Abstract: The geological environment of coal mine goaf is complex and varied, and the geological structures such as faults, fractures, and folds developed inside are not only difficult to accurately grasp, but also significantly change the migration path and water pressure distribution pattern of groundwater. When facing multiple uncertain factors such as large-scale groundwater flow, complex geological structure coupling, and dynamic changes in water pressure, traditional monitoring techniques have prominent problems such as insufficient adaptability and limited warning accuracy. Therefore, a water pressure limit warning method of goaf catchment flow field based on bidirectional fluid structure coupling theory is proposed. A bidirectional fluid solid coupling model is established for the catchment flow field in goaf, and the precise simulation of the coupling effect between fluid flow and rock deformation is achieved through the COMSOL Multiphysics multi physics simulation platform. Key parameters including fluid pressure field, velocity field, and fluid solid interaction force field are obtained. To address the issue of information redundancy in massive monitoring data, a local sparse unsupervised feature selection algorithm is used to reduce the dimensionality of the original data, effectively extracting the key feature dataset that best reflects changes in water pressure status. A water pressure limit warning model is established based on support vector machine. Nonlinear problems are transformed into linearly separable problems in high-dimensional space through kernel function mapping, achieving intelligent prediction and graded warning of water pressure status in goaf. The experimental results show that the proposed bidirectional fluid structure coupling model performs well in terms of data accuracy, and the simulated and measured fluid pressure, flow velocity, and fluid structure interaction force data are highly consistent. The feature selection process significantly reduces data redundancy through local sparse unsupervised methods, effectively improving the independence of the feature set. In terms of predictive performance, the bidirectional fluid structure coupling SVM prediction method proposed in this study has higher overall warning accuracy, higher high-risk state recognition rate, and lower false alarm rate compared to traditional methods, meeting the real-time warning requirements. This data fully validates the reliability and engineering applicability of the method in water pressure warning of goafs.