Study On Mechanical Properties Of Structures Based On Improved Stacking Ensemble Learning

With the upgrading and transformation of the manufacturing industry to high-end manufacturing,the demand for new engineering materials and structures is increasing.Steel structures are one of the most widely used engineering structures.However,experimental-based mechanical properties studies of steel structures require a large amount of resources and time,and the computational process of mechanical properties studies based on the finite element approach is unstable and time-consuming,etc.Both traditional methods for evaluating the mechanical properties of steel structures are not favorable to the design of new steel structures that are stable,efficient,and inexpensive.In recent years,the discovery and design of novel materials and structures based on machine learning techniques has received a lot of attention,and it has made considerable strides in terms of efficiency and prediction accuracy.However,most datasets for novel materials and structures are based on small sample sizes derived from experimental tests,which can easily induce overfitting and lead to low prediction accuracy for machine learning models.To address the existing problems,this paper employs machine learning techniques to model and analyze two steel structures,and also conducts reliability analysis and interpretability analysis based on the established prediction models.The innovative points and main contents of this paper are as follows:(1)A Grid Search-Stacking Algorithm(GSSA)algorithm based on the improved Stacking algorithm is suggested to successfully overcome the problem of low prediction accuracy owing to training on small sample sets.Firstly,the parameters of the base models in standard Stacking algorithm are adjusted using a Bayesian optimization algorithm;then,the optimal combination of base models is chosen from the alternative base models using a grid search algorithm;finally,the base models are trained on the original dataset using a leave-one-out cross-validation method to generate new data.(2)A buckling strength prediction model is established for high-strength steel Y-section columns by using the GSSA algorithm with initial geometry and initial defects as input variables and buckling strength as output variable.Firstly,the prediction accuracy of GSSA model,traditional Stacking algorithm model and several common machine learning algorithm models are compared and analyzed by three regression evaluation criteria,and different evaluation criteria show that the GSSA model has the highest prediction accuracy,with the three regression evaluation criteria being 0.9752,31.9545 and 41.6456;then,Bland-Altman method is adopted to evaluate the consistency of the GSSA model,and the results show that the deviations of all the tested samples fall within the 95%consistency limits,indicating that the GSSA model has good reliability;finally,the SHAP method is employed to assess the interpretability of the prediction results,and the results reveal that the width of the steel plate _bb is the most essential feature affecting buckling strength,which has a positive relationship with it.(3)A tensile strength prediction model based on GSSA algorithm is established for industrially produced steel with chemical composition and process parameters as input variables and tensile strength as output variable.Firstly,the prediction accuracy of the GSSA model is analyzed,and the three regression evaluation criteria of GSSA model are calculated to be 0.9751,14.5159 and 19.6683,which are better than the traditional Stacking algorithm model and several common machine learning algorithm models;then,the GSSA model is analyzed for consistency through the Bland-Altman method,and the results show that the deviation of 95 out of 100 test samples fall within the 95%consistency limits,so that the GSSA model had good reliability;finally,the interpretable analysis of the predicted results by SHAP method show that the carbon equivalent 2 Pcm is the most critical factor affecting the tensile strength,and the tensile strength increase with the increase of carbon equivalent 2.In this paper,a theory model based on GSSA algorithm is proposed to predict the mechanical properties of two steel structures,and the prediction results are analyzed by Bland-Altman method and SHAP method.It realizes the high-precision prediction of mechanical properties and evaluation of influencing factors,which provides guidance for the rapid analysis and design of new steel structures.