Feature kernel support vector machines for signal detection with reverberation

Because kernel functions in a support vector machine influence detection performance,analysis of the effects of kernel features is needed to effectively design feature kernel functions.A new kernel was developed by the product of original kernel functions and the geometric mean value of the high order statistics of two arbitrary samples.The new kernel function is adaptively adjusted by using the non-Gaussian differences between reverberation and target echoes for improving classification performance.It was also proven that it is possible to enlarge the differences between two kinds of samples when using the feature kernel.The suggested kernel also satisfies the Mercer theorem.The proposed feature kernel support vector machine was used for signal detection of Gaussian and non-Gaussian reverberation.The training and detecting algorithms used in testing are given.The results of experiments and simulations showed that when the data tested has significant differences in features,and the reverberation has a non-Gaussian distribution,the new algorithm's performance is better than matching filters and support vector machines based on traditional kernel functions.