A Deep Learning Framework for Noise Elimination of Partial Discharge Signals

Noises can contaminate acoustic partial discharge (PD) signals, resulting in inaccurate detection and interpretation. As a result, denoising acoustic PD signals has achieved significant attention to enable precise diagnosis and analysis. This manuscript presents a deep learning (DL) method for denoising acoustic PD signals obtained from a high-voltage power apparatus. The obtained signals are contaminated with various levels of white Gaussian noises followed by two-dimensional time-frequency transformation via continuous wavelet transform (CWT). The CWT images are then fed to a fully convolutional network (FCN) based autoencoder (AE), that supports skip connections, to perform denoising on the acoustic PD signals. The proposed FCN based AE is validated upon four mother wavelets, and it has been observed that the best performance is achieved for the ‘Daubechies’ wavelet. Moreover, the denoising performance has been assessed using various evaluation metrics, where different levels of white Gaussian noises are added. Experimental results reveal that the proposed denoising method outperforms the existing DL quantitatively. The proposed scheme can also be used to monitor and diagnose other signal modalities in real time.