Graph-Based Framework with Waveform-Informed Connectivity for Multi-Label Partial Discharge Source-Type...

Area of Science:

• Electrical Engineering
• Materials Science
• Data Science

Background:

• Partial discharge (PD) source-type classification is crucial for maintaining high-voltage apparatus.
• Current methods struggle with stochastic interference and multiple PD sources.

Purpose of the Study:

• To develop a robust graph-based framework for multi-label PD source classification.
• To integrate waveform morphology and event characteristics for improved accuracy.

Main Methods:

• A multi-task neural network extracts pulse embeddings and confidence scores.
• Graph construction uses spatial proximity and morphological similarity for connectivity.
• An edge-conditioned graph neural network classifies PD sources using weighted message passing.

Main Results:

• The framework achieved a Matthews correlation coefficient (MCC) of 0.98 and an exact match ratio of 0.97.
• It significantly outperformed baseline methods in single-source, noisy, and multi-source conditions.
• Robustness confirmed on an independent dataset (MCC of 0.93) without retraining.

Conclusions:

• The proposed graph-based framework offers superior performance for PD source classification.
• Integrating waveform morphology and relational structure is key to its effectiveness.
• The method demonstrates high accuracy and robustness, particularly in challenging multi-source scenarios.