Abstract: In the grid-connected operating environment of neutral point clamped (NPC) three-level inverters, IGBTs face the challenge of single-tube and dual-tube simultaneous faults, which makes the differences between fault features extremely subtle and consequently hinders the effective improvement of dual-tube fault identification accuracy. To address this issue, this paper proposes a novel fault diagnosis method that combines multi-channel two-dimensional recurrence fusion plots and a lightweight multiscale convolutional residuals (LMCR) network. Firstly, three-phase current signals are obtained through simulation as fault signals. Then, recurrence plots (RPs) are utilized to convert the three-phase current signals into two-dimensional images, respectively, which are subsequently fused in a multi-channel manner to capture features such as periodicity, abrupt changes, and trends in the time series. Finally, the recurrence fusion plots are input into the LMCR model for fault identification. The LMCR model integrates multi-level Inception structures and residual networks to extract and fuse features at different scales, ensuring the mitigation of vanishing and exploding gradients in the network. Experimental results demonstrate that this method excels in IGBT fault identification, achieving an average accuracy of 100% in a noise-free environment and 92.53% in a noisy environment, fully proving its strong feature extraction capability and excellent noise resistance.