To address the low data processing efficiency and insufficient identification accuracy of traditional bridge risk identification methods when dealing with complex risk scenarios and high-dimensional monitoring data, an integrated bridge risk identification method based on XGBoost and RF-MLP is proposed. Typical risk scenarios that may occur during the actual operation of a three-span continuous girder bridge are considered, and a risk scenario database is constructed from three dimensions, including risk type, risk location, and damage severity. Structural response data under multiple risk scenarios were generated based on finite element simulations, and highdimensional training samples were established. Key features were extracted using a random forest algorithm, and XGBoost was introduced to optimize the network weight update strategy, thereby improving the nonlinear modelingcapability and convergence accuracy of the multilayer perceptron. The trained model was finally applied to experimental data from an actual bridge to verify its identification performance and engineering applicability. The experimental results show that the proposed XGBoost–RF–MLP model achieves training accuracies of 98.07% for classification, 83.02% for localization, and 96.67% for quantification, which are significantly higher than those of the traditional RF-MLP model in terms of accuracy and robustness. In the full-scale bridge experiment, identification accuracies of 97.4%, 87.5%, and 100% are achieved, respectively. The proposed method effectively addresses bridge safety identification tasks under complex risk scenarios and multiple operating conditions, and provides a feasible technical approach and theoretical support for intelligent monitoring and risk management of bridge structures.