独立分量分析的算法分析与改进

Fast ICA算法是基于一批已取得的样本数据进行处理，它不适用信道矩阵变化的情况；虽基于自然梯度的Info max法是根据单次观测的样本值来调整分离矩阵，但它仅适合单类信源情况。在信道恒定和变化情况下，仿真比较上述算法的优缺点，同时为解决在线算法中收敛速度和稳态误差的矛盾，提出一种改进的变步长算法。该算法将步长变化与信号的分离程度相联系，根据信号之间的相似性测度变化量自适应地控制步长，最后仿真验证该算法的实用性。

Analysis and Improvement of Independent Component Analysis Algorithm

The algorithm of FastlCA for processing the acquired sample data is unsuitable for the case of channel matrix changing. Although the Infomax algorithm based on the natural gradient can adjust the separation matrix according to the sample data of single observation, it is only appropriate for processing single type source signal. In this paper, under the condition of constant and variable information channels, advantages and disadvantages of the above-mentioned algorithms were analyzed respectively by simulation. Meanwhile, an improved algorithm was proposed in order to solve the contradiction between convergence speed and steady-state error of the online algorithm. In this improved algorithm, the relation between the variable step-size and the degree of separation was established, and the step-size could be adjusted adaptively according to the similarity measure of the separation signals. The simulation showed that the algorithm has a good practical performance.