Abstract: Independent component analysis is a statistical approach for representing an observed multi-dimensional sensor vector into several components which are as mutual independent as possible. In this paper, a new independent component analysis method is proposed. The presented method exploits the conjugate gradient searching algorithm rather than the nature gradient algorithm to derive the learning equations for training the transforming matrix. The objective function is obtained based on the theory of maximum information. In addition, the score functions included in the learning equation are estimated adaptively by a kernel density estimation method rather than replaced by choosing certain non-linear functions empirically. Several simulation results have shown the effective behavior of the proposed conjugate gradient based independent component analysis method with the application in the blind source separation problem.

Key words: independent component analysis, conjugate gradient, maximum information entropy, objective function

摘要： 独立分量分析是一种将观测向量分解为若干个独立统计的分量的一种统计学方法。提出了一种新的独立分量分析方法，该方法在最大信息理论的基础上引入目标函数，并利用共轭梯度搜索算法替代自然梯度算法，推导出用于训练转换矩阵的学习方程。运用核密度函数估算方法自适应地估算学习方程中包含的评价函数项。仿真结果表明，提出的基于独立分量分析的共轭梯度算法在求解盲源分离问题中切实有效。

关键词: 独立分量分析, 共轭梯度, 最大信息熵, 目标函数