**4. Case Studies**

In this section, various transformer case studies are presented to corroborate the efficacy of the proposed BCSVM algorithm in identifying faults of an unknown dataset to the proposed algorithm learning process. The training and testing of the datasets were not examined from the same utility. The training data adopted from service field oil sample data and the testing oil sample dataset were established by physical unit inspection, as shown in Table 6. The latter will be an opportunity for examination of the proposed algorithm in a more indubitable approach and realizing the capability to development of an efficacious ML algorithm for field data enactment. In the proposed approach, six ML algorithms are assessed for examining the correlation between the response and predictors.

The proposed BCSVM is evaluated in Table 8 against the actual, ANN and IECGR techniques on a set of transformers' DGA data that was not included in the training of the proposed algorithm. A major strength of BCSVM is the 30-fold cross-validation performance to circumvent overfitting problems.


**Table 8.** DGA fault classification case studies.

\* ND—not detectable.

From Table 6, it is observed that the fault class tag T2 has low accuracy against the actual data. Intriguingly, all other fault class tags performed well in the context of accuracy. Additionally, the IECGR was unable to conclusively diagnose some of the faults due to the limitations of the code ratios. Further, the proposed technique constitutes proof that it can be reliably applied in the prediction of unknown oil sample datasets, predicting all the case studies accurately.

#### **5. Conclusions**

The BCSVM construction for fault diagnosis of power transformers has been represented. Several faults are unpredictable by the IEC gas ratio (IECGR) method, which results in an undetectable conclusion. When using ANN, given that it is excellent at learning, the restraint of the inability of fuzzy logic to adapt the created rule base with the varying system for indivisible and irregular data is eradicated. Nonetheless, ANN has the hindrance of overfitting; therefore, it has lower generalization capability and provides circumscribed precision to fault identification. To circumvent all these challenges, in this work, power transformer fault identification was conducted by employing a binary classification support vector machine (BCSVM). The case study results demonstrate that the proposed BCSVM technique has a higher degree of diagnostic accuracy than the IECGR and ANN methods owing to its enhanced generalization capability, and it can classify indivisible DGA datasets by utilizing the kernel function.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The author declares no conflict of interest.
