Aiming at the traditional time-frequency domain fault diagnosis methods that cannot realize adaptive identification and classification of faults with low accuracy, a long-short time memory network (LSTM) method based on improved information entropy (IIE) was proposed. Firstly, the original signal was subjected to ensemble empirical mode decomposition (EEMD) and variational mode decomposition (VMD). All the intrinsic mode functions (IMF) containing fault information were subjected to information entropy. The information entropy was used to reflect the information amount and kurtosis index of IMF to describe the advantages of impact components, and the information entropy was improved to form the feature vector. Finally, combining with the advantages of LSTM in dealing with nonlinear data, the combination of features were used to train the LSTM network to establish a diagnostic model. The experimental results show that the method can identify multiple faults accurately and efficiently, and the accuracy rate is higher than the single EEMD-LSTM, VMD-LSTM, artificial neural network and other traditional methods.