Abstract: Bias removal techniques usually remove the bias which is caused by noise in the method of correcting noiseless ICA. Nevertheless, the demixing matrix is still identifiable by using the noiseless ICA algorithms in the presence of additive Gaussian noise. So it is preferable to perform denoising with the vector of observed random variables, rather than to make modification to the demixing matrix. Then the QR-decomposition-based Recursive Least Squares（RLS） adaptive noise cancellation（QRRLS） is introduced to combine with Fast-ICA algorithm. To test performance of the proposed approach, two experiments for it and the LMS-ICA algorithm are conducted on the conditions of identical noise and correlation noises respectively. By comparison, it shows that the proposed approach outperforms the latter. Moreover, in order to measure the performance availably, the least-squares method is adopted to calculate the Signal to Noise Ratio（SNR） of recovery signals.

Key words: noisy independent component analysis, bias removal techniques, identifiability of demixing matrix, Recursive Least Squares（RLS） adaptive noise cancellation, least-squares method

摘要： 偏差去除算法通常假设高斯噪声条件下对普通ICA算法进行修正来消除噪声带来的影响。但是存在高斯噪声条件时，普通ICA算法对解混矩阵仍然可以辨识。故引入基于QR分解的RLS自适应噪声抵消算法和Fast-ICA算法相结合，只需对观测信号进行去噪处理，不需要对解混矩阵修正。并分别在同一噪声和相关噪声条件下做了仿真实验，与LMS-ICA算法进行了比较。仿真实验证明，该方法比后者效果显著。提出了用最小二乘算法计算分离信号的输出信噪比，作为评价算法的性能指标。

关键词: 噪声独立分量分析（ICA）, 偏差去除技术, 解混矩阵的可辨识性, 递归最小二乘（RLS）自适应抵消, 最小二乘算法