In order to realize the fault diagnosis of the tennessee eastman (TE) process, the fault diagnosis method of the deep belief network (DBN) is improved. The traditional DBN will generate redundant features in the training process, weaken the feature extraction ability of the network, improve the DBN to add the penalty regular term in the likelihood function of the unsupervised learning phase, obtain the sparse distribution of the DBN training set through the sparse constraint, and then use the Laplace function. The distribution guides the sparse state of the DBN node, and uses the positional parameters in the Laplace function to control the sparse strength, so that the unlabeled data features can be more intuitively represented. Finally, the improved DBN and traditional DBN and BP neural network simulation results are compared. The experimental results show that the improved DBN is superior to the traditional DBN and BP neural network in fault diagnosis, achieving the best diagnostic accuracy and high Theoretical research value.