In order to realize the recognition and diagnosis of process faults, this paper uses CapsNet model to train data. Firstly, using the spatial characteristics of the network model，the training data was characterized and normalized in the form of vectors. Then，a convolution operation was performed to classify the faults on the dynamic consistent routing update. Finally，the reconstitution module was added to modify the input data matrix， reduce the loss error and make the network converge quickly. At the same time, feature visualization wasper formed on each layer of the network, and the changes in the feature map of each layer were clearly seen. The experimental results show that the process fault recognition performance of this model is better than other neural network models.