A three-branch heterogeneous fusion diagnostic method based on GAF-RP-LSTM-Transformer is proposed for inverter open-circuit faults. First, the output current signals are processed to extract localized fault characteristics through complementary ensemble empirical mode decomposition and phase randomization techniques (CEEMD-PRT). Then, the one-dimensional temporal signals are converted into two-dimensional images via Gramian angular field (GAF) and recurrence plot (RP) transformations, effectively leveraging global trend features (GAF) and nonlinear dynamic characteristics (RP) embedded in the temporal sequences. To overcome the limitations of conventional one-dimensional feature extraction in spatial correlation representation, a long short-term memory (LSTM) network is employed to capture dynamic temporal features, while a dual-branch GAF-RP-Transformer model extracts spatial features from the two-dimensional images. To enable multidimensional fusion of temporal and spatial characteristics, a novel heterogeneous feature fusion module is proposed, leveraging the complementarity of multi-modal images to enhance the model’s ability to capture subtle fault differences. Experimental results demonstrate that the proposed model achieves a classification accuracy of 99.3% on the test sets, significantly outperforming comparative models while maintaining high diagnostic accuracy under varying noise conditions. In particular, under 30 dB and 20 dB noise levels, the accuracy degradation remains small, indicating strong robustness. Simulation results validate the effectiveness and superiority of the GAF-RP- LSTM-Transformer three-branch heterogeneous fusion framework in inverter fault diagnosis.