Background Discharge faults in the insulating support plate of Tesla-type pulse sources threaten system reliability and safety. Traditional fault identification relies heavily on manual inspection, which suffers from low efficiency, poor consistency, and strong dependence on operator experience.

Purpose This study aims to develop a data-driven intelligent fault identification method for discharge faults in Tesla-type pulse source insulating support plates, overcoming the limitations of manual diagnosis and providing a reliable technical solution for practical applications.

Methods Simulation data of discharge faults were generated to train two types of models: a feature engineering-based classification model using Support Vector Machine and an end-to-end fault identification model based on a Wide Deep Convolutional Neural Network. Both models were systematically compared in terms of accuracy and recall rate across multiple fault categories. The selected superior models were then validated using independent experimental data to assess their generalization capability.

Results SVM and WDCNN achieved optimal and stable identification performance for most fault categories. Independent experimental validation further confirmed that both models accurately identified discharge fault states, demonstrating strong generalization ability and practical effectiveness.

Conclusions The proposed data-driven method provides a reliable and efficient approach for intelligent fault identification of Tesla-type pulse source insulating support plates. The validated SVM and WDCNN models are recommended for practical deployment, offering a viable pathway to replace manual diagnosis in pulsed power systems. Future work may extend this method to real-time online monitoring.