Fault diagnosis of three-phase inverter based on multi-source data fusion and transfer learning with small samples

Abstract As the core component of new energy electric vehicles, the operation state of three-phase inverter is directly related to the stability and safety of the system. In addition, the strong suddenness of the open-circuit fault of the inverter often leads to the problem that the samples are scarce and the fault features are difficult to extract, and the traditional data-driven method under small samples is difficult to ensure the diagnostic accuracy and generalization performance. Aiming at the above problems, a set of fault diagnosis methods for three-phase inverters based on multi-source data fusion and transfer learning is proposed. The three-phase current signal is mapped to the frequency domain by fast Fourier transform (FFT), and a two-dimensional image sample with multi-channel coupling characteristics is generated by SDP fusion. Then, Pre-Norm encoder, explicit Q/K/V attention mechanism, SiLU activation function and optimized classification head structure are introduced on the basis of MobileViT backbone network, and combined with transfer learning and phased fine-tuning strategy to improve feature extraction efficiency and small sample generalization ability. The experiment is carried out on seven types of fault data sets. SDP fusion has been verified to have higher separability and accuracy. Under the same data and training configuration, the classification accuracy of the proposed method is about 2% higher than that of the sub-optimal model. The ablation study demonstrates that transfer learning, fine-tuning, and the improved classification head and encoder contribute significantly to performance, while the k-shot experiments further confirm the robustness of the model under very limited samples.