Sparse LMS method for combination of zero attraction penalty and attraction compensation

Abstract

The invention relates to a sparse LMS method combining zero attraction penalty and attraction compensation, belonging to the field of signal processing. The method combines zero attraction penalty and attraction compensation, divides the coefficients of the estimated filter into near zero coefficients, small coefficients and large coefficients, and then adopts different attraction methods. In each iterative update, for the near-zero coefficients of the estimated filter, only the product term in the iterative update formula is used to calculate; for the large coefficients of the estimated filter, a slight attraction compensation is performed to speed up the estimated filtering The coefficient of the filter is used to approximate the convergence speed of the large coefficient of the channel; for the small coefficient of the estimated filter, if the coefficient approximates the zero coefficient value of the channel or the large coefficient value of the channel during the iteration process, then the estimation filter is approximated as described above. method with zero and large coefficients, otherwise, a simple zero-attraction penalty is applied to the coefficient. The method has fast convergence speed, low complexity and wide range of tuning parameters.

Description

technical field [0001] The invention belongs to the field of signal processing and relates to a sparse LMS method combining zero attraction penalty and attraction compensation. Background technique [0002] Many channels are sparse, and identifying such sparse channels requires specific adaptive filtering algorithms. At present, the types of algorithms for sparse system identification are l 0 -norm, l 1 -norm and l p -norm, where l 0 -norm is to perform zero-attraction penalties on the coefficients of the estimated filter within a certain small threshold, l 1 -norm is a zero-attraction penalty for all coefficients of the estimated filter, l p is a zero-attraction penalty involving division and exponents for all coefficients of the estimated filter. l p -norm achieved better than l 0 -norm and l 1 The -norm type algorithm has better performance, but due to the high complexity, this method is difficult to implement in hardware. l 1 The typical algorithm in the -norm...

AI Technical Summary

Problems solved by technology

l p -norm achieved better than l 0 -norm and l 1 The -norm type algorithm has better performance, but due to high complexity, this method is difficult to implement in hardware

l 1 The typical algorithm in the -norm type is Zero-attracting Least MeanSquare (ZA-LMS), which gives the same zero-attracting penalty to all channel coefficients, and does not distinguish between zero and non-zero channel coefficients, resulting in Its steady-state mean square deviation (Mean Square deviation, MSD) is not excellent

Y.Chen also proposed a reweighted ZA-LMS (Reweight Zero-attracting Least Mean Square, RZA-LMS). The zero-attracting function of this method cleverly reduces and enlarges the large and small coefficients, making It is more reasonable for channel estimation, but this method requires division

Y.Gu proposed a l 0 -LMS method, the coefficient of the estimated filter is lower than a certain threshold before the zero-attraction penalty is performed, but this method has great limitations on the optimal parameter selection and the accuracy of the estimated coefficient

In order to obtain a lower mean square steady-state difference and reduce the parameter limit, LeiLuo proposed a lower parameter limit and a lower MSD l 0 -ILMS method, which also does not process the larger coefficients of the estimated filter

Images