In order to solve the problem that the accuracy of traditional fault diagnosis model is limited by the scarcity of fault samples and the coupling of time dimension and variable dimension of fault data, a fault diagnosis method based on deep learning is proposed for complex industrial systems such as moisture separator and reheater system. Firstly, the digital twin system of moisture separator and reheater is constructed to establish the fault diagnosis data warehouse and solve the problem of scarcity of data samples. Secondly, based on the previous step, a fault diagnosis model based on deep residual network is constructed to diagnose the typical faults of steam water separation and reheat system, including uneven flow, break, deterioration of heat transfer and change of valve characteristics, so as to solve the problem of timevarying and multidimensional data variables. The simulation results show that the digital twin system can realize the accurate simulation of the steadystate, dynamic and fault conditions of the steam water separation and reheat system, and meet the data requirements of the subsequent indepth learning model; the fault diagnosis model based on deep residual network can realize the fault diagnosis of timevarying and multidimensional industrial data. The Tdistributed stochastic neighborhood embedding (TSNE) methodis used to visualize the model and verify that the suggested diagnostic model distinguishes significantly between different fault types of data.