Research About Abnormal Sound Multi-class Recognition Algorithm Based On EEMD

In order to maintain social public safety and national stability,it is very necessary to study the identification method of abnormal sounds in public places.At present,traditional voice signal processing methods are also used to process the recognition of the happened abnormal sounds in public places.These methods used for abnormal sound recognition will lead to the low recognition rate of abnormal sounds.In order to solve the problem of low recognition rate of abnormal sounds,this thesis researches from signal feature extraction,recognition and support vector machine(SVM)parameter optimization in three aspects,and puts forward the corresponding improved algorithms respectively.In order to improve the recognition rate of abnormal sound signals,an improved feature extraction algorithm of abnormal sounds based on Ensemble Empirical Mode Decomposition(EEMD)is proposed firstly.According to the characteristics of nonlinear and non-stationarity of abnormal sound signals,EEMD is carried out first for abnormal sound signals and several layers of intrinsic mode functions(IMF)are obtained.Then mel-frequency cepstrum coefficient(MFCC),short-time energy,the ratio of energy of this layer IMF to the energy of original signal and the average of MFCC of each layer IMF are calculated.Finally,the extracted MFCC,short-time energy and energy ratio are combined as the eigenvector of abnormal sounds.The simulation results show that the improved feature extraction algorithm can improve the recognition rate of abnormal sound signals compared with the existing feature extraction algorithms.Aim to the problem of big error accumulation of the traditional decision directed acyclic graph support vector machine(DDAGSVM)algorithm,the DDAGSVM algorithm is improved in the thesis.Before constructing the decision directed acyclic graph,the recognition rate of each support vector machine is calculated.Then,the support vector machine with the highest recognition rate is used as the root node of decision directed acyclic graph,so as to reduce the error accumulation in the classification process and achieve the goal of enhancing the recognition rate.The simulation results show that,the improved DDAGSVM algorithm can effectively improve the recognition rate of abnormal sound signals compared with one against one support vector machine algorithm and the traditional DDAGSVM algorithm.Because the classification performance of support vector machine is greatly influenced by nuclear parameter and penalty factor,the parameters of support vector machine are optimized firstly before the abnormal sound signal identification.In order to solve the problem of parameter optimization of support vector machine,combining with the characteristics of particle swarm optimization(PSO)algorithm and particle filter(PF)algorithm,the support vector machine parameter optimization method based on PSO-PF algorithm is proposed in the thesis.The sampling process of PF algorithm is optimized by PSO algorithm in the method.It makes the particle set move to the region with the higher posterior probability density through continuously updating the speed and position of particles.In order to adjust the global and local search capabilities in the parameter optimization process,the linear weight is established in the parameter optimization method in the paper.The goal is to make the algorithm be able to adjust the global and local search capability of the algorithm flexibly in the process of optimization.The simulation results show that,the proposed support vector machine parameter optimization method based on PSO-PF algorithm in the thesis can improve the classification performance of support vector machine compared with PSO algorithm and genetic algorithm,so as to enhance the recognition rate of abnormal sounds.