*5.3. Results of the Diagnosis by Optimized SVM*

5.3.1. Diagnostic Precision with Supervised Mutual Information Feature Selection Method versus Unsupervised Approach

According to the framework described in Section 2, a PSO-optimized SVM model is applied to classify cases to reflect the fitness of the selected features.

The feature sequences selected by the supervised mutual information feature selection algorithm and the unsupervised mutual information feature selection algorithm are applied, respectively. Different numbers of features are selected from feature sequences obtained by both methods, and the diagnostic accuracy of the both method is shown in Figure 4, in which the red point is where the best fitness is obtained.


**Table 3.** Extracted feature sequence by unsupervised MI V.S. supervised MI.

**Figure 4.** Diagnostic precision with supervised mutual information feature selection vs. unsupervised approach.

As can be seen from Figure 4, features selected from both unsupervised and supervised feature selection methods have good performance as input to diagnostic accuracy in cases, and both of the models achieve greatly increased diagnostic accuracy in the second feature. 5.3.2. Diagnostic Precision by Features of the Unsupervised Approach with Best Fitness vs. Other Classical Feature Set

Feature set with the highest diagnostic accuracy obtained by unsupervised mutual information feature selection method is used as input to optimized SVM diagnosis model for main transformer cases, other typical feature sets are used in contrast as shown in Table 4.

**Table 4.** Diagnostic precision by features of the unsupervised approach with best fitness vs. other classic feature set (percent).


Note: 1 refers to the features of the unsupervised approach with the best fitness obtained as shown in Figure 4 as input to the optimized svm model; 2 refers to the features used in the three ratios method [1]; 3 refers to the features used in some intelligent methods [16]; 4 refers to the features used in the Duval Triangle method [2].

As shown in Table 4, the feature set obtained by the unsupervised mutual information feature selection algorithm, is used as the input of the optimized SVM diagnosis model and performs better than other inputs of the feature set in the case of diagnosis of the main transformer condition diagnosis. Therefore, the algorithm has high applicability.
