快速RLS算法

版权声明：本文为作者原创文章，如转载请注明出处。

前言

RLS算法需要的运算量比较大，在实际应用中会受到一些限制，因此出现了许多快速算法。比较常见的两种算法为阶递归自适应滤波器(Lattice-Based RLS)和快速横向滤波器(Fast Transversal RLS)。所有的快速算法都探索了信息数据的一些结构特性，以实现较低的计算复杂度。快速算法的模式是相似的，因为前向和后向预测滤波器是快速算法的基本组成部分。

1 阶递归自适应滤波器

对于一个预测过程，前向后验预测误差为
$${\epsilon }_f(k,i+1)=x(k)-{\mathbf{w}}_f^T(k,i+1)\mathbf{x}(k-1,i+1)$$式中$${\mathbf{w}}_f^T(k,i+1)=[w_{f0}(k),w_{f1}(k),w_{f2}(k),...w_{fi}(k){]}^T$$$$\mathbf{x}(k-1,i+1)=[x(k-1),x(k-2),...x(k-i-1){]}^T$$
定义加权误差向量
$${\mathbf{\epsilon }}_f(k,i+1)=\hat{\mathbf{x}}(k)-{\mathbf{X}}^T(k-1,i+1){\mathbf{w}}_f(k,i+1)$$
式中$$\hat{\mathbf{x}}(k)=[x(k),{\lambda }^{1/2}x(k-1),,\lambda x(k-2),...,{\lambda }^{k/2}x(0)]$$
写为矩阵形式有
$${\mathbf{\epsilon }}_f(k,i+1)={\mathbf{X}}^T(k,i+2)\left[\begin{array}{c}1\\ -{\mathbf{w}}_f(k,i+1)\end{array}\right]$$
均方误差为
$${\xi }_f^d(k,i+1)={\mathbf{\epsilon }}_f^T(k,i+1){\mathbf{\epsilon }}_f(k,i+1)$$
$$=\sum _{l=0}^k{\lambda }^{k-l}[x(l)-{\mathbf{x}}^T(l-1,i+1){\mathbf{w}}_f(k,i+1){]}^2$$
令导数为0可以得到
$${\mathbf{w}}_f(k,i+1)={\left[\sum _{l=0}^k{\lambda }^{k-l}\mathbf{x}(l-1,i+1){\mathbf{x}}^T(l-1,i+1)\right]}^{-1}\sum _{l=0}^k{\lambda }^{k-l}\mathbf{x}(l-1,i+1)x(l)$$
$$=[\mathbf{X}(k-1,i+1){\mathbf{X}}^T(k-1,i+1){]}^{-1}\mathbf{X}(k-1,i+1)\hat{\mathbf{x}}(k)$$
$$={\mathbf{R}}_{Df}^{-1}(k-1,i+1){\mathbf{p}}_{Df}(k,i+1)$$
将最佳权系数代入可以得到最小均方误差
$${\xi }_{fmin}^d(k,i+1)=\sum _{l=0}^k{\lambda }^{k-l}x(l)[x(l)-{\mathbf{x}}^T(l-1,i+1){\mathbf{w}}_f(k,i+1)]$$
$$=\sum _{l=0}^k{\lambda }^{k-l}{x}^2(l)-{\mathbf{p}}_{Df}^T(k,i+1){\mathbf{w}}_f(k,i+1)$$
$$={\sigma }_f^2(k)-{\mathbf{w}}_f(k,i+1{)}^T{\mathbf{p}}_{Df}(k,i+1)$$
将上面两式组合可以得到
$${\mathbf{R}}_D(k,i+2)=\left[\begin{array}{cc}{\sigma }_f^2(k)& {\mathbf{p}}_{Df}^T(k,i+1）\\ {\mathbf{p}}_{Df}^T(k,i+1）& {\mathbf{R}}_{Df}(k-1,i+1)\end{array}\right]$$

$${\mathbf{R}}_D(k,i+2)\left[\begin{array}{c}1\\ -{\mathbf{w}}_f(k,i+1）\end{array}\right]=\left[\begin{array}{c}{\xi }_{fmin}^d(k,i+1)\\ 0\end{array}\right]$$
同理可得，后向预测模型为
$${\mathbf{R}}_D(k,i+2)\left[\begin{array}{c}-{\mathbf{w}}_b(k,i+1）\\ 1\end{array}\right]=\left[\begin{array}{c}0\\ {\xi }_{b_{min}}^d(k,i+1)\end{array}\right]$$
引入新变量$\delta$
$${\mathbf{R}}_D(k,i+2)\left[\begin{array}{c}1\\ -{\mathbf{w}}_f(k,i）\\ 0\end{array}\right]=\left[\begin{array}{c}{\xi }_{fmin}^d(k,i)\\ 0\\ {\delta }_f(k,i)\end{array}\right]$$

式中$${\delta }_f(k,i)=\sum _{l=0}^k{\lambda }^{k-l}{\epsilon }_f(l,i)x(l-i-1)$$
同样的有
$${\mathbf{R}}_D(k,i+2)\left[\begin{array}{c}0\\ -{\mathbf{w}}_b(k-1,i）\\ 1\end{array}\right]=\left[\begin{array}{c}{\delta }_b(k,i)\\ 0\\ {\xi }_{bmin}^d(k,i)\end{array}\right]$$
式中$${\delta }_b(k,i)=\sum _{l=0}^k{\lambda }^{k-l}{\epsilon }_b(l-1,i)x(l)$$
后验误差迭代为
$${\xi }_{b_{min}}^d(k,i+1)={\xi }_{b_{min}}^d(k-1,i)-\frac{{\delta }^2(k,i)}{{\xi }_{f_{min}}^d(k,i)}$$
后验估计权系数迭代为
$${\mathbf{w}}_b(k,i+1)=\left[\begin{array}{c}0\\ {\mathbf{w}}_b(k-1,i)\end{array}\right]-\frac{\delta (k,i)}{{\xi }_{f_{min}}^d(k,i)}\left[\begin{array}{c}-1\\ {\mathbf{w}}_f(k,i)\end{array}\right]$$
同样的，前验误差迭代为
$${\xi }_{f_{min}}^d(k,i+1)={\xi }_{f_{min}}^d(k,i)-\frac{{\delta }^2(k,i)}{{\xi }_{b_{min}}^d(k-1,i)}$$
前验估计权系数迭代为
$${\mathbf{w}}_f(k,i+1)=\left[\begin{array}{c}{\mathbf{w}}_f(k,i)\\ 0\end{array}\right]-\frac{\delta (k,i)}{{\xi }_{b_{min}}^d(k-1,i)}\left[\begin{array}{c}{\mathbf{w}}_b(k-1,i)\\ -1\end{array}\right]$$
后验前向误差为
$${\epsilon }_f(k,i+1)={\epsilon }_f(k,i)-{\kappa }_f(k,i){\epsilon }_b(k-1,i)$$
式中${\kappa }_f(k,i)$称为前向反射系数
$${\kappa }_f(k,i)=\frac{\delta (k,i)}{{\xi }_{b_{min}}^d(k-1,i)}$$
后验后向预测误差为
$${\epsilon }_b(k,i+1)={\epsilon }_b(k-1,i)-{\kappa }_b(k,i){\epsilon }_f(k,i)$$
式中${\kappa }_b(k,i)$称为后向反射系数
$${\kappa }_b(k,i)=\frac{\delta (k,i)}{{\xi }_{f_{min}}^d(k,i)}$$
在初始阶段
$${\epsilon }_b(k,0)={\epsilon }_f(k,0)=x(k)$$
$${\xi }_{f_{min}}^d(k,0)={\xi }_{b_{min}}^d(k,0)=x^2(k)+\lambda {\xi }_{f_{min}}^d(k-1,0)$$
$\delta$的迭代为
$$\delta (k,i)=\lambda \delta (k-1,i)+\frac{{\epsilon }_b(k-1,i){\epsilon }_f(k,i)}{\gamma (k-1,i)}$$
式中
$$\gamma (k,i+1)=\gamma (k,i)-\frac{{\epsilon }_b^2(k,i)}{{\xi }_{b_{min}}^d(k,i)}$$
加权均方误差为
$${\xi }^d(k,i+1)=\sum _{l=0}^k{\lambda }^{k-l}{\epsilon }^2(l,i+1)$$
对应的
$${\delta }_D(k,i)=\lambda {\delta }_D(k-1,i)+\frac{\epsilon (k,i){\epsilon }_b(k,i+1)}{\gamma (k,i)}$$
前向乘子为
$$v_i(k)=\frac{{\delta }_D(k,i)}{{\xi }_{b_{min}}^d(k,i)}$$
后验输出误差为
$$\epsilon (k,i+1)=\epsilon (k,i)-v_i(k){\epsilon }_b(k,i)$$
由此获得阶递归RLS算法的全过程。

2 快速横向滤波器

后验前向预测误差为
$${\epsilon }_f(k,N)=x(k)-{\mathbf{w}}_f^T(k,N)\mathbf{x}(k-1,N)$$
$$={\mathbf{x}}^T(k,N+1)\left[\begin{array}{c}1\\ -{\mathbf{w}}_f(k,N)\end{array}\right]$$
后验与先验前向预测误差之间的关系为
$$e_f(k,N)=\frac{{\epsilon }_f(k,N)}{\gamma (k-1,N)}$$
式中
$$\gamma (k,N+1)=\frac{\lambda {\xi }_{fmin}^d(k-1,N)}{{\xi }_{fmin}^d(k,N)}\gamma (k-1,N)$$
前向最小加权均方误差为
$${\xi }_{f_{min}}^d(k,N)=\lambda {\xi }_{f_{min}}^d(k-1,N)+e_f(k,N){\epsilon }_f(k,N)$$
前向预测权系数迭代
$${\mathbf{w}}_f(k,N)={\mathbf{w}}_f(k-1,N)+\varphi (k-1,N)e_f(k,N)$$
式中
$$\varphi (k,N+1)=\left[\begin{array}{c}0\\ \varphi (k-1,N)\end{array}\right]+\frac{1}{{\xi }_{f_{min}}^d(k,N)}\left[\begin{array}{c}1\\ {\mathbf{w}}_f(k,N)\end{array}\right]{\epsilon }_f(k,N)$$
利用$\hat{\varphi }(k,N+1)$代替$\varphi (k,N+1)$有
$$\hat{\varphi }(k,N+1)=\left[\begin{array}{c}0\\ \hat{\varphi }(k-1,N)\end{array}\right]+\frac{1}{\lambda {\xi }_{f_{min}}^d(k-1,N)}\left[\begin{array}{c}1\\ -{\mathbf{w}}_f(k-1,N)\end{array}\right]e_f(k,N)$$
则前向预测权系数迭代可以改为
$${\mathbf{w}}_f(k,N)={\mathbf{w}}_f(k-1,N)+\hat{\varphi }(k-1,N){\epsilon }_f(k,N)$$
后验与先验预测误差之间的关系为
$${\epsilon }_b(k,N)=e_b(k,N)\gamma (k,N)$$
转换因子
$$\frac{\gamma (k,N+1)}{\gamma (k,N)}=\frac{\lambda {\xi }_{b_{min}}^d(k-1,N)}{{\xi }_{b_{min}}^d(k,N)}$$
后向先验预测误差为
$$e_b(k,N)=\lambda {\xi }_{b_{min}}^d(k-1,N){\varphi }_{N+1}(k,N+1)$$
FTF中转换因子的迭代可以表示为
$${\gamma }^{-1}(k,N)={\gamma }^{-1}(k,N+1)-{\hat{\varphi }}_{N+1}(k,N+1)e_b(k,N)$$
后向最小加权均方误差为
$${\xi }_{b_{min}}^d(k,N)=\lambda {\xi }_{b_{min}}^d(k-1,N)+e_b(k,N){\epsilon }_b(k,N)$$
则后向预测权系数迭代为
$${\mathbf{w}}_b(k,N)={\mathbf{w}}_b(k-1,N)+\hat{\varphi }(k,N){\epsilon }_b(k,N)$$
式中
$$\left[\begin{array}{c}\hat{\varphi }(k,N)\\ 0\end{array}\right]=\hat{\varphi }(k,N+1)-{\varphi }_{N+1}(k,N+1)\left[\begin{array}{c}-{\mathbf{w}}_b(k-1,N)\\ 1\end{array}\right]$$
滤波先验误差为
$$e(k,N)=d(k)-{\mathbf{w}}^T(k-1,N)\mathbf{x}(k,N)$$
则后验误差为
$$\epsilon (k,N)=e(k,N)\gamma (k,N)$$
滤波权系数迭代为
$$\mathbf{w}(k,N)=\mathbf{w}(k-1,N)+\hat{\varphi }(k,N)\epsilon (k,N)$$
由此获得快速横向RLS算法的全过程。