Implementation of Blind Source Separation Based on Improved Variable Step-Size Natural Gradient Algorithm
Shaofeng Peng, Dongmei Zhou, Peng Wang, Chao Wang · 2022 7th International Conference on Signal and Image Processing (ICSIP) · 2022
The traditional blind source separation algorithm suffers from the problems of large operation, and the convergence speed and the separation effect are constrained by each other. In this paper, the blind source separation algorithm under the natural gradient criterion is studied, and the natural gradient blind source separation algorithm is improved by combining the idea of variable step length. The method determines the separation phase of the signal by the correlation coefficient between the separated signals, and controls the step length factor according to the separation phase, and introduces the momentum term in the algorithm. Experiments prove that the improved variable step size natural gradient algorithm has the best separation performance and the fastest convergence speed compared with the adaptive step size algorithm and the fixed step size algorithm, and the separation recognition rate reaches 78.4%.