Research On Detection Technology Of Noise Signal Abrupt Point Based On GSP In Impulsive Noise Environment

Graph Signal Processing(GSP)is currently a popular research direction in the field of signal processing.At present,the GSP processing framework has a lot of research and applications such as Wireless Sensor Networks(WSN),Biomedical Science,Artificial Intelligence(AI),Image Processing,Machine Learning(ML),Big Data and other scenarios.It aims to process a series of data with high dimension and complex structure that cannot be solved by traditional digital signal processing methods.Based on Digital Signal Processing(DSP),researchers have proposed related basic concepts such as Graph Fourier Transform(GFT),Graph Laplacian Matrix(GLM),Graph Time Shift Operator(GTSO),and Graph Filter(GF)in the GSP processing framework.These concepts lay the theoretical foundation for many specific applications in GSP.Among many applications of graph signal processing,the detection of signal abrupt points in wireless sensor networks is an important concern.When we need to judge the accurate position of TIN(Transient Impulse Noise),this topic will be of great value.Doctoral students in our team have also conducted research on this topic with compressed sensing technology.In the field of signal abrupt points detection,methods based on Cumulative Sum(CUSUM)are commonly used.Besides,approximate likelihood ratio,joint estimation and detection methods are also proposed.These methods are mainly based on the prior knowledge of mean and variance related to the data before and after the signal abrupt,algorithm complexity,genres of detectable signals,etc.,but there are still some restrictions on performance.For example,the joint detection estimation method can replace unknown parameters with estimation,but the detection performance is not as good as the pure CUSUM method,so there is no best solution.At the same time,this type of classic digital signal processing method needs to know the statistics information associated with the data which after signal abrupt.In the real scenes,these statistics are often unknown because of the variety of TIN(Transient Impulse Noise).Therefore,some professors have tried to conducted the research on the graph signal abrupt points detection under the GSP theoretical framework[4].The GSP method mainly only uses the prior knowledge of the data that before signal abrupt,and then calculates the cumulative sum,finally maximizes and corrects the statistics.A series of experiments have verified the feasibility of the GSP processing scheme.However,the topic of the accuracy and stability of the graph signal abrupt points detection remains to be studied and improved.Therefore,based on the performance of graph signal abrupt points detection,the main contributions of this thesis are as follow:(1)One of the key issues in graph signal processing technology is to construct an effective graph.Under the premise of adopting the random graph structure,the scheme of using the random step size normalized Laplace matrix in the steps of calculating statistics and using the correlation coefficient to define the weights in the steps of fusion statistics is proposed to improve the detection accuracy and stability of the baseline system.The original graph signal abrupt points detection baseline system uses the combined graph Laplace matrix in the steps of calculating statistics and uses the reciprocal of the dimension to define the edge weights in the steps of fusion statistics.In the graph signal processing framework,the edge connecting the two vertexes represents the connection between them,and the weights on the edges indicate the degree of correlation between two vertexes,so theoretically there should be better detection performance by defining the weight with the correlation coefficient.This thesis uses commonly used graph Laplacian matrix: combined Laplacian matrix,random step size normalized Laplacian matrix,symmetric normalized Laplacian matrix with the weight defined by dimension reciprocal definition or correlation coefficient.Take simulation experiments to verify that the weight defined by correlation coefficients combined with random step size normalized Laplacian matrix is more accurate and the detection results are more stable.(2)In the steps of fusing statistics,this thesis proposes a self-spin graph structure to eliminate isolated vertices,and the optimal weight distribution scheme is discussed to further improve the detection accuracy of the graph signal abrupt points detection system.The random graph used by the original baseline detection system may make some vertexes become isolated vertexes.When the signal on the vertexes of the graph changes suddenly,the statistics on the isolated vertexes will not be fused and calculated because they don't have any connections with other vertexes which will cause more detection delay.Based on this issue,a self-spin graph structure is proposed.Each vertex in the graph structure has three edges:one is connected to itself as a self-spin edge,and the other two edges are connected to the front and back vertexes around it.This kind of graph structure eliminates the isolated vertexes and achieve full coverage of statistics on all graph vertexes.Through simulation experiments,it is first verified that the directionless self-spin graph structure has better detection accuracy than the directed self-spin graph structure;secondly,different weight distribution schemes are discussed to verify that when the self-spin edge weight is 0.9,other two non-spin edge weight is 0.05,the performance is better;and then compared with the classic DSP method which proposed not long ago-Bayesian fastest detection method[5],it is found that the improved GSP scheme proposed in this thesis has better performance;finally,the scene of multiple abrupt points has been considered and tested by simulation which verifies that each abrupt points can be successfully detected independently.(3)The improved graph signal abrupt points detection scheme is applied to detect the start and end points of the real speech signal mixed with TIN,which verifies the effectiveness and the robustness of the detection system.In the real environment,thunder,engine start,and rapid door knocking and so on are all transient impulsive noises,which will seriously affect the quality of speech signals when they suddenly appear in the speech.Although there has been much research on deleting noise or speech enhancing.However,in the case of TIN,due to the complexity and variability of TIN noise and its characteristics are quite different from stationary noise,if we directly perform normal methods,it may inevitably cause the distortion of the normal speech segment in a certain extent.Considering the short duration of TIN,for the speech segment mixed with impulse noise,we can first use the improved graph signal abrupt points detection scheme proposed in this thesis,and then carry out specific follow-up processing such as speech enhancement for TIN.Therefore,the detection of robust TIN location is very important.In this thesis,on the basis of the detection of the abrupt change points of the noisy signal,the end point of the pulse is further detected,and then the noisy speech is extracted from the entire speech,and the method of minus the mean is used to eliminate noise.The effectiveness of the scheme is verified through experiments.