多时延解相关准则下的非平稳信号在线盲分离

利用一种基于多时延解相关准则的代价函数和自然梯度原则,分别推导出非平稳信号在瞬态线性混合及卷积混合情况下的在线盲分离算法,仅通过对有限个时延样本的相关矩阵进行Frobenius范数最小化运算得到分离矩阵,计算工作量小,无需对样本进行分块处理,可实现对信号的连续跟踪操作。仿真结果表明算法分离精度高,分离过程平稳,在任意混合情况下均能获得良好的分离性能。

On-line blind source separation of non-stationary signals by multiple time-delayed decorrelation rule

An on-line blind source separation algorithm of non-stationary mixed signals is presented. The algorithm uses the multiple time-delayed decorrelation and natural gradient rule to obtain the stationary point of the cost function, only a small number of observed samples is needed to minimize the Frobenius norm of the correlation matrices, so the computing cost is low and continuous separation calculation could be realized by on-line operation. Simulations show that the algorithm can get high separation performance for both non-stationary temporary and convolved mixing signals.