Research On Abnormal Audio Event Detection Based On Convolutional Neural Networks

Audio signal,as a common carrier of specific information in life,has become one of the most effective ways to obtain and disseminate information for human,so it is widely used in the fields of biomedicine,industrial production and agricultural supervision.With the improvement of economic level,it is necessary to construct a mature social security system in order to create a more stable and safe social environment.The surveillance system has become an important means to guarantee and maintain social security.Traditional video surveillance,due to its own limitations,is often not ideal for its actual monitoring.Therefore,audio surveillance has received great concern,the information carried in the audio signal can assist video surveillance to improve the existing monitoring system.Audio event detection is the core and key of audio surveillance.Because the environment sound is more complex,diverse and irregular,traditional acoustic models(Gaussian mixture model,support vector machine and hidden Markov model,etc.)lack the ability of modeling,resulting in obvious defects.In recent years,the concept of deep learning has been successfully introduced into the field of audio event detection,which effectively improves the detection effect.Therefore,based on the analysis of abnormal sound characteristics and the modeling of environmental noise in public places,the applicability and recognition performance of convolutional neural network in abnormal acoustic event recognition are studied in this paper.Moreover,the influence of different dimension of the network model on the noise robustness and error convergence rate is compared in detail.The main contents of the work are as follows:(1)Research on audio signal processing front-end algorithm in the acoustic environment of public places.Through the analysis of the environment noise components in public places,the distribution and time-frequency characteristics of the noise under the acoustic environment are obtained,and a hybrid model of noise and abnormal sound in public places is established.On the basis,endpoint detection algorithm and audio denoising algorithm are determined respectively.In the part of sound denoising,the abnormal sound denoising method based on the improved adaptive filter is studied,and experiments show that the proposed method can effectively suppress the noise even under low SNR.In the part of endpoint detection,two kinds of endpoint detection methods,short-time energy and zero crossing rate two-threshold method and adaptive sub-band spectral entropy method,are analyzed experimentally.It is concluded that the adaptive sub-band spectral entropy method has better stability and accuracy in noisy environment.(2)Research on the solution.to the problem of relative scarcity of training data with good annotation.The acquisition of efficient and accurate acoustic models requires a large amount of training data to support it.The study of acoustic models is greatly restricted because of the lack of training data.In this paper,data augmentation is proposed to solve the problem of the scarcity of abnormal sound samples with labels.Under the premise that the actual meaning of the sample label is consistent,new data will be amplified from the existing training samples through a variety of appropriate transformation.The acoustic model that uses training samples after data augmentation to learn has better robustness and generalization ability to unknown factors so that it can be better promoted in the complex environment of public places.(3)Research on abnormal sound recognition based on convolutional neural network.The applicability and recognition performance of convolutional neural network in abnormal sound are analyzed and verified by comparing it with Gaussian mixture model and BP neural network.At the same time,the structure of convolutional neural network is changed and adjusted according to the one-dimensional characteristics of audio signal,and compares it with the traditional convolution neural network in recognition ability,noise robustness and error convergence speed,etc.The experimental results show that the simplified convolutional neural network produces higher error convergence rate and average accuracy with the relative increase of 2.91%in the noiseless environment.Nevertheless in noisy context,the traditional convolution neural network performs slightly better.(4)Research and implementation of abnormal audio event detection system based on convolution neural network.Based on the summary and induction of audio signal front-end processing and audio event detection algorithm,an abnormal audio event detection system which can select dimension of convolutional neural network model through different SNR conditions is implemented.The system was developed on Matlab platform,which mainly has the function of audio event detection audio acquisition and input,front-end processing,model training and so on.The actual test results show that the system can achieve a good recognition effect.