There have been confront with a low identification accuracy problem due to the poor repeatability and high data residual value of laser-induced breakdown spectrum. In order to solve such problems, an distinguishing method of abnormal value based on Grubbs criterion (3δ-Grubbs) was proposed. The method can effectively replace the data of large residual values to reduce the probability of over-fitting in the classification recognition algorithm. Finally, by using three classification recognition algorithms: linear discriminant analysis, random forest classification and support vector machine, we identified the LIBS spectrum of rocks. Before the data noise reduces, the recognition accuracy of the three methods were: linear discriminant analysis 79.6%, random forest classification 75.2%, support vector machine 94.5%.After data noise is reduced,the recognition accuracy of the three methods is as follows: linear discriminant analysis 92%, random forest classification 97%, support vector machine 99.4%.