基于改进的ISODATA的超球覆盖仿生模式分类算法  

Bionic pattern classification algorithm for hypersphere coverage based on improved ISODATA

作　　者：刘莉萍 冯清贤 余志斌[1] Liu Liping;Feng Qingxian;Yu Zhibin(School of Electrical Engineering,Southwest Jiaotong University,Chengdu 611756,China;The 29th Research Institute of China Electro-nics Technology Corporation,Chengdu 610036,China)

机构地区：[1]西南交通大学电气工程学院,成都611756 [2]中国电子科技集团公司第二十九研究所,成都610036

出　　处：《计算机应用研究》2023年第3期689-695,共7页Application Research of Computers

基　　金：装备发展部领域基金资助项目;电磁应用重点实验室基金资助项目。

摘　　要：现有仿生模式识别分类器难以解决含有多个聚集点、非线性和稀疏性样本的分类问题。因此,引入特征分类贡献度,提出了基于改进的迭代自组织数据分析(M-ISODATA)的超球覆盖仿生模式识别算法。首先引入马氏距离对自组织数据分析方法(ISODATA)的欧氏距离替换,并引入熵权法对马氏距离进行加权以赋予各特征不同的贡献度;同时为了去除干扰样本点,引入改进的局部离群因子检测方法(M-LOF)对样本进行训练,减少了不同类别流形之间的重叠区域。再利用改进的自组织数据分析方法(M-ISODATA)对每类训练样本点动态聚类,寻找到同一类的多个小类覆盖区中心后,用超球进行该类的有效覆盖,并对落入重叠区域的测试样本点进行二次划分,实现测试样本的正确分类。最后在iris数据集上验证该算法的有效性,并将该算法应用于雷达辐射源信号的分类识别。实验结果表明,该算法具有很好的拒识、免重训能力,对于雷达信号的识别率能达到97.29%,相比于传统典型模式识别算法具有更好的识别能力。The existing biomimetic pattern recognition classifiers cannot solve the problems of classification with multiple aggregation points, nonlinear and sparse samples. Therefore, this paper introduced the contribution degree of feature classification, and proposed a hypersphere coverage bionic pattern recognition algorithm based on improved iterative self-organizing data analysis(M-ISODATA). Firstly, this paper introduced Markov distance to replace Euclidean distance of the self-organized data analysis method(ISODATA), and introduced entropy weight method to give each feature different contribution degrees. At the same time, in order to remove the interference sample points, this paper introduced an improved local outlier detection method(M-LOF) to train the samples, which reduced the overlapping area between different classes of manifolds. Then, it used the improved self-organizing data analysis method(M-ISODATA) to dynamically cluster each type of training sample points. After finding the center of the coverage area of multiple subcategories of the same class, this paper used the hypersphere to effectively cover the class, and divided the test sample points falling into the overlapping area twice to achieve the correct classification of test samples. Finally, it verified the validity of the algorithm on iris dataset, and applied the algorithm to the classification and recognition of radar emitter signals. The experimental results show that the algorithm has a good ability to reject recognition and avoid repeated training, and the recognition rate for radar signals can reach 97.29%, which is better than the traditional typical pattern recognition algorithm.

关 键 词：超球覆盖 加权马氏距离 局部离群因子 自组织数据分析 免重训 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]