Projection pursuit is a dimension reduced data treatment method. Multi dimensional question can be converted into low dimension question according to projection direction. The key of using this method is to search the best projection direction, which can be solved by transforming a nonlinear optimization question, combined with Matlab genetic algorithm toolbox for the optimal solution. In this paper, Xinzhuangzi Mine was taken as an example, establishing projection pursuit model to discriminate water inrush source, compared with fuzzy comprehensive evaluation, artificial neural network and gray clustering model. The result indicated that projection pursuit discrimination model has higher discrimination accuracy than fuzzy comprehensive evaluation, artificial neural network and gray clustering model, which provides a new way to discriminate water inrush source.