Research On Source Enumeration And Source Separation Algorithm Under Underdetermined Conditions

With the rapid development of artificial intelligence and digital communication,more emerging industries such as intelligent transportation,smart home and biological monitoring have gradually embedded sound sensors into smart devices.They have played a role in characterizing the surrounding environment and identifying sound scenes by capturing and perceiving of sound signals,thereby realizing the detection of biological signs and the early warning of safety events.However,in the real environment,all kinds of sound sources are often overlapping and randomly mixed,and the sensors only receives the sound signal mechanically,making the received observation signals complex and difficult to distinguish.In addition,due to the limited number of sensors embedded in smart devices,and the number of sound sources in the environment cannot be predicted,making it more difficult to achieve sound source separation under such an underdetermined condition.It is particularly significant to find an effective sound source separation algorithm and source number estimation algorithm under such underdetermined conditions.Therefore,in this paper,the source separation based on Independent Component Analysis(ICA)is deeply studied.In order to solve the problem of source enumeration under the underdetermined conditions,the Empirical Mode Decomposition(EMD)is introduced,and the improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(ICCEMDAN)and Singular Value Decomposition(SVD)are combined to solve the problem of source enumeration.ICCEMDAN and FastIndependent Component Analysis(FastICA)are integrated to separate sources on the premise of correctly predicting the number of sound sources.Aiming at the problem of blind source separation and source enumeration under the underdetermined conditions,the research content of this paper is as follows:(1)In order to solve the problem of source enumeration,ICCEMDAN is introduced.First of all,the limited observation signals are decomposed to get a set of intrinsic mode function(imf)and a remaining component;Then,the original observation signals and the obtained components were re-tensioned into a new vector,the covariance matrix of the new vector was calculated,and singular value decomposition was performed on the new vector;Finally,the singular values are arranged successively,the difference of adjacent singular values is calculated,the boundary between signal subspace and noise subspace is traced,and the number of source signals is determined.Simulation results show that this method can predict the number of sources effectively and accurately.(2)In order to solve the problem of source separation,firstly,the observed signals are divided into a series of imf by ICEEMDAN.Then,imf is selected and restructured based on correlation and kurtosis to expand signal dimension and construct virtual multi-channel signals.Finally,FastICA was used to analyze the newly constructed observation signals to realize the separation of sound sources.Simulation experiments on the mixing of analog signals and the mixing of ambient acoustic signals show that the proposed method can improve the correlation coefficient,signal distortion ratio and global mean square error.