基于贝叶斯优化XGBoost的建筑施工事故类型预测

Prediction on accident types of building construction based on Bayesian optimized XGBoost

为探究建筑工程施工过程中不同风险特征与施工事故类型之间的关联性,提出1种融合特征选择算法和机器学习算法的建筑事故类型预测模型。基于619项国内建筑事故报告建立建筑施工风险特征体系,通过条件互信息最大化(CMIM)-Boruta方法筛选出26个关键风险特征,将其作为贝叶斯优化极限梯度提升(XGBoost)预测模型的输入变量并在测试集上评估该模型的预测精度。研究结果表明:XGBoost模型的预测性能优于其他机器学习模型;CMIM-Boruta方法和贝叶斯优化方法能够有效提升机器学习模型的预测性能;通过2个实际事故案例验证得到该模型具有一定实用性。研究结果对相关企业安全管理人员更准确地识别施工现场潜在危险、采取更具针对性的预防措施具有一定参考意义。

In order to explore the correlation between different risk characteristics and construction accident types in the construction process of construction engineering,a prediction model of construction accident types combining the feature selection algorithm and machine learning algorithm was proposed.Based on 619 domestic construction accident reports,a construction risk feature system was established,and 26 key risk features were screened out by the conditional mutual information maximization(CMIM)-Boruta method,which were used as the input variables of Bayesian optimized extreme gradient boosting(XGBoost)prediction model.The prediction accuracy of the proposed model was evaluated on the test sets.The results show that the XGBoost model has better prediction performance than other machining learning methods.Moreover,the CMIM-Boruta method and Bayesian optimization method can effectively improve the prediction performance of machine learning models.Through two actual accident cases,it is proved that the model has certain practicability.The research results have reference significance for the safety management personnel to more accurately identify the potential hazards at the construction site and then take more targeted prevention measures.