methods, which validates its effectiveness in enhancing diagnostic accuracy and stability. Validation on an additional DGA dataset further demonstrated the generalization capability of the proposed method, achieving a mean accuracy of 94.24%.