The fault classification method based on long short-term memory recurrent neural network (LSTM-RNN) model is improved based on multiple labels in light of higher accuracy and less training samples when classify rolling bearing faults based on the traditional algorithm. First, a simulation model of rolling bearing fault signal is established, and the spectrum and classification of the rolling bearing fault simulation signal are analyzed. Second, the spectrum feature vectors of the rolling bearing are coded based on the structural characteristics of the multi-label LSTM-RNN model. The effectiveness of a multi-label classification method based on LSTM-RNN is verified by the simulated fault signal. Finally, a test platform is established to collect rolling bearing faults at three different speeds. The multi-label LSTM-RNN classification method and single-label method are compared based on nine groups of data extracted by three methods. The experimental results show that the average classification accuracy of multi-label LSTM-RNN classification increases to 99.21% from 69.07% of the single-label method. The sample size reduces by 69.55% compared with that of the single-label method when the correctness rates of the two classification methods are similar. The multi-label classification method based on LSTM-RNN is suitable for complex vibration signals, which is practically useful in realizing fast and accurate fault diagnosis of rotating machinery.