In order to reasonably choose a sample condition which supports efficient intelligent diagnosis, and to overcome the problems of too many weights and weak local information extraction capability of intelligent traditional BP (back propagation) network, an open-circuit faults diagnosis method based on CNN(convolutional neural network) was studied. Moreover, by taking the typical three-phase two-level inverter as the specific object, the advantages of the CNN method on network weights number, network training stability and diagnosis effects under different conditions of sample durations and training sample numbers over the BP network method were analyzed quantitatively. Results show that the CNN method can build a deeper network model with much less weights than the BP network method, and it can achieve efficient and accurate model training and diagnosis with shorter and less samples.