Classification of abnormal sounds around power system based on multi-feature fusion and deep learning

Abnormal sound monitoring of power system operating environment is one of the key technologies to ensure the safe and stable operation of equipment. In this paper, a classification method of monitoring equipment in power system based on multi-dimensional acoustic feature fusion and improved Visual Geometry Group 16 (VGG16) deep learning model is proposed to realize efficient recognition of abnormal sounds. This method first extracts Mel Spectrum, Mel Frequency Cepstral Coefficients (MFCC) and Chromagram (Chroma) features from the original audio, and fuses them into three-channel image input through standardization and space-time alignment, combining with independently processed MFCC time series features. Then, an improved VGG16 model is designed for the time series characteristics of sound in power system monitoring scenes, which retains its convolution feature extraction ability, and introduces Bidirectional Gated Recurrent Unit (BiGRU) and attention mechanism to enhance the modeling and fusion ability of timefrequency features. The method is verified on the data set of five typical environmental sounds of power system, including birdsong, construction noise, background noise, thunder and train noise. The experimental results show that the classification accuracy of the method on the test set reaches $\mathbf{9 4 \%}$, and the F1 score is $\mathbf{0 . 9 4 1}$. It provides an efficient scheme for monitoring power system.