Research On Complex Fast Independent Component Analysis Algorithm

Independent component analysis algorithm (ICA) is an effective blind signal processing technique recenly developed. Along with the development of it, the complex value ICA algorithm was produced. The complex ICA has important theory and application value, and it has a wide application in the speech signal, image processing, antenna array and many other areas. In the past ten years, its theory had been developed quickly and a lot of effective algorithms had been made. Currently, the complex ICA algorithm is a hotspot in international signal processing areas.Fristly, this paper introduces the basic theory of ICA, the typical ICA algorithms and its performance analysis. The basic theories mainly include information theory, the mathematical model of ICA and its solvability analysis etc. Typical ICA algorithms mainly include maximum entropy algorithm, H-J algorithm, minimum mutual information algorithm, stochastic gradient algorithm and the natural gradient algorithm etc.Secondly, this paper puts forward the improved complex value fast independent component analysis algorithm (CFICA). Based on the three order convergence complex ICA algorithm, the five order convergence CFICA algorithm is proposed. The simulation result shows that the separation efficiency of improved algorithm is superior to the traditional manner. Furthmore, CFICA can significantly reduce the times of iterations and the running time, and can also increase the convergence rate and operation efficiency.Finally, in order to overcome the sensitivity of initial value of algorithm, this paper presents the high order convergence CFICA algorithm based on relaxation factors, which contains three order convergence CFICA algorithm based on the relaxation factors and five order convergence CFICA algorithm based on the relaxation factors. Simulation experiments show that the improved algorithms do not depend on the choice of initial value, effective overcome the sensitivity problem of initial value, and improve the algorithm convergence performance.