In order to improve the accuracy of fault diagnosis and shorten the training time of the neural network, a new method of bearing fault diagnosis is presented that combines the period-energy feature with the LMD optimization feature. First, morphological filtering was used to remove noise from signals. Second, as the standard by one rotation period sampling points, all kinds of fault signals were intercepted. The period-energy feature and LMD feature were extracted from the intercepted signal. Third, the features were processed by u-law compression expansion and moving-average processing. Finally, two Rolling bearing fault (RBFs) were designed with the same precision. The first RBF was trained with the period-energy feature and LMD feature, and the second was trained with period-energy of optimization and LMD feature of optimization. Then, the bearing fault was diagnosed with the well-trained neural network. The experimental results showed that the diagnostic accuracy improved by 10%, and the convergence iteration times of the training neural network were reduced by 50%, thus indicating improved diagnostic accuracy of bearing fault diagnosis and shortened convergence training time of the neural network.