To solve these two problems that it is difficult to extract the characteristics of weak periodic impact caused by the strong background noise of the signal generated during the rolling bearing fault, and the general diagnosis model does not have a strong recognition effect on the timing characteristics of the fault vibration signal during the intelligent diagnosis of the bearing fault mode, this paper put forward a fault diagnosis method based on the maximum correlation kurtosis deconvolution(MCKD)algorithm, Teager energy operator and long short-term memory(LSTM). Firstly, the rolling bearing vibration signal is denoised by MCKD algorithm, the periodic impact characteristics of the signal which are covered by noise are extracted, the Teager energy operator is used to detect the transient impact of the signal and the Teager energy sequence is obtained. The results are then divided into training sets and test sets, the training set is input into the established LSTM fault diagnosis model for learning. Finally, the LSTM model with appropriate parameters is applied to the test set to output fault diagnosis results. The experimental results show that the proposed method can diagnose faults of various types and sizes at one time and has high identification accuracy. It is a fault diagnosis method that can effectively utilize the timing characteristics of strong background noise signals.