Research On Detection Technology Of Abnormal Sound Events In Prison Based On Pattern Recognition

At present,speech signal processing is mainly aimed at the synthesis,separation,tracking and recognition of human speech.Researchers use computers to extract the features of speech signals,construct a high-dimensional feature space for these features and put them into a speech classifier designed by the machine.According to the classification results,the machine makes different responses,which is to let the machine understand human speech.However,human speech signal is only a part of the sound signal,and it has become a new research hot spot in recent years to recognize the happening events from the sound signal.In addition,security is a project that the state attaches great importance,among which the prison as a place to hold prisoners,its security work needs to be more comprehensive and meticulous.The domestic prison security system is mainly based on single video surveillance,which cannot give full warning to the emergencies in the prison.Therefore,the research on abnormal sound recognition technology in prison is a beneficial exploration in this paper.The main work is as follows:Firstly,according to the prison environment,this paper selects 11 kinds of abnormal events which may appear in the prison and constructs the sound repository based of abnormal events.In addition,using the method of data enhancement to solve the problem of data shortage.The algorithm of abnormal sound event recognition in prison based on GMM is studied.By preprocessing the prison abnormal sound repository,five kinds of sound characteristics including MFCC,?MFCC,??MFCC,BFB and short-time energy were extracted,and experiments were conducted.The experimental results show that the accuracy of GMM model using MFCC characteristic parameters is higher.In addition,when the order of the Gaussian mixture model reaches 32,the recognition effect does not change.Among them,the identification accuracy of siren is the highest.When extracting 12-dimensional MFCC characteristic parameters,the accuracy can reach 70%.After adding the difference coefficient of MFCC,the identification rate can be further improved,reaching 74.4%.Several kinds of neural network structures such as multilayer perceptron,convolutional neural network and recurrent neural network for abnormal sound event recognition are studied.A bilinear structure neural network with the fusion of convolutional neural network and recurrent neural network is proposed,and the recognition accuracy rate of the prison abnormal sound is 89.17%.The recognition rate of the bilinear neural network is more than 5% higher than the common convolutional neural network and recurrent neural network.