PMSNet: Multiscale Partial-Discharge Signal Feature Recognition Model via a Spatial Interaction Attention Mechanism

Area of Science:

• Electrical Engineering
• Materials Science
• Signal Processing

Background:

• Partial discharge (PD) is a critical indicator of insulation degradation in electrical equipment.
• Traditional PD signal identification methods struggle with noise interference, limiting their effectiveness.
• Accurate PD identification is essential for ensuring the safe operation of electrical systems.

Purpose of the Study:

• To develop a robust and accurate method for identifying partial discharge (PD) signals.
• To overcome the limitations of traditional methods in handling noisy PD data.
• To improve the assessment of electrical equipment insulation status.

Main Methods:

• A novel PD signal identification method based on multiscale feature fusion using PMSNet.
• Utilizing a CNN backbone with a multiscale feature fusion pyramid and DSFB modules.
• Incorporating a Transformer encoder with spatial interaction-attention for enhanced feature interaction.
• Employing a final classification feature generation module (F-Collect) for PRPD map recognition.

Main Results:

• PMSNet achieved over 80% validation accuracy and exceeded 85% training accuracy on the PRPD dataset.
• The method demonstrated a 10% improvement in recognition accuracy compared to traditional high-frequency current and current pulse detection methods.
• Validation loss remained low at approximately 0.3%.

Conclusions:

• PMSNet significantly enhances the recognition accuracy and robustness of partial discharge signal identification.
• The proposed method offers a practical and effective solution for assessing insulation status and preventing electrical equipment failures.
• The multiscale feature fusion approach shows strong potential for real-world applications in electrical safety.