Artificial Intelligence-based Partial Discharge Pattern Recognition for Enhanced Diagnostic Monitoring of HV Cable Systems

Partial discharge (PD) analysis is a critical diagnostic tool for ensuring the reliability and longevity of high-voltage (HV) cable systems. Conventional PD recognition techniques are often hampered by limited robustness to noise, reduced sensitivity to weak discharges, and an inability to generalise across various insulation defects. This paper presents a AI-based intelligent system that will identify and categorise the PD patterns with a level of accuracy in real time. Using deep learning frameworks such as Convolutional Neural Networks (CNNs), Gated Recurrent Units (GRUs), and Transformer encoders, the system takes time-frequency representations of PD signals. It classifies them into major discharge categories: internal, surface, corona, and noise. The dataset was filtered from a variety of sensor sources, preprocessed, denoised, normalised, and labelled for supervised training. The model was shown to perform exceptionally well, with average classification measures across five evaluation sets: accuracy of $97.88 \%$, precision of $94.22 \%$, and recall of $94.96 \%$. Branches of CNNs based on spectrograms have been effective at capturing spatial emission patterns, whereas transformer attention modules have been effective at capturing temporally complex discharges. The proposed system will enable monitoring of HV networks at any point and provide an opportunity to detect faults early, significantly reducing maintenance costs and unexpected failures. The strong performance of the architecture across various PD categories highlights its level of generalizability and implementability.