Abstract
Accurate classification of transformer partial discharge (PD) severity is essential for insulation diagnostics yet remains challenging due to nonlinear feature relationships and class imbalance. This study evaluates whether feature extraction improves PD severity classification and compares the effectiveness of linear and nonlinear extraction methods. A dataset of 294 samples was categorized into four IEC-aligned severity classes. Two raw measurements (discharge magnitude and applied voltage) were expanded into a 15-dimensional feature space. Principal Component Analysis (PCA) and a bottleneck Autoencoder (AE) were used for linear and nonlinear feature extraction, respectively. Extracted features were classified using an identical Multilayer Perceptron (MLP). Both feature extraction methods improved classification performance over raw and full-feature baselines (96.6%). PCA+ANN achieved 100.0% accuracy (k = 9), while AE+ANN achieved 98.3% (k = 8). The AE misclassified one minority "Normal" sample due to poor latent boundary representation. Reconstruction analysis showed the highest error for the Normal class, reflecting imbalance-driven optimization bias. Feature extraction enhances PD severity classification, with linear PCA outperforming nonlinear AE in this near-linearly separable dataset. PCA's deterministic projection preserves minority class boundaries more effectively, whereas AE performance is limited by class imbalance. These findings suggest that nonlinear methods provide advantages only in more complex feature spaces.