A method for domestic audio event recognition based on attention-CRNN

To monitor the equipment in the power grid's communication room intelligently, so as to find the equipment operation failure in time, a method for domestic audio event recognition based on attention-CRNN is proposed. The method is divided into 3 steps: (1) Extracting Mel Frequency Cepstral Coefficients (MFCC) from audio data; (2) Using convolutional neural network (CNN) to extract audio features, and using this feature as the input of a recurrent neural network Bi-GRU (Bidirectional Gated Recurrent Unit), and assigning attention weights to the features in time series; (3) Putting the time series features into the fully connected layer for forward propagation, using the Softmax function as classifier and the cross-entropy as loss function, and using adam gradient descent method to reversely update the model weights. We use the international competition DCASE 2018 Task 5 domestic audio data set for experiments. The experimental results show that the f1_score value of this method is 89.6%, which is an increase of 3.9% compared to the 85.7% of the official baseline given by the competition.