A Gas Production Classification Method for Cable Insulation Materials Based on Deep Convolutional Neural Networks

Area of Science:

• Electrical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Evolved gas analysis (EGA) is crucial for assessing power cable insulation health non-invasively.
• Current methods struggle with simultaneous aging mechanisms and recognizing multiple fault patterns in mixed-gas data.

Purpose of the Study:

• To develop an intelligent analytical method for accurate insulation condition assessment.
• To propose a deep convolutional neural network (DCNN)-based multi-label classification framework.

Main Methods:

• Utilized concentration data of six characteristic gases from five insulation materials (EPDM, EVA, SR, PA, XLPE).
• Applied data analysis techniques (logarithmic transformation, Z-score normalization) and DCNN with multi-scale convolution, residual connections, and attention mechanisms.
• Employed weighted binary cross-entropy loss for multi-label classification of degradation states.

Main Results:

• The DCNN model effectively learned material-specific gas generation patterns.
• Accurately identified complex co-occurring fault patterns and multiple degradation states simultaneously.
• Demonstrated superior performance in recognizing concurrent fault scenarios.

Conclusions:

• The proposed DCNN framework enhances the accuracy and comprehensiveness of power cable insulation condition assessment.
• Provides a robust intelligent method for diagnosing complex fault conditions in critical electrical equipment.
• Offers technical support for improving the reliability of power cable infrastructure.