Abstract: A novel method for power transformer fault diagnosis based on the C-SVC(support vector classification with the optimized penalty parameter C) and cross-validation algorithm is presented, which can monitor and detect latent transformer faults timely and accurately. The training and testing sets of the C-SVC algorithm are built upon the data about the dissolved gases including hydrogen, methyl hydride, ethane, aethylenum and acetylene produced from transformer faults. Through the optimizing process of the penalty parameter and kernel function parameter γ in the training set, the optimal support vector machine model can be gotten, with which the classification of data in the testing set can be conducted to determine fault features. The method has been validated by many practical examples to be feasible and efficient with high fault diagnosis accuracy.

Key words: transformer, fault diagnosis, support vector machine (SVM), C-SVC algorithm, cross-validation, kernel function