:For photovoltaic inverter circuit in the photovoltaic system is complex, failure time is short, this paper puts forward a kind of based on the improved variational mode decomposition and the convolution of the neural network fault diagnosis methods, which can effectively solve the fault feature extraction is difficult, characteristic parameters of singularity is poor, and the poor because of the characteristic parameters, low caused by the fault diagnosis problem.Firstly, the software SIMULINK was used to establish the photovoltaic inverter soft fault model, and relevant parameters were collected as samples.Then, VMD is used for variational modal decomposition of the parameters to obtain some components, and wavelet transform is used to extract the wavelet energy of each modal component to obtain the fault characteristic value and reduce the dimension.Finally, the CNN was used for fault diagnosis, and the results were compared with the traditional VMD-CNN neural network and VMD-BP neural network, which verified the correctness and accuracy of the soft-fault diagnosis of photovoltaic inverter using the network, and had certain advantages.