Abstract: As the most important equipment in pulverizing system of thermal power plant, the running state of coal mill directly affects the running performance of thermal power unit. Therefore, fault diagnosis of coal mill is of practical significance to ensure the safety of power plant production. Aiming at problems such as uncertain failure type and time lag in fault diagnosis during the actual operation of the coal mill, this paper proposed a fault diagnosis method for the coal mill based on wavelet packet-LSTM neural network. First, a prediction model for the outlet pressure and outlet temperature of the coal mill was established by utilizing the LSTM neural network. The normal data and fault data during the operation of the coal mill were combined into mixed data which be used as the input of the LSTM neural network prediction model to predict the outlet pressure and temperature of the coal mill and obtain the residual signa. Secondly, the wavelet packet decomposition method was used to distinguish and identify the sudden abnormal points of the residual signal. The correlation degree method was used to diagnose the fault types of coal mills by analyzing the trend of changes in all variables of partial data before and after the failure in the coal mill fault library and the mixed data. The results show that the average relative error of the outlet pressure and outlet temperature of the coal mill predicted by the LSTM neural network is not more than 1%. The Wavelet packet decomposition method is used for the residual signal, which can more accurately confirm the time point of the fault. The correlation coefficient method is used to analyze the change trend of all variables, which can identify the type of failure of the coal mill.