A recognition method for abnormal substation noise based on non-intrusive decomposition technology

This study proposes a recognition method for abnormal substation noise based on non-intrusive decomposition technology. The aim is to achieve efficient separation and accurate identification of substation noise through advanced signal processing and deep learning technologies. Firstly, this method uses non-contact sound sensors to collect noise data at the plant boundary, including the operating noise of transformers and reactors. Subsequently, after data preprocessing and multi-scale decomposition of wavelet transform, combined with short-time Fourier transform (STFT) and Mel-frequency cepstral coefficients (MFCC) to extract time-frequency features, and then conduct joint analysis of time series and spectrum features through convolutional neural network (CNN). Finally, the model judges the noise state and distinguishes normal, early fault and fault states. Compared with traditional methods, this study has significant advantages in reducing monitoring costs, adapting to complex environments and accurately locating abnormal noise, providing an effective solution for fault detection of substation equipment.