X-PDResNet: An Explainable Residual Neural Network for Discrimination of Partial Discharge Defects in High Voltage Apparatus

Localization and categorization of partial discharge (PD) defects are essential for understanding insulation degradation and preventing failures in high-voltage electrical apparatus. Although numerous machine learning and deep learning approaches have been successfully applied for PD diagnosis, most existing models operate as black-box frameworks, offering limited transparency regarding their classification decisions. To address this limitation, this paper proposes an explainable residual neural network (X-PDResNet) for multiclass PD defect diagnosis. A customized residual network has been proposed to discriminate PD defects by employing time–frequency (TF) representations for effective extraction of both temporal and spectral characteristics. Further, an optimized explainable architecture is developed to generate fine-grained TF representations by modulating internal parameters of PDResNet. The proposed approach depicts its interpretability in terms of intermediate feature maps and relevance maps across different residual layers of the customized network. Feature activations extracted from shallow and deep layers reveal the progressive evolution from low-level texture patterns to highly discriminative representations specific to PD, allowing visual insights into the model’s decision-making process. Extensive experiments and ablation studies demonstrate the superiority of the proposed X-PDResNet with 98.67% accuracy, 98.68 % precision and 99.33 % specificity. The proposed explainable framework also shows strong potential for non-invasive condition monitoring applications in electrical insulation systems.