基于机器学习的CFST柱高温剩余强度系数预测方法

Research on Prediction Method of Residual Strength Index of CFST Columns at High Temperature Based on Machine Learning

目的 为提高计算精度和速度,利用机器学习模型泛化数据,以预测CFST柱的高温剩余强度系数。方法 利用生成对抗网络将搜集到的110个试验结果泛化生成407组数据,据此训练机器学习模型,并使用试验结果评估其性能,以确定最优模型;然后使用生成数据输入建立的模型预测CFST柱高温剩余强度系数,并和现有计算方法进行对比。结果 建立的随机森林模型在性能度量上表现最好,拟合优度达到0.947 7,均方误差为0.001 8,精度为94.7%;预测结果误差在±10%内的数据为83%,在±20%内的数据为100%;剩余强度系数主要影响因素依次为温度、钢材屈服强度、混凝土抗压强度和横截面积,钢管厚度影响很小。结论 提出的预测方法优于现有计算方法,具有更快的计算速度、更小的结果误差以及更强的模型可解释性,该方法可为CFST柱抗火设计提供参考。

The purpose of this study is improving calculation accuracy and speed,machine learning models were employed to generalize data for predicting the residual strength index of CFST columns under high temperatures.A Generative Adversarial Network was used to generalize and generate 407 datasets from the 110 collected experimental results.These were then used to train machine learning models,with the experimental results used to evaluate their performance and determine the optimal model.The generated data were input into the established model to predict the high-temperature residual strength index of CFST columns,and the results were compared with existing calculation methods.Comparative analysis with available experimental results showed that the Random Forest model had the best performance in terms of metrics,achieving a goodness-of-fit of 0.947 7,a mean squared error of 0.001 8,and an accuracy of 94.7%.The prediction error for 83% of the data was within the ±10% range,and for 100% of the data,it was within the ±20% range.The main influencing factors for the residual strength index,in order of importance,were temperature,steel yield strength,concrete compressive strength,and cross-sectional area,with steel tube thickness having a minimal impact.The findings demonstrate that the proposed prediction method outperforms existing calculation methods,offering faster computation speed,smaller result errors,and stronger model interpretability.This method can provide a reference for fire-resistant design of CFST columns.