The graph neural network models have been widely used in the field of fault diagnosis due to the advantage of abundant fault characterization capabilities. However， the existing models only utilize the local information among neighboring nodes when dealing with fault data， and fail to fully extract the global feature information. Meanwhile， in order to overcome the problems of low accuracy and insufficient generalization ability of single model. This paper proposes an ensemble method with multi-scale graph pooling feature fusion and graph convolutional network （MSGP-GCN）. The graph model is constructed from the original signal， and global information is obtained using graph pooling coarsening. Then weights are assigned at different scales based on the degree of the nodes， and the global information is used to update the node features in combination with the weights. The updated node fea? tures are input into different classifiers respectively， and the intelligent fault diagnosis result is obtained by majority voting strategy among these classification results. The proposed approach is fully verified by two fault datasets， the SEU simulation dataset and the real coal mill dataset. The experimental results show that the proposed model can effectively improve fault diagnostic accuracy and generalization ability in aforesaid two real datasets， and the average diagnostic accuracy reaches 98.31% and 97.21%， respectively.