Abstract: Taking fault diagnosis for bearing as application background, a method of feature selection based on projection was proposed on basis that low-dimensional projection can reflect some features of high-dimensional data. The method uses genetic algorithm to search the projection direction which can reflect classificatory characteristics of samples, and removes those features which unrelated to the projection by the projection direction. The method overcomes some shortcomings of complex calculations of conventional methods, and avoids “the curse of dimensionality” effectively. The simulation result shows that the method can improve efficiency and accuracy of classification by decreasing the dimension of the data without reducing the classificatory characteristics.