ERB-EWT For The Feature Extraction Of Abnormal Sounds In Public Places

The safety of public places is related to the social stability,the lives and property of the persons.The security and monitoring of public places has always been the top priority of the national security strategy.At present,the safety of the public places are mainly based on video surveillance.However,the effect of the video surveillance may be decreased when there is a lack of light or blind spots.We know that the occurrence of anomalous events is usually accompanied by the abnormal sounds.Therefore,the monitoring of abnormal sound in public places is an important development direction of safety monitoring in public places.The research on the extraction of abnormal sound features in public places involved in this paper is a core theory and technical problem in this field.In this dissertation,an empirical wavelet filter bank is constructed by simulating the equivalent rectangular bandwidth.The feature extraction method of Empirical wavelet transform(ERB-EWT)based on equivalent rectangular bandwidth is proposed.The relevant experiments prove that the method proposed in this paper is effective and feasible.The main work of this paper is as follows:(1)The research status of research methods for abnormal sound features extraction at home and abroad.We focus on the relevant methods and applications of sound signal feature extraction at home and abroad.(2)Analysis of basic processing methods for abnormal sound in public places.By analyzing the current feature extraction method of abnormal sound in public places,the one is speech signals analysis method,and the other is the time-frequency analysis and decomposition method.We concluded that the methods based on voice signal analysis method can quickly extract the characteristics of the sound signal,but its robustness to environmental noise is poor.The methods based on time-frequency analysis are more robust to the environment,but the drawbacks of such methods are the lack of theory and the time-consuming is longer.(3)Equivalent rectangular bandwidth empirical wavelet transform is used to extract the abnormal sound features in public places.It is a new method of modal decomposition.It has a reliable mathematical derivation process,which is more stable in signal decomposition,easier to calculate,and faster.However,when EWT is dealing with complex spectrum signals such as abnormal sound in public places,the inherent mode of the signal is not easily determined,resulting in difficult boundary segmentation,and the established filter bank cannot reflect the amplitude-frequency characteristics of the signal well.In this paper,the Equivalent Rectangular Bandwidth(ERB)can reflect the characteristics of the human auditory mechanism,and an improved empirical wavelet transform(ERB-EWT)based on the simulated equivalent rectangular bandwidth is proposed for extracting the future of the abnormal sound in public places.Firstly,the Fourier spectrum of the signal is divided based on the equivalent rectangular bandwidth,rather than the method of dividing the Fourier spectrum of the signal according to the extreme point of the signal spectrum in the empirical wavelet transform.The resulting boundary is substituted into the empirical wavelet transform to construct the ERB-EWT filter bank.The ERB-EWT proposed in this paper has nothing to do with the abnormal sound signal in the public place,and does not need to preset the number of intrinsic modalities of the abnormal sound signal.Then,the ERB-EWT is used to decompose abnormal sounds in public places to obtain a series of different modes.The logarithmic energy is calculated from each modal and the logarithmic energy is composed of a one-dimensional vector as the characteristics of the abnormal sound.Vectors are used for classification recognition.(4)The relevant verification experiments are designed and conducted.The experiment is divided into two parts.First,the performance test of the proposed method is proposed.The equivalent rectangular bandwidth empirical wavelet transform proposed in this paper decomposes the signal.It takes less time,and the reconstruction error is little.The second is comparison of experiments.Compared with methods based on speech signal analysis,the robustness to noise is better and the recognition result is better.Compared with methods based on time-frequency analysis,feature extraction of our method takes less time and better recognition results.