Study On Collapse Resistance Of Concrete Frame Structures Based On Machine Learning Algorithms

With the development of science and technology in the field of engineering,experts promote the development of building construction sector with new breakthroughs and progress.In recent years,the various progressive collapse incidents occurred within buildings from time to time reveal that the potential risks of buildings to resist progressive collapse are still large,which also shows that the existing relevant design codes are still not enough to fully cover the design of buildings to resist progressive collapse.Therefore,it is meaningful and desirable to put much more research efforts in the filed buildings on how to resist progressive collapse.At present,researchers at home and abroad have carried out a lot of research work on the progressive collapse resistances of structures by means of experimental research,theoretical analysis and numerical simulation.With the emergence of machine learning methods broadly applied in all aspects of engineering design and practice,there exists a new possibility to examine the progressive collapse resistance of concrete frame structures.This work aims to provide a basis for the design of progressive collapse resistance of concrete frame structures by use of the machine learning related algorithms.In this work,the appropriateness of three kinds of machine learning algorithms for the progressive collapse resistance of concrete frame structures were determined and a large database regarding tests of concrete frame structures to resist progressive collapse was setup.Then the models based on the selected three machine learning algorithms to evaluate the load bearing capacities with respect to the compressive arching action and catenary action in concrete frame structures were established followed by the comparison among the three evaluation models from the perspective of accuracy and effectiveness.Finally,a prediction model with explicit expressions was proposed to predict the load bearing capacities with respect to the compressive arching action and catenary action for engineering design.The main research contents of this work are as follows:(1)By exploring the machine learning related theories and their practical applications and the actual needs of this specific work,three machine learning algorithms i.e.back-propagation neural network,support vector machine regression algorithm and gene expression programming were proposed to be feasible to evaluate and predict the load bearing capacities regarding the compressive arching action and catenary action in concrete frame structures.(2)The literature review work of this study reveals that the loading bearing capacities in the compression arching action and catenary action stages can be viewed as the key indicators for evaluating the progressive collapse resistance of concrete frame structures.After analyzing the main factors that affect the load bearing capacities regarding the compressive arching action and catenary action,a relatively large experiment database consisting of 158 concrete frame substructure specimens was established from 42 published journal papers.The database was then subdivided into training set,validation set and test set through pre-training.Also,the statistical parameter distribution characteristics and correlation characteristics of each variable in the database are studied to provide a basis for the establishment of machine learning models.(3)Aiming at evaluate loading bearing capacities in the compression arching action and catenary action stages of the concrete frame structures,the hyperparameters of the three machine learning models were debugged by the trial and error method and finally optimized three evaluation models respectively based on the back-propagation neural network,support vector machine regression algorithm and gene expression programming were established.The direct evaluation fitting results as well as the regression results were used to qualitatively describe the feasibility of the three proposed models and the learning error index was used to quantitatively evaluates the learning accuracy as well as the generalization ability of the three learning models.The results show that,compared with the evaluation model based on back-propagation neural network and support vector machine,the evaluation model based on gene expression programming shows better merits in accuracy and practicability in evaluating the loading bearing capacities in the compression arching action and catenary action stages of the concrete frame structure.(4)On the basis of the evaluation model based on the gene expression programming algorithm,the prediction models with explicit expressions on the load bearing capacities with respect to the compressive arching action and catenary action for the concrete frame structures were proposed based on the gene expression programming algorithm.It is found that the proposed models exhibit better good accuracy and effectiveness to predict the load bearing capacities with respect to the compressive arching action and catenary action when compared with the currently widely accepted analytical models.The results of local and global sensitivity analyses also show that the prediction models based on gene expression programming algorithm are capable of grasping the influences of the properties of steel reinforcement,concrete as well as external constraints on the load bearing capacities with respect to the compressive arching action and catenary action.Therefore,it is believed that the proposed prediction models based on gene expression programming algorithm can be used for engineering designs regarding the progressive collapse of concrete frame structures.