Wavelet-Based Convolutional Neural Network for Denoising Partial Discharge Signals Extracted via Acoustic Emission Sensors

This letter aims to propose a wavelet-based convolutional neural network architecture to eliminate noise from partial discharge (PD) signals produced in an electrical apparatus. The PD signals are recorded via three acoustic emission (AE) sensors placed on the surfaces of an emulated cubical tank developed in the laboratory. For this proposed research, an encoder–decoder framework has been modeled to produce noise-free signals. The advantages of the proposed denoising method are twofold. First, various mother wavelets are employed as convolutional filters in the convolutional block of the encoder and the decoder. Second, discrete-wavelet-transform-based downsampling and upsampling operations have been incorporated to reconstruct the PD signals more effectively than the classical pooling operation. Experiments have been performed taking both the measured and simulated PD signals with various signal-to-noise-ratio levels. Moreover, ablation analysis has been carried out by varying various wavelets during sampling operations. The proposed denoising scheme can also be implemented for condition monitoring in other signal modalities.