基于机器学习的矩形钢管混凝土柱偏压承载力预测

Eccentric compression capacity prediction of rectangular CFST columns based on machine learning

矩形钢管混凝土偏心受压构件作为主要受力构件,准确预测其承载力是保障结构安全的重要前提。采用机器学习方法对矩形钢管混凝土柱偏压承载力进行建模预测。建立了包含804个构件的试验数据库,并通过受力机制与数据相关性分析相结合的方法确定机器学习模型的输入参数;采用BP神经网络、RBF神经网络、高斯过程回归等机器学习算法和多元线性回归方法,建立了其偏压承载力预测模型,并对预测结果进行对比分析。结果表明:几何和荷载参数与偏压承载力之间可呈正相关性或负相关性,而材料参数与偏压承载力之间仅呈正相关性,上述相关性同时受到荷载偏心方向的影响;所建立的3种机器学习模型基于总数据库的预测精度与按偏心方向划分数据库的预测精度相近,表明模型能够反映偏心方向对承载力的影响,避免了进一步划分数据库的需求;高斯过程回归模型的预测精度最高,对数据库中超过60%构件的预测误差小于5%;与设计规范和以往研究中的方法相比,所建立的机器学习模型整体上具备更高的预测精度和更广的适用范围。

Rectangular concrete-filled steel tube(CFST)is widely utilized as a main load-carrying member,for which the accurate prediction of its eccentric compression capacity is crucially important for the structural safety.The eccentric compression capacity of rectangular CFST columns was predicted through machine learning(ML)modeling methods.An experimental database containing 804 CFSTs was established,and input parameters of ML models were determined through the combination of force mechanism and data correlation analysis.Prediction models of the eccentric compression capacity were established using the BP neural network,RBF neural network,Gaussian process regression and multiple linear regression methods,and their prediction results were compared and analyzed.The results reveal that geometric and loading parameters may present positive or negative correlations with the eccentric capacity,whilst the material parameters only exhibit positive correlations.The eccentricity direction of loading also affects these correlations.The accuracy of the three ML models trained by the overall database is close to those trained by the database divided according to eccentric directions,which indicates that the models can effectively reflect the impact of eccentricity direction on the capacity,thus avoiding the need for further subdivision of the database.Gaussian process regression shows the best prediction accuracy,giving less than 5%prediction errors for more than 60%samples.Compared with methods in the design standards and previous studies,the ML models established herein generally exhibit higher accuracy and wider applicable ranges.