Fault diagnosis of rolling bearings is essential to ensure the proper functioning of the entire machinery and equipment. Variational mode decomposition (VMD) and neural networks have gained widespread attention in the field of bearing fault diagnosis due to their powerful feature extraction and feature learning capacity. However, past methods usually utilize experiential knowledge to determine the key parameters in the VMD and neural networks, such as the penalty factor, the smooth factor, and so on, so that generates a poor diagnostic result. To address this problem, an Adaptive Variational Mode Decomposition (AVMD) is proposed to obtain better features to construct the fault feature matrix and Sparrow probabilistic neural network (SPNN) is constructed for rolling bearing fault diagnosis. Firstly, the unknown parameters of VMD are estimated by using the genetic algorithm (GA), then the suitable features such as kurtosis and singular value entropy are extracted by automatically adjusting the parameters of VMD. Furthermore, a probabilistic neural network (PNN) is used for bearing fault diagnosis. Meanwhile, embedding the sparrow search algorithm (SSA) into PNN to obtain the optimal smoothing factor. Finally, the proposed method is tested and evaluated on a public bearing dataset and bearing tests. The results demonstrate that the proposed method can extract suitable features and achieve high diagnostic accuracy.