Aiming at the problems of insufficient signal feature extraction, insufficient recognition ability of deep learning network and low recognition rate under high noise condition in inverter fault diagnosis in traditional voltage source converter static synchronous compensator, an inverter fault diagnosis method based on the combination of dual-mode decomposition, multi-channel input(MCI), parallel convolutional neural network(PCNN), bi-directional long and short-term memory (BiLSTM) neural network and self-attention(SA) mechanism is proposed. Firstly, the three-phase current output of the inverter is decomposed by variational mode decomposition and time-varying filter empirical mode decomposition, which reduces the complexity of the original signal and realizes the law complementation between different modal components. Secondly, MCI-PCNN-BiLSTM-SA combined model is used to extract, learn and recognize the feature matrix. Finally, the proposed method is validated by simulation, and the results show that the proposed method has strong feature extraction ability, with an average recognition rate of 99.48% in the case of no noise and 95.59% in the case of high noise.