Partial Discharge (PD) Signal Recognition in Fiber-Optic Acoustic Emission Using a Lightweight Attention Mechanism (ECA)

Abstract

Partial discharge (PD) is a primary cause of insulation degradation in electrical equipment. Consequently, the investigation of PD signals is of paramount importance for enhancing the safe operation of power devices. This paper delves into the principles of four typical PD signals and designs experiments to simulate four common PD mechanisms: Along Surface, Igniter, Suspension and Tip. Utilizing the latest fiber-optic acoustic emission system, we constructed an experimental platform for PD signal acquisition to capture acoustic emission signals under various voltage gradients. These signals are then denoised using a wavelet threshold algorithm. Subsequently, the denoised acoustic emission signals are segmented into training set samples. Based on the ResNet model, we incorporate an ELU activation function to enhance the model and introduce both CA and ECA attention mechanisms. Through extensive comparative experiments, we select the most effective model. The experimental results demonstrate that the ELU activation function significantly improves the model’s recognition performance. Additionally, the ECA attention module, added after stage13, exhibits excellent recognition capabilities for the four typical PD signals, achieving an accuracy rate of 98%.

Keywords

• Deep learning
• attention mechanisms
• partial discharge (PD)
• fiber-optic acoustic emission sensor (FOAES)
• signal recognition