Abstract: In long-distance large-diameter EPB shield tunnelling in composite strata, screw conveyor shaft fracture accidents can lead to severe safety and construction schedule risks. To quantitatively predict the spatial distribution characteristics of screw conveyor shaft fractures, a combined numerical simulation and statistical modeling approach was used, combining nonlinear regression algorithms and BP neural networks. A prediction model for screw conveyor shaft fracture risk was developed and tested based on actual engineering projects. The results show that: (1) The screw conveyor fault prediction model achieves high accuracy in predicting shaft fracture risks and demonstrates reliability in risk probability predictions, providing non-parametric interval estimates for risk ring numbers; (2) Measures such as using fully solid shafts or double-head solid shafts, controlling the welding quality of stressed components like cutter heads and screws, regulating the cutter head opening size, and setting reasonable torque limits for screw conveyors can reduce the risk of screw conveyor shaft fractures.