In order to solve the problems of current belt conveyor roller fault diagnosis methods, such as contact measurement, low accuracy and difficulty in large-scale underground detection, a fault diagnosis method for roller was proposed based on MFCC feature and parameter optimization SVM. The collected roller sound signals were decomposed into several intrinsic mode function (IMF) by using variational mode decomposition (VMD), and the IMF component was optimized based on the composite index of envelope entropy and kurtosis. Mel frequency cepstrum coefficients (MFCC) of selected components were extracted as features, and grey wolf optimizer (GWO) was used to optimize SVM parameters. The sample feature vectors were input into GWO-SVM for fault classification. The results show that the average accuracy of the method is more than 95% for the four working conditions of normal roller, roller with inner ring fault, roller with outer ring fault and roller stuck. Compared with single index, the fault characteristics of IMF components extracted by composite index are more representative. Compared with other optimization algorithms, this method has higher recognition accuracy and faster classification speed.