Multi-component Signal Parameter Estimation Based On Manifold Clustering And Classification Methods Research

As Machine Learning develops, it has been widely used in various fields, such as face recognition, identity recognition, personalized recommendation, and search engine page rank, etc. Apart from those fields, Machine Learning can also be applied to processing and recognizing radar emitter signals, the recognition of which includes how to separate multi-component signals and estimate the parameters, how to classify and identify radar emitter signals, how to confirm their types, and how to partition their threat levels. This thesis, to meet the needs of electronic countermeasure, focuses on clustering techniques in Machine Learning and the application of those techniques to separate multi-component signals and estimate the parameters.First, this thesis provides a new algorithm of linear manifold clustering. According to geometric properties of overlapping linear manifold clustering, the algorithm, with the method of looking for the center of ordinary clustering, finds the overlapping area of all kinds of linear manifold distribution data, obtaining the intersection set. Then the feature vector which equivalently represents each sample point in the linear manifold structure is gained by processing the intersection set and each sample point. The next step is to conduct spectral clustering of each feature vector, and clustering results are thus obtained. The algorithm is finally applied to parameter estimation of multi-component signals.Second, as for overlapping nonlinear manifold clustering, the thesis gives an algorithm only for parameter estimation of multi-component signals. That is, the algorithm cannot be applied to ordinary nonlinear manifold clustering. Based on the premise that the modulation mode of each single component signal has been known, the algorithm obtains the intersection set by finding the overlapping area of all kinds of nonlinear manifold distribution data. Then through fitting each sample point and intersection set, parameter estimation of each sample point's modulation mode is thus gained, for the modulation mode of each single component signal has been known. And clustering results are obtained by clustering parameter set as feature vectors. With respect to the instability of SCC algorithm, certain improvement has been made about initial sampling. Specifically, those sample points in the same linear manifold distribution should be selected as initial sampling points in case that the algorithm is trapped in local optimal solution, improving the stability of the algorithm.Third, the thesis studies the processing and recognizing of the ultra-high sound velocity of the airborne radiation source signals. Owing to the high mobility of the target radiation source, received signals tend to be influenced by Doppler effecting. However, the recognition rate of the target radiation source can be improved by eliminating the recognition error caused by Doppler effecting based on the dynamic characteristics of the target radiation source. And for the radar emitter recognition method based on Mahalanobis distance, given a online radar emitter recognition method based on the weight Mahalanobis distance,which can well reflect the geometric distribution of radar emitter sample points, so improving the recognition rate.