Abnormal Sound Feature Extraction Method In Public Places Based On CELMDAN

Currently, the public security monitoring field is mainly composed of video surveillance, and because of having no more breakthrough in core theory and technology of audio monitoring, the function of existing audio monitoring systems only include simple sound acquisition, transmission and so on, being lack of abnormal sound intelligent recognition. The accompanying abnormal sounds when abnormal event occurs always contain a large number of relevant information, so audio monitoring can be used as a reasonable supplement of video surveillance. This has become the development direction of public security monitoring field research. Public places' abnormal sound feature extraction involved in this project is the core technology of intelligent audio monitoring. Therefore, our project has great social significance and theoretical research value.Abnormal sound feature extraction methods mostly use typical parameter or a combination of several parameters in speech signal processing, such as short-time zero crossing rate, short-time average energy and Mel-Frequency Cepstrum Coefficient(MFCC), and these feature extraction methods get pretty effect in a certain range. But due to the particularity of abnormal sounds, the feature extracting effects of above parameters have great limitations. Local Mean Decomposition(LMD) is a better processing method for nonlinear and non-stationary signal, having been successfully applied in machinery fault diagnosis, brain wave analysis, seismic signal analysis, etc. This dissertation puts forward to improve LMD and then use it in public places' abnormal sound feature extraction. According to the characteristics of abnormal sounds, firstly, this paper analyzes the existing problems of original LMD in theory, and improves its inherent problems such as endpoint effect and high time-consumption; secondly, we propose a way of adaptively adding noise, which alleviates the mode mixing problem of LMD by introducing noise, and this paper proposes Complete Ensemble Local Mean Decomposition with Adaptive Noise(CELMDAN) and then uses it in public places' abnormal sound feature extraction. Related validation experiments show that our proposed approach outperforms traditional MFCC and other time-frequency analysis methods, while having higher abnormal sound recognition rate.The main work is as follows:(1) Doing theoretical research on endpoint effect and decomposition's high time-consumption of LMD, to improve its decomposition effect.1) The extreme points of abnormal sound signal has small spacing with each other and close distribution, while LMD did not consider the extremum condition on both ends of the target signal, causing the component distortion phenomenon spread from both ends to middle of signal, which is called end effect. For this problem, this paper presents a boundary processing method to accurately estimate the extreme point information about the signal endpoints, which is aimed to avoid the decomposition results' endpoint effect caused by the extreme information distortion in that position from the source. Final results are verified through the experiment.2) Abnormal sound signal fluctuates frequently and has rich local information, while the moving average process of LMD is time-consuming and will lose some local information of signal. For this problem, this paper uses linear interpolation method to replace the nagging moving average process in LMD, which ensures the information integrity and at the same time reduces its computational complexity. Besides, abnormal sound signal usually lasts long and the main information is contained in high frequency components, while the uncertainty of Product Function(PF) components' order and screening time in LMD can cause high time-consumption and affect the decomposition effect. For this problem, this paper solves the unsure problem of PF components' order through statistical analysis of LMD decomposition results, and then regards decomposition result as the feedback assessment of screening number, where we choose the screening time corresponding to the best decomposition result, to reduce decomposition time and also avoid the over-sifting or under-sifting phenomenon. Final results are verified through the experiment.(2) Proposing abnormal sound feature extraction method in public places based on CELMDAN.1) The frequency components of abnormal sound signal are complex, while mode mixing problem of LMD would certainly influence the feature extraction effect. The Ensemble Local Mean Decomposition(ELMD) method can effectively alleviate the mode mixing problem; however it brings some new ones such as non-negligible reconstruction error and residue noise. For this problem, this paper uses the basic idea of ELMD, combined with the improvement measures of LMD's endpoint effect and high time-consumption in section(1), and our proposed method is called CELMDAN. Its characteristic is the introduction of nested decomposition, which means that in the ith link of adding noise, we add the(i-1)th PF component of Gaussian noise onto residue, and then obtain the first PF of that mixed signal, and then do LMD to get its first PF, after many repeated times, we take the average as the ith component of our method. This paper theoretically proves that our proposed method is complete, which means that the reconstruction error is zero.2) In order to verify the effectiveness of the proposed CELMDAN method, this paper conducts feature extraction and recognition experiments on analog signal and on public places' typical abnormal sounds. Firstly, the result of analog signal shows that, while guaranteeing the ideal decomposition effect, our method can effectively solve the problem of endpoint effect and mode mixing, and the reconstruction error of CELMDAN is far less than that of ELMD. Secondly, the feature extraction and recognition result of public places' abnormal sound database shows that, our proposed CELMDAN has better ability of feature description than MFCC and other time-frequency analysis method.(3) According to the proposed method, this paper designs and implements a demonstration system of abnormal sound detection and recognition. It can realize many functions, such as test signal sequence synthesis, endpoint detection and recognition of abnormal sounds, visualization of recognition results and so on.