Partial Discharge Recognition with a Multi-Resolution Convolutional Neural Network.

Area of Science:

• Electrical Engineering
• Materials Science

Background:

• Partial discharge (PD) is a critical indicator of insulation system degradation in gas-insulated switchgear (GIS).
• Monitoring ultra-high-frequency (UHF) signals is an effective method for assessing insulation health and classifying PD events.
• Existing methods require robust algorithms for accurate PD classification.

Purpose of the Study:

• To propose and realize a novel deep learning-based algorithm for classifying PD in GIS.
• To enhance PD classification accuracy by utilizing multi-resolution UHF spectra.
• To validate the proposed algorithm's effectiveness against established methods.

Main Methods:

• A multi-column convolutional neural network (CNN) with specialized temporal, frequency, and texture filters was developed.
• A long short-term memory (LSTM) network was employed for fusing multi-sensor information.
• A transfer learning approach inspired by manifold learning was used to mitigate overfitting.

Main Results:

• The proposed algorithm achieved superior performance in classifying 13 simulated GIS defect modes compared to 18 baseline methods.
• Advanced visualization techniques provided qualitative interpretations of learned features.
• The effectiveness of the architecture was explained through a matrix projection framework.

Conclusions:

• The developed deep learning algorithm offers a highly effective solution for PD classification in GIS.
• The integration of multi-resolution UHF spectra and advanced neural network architectures improves diagnostic accuracy.
• The study provides a robust framework for enhancing the reliability of insulation monitoring in critical electrical infrastructure.