Applying Momentum Technology Of Blind Source Separation Algorithm Research

A priori knowledge of the source signal and hybrid systems are unknown or poorlyunderstood, and only using the sensor to receive the signal observed recoverytechnology called blind separation of source signals. Now blind source separation iswidely used in biomedical signal processing, seismic signal processing, wireless signalprocessing, speech signal processing, image enhancement, etc. Which attracted thefocus of scholars in the field, including signal processing, neural networks.This article focuses on mixed-signal adaptive (online) elaborated on the basis oftheoretical knowledge of blind source separation, simulate and analysis the performanceof natural gradient algorithm, symbol natural gradient algorithm, variable step sizesymbol natural gradient algorithm etc, resolve the corresponding problems for existingalgorithms to effectively improve the overall performance of the algorithm, proposingimprovement programs and strategies. Article is organized as follows:First of all, discourse the basic theory of blind source separation, including themathematical model of blind source separation, uncertainty separable theory, commoninformation theory of knowledge (information entropy,mutual information, etc.),optimization criteria for the cost function of algorithm as well as the evaluation criteriaevaluating the performance merit of algorithms;Secondly, given the basic algorithm for blind source separation, including thenatural gradient algorithm, the symbol natural gradient algorithm, variable step sizenaturalgradient algorithm, and simulate their separation performance, combined withsimulation results and on the basis of the algorithm to understand in the deepening of avariety of blind sources separation, and seek to improve the performance of thealgorithm;Then, drawing on the momentum thinking of the neural network, propose theimproved algorithm, including the integration of the double momentum variable stepsize symbol gradient algorithm, as well as variable momentum factor natural gradient algorithm: adding step momentum into symbol natural gradient algorithm, so that stepreach its optimal value through fewer iterations, to accelerate the convergence speed;separation matrix momentum accelerate convergence of the performance of algorithmfurther; proposed variable momentum factor natural algorithmis mainly based on thestochastic gradient method, the momentum factor can adaptively adjust its value withthe external environment changes to speed up the convergence rate for smaller steadystate error. Simulation results confirmed that the algorithm proposed in this papereffectively accelerate the convergence speed under the premise of maintainingsteady-state performance;Finally, a summary and a prospect of blind source separation technology werediscussed.