Underwater Target Classification Based On Time-Frequency Feature Extraction And Support Vector Classifier

Underwater active target classification has been a crucial problem of underwater acoustic signal processing, which is important for both martial and civil applications. However, its development seems to be relatively slow. This graduate thesis concentrates on the underwater active target classification problem. Generally speaking, the received echo has a huge dimension of data. So, the key to this problem is to find the classes from the high-dimensional data space. The underwater active target classification is divided into two parts, one of which is the extraction of features from the original data. In detail, this process is to use the signal analysis methods or some other algorithms of data compression to map the high-dimensional original data into a relative low dimension - feature space. The other part of this problem is the classifier design, which maps the feature space into the label space and totally solves the whole underwater active target classification problem.This graduate thesis has researched these two parts to some extent. In the feature extraction part, it focuses on two kinds of signal analysis methods, first of which is based on the time-frequency analysis and strength of echo from target. The time-frequency analysis is a typical method to analyze the non-stationary signal, which performs well in the processing of various kinds of non-stationary signals. This thesis is mainly concentrated on the use of spectrogram to extract the physical information of strength of echo from target. The other one is to apply multiple sorts of orthogonal signal transformation to extract the information of the targets. The signal transformation is to transform the original data in time domain into other domains, which might be able to represent the essential information of targets and helpful for the underwater active target classification problem. This thesis adopts the Fourier transformation, wavelet transformation and the discrete cosine transformation to extract the features together with a certain criterion. In the classifier design part, the newly-developed theory—support vector machine (SVM) is implemented. With the analysis of its theoretical foundation and combining with the characteristics of underwater active target classification problem, the graduate thesis adopts the C-SVM and the relevance vector machine as the final classifiers.The main creative points of this graduate thesis include:(1) Feature extraction algorithm based on the time-frequency analysis and target's physics characteristics—strength of echo from target. With the research on various kinds of the time-frequency distributions, the physical characteristics of target and the properties of transmitted signal, the algorithm based on the spectrogram and strength of echo from target has been proposed. For the spectrogram, the pattern of window and its parameters have been specially chosen. Moreover, to suppress the non-Gaussian reverberation, the robustness of the spectrogram has been studied in this algorithm.(2) Feature extraction algorithm based on the multiple signal transformations and the combination of their transformation coefficients. By the exploration of the physical and mathematical interpretation of multiple signal transformations, the algorithm based on the combination of transformation coefficients has been presented with the discriminant criterion in the non-Gaussian noise, which is also proposed in this graduate thesis.We adopt two sets of data to verify the effectiveness of the proposed methods in this thesis. One is acquired in the laboratory waveguide and the other in a lake. The data set from the laboratory waveguide has proved the proposed method useful to identify the true target from probably the false target. And the data set from a lake validates that the different types of the true targets could be classified by the proposed methods while the reverberation could also be differentiated from the targets no matter whether they are true or false targets. Additionally, when the simulated non-Gaussian noise and the practically collected reverberation is added into the echoes, the classification results remain acceptable. So, it is true that under some circumstance, our method could effectively recognize the true target and false target under the strong reverberation background.