Keywords: 
• 人工蜂群算法 /
• 盲分离 /
• 混沌信号 /
• 相空间重构

Abstract: The inherent features, such as non-periodic, wide band spectrum, and extremely sensitive to initial values etc. make it quite a challenge to blindly separate the mixed chaotic signals. A new blind source separation method based on the artificial bee colony algorithm is proposed in this paper. This method can recover chaotic sources from noisy observations on their linear mixtures without any prior information about the source equations. The proposed method is structured in the phase space of the demixed signals, which is reconstructed from the observations by using delay-embedding method. An objective function in the reconstructed phase space is designed so that the blind source separation problem is transformed into an optimization problem. The optimal demixing matrix is obtained by maximize the objective function with an artificial bee colony optimizer and the chaotic sources are then recovered by multiplying the observed mixtures and the optimal demixing matrix. Before the optimization procedure is made, a pre-whitening should be employed. Additionally, the parameterized representation of orthogonal matrices through principal rotation is adopted to reduce the dimension of the optimization procedure so that the proposed blind source separation algorithm can converge quickly. Different from the traditional independent component analysis approaches which concern mainly the statistical features, the proposed blind source separation method utilizes the dynamics in the observed mixtures by means of phase space reconstruction. Therefore, better performance can be achieved when it is used to deal with chaotic signals. In computer simulation, two cases are taken into consideration: namely, the mixture is noiseless or not contaminated by noise. The correlation coe?cient criterion and the performance index criterion are adopted to evaluate the separation performances. Simulation result shows that in most cases the proposed approach converges within a few tens of iterations and the chaotic sources can be accurately recovered. The impact of noise level and signal length on the separation performance is investigated in detail. The overall performance of the proposed approach is much better than the traditional independent component analysis approaches. Moreover, the capability of separating the mixed chaotic and Gaussian signals revealed in the simulation indicates that the proposed approach has the potential to be applied in a wider range of applications.