Recently, we have proposed an algorithm for construction of a hierarchy of neural network classifiers based on a modification of error backpropagation. It combines supervised learning with self-organization. Recursive use of the algorithm results in creation of compact and computationally effective self-organized structures of neural classifiers. The algorithm is applicable for unsupervised analysis of both static objects and dynamic objects, described by time series. In the latter case, the algorithm performs segmentation of the analyzed time-series into parts characterized by different types of dynamics. The algorithm has been successfully tested on pseudo-chaotic maps. In this paper the above algorithm is applied to Solar wind data analysis. Preliminary results indicate that new structural classes in the Solar wind could be distinguished aside from the traditional two- and three-state concepts.