WithCracks distribution patterns the deepening of the development of urban underground space, the construction qualityMonitoring construction quality and crack distribution of underground diaphragm wallDiaphragm walls, as a key component of foundation pit support and underground structure seepage prevention, have a decisive impact on engineering safety. In this paper, we propose an improved multi-source data fusion monitoring model, Hybrid-GSDNet, which is deeply extended on the basis of the existing Kalman filter-based displacement prediction and graph neural network crack recognition framework. Hybrid-GSDNet integrates three types of monitoring data: distributed optical fiber sensing (DOFS), infrared thermal imaging (IRT) and geological radar (GPR), and models the wall-soil interaction through a multi-scale graph convolution module, and introduces a temporal attention encoder to identify the time-varying propagation mechanism of cracks. In addition, the model introduces an adaptive sensing fusion mechanism to dynamically adjust the weights according to the confidence level of the sensor data, which effectively improves the robustness to interference and data loss. The model innovatively combines the physical evolution mechanism of fractures with the extraction of deep structural features to achieve high-precision identification of the fracture initiation zone and prediction of the expansion path. Experimental verification on underground diaphragm wallDiaphragm walls engineering shows that Hybrid-GSDNet achieves 92.6% crack location accuracy.