Abstract:  Independent component analysis (ICA) is a typical and important blind source separation(BSS) method. It is to recover source signals which are considered statistically independent given only the outputs of a number of sensor mixed signals. How to choose a model of the score functions of the unknown source signals is a basic and important problem in most ICA algorithm.This is usually achieved based on the parametric and nonparametric density estimation approaches. Classical ICA algorithms fit parametric models for the hidden sources. Moreover, most lCA algorithms always assume that the distribution of source signals has a priori knowlcdge，but this difficult to obtain in many situations, It has been realized that the unknown distributions of hidden sources can be estimated by nonparametric methods，some authors have discussed the nonparametric lCA method，but the kernel they used was fixed which made the algorithm does not robust in some conditions. How to develop unified nonparametric lCA algorithms, which enable to separate hybrid source signals including symmetric and asymmetric distributions with self-adaptive score functions, is very important, In this paper, efficient lCA algorithm based on generalized Gaussian density (GGD) kernel is proposed using a linear ICA neural network which can separate hybrid mixture of source signals which include both symmetric and asymmetric distribution sources.The proposed lCA algorithm is able to separate the hybrid mixtures of source signals using only a flexible model. Moreover, the GGD kernel can adaptively adjust its shaping and scaling parameter according to the high order statistics of the source signals. It paves the way to wider applications of ICA methods to real world signal processing. Simulations confirm the effectiveness of the proposed algorithm.