An efficient and robust varying-parameter projection neural network for sparse signal reconstruction

Abstract

Sparse signal reconstruction plays an important role in many practical applications, and therefore, many scholars strive to study its various effective (approximate) solution approaches, including some projection neural network models with finite-time or fixed-time convergence due to their parallel nature and convenience for hardware implementation. However, we experimentally find that noise has a significant influence on the convergence performance of these models. To overcome this problem, in this paper, we propose a new efficient varying-parameter projection neural network model (VPPNN) using the sliding mode control technique and a novel time-varying parameter for sparse signal reconstruction. Theoretical analysis shows that this new model not only has shorter fixed convergence time under certain conditions but also has noise-tolerance when compared to the existing projection neural network models. Numerical simulation experiments (via sparse signal/image reconstruction) are performed to illustrate the superiority of the new proposed model over the existing ones in terms of the convergence rate and robustness.