“…The main drawback of these approaches is that the observed gas ratios do not match all the fault types [9,10]. To improve the drawback and to dig the potential fault law from the gases content, some new fault diagnosis methods, including fuzzy logic [11], Dempster Shafer theory [12], grey clustering [13] and rough set [14], self-adaptive RBF neural network [15], auto-associative neural networks and mean shift [16], evolutionary wavelet neural network based on genetic algorithm [17], and support vector machine (SVM) based methods [18][19][20], have been proposed in the recent years. In these methods, SVM is an effective method and is wildly applied in classification problems (fault diagnosis of transformers based on DGA is often considered as a classification problem) [21][22] because its advantages of small sample learning, global optimization, and structural risk minimization [23][24].…”