Summary: Key variable identification is related to many trouble-shooting problems in process industries. Linear kernel support vector machine based recursive feature elimination is proved to be an effective algorithm to do key variable identification. But when variables are nonlinear relative, it is difficult to work. A nonlinear kernel support vector machine based recursive feature selection is proposed to select key variables with nonlinear relation or linear relation. The experimental results on dataset from Tennessee Eastman process simulator proved the effectiveness of our method.