提高雷达HRRP目标识别和拒判性能的核学习算法

New kernel learning method to improve radar HRRP target recognition and rejection performance

雷达高分辨距离像（HRRP）数据具有明显的多模分布特性．在雷达HRRP识别和拒判中，采用单个高斯核很难准确地描述HRRP数据的多模分布．针对该问题，将单个高斯核扩展到多个高斯核线性组合的形式，并将该组合形式应用到支持向量域描述（sVDD）中来处理识别和拒判问题,根据对组合系数自由度的不同限制，扩展后的多核支持向量域描述（Multi-kernel SVDD）方法可以分别表述为不同的凸优化形式：二阶锥规划（SOCP）和半正定规划（sDP），它们都可以收敛到全局最优解．新方法采用了更加复杂的核函数形式，能够更加灵活地描述HRRP数据在高维特征空间的多模分布，从而提高雷达HRRP的识别和拒判性能．仿真实验结果显示该方法的损失值仅为单核SVDD的88．6％～93．2％．

Radar high-resolution range profiles （HRRP） satisfy typical multimodal distribution. In radar HRRP target recognition and rejection, it is difficult to utilize a singe Gaussian kernel to describe the multimodal distribution. According to this, support vector data description （SVDD） was expanded from a single Gaussian kernel to a linear combination of multiple Gaussian kernels and then this combination is used to treat the recognition and rejection problem. Based on different degrees of freedom on the combinational coefficients, the resulting Multi-kernel SVDD could be expressed as different convex optimization problems; SOCP or SDP, and both of them could be solved with global optimal solutions. The proposed method employs more complicated kernel formations, and it can describe the multimodal distribution of HRRP data more flexibly in the high-dimensional feature space so as to improve the recognition and rejection performance, Experimental results show that the loss value of the new method is just 88.6%～93.20/60 that of the single kernel SVDD.