Artificial intelligence algorithms are widely used in fault diagnosis of modular multilevel converter ( MMC). However, the existing algorithms require a large number of target domain samples to train the model. To address the problem that it is difficult to diagnose accurately under small samples, a MMC small sample discrete fault diagnosis method based on a multi-source fusion graph and SE-BiGRU-ResNet model is proposed. Firstly, according to the characteristics of an open-circuit fault, the output phase current and bridge arm voltage is selected as the key fault parameters. Secondly, the 1D fault parameters are mapped into the corresponding 2D feature images by using the recurrence plot, Markov transition field, and the Gramian angular field algorithm. To comprehensively strengthen the feature saliency of the image, the multi-source fusion graph is obtained by adding each graph according to the channel dimension. Finally, based on the residual network (ResNet), to improve the ability of the model to capture key spatiotemporal features, the squeeze-excitation (SE) module and the bidirectional gated recurrent unit (BiGRU) module are introduced. The SE-BiGRU-ResNet model is formulated to train and test the multi-source fusion graph. Compared with other methods, the experimental results show that the accuracy of fault diagnosis of IGBT in the fault bridge arm and positioning sub-module reaches 98. 10% and 99. 13% in the case of small samples, and the diagnostic accuracy is high. The test process has a second-level response time. It still has good diagnostic performance and strong generalization ability under extreme conditions. Keywords:modular multilevel converter; fault diagnosis; sm