The modular multilevel converter (MMC) is a key power conversion component in flexible DC transmission and distribution systems. However, its cascaded submodule topology, which incorporates a large number of switching devices, presents reliability challenges and contributes to a higher failure rate. Traditional open-circuit fault diagnosis methods for MMC switching devices often rely on additional sensors and are susceptible to interference due to threshold sensitivity. To address these limitations, this paper introduces a novel open-circuit fault diagnosis and localization approach based on tensor feature extraction and a two-dimensional broad learning system (2D-BLS), enabling fast and highly accurate fault identification.The proposed method constructs a third-order tensor from submodule capacitor voltage data to efficiently handle multi-channel MMC signals. Through Tucker decomposition, the method separates fault-type classification from fault-location identification while extracting meaningful tensor features. Each subtask′s tensor features are then processed using dedicated sub-classifiers built on the 2D-BLS framework. The 2D-BLS employs a bilinear transformation to maintain structural information while significantly reducing the number of parameters. The outputs of all sub-classifiers are subsequently fused to accomplish fault diagnosis and localization.This approach eliminates the need for additional sensors and empirical thresholds, reduces the model′s class complexity, and enhances both diagnostic accuracy and computational efficiency. It is particularly well-suited for handling multiple open-circuit faults in switching devices. Simulation and experimental results confirm that the proposed method achieves a diagnosis and localization time of less than 15 ms with an accuracy exceeding 98.5%, demonstrating its effectiveness and superiority.