Washing Machine's Working States Recognition Based on Audio Fusion Feature

In order to effectively determine the normal working status of smart home appliances, this paper realizes the recognition of washing machine's working states based on audio signals of microphone array (6 sensors). The paper proposes a fusion classification method combining optimizing STFT features and LSTM, and compares multi-channel feature fusion and voting classification methods. Determine the relatively optimal frequency resolution feature by changing the frequency resolution of STFT. The multi-channel feature fusion classification method is to concatenate the features of six different channels, input them into the LSTM network model for training, and obtain a comprehensive LSTM classifier. Voting fusion classification method: Train a LSTM network model independently for each channel's feature, and obtain 6 corresponding LSTM sub-classifiers, and then vote for the predicted classes of 6 channels corresponding to the same sample to obtain the fusion classification result. A vertical axis washing machine is taken to verify the proposed method, the experiment results show that both fusion classification methods have better performance than single channel methods. Compared with the multi-channel feature fusion method, the recognition rate of the multi-channel feature fusion method is slightly higher than that of the voting classification method, with a maximum of 99.7%.