Abstract: Aiming at the problems that it is difficult to obtain a large amount of coal mill fault data from the actual operating data of thermal power plants and the precise mathematical model of coal mill is difficult to obtain, which affects the formulation of its fault diagnosis strategy, this paper proposes a deep learning based on simplified mechanism model The fault diagnosis algorithm is used to effectively detect the operating status of the coal mill. Based on the mechanism model of the coal mill system, a large amount of fault data has been obtained through analysis and simulation of different fault types. And through the improved stacked auto-encoder (ISAE) to combine the simulated fault data with the deep learning algorithm to establish a deep learning fault diagnosis strategy, ISAE extracts the essential characteristics of the fault data layer by layer in an unsupervised way, with good learning and The ability to identify fault characteristics. The simulation results also show that the proposed ISAE can well detect the faults of the coal mill, the fault diagnosis accuracy rate is as high as 98.46%, and the early warning can be issued. Key words: Coal mill; Mechanism model; Deep learning; Fault diagnosis; Improved stacked auto-encoder