Kohonen网络在烟叶动态分类中的应用

Application of the Kohonen Neural Network to Tobacco Classification

针对输入为高维化学指标数据的烟叶分类问题 ,提出 1种改进的 Kohonen自组织特征映射神经网络的聚类方法。在数据预处理时 ,加入了领域专家经验 ,对输入特征向量中的各个分量分配不同的分类参与度 ;用 Gauss邻域函数替代了标准 Kohonen网络的方形邻域 ;在 2个学习阶段学习率和邻域宽度采用了不同的递减函数。通过应用证明了改进后的 Kohonen网络的收敛效果和聚类精度比 K- means聚类方法和标准的 Kohonen网络都有较大的提高。

In this paper, an improved Kohonen Self-Organizing Feature Map neural network is presented, in which domain expert experience is added in the data processing course, the components of input vectors are given different classification-participating level values, the Gauss neighborhood function replaces the square function and different descending functions of learning rate and neighborhood width are used in two learning periods. Compared with the traditional K-means and standard Kohonen network clustering algorithm, the convergence speed and the clustering precision of the improved Kohonen network are both increased.