Acoustic Partial Discharge Signal Denoising Using a Residual Convolutional Neural Network

Partial discharge monitoring is a crucial diagnostic tool for assessing insulation health in high-voltage power equipment. However, effective PD detection is often hindered by noise contamination, including repetitive pulse-shaped noise, discrete spectral interferences, and broadband white noise. Traditional denoising techniques, such as wavelet transform (WT), empirical mode decomposition (EMD), and singular value decomposition (SVD), struggle with maintaining signal integrity while effectively removing noise. This study proposes a novel residual convolutional neural network (R-CNN) for acoustic PD signal denoising. The model leverages residual learning principles to predict and subtract noise components while preserving critical PD features. Experimental validation uses acoustic PD signals collected from a transformer-based experimental setup. The proposed residual-CNN is evaluated against conventional methods, including SVD and variational mode decomposition (VMD), under different input signal-to-noise ratio (SNR) conditions. Performance metrics, such as SNR improvement, root mean square error (RMSE), and noise reduction level (R<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nl</inf>), demonstrate that the R-CNN significantly enhances denoising effectiveness. The results highlight the model's superior ability to reconstruct noisy PD signals, making it a promising solution for real-time PD monitoring in power transformers.