This paper firstly extends the self-organizing neural networks to general case. Then the self-organizing process is applied to classify the dynamic process of nonlinear control systems. The nonlinear system is self-organized according to the probability density of the input and output samples and is approximated by many linear sub-spaces with different classifying centers and receptive fields. The self-organizing least squares identification of nonlinear systems is constructed based on the general least squares algorithms, the linearization errors of sub-spaces, and the cooperation and competition mechanism. The simulation results have shown the efficiency of the suggested algorithm.