Recently, graph neural network (GCN)-based fault diagnosis methods have been received increasing attention. However, regarding the obtained results, existing methods are often lack of interpretability. To solve this problem, in this paper, a graph convolutional network-class activation mapping (GCNCAM) model is established, which can interpret the results of GCN-based fault diagnosis method. Specifically, the definition of interpretability in GCN is firstly clarified, then the model uses GCN layer to extract features Z GCN of input measurements, it uses Z GCN to obtain the diagnostic results through the global average pooling (GAP) layer and fully connected layer. Next, the trained fully connected layer and a softmax layer are used to process the output value Z GCN directly, and the normalized interpretability vectors of measurements are obtained. To verify the effectiveness of the method, an experiment on a hardware-in-the-loop platform of the rectifier of a high-speed train is carried out. The experimental results show that the proposed GCN-CAM model can not only complete the task of fault diagnosis, but also provide a convincing explanation for the diagnosis results