Intelligent fault diagnosis of electric motor bearings can be achieved by analyzing their vibration signals. Deep learning methods have been widely used in fault diagnosis because of their ability to fit the behavior of complex systems better. However, the limitations of the single-channel time-series vibration signal and the randomness of the selection of hyperparameters for the deep learning model significantly impact the fault diagnosis results. This paper proposes a novel approach for fault diagnosis of motor bearings based on multiscale multivariate dispersive entropy (MMDE) and Bayesian-optimized long short-term memory (BOLSTM) networks. At first, MMDE is used to extract features from vibration signals of motor bearings, which are then fed into the LSTM networks for classification. Secondly, to optimize the performance of the LSTM networks, a Bayesian optimization algorithm is applied to find the optimal hyperparameters. Finally, the proposed method uses a bearing fault diagnosis dataset to validate the performance.