Research On Analysis Method Of Mutant Components In Time-varying Signal

Nowadays,information has become one of the most important resources in the information age and the information usually corresponds to the components in the signal.Whether the required signal components can be effectively and accurately obtained from the signal is an important research problem.Among the many types of signals,time-varying signals have always been the focus of research,such as EEG and audio,which are typical representatives of time-varying signals,with nonlinear and non-stationary characteristics.In most cases,it can be considered that the timevarying signal is composed of two components,the steady component and the mutant component,where the mutant component is the required signal component and contains key information.Thus,this thesis proposes a set of methods for analyzing mutant components in time-varying signals,specifically divided into three parts: mutant component detection method,mutant component representation method and mutant component identification method.The mutant component detection method can effectively and accurately detect the mutant component in the time-varying signal.In the implementation of the method,the meaning and difference of the steady component and the mutant component are firstly clarified,and then the heuristic empirical mode decomposition with a masking signal method is proposed for the decomposition of time-varying signals,which can effectively avoid the mode mixing phenomenon and the problem of inconsistent dimensions,combined with the Hilbert transform,a detection feature composed of instantaneous frequency and instantaneous energy can be obtained.The detection feature is used as the feature vector to train support vector machine and combined with the window adaptive update strategy,it can effectively detect the mutant component in the time-varying signal.Experiments on EEG and audio verify that the detection method is effective and feasible,at the same time,it can achieve a high detection precision and a high detection recall.The mutant component representation method can effectively and accurately represent the mutant component in the time-varying signal.In the implementation of the method,the advantage of sparse modeling in the signal with limited components is used and a sparse dictionary initialization method based on double-sparsity model is proposed,which can accelerate the optimization of dictionary learning while enhancing the interpretability of dictionary structure.The steady component and the mutant component in the time-varying signal are separately trained to obtain the corresponding sparse dictionary,then combined with a certain sparse representation coefficient matrix to obtain an initial dictionary for learning and optimizing.When training structured dictionary,the traditional dictionary learning method based on singular value decomposition is adjusted,the penalty term is introduced into the sparse performance index.After the dictionary learning is completed,a sparse dictionary and sparse representation coefficient matrix that can accurately represent the mutant component can be obtained.Experiments on EEG verify that the representation method is effective and feasible,at the same time,it can achieve small signal reconstruction errors.The mutant component identification method can effectively and accurately identify the mutant component in the time-varying signal.In the implementation of the method,the application and implementation of the sparse representation classifier are studied.For the extraction of identification features,an extraction method based on local sparse representation coefficients is proposed.It is necessary to separately train the corresponding sparse dictionary for each sample set of the existing categories and simply merge them,then calculate the sparse representation vector by using the orthogonal matching tracking algorithm.The vector,through the processing of the local retention function,obtains a sparse representation vector that only retains a certain part and reconstructs it with the merged dictionary.The corresponding component with a minmal reconstruction error can determine the classification.In addition,in order to avoid the negative effects of the wrong label information in the training dataset on the classifier,the removal threshold and the insertion threshold are set.The experiments on the audio signal verify that the identification method is effective and feasible.At the same time,this identification method can reduce the interference caused by the wrong label information to the classifier to a certain extent.The thesis presents a set of analysis methods that apply mutant components in time-varying signals,specifically mutant component detection method,mutant component representation method and mutant component identification method.The methods comprehensively utilize the characteristics and differences of steady components and mutant components in time-varying signals and has achieved good results in EEG and audio signals.What's more,on the basis of the proposed mutant component analysis method,the analysis method is systematically integrated to realize an automatic and intelligent system,which lays theoretical and practical foundation for the subsequent processing and method development of the mutant component in the time-varying signal.