The accuracy of data-driven open-circuit fault (OCF) diagnosis methods is affected by varying operating conditions. This issue is often ignored. In this article, an improved convolutional neural network (CNN) and a sample amplification method are proposed to eliminate the influence of varying operating conditions on the online OCF diagnosis for neutral point clamped (NPC) inverter. First, 73 types of OCF sample collection can be greatly reduced to 14 by following the sample amplification method. The signals of any phase can be generated by a single fundamental period signal. This provides a significant savings in sample collection time. Second, the spatial attention mechanism (SAM) is added after the first convolutional layer of the CNN model. The feature extraction capability of the model is enhanced for time-domain waveform scaling under variable operating conditions. Simultaneously, the last full connection (FC) layer of the CNN model is retained and the other FC layers are substituted with a global maximum pooling (GMP) layer. This has the advantage of reducing the number of network parameters and further conserving the effective feature information. In conclusion, the experimental results show that the sample amplification method and the improved CNN model for online fault diagnosis under varying operating conditions exceed 99% accuracy. The CNN with spatial attention mechanism and global maximum pooling (SAMCNN-GMP) is more effective and stable than the CNN model.