In view of the harsh operating environment of the coal mills of thermal power unit and the frequent occurrence of coal mills defects, this paper evaluated the operating status of the coal mills and command whether it was in a fault condition or not. The evaluation and judgement was based on the SIS system of thermal power plant to collect the pulverized coal temperature, primary wind speed, coal mill outlet temperature and other variables. In this paper, the relevant parameters of the coal mill are analyzed and the main variables affecting the vibration of the coal mill are determined. Then based on BP neural network, this paper carried out fast implementation through FPGA and verified its feasibility and correctness in Modelsim-altera simulation software. This paper also predicted the state of the coal mill’s operating, and the sliding window method (deviation degree) is used to diagnose and warn the fault of mill outlet temperature, outlet pressure, motor current and vibration. With the simulation and verification of the historical operation data of coal mills in a thermal power plant, the proposed fault diagnosis model of coal mill based on FPGA self-learning has high precision and is easy to implement in engineering.