*3.1. Training and Testing of SVM 1*

The training and testing of SVM 1 are designed to categorize between the transformer's normal and faulty conditions. The training stage was carried out using 70 oil samples with fault and normal conditions. A total of 14 oil data samples were then used in testing SVM 1. The results indicated that all the new oil data samples were pigeonholed accurately. The selection response and predictors selected by SVM1 using the ingested data are demonstrated in Figure 5. A 30-fold cross-validation has been selected.


**Figure 5.** The dataset variables were selected by the SVM 1 classifier.

The dissolved gases were automatically placed as the predictors and the fault type as the response. In the SVM 1 algorithm H2 and C2H2 was considered to acquire the desired output. The characterization of H2 versus C2H2 is illustrated in Figure 6.

Further, various ML algorithms, i.e., linear SVM, quadratic SVM, cubic SVM, fine gaussian SVM, medium Gaussian SVM, and coarse Gaussian SVM, are compared in terms of accuracy (%), prediction speed (objects/sec) and training time (sec), as tabulated in Table 4. The testing and calculation of the accuracy of referenced algorithms tested on the same data were evaluated based on the percentage of accuracy to classify fault type, the prediction speed and the training time of the respective algorithm.



In the SVM 1 algorithm, the quadratic SVM provides the highest degree of accuracy, and the corresponding configuration matrix shown in Figure 7 was imported to the workspace to predict the new testing data samples.

**Figure 6.** H2 versus C2H2. Blue and orange (•)—raw data, orange and blue (×)—predicted data.

**Figure 7.** SVM 1 confusion matrix.

## *3.2. Training and Testing of SVM 2*

The training and testing of SVM 2 were purposed to categorize between PD and T faults. The training stage was carried out by using a total of 15 oil data samples with T fault and with PD fault conditions respectively. Subsequently, a total of 15 oil data samples were utilized in testing SVM2. The results indicated that a total of 14 oil data samples were classified accurately. The selection of the response and predictors selected by SVM 2 using this ingested data is demonstrated in Figure 8. A 30-fold cross-validation was selected.


**Figure 8.** The dataset variables selected by the SVM 2 classifier.

The dissolved gases were automatically placed as the predictors and the fault type as the response. In the SVM 2 algorithm, CH4 and C2H2 were considered to acquire the desired output. The characterization of CH4 versus C2H2 is illustrated in Figure 9.

Further, the various ML algorithms were compared in terms of accuracy (%), prediction speed (objects/sec) and training time (sec), as tabulated in Table 5.


**Table 5.** SVM 2 DGA classification outcomes.

In the SVM 2 algorithm, the linear SVM, fine Gaussian SVM, medium Gaussian SVM and coarse Gaussian SVM provided the highest degree of accuracy and the coarse corresponding configuration matrix shown in Figure 10 was imported to the workspace to predict the new testing data samples.

**Figure 10.** Confusion matrix for SVM 2 classifier.
