Research On Multi-class Recognition Method Of Abnormal Sound Based On Time-frequency Analysis

Public places are part of people's lives and work,the level of protection is related to everyone's vital interests.The classification and recognition of abnormal sounds in public places is very important practical significance.At present,methods similar to processing speech signals are often used to identify abnormal sounds.Due to the particularity of abnormal sounds,the classification effect obtained by these methods is not ideal.In order to explore better methods for identifying and classifying abnormal sounds,research has been carried out from the time-frequency domain analysis of the signal,combined feature extraction,time-frequency spectrum characteristics extraction,and deep convolutional neural network construction.In order to clarify the characteristics of the abnormal sound signal in the time domain,frequency domain,and time-frequency domain.The time domain features short-term energy and short-term average zero-crossing rate,frequency domain features Mel frequency cepstrum coefficients,inverted Mel frequency cepstrum coefficients and linear prediction coefficients,time-frequency domain features Hilbert time-frequency spectrum and marginal spectrum are discussed.The results show that time domain features are the most intuitive classification feature,frequency domain analysis can highlight certain features that cannot show in the time domain,time-frequency domain features can intuitively express the relationship between the signal sample time point and frequency.Start with the conclusions of time domain analysis and frequency domain analysis,in order to optimize the efficiency of combined features in abnormal sound recognition and classification,an abnormal sound validity detection algorithm and a multi-layer feature extraction algorithm are proposed based on ensemble empirical mode decomposition.Firstly,the ensemble empirical mode decomposition is performed on the abnormal sound frame signal to obtain the intrinsic mode function.Then,according to the given intrinsic mode function layer number threshold,the validity detection of the frame signal is performed.Secondly,Mel frequency cepstrum coefficients,inverted Mel frequency cepstrum coefficients,linear prediction cepstrum coefficients,short-term energy and energy ratio for each layer of intrinsic modal function of the effective frame signal are extracted.Then,normalized these features and spliced into multiple layers feature.Finally,a deep convolutional neural network based on the Inception-v3 model is built to conduct simulation experiments.The simulation results show that the multi-layer feature as a combination feature that is effective,fast convergence,and robust to noise,and the recognition rate can reach 98.65% in the four types of abnormal voice recognition.In order to improve the efficiency of the Hilbert time-frequency spectrum with high frequency resolution in the recognition and classification of abnormal sounds,a time-frequency spectrum stabilization method based on the idea of convolution operation and an adaptive enhancement method are proposed.Firstly,the time-frequency spectrum and the marginal spectrum with high frequency resolution are obtained by the Hilbert-Huang transform,and the corresponding kernel matrix is obtained by setting the time-domain and frequency-domain stationary factors and weights,and the time-frequency spectrum is smoothed according to the process of convolution operation.Then,the marginal spectrum values are used as the enhancement factor to adaptively enhance the spectral lines in the corresponding time-frequency spectrum.Finally,the time-frequency spectrum is used as the classification feature,and the deep convolutional neural network is used to realize the classification and recognition of abnormal sounds.The simulation results show that the proposed convolution stabilization method can effectively change the expression form of the Hilbert time-frequency spectrum,and the adaptive enhancement method can make the overall expression effect of the time-frequency spectrum becomes clear,and both can improve the recognition rate of abnormal sounds.