The paper proposed a gear fault diagnosis method that is based on variable predictive model based class discriminate (VPMCD) and empirical mode decomposition (EMD) and applied it into the analysis of the steady state signals of gears. The method of VPMCD is a new way to pattern recognitions, which is especially appropriate for the classification of nonlinear problems. It made full use of the inner relations among characteristic values extracted from those original data to recognize models. In the gear fault diagnosis methods of VPMCD and EMD; firstly, gear vibration signals would be adaptively decomposed into several single-component signals by EMD; secondly, the sample entropy of each component would be extracted as characteristic values. Finally, the VPMCD classifier would be used to recognize and classify the faults. The result shows that this method can effectively highlight the fault features of gear fault vibration signals, and advance the accuracy of gear fault diagnosis.