Research On Intelligent Signal Classification

Nonstationary signal detection and classification is the very important research subjects in intelligence signal processing. Time-frequency representation(TFR) and wavelet transform(WT) can provide a direct and effective way to process nonstationary signals. Support vector machine (SVM) is a novel type of learning classification algorithm. The paper studies the key technologies and methods in the three domain, and emphatically probes into their fusion and the applications to the nonstationary signal classification. The major contents include the optimal kernel criterion of quadratic time-frequency distributions, the selection of the best wavelet-packet basis, the improvement of support vector algorithm with respect to nonstationary signal classification, jointly optimization of the TFR kernel and support vector kernels based on minimizing the probability of classification error and SAR image classification combining the wavelet multi-feature fusion with the SVM.The following is the summarization of the main works:Any specific TFR from Cohen's class is uniquely characterized by its kernels. Under the certain restricts, the kernel design to make some measure maximum is the research direction in Time-frequency analysis. To the questions of accurate classification of signals, this paper discusses the selection of TFR kernel and distance definition using TFR, develops the optimal criterion for designing kernel by which the similarities of same class and the differences between the classes are emphasized, then gives a jointly optimal method to both distance measurement and TFR kernel under minimizing the probability of classification error. Applied the method to multi-component linear frequency-modulated signals, a typical non-stationary signal, it confirms the method validity and feasibility.For a great number of nonstationary or time-varying signals, the classification features are often localized both in time and frequency domain. Because wavelet-packet transform can optimally divide whole frequency band and provide a arbitrary time-frequency decomposition, it can extract more important features than those extracted by the othertransforms. The paper firstly studies the selection of the best wavelet-base according to the distance and entropy criteria,proposes a way of feature extraction and classification under the optimal wavelet-packet decomposition, and then applies it to the one-dimension range profiles of radar target, which shows it can get higher classification precision. After the discussion of the wavelet characteristics of speckle and the feature extraction scheme of SAR images, this paper presentes a method of de-noisy and classification that combines the texture feature with the grey features of filtered images. The experiment result shows that the method can improve the performance of de-noisy and classification of SAR image.Support vector machine is an effective type of learning classification algorithm but it needs solving a quadratic programming problem. Proximal SVM can speed the convergence rate because it only solves the sample linear equations. We construct a radar target classifier. The classification process consists of two stages. Firstly the kernel principle component analysis is used to select the nonlinear feature of range profiles. Secondly proximal SVM is used as a classifier. The simulation results indicate that the method proposed has comparable recognition correctness but the computational time is shorter than that by using the standard SVMAlthough the time-frequency representations and time-scale representations can work better in signal classification, they have some limitations such as the high dimensionality feature etc. Combining the advantages of TFR and SVM, we can obtain a universal, better-performance classification algorithm. The selection of the kernel is the key in the time-frequency analysis and support vector algorithm, therefore, an efficient hybrid optimization of kernel function for TFR and SVM is proposed, which adaptively adjusts the kernel parameters to minimize the clas...