For the problem that the decomposition parameters (K，α) of VMD could not be mastered in advance during actual signal processing and therefore its use was prohibited, as well as for the selection problem of the characteristic parameters containing fault information, an adaptive variational modal decomposition (AVMD) algorithm was proposed. In this algorithm, the average envelope entropy and envelope kurtosis of decomposed modes were fused as the objective function, and the decomposition parameters (K_op，α_op) of VMD were obtained by using the search and optimization of grasshopper algorithm (GOA). Then the original vibration signal was decomposed by VMD, the sensitive modes with 90% or more energy were selected through the calculation of energy percentage, and the feature vectors of their multidomain joint feature parameters were constructed. Finally, the support vector machine (SVM) was used to identify the four states of the rolling bearing. The rolling bearing data set analysis showed that the 99.166 7% accuracy of fault pattern recognition for fault diagnosis features extracted by AVMD method is higher than that extracted by EMD, EEMD, traditional VMD and PSO-VMD methods.