In view of the shortcomings of deep learning methods in modeling small sample fault data, this paper proposes a fault classification method based on Graph Attention Convolutional Neural Network (GATCN) combined with Transformer, which can be well adapted to small sample fault data sets. Firstly, the multi-dimensional sensor data is converted into a multi-component graph representation, and the weights of the edges are updated by the attention mechanism between the nodes of the graph to represent the topological structure between different nodes, to extract the correlation between the multi-component parameters. Then, the Transformer module learn the extracted spatial features to capture the temporal feature relationships of the time series data. Finally, the classification layer based on Softmax is used to classify faults. The model is demonstrated on the TEP data set and the fault datasets from three satellite subsystems. The results show that our model is superior for small sample fault data sets.