Early detection of transformer faults is essential for ensuring power system reliability, reducing maintenance costs, and preventing unexpected outages. Traditional dissolved gas analysis (DGA)-based diagnostic approaches often exhibit limited accuracy when handling complex fault conditions. To overcome these limitations, this study investigates machine learning techniques for transformer fault diagnosis using DGA data. Several preprocessing methods were evaluated, and logarithmic transformation was identified as the most effective approach for improving classification performance. Four supervised learning algorithms Fine Tree, Optimizable Tree, Boosted Trees, and Optimizable Ensemble were implemented and evaluated in MATLAB. Performance was assessed using 5-fold and 10-fold cross-validation based on accuracy, precision, recall, and F1-score metrics. The results showed that the Optimizable Ensemble model combined with logarithmic transformation and 10-fold cross-validation achieved the highest performance for fault type detection, reaching an accuracy of 82.1%. For fault severity classification, the same model with logarithmic transformation and 5-fold cross-validation produced the best results, with an accuracy of 82.0%. Comparative analysis demonstrated that the proposed approach consistently outperformed conventional diagnostic methods. These findings confirm the effectiveness of optimized ensemble learning with logarithmic preprocessing for transformer fault diagnosis and severity assessment.