Aiming at the problems that the existing deep learning methods have poor fault diagnosis accuracy in the case of small samples and the way of constructing graphs of graph neural networks depends on other algorithms, a graph construction method is proposed, and based on this method, a method based on the prior knowledge-graph attention network(PGAT) model of the graph attention mechanism and the prior knowledge base is proposed. The labeled samples and unlabeled samples are connected together in a fixed way, and the similarity between the samples is calculated by introducing the graph attention mechanism, so that the newly added samples do not depend on the topology of the graph, it also solves the problem that graph convolutional neural networks are not easy to expand. Experiments on the benchmark dataset and the oxygen top-blown converter dataset show that with only a small amount of valid data, it has better fault diagnosis accuracy than other models.