Improved support vectors machine for signal detection in non-reverberation

Active sonar system have developed using large arrays and broadband transmitted signals for improving the signal to reverberation power ratio after beam forming and matched filter. While the size of the range bearing resolution cell is reduced resulting in the reverberation becomes non-Gaussian distribution. The performance of matched filter is reduced. We proposed two methods to modify the kernel function of SVM for improving the performance of detection in non-Gaussian reverberation. The first method is to use mixture kernel function. And the experiment results show the performance is as good as SVM with Gaussian kernel and better than RBF and matched filter in non-Gaussian reverberation. The second method is to use the kurtosises of reverberation and targets echo to drive the kernel. The experiment results show the performance is better than matched filter and SVM with Gaussian kernel in non-Gaussian reverberation. When detection probability is 0.5, the input signal to reverberation power ratio is respectively reduced 1.5 dB and 7 dB than conventional SVM and matched filter.