Fault diagnosis is one of the most important issues for a modular multilevel converter (MMC). However, conventional solutions are deficient in two aspects. Firstly, they lack the necessary feature information. Secondly, they are incapable of performing open-circuit fault diagnosis of the modular multilevel converter with the requisite degree of accuracy. To solve this problem, an intelligent diagnosis method is proposed to integrate the modal time–frequency diagram and FFT-CNN-BiGRU-Attention. By selecting the phase current and bridge arm voltage as the core fault parameters, the particle swarm algorithm is used to optimize the Variational Modal Decomposition parameters, and the fault signal is decomposed and reconstructed into sensitive feature components. The reconstructed signals are further transformed into modal time–frequency diagrams via continuous wavelet transform to fully retain the time–frequency domain features. In the model construction stage, the frequency–domain features are first extracted using the fast Fourier transform (FFT), and the local patterns are captured through a combination with a convolutional neural network; subsequently, the timing correlations are analyzed using bidirectional gated loop cells, and the Attention Mechanism is introduced to strengthen the key features. Simulations show that the proposed method achieves 98.63% accuracy in locating faulty insulated gate bipolar transistors (IGBTs) in the sub-module, with second-level real-time response capability. Compared with the recently published scheme, it maintains stable performance under complex working conditions such as noise interference and data imbalances, showing stronger robustness and practical value. This study provides a new idea for the intelligent operation and maintenance of power electronic devices, which can be extended to the fault diagnosis of other power equipment in the future.