Research On Shear Load Capacity Of Reinforced Concrete T-Beams Based On Machine Learning

Accurate prediction of the shear capacity of reinforced concrete beams has been a f ocal point of research for scholars over the past century.Although numerous shear analy sis theories have been proposed,a universally accepted unified model has yet to emerge.Most calculations for shear capacity in various national codes rely on semi-empirical an d semi-theoretical formulas,with differing theoretical foundations,experimental data,an d engineering experiences leading to significant discrepancies between the formulas of d ifferent countries’ codes.Among these,the Modified Compression Field Theory(MCFT)is a widely recognized and adopted shear theory in some national codes.However,its in itial version only targeted pure shear panels,while in reality,the shear zone of a beam is subjected to the combined effects of bending moments and shear forces.Therefore,it is essential to consider the shear capacity of reinforced concrete beams under flexural-shea r conditions.Additionally,the calculations based on the MCFT model are relatively com plex and involve numerous assumptions when simplified,which inevitably affects the ac curacy of the model’s predictions.These formulas are primarily based on experimental r esults for rectangular section beams,while there is limited research and experimentation on the shear capacity of T-shaped section beams,which are usually simplified into equi valent sections for calculations.At the same time,it is challenging to accurately account for the randomness of various parameters affecting the shear strength,making it essentia l to conduct a probabilistic analysis of shear capacity.This study,based on the classic Mo dified Compression Field Theory,focuses on the shear capacity of T-beams under flexur al-shear conditions using explicit calculations and probabilistic modeling methods.The main research contents are as follows:(1)Various deformation theory models originating from the classic truss theory(inc luding variable angle trusses,softened trusses,and tension-compression bar models)and other shear resistance models(plastic theory,limit equilibrium theory,and stress path m ethod)were reviewed and summarized.The evolution process of the shear application in Compression Field Theory(CFT),Modified Compression Field Theory(MCFT),and D isturbed Stress Field Model(DSFM)was detailed.The shear design methods in four rep resentative codes were summarized and compared,with the advantages and disadvantag es of each analyzed by comparing the influencing parameters.By analyzing the characte ristics of various failure theories,this study provides a theoretical foundation for subseq uent shear design.(2)A total of 213 sets of experimental data of reinforced concrete T-shaped section simply supported beams subjected to shear were collected from domestic and internation al literature.The constitutive relations,strain coordination equations,average stress,and local force-physical equations of cracks in the Modified Compression Field Theory wer e summarized.The enhancement effect of the T-beam flange on the shear capacity was c onsidered,and the basic equations of MCFT were modified using the bending moment it eration condition.A calculation method for the shear capacity of concrete T-beams under flexural-shear coupling was proposed,and the corresponding nonlinear iterative progra m was developed.The iterative MCFT model presented in this study was verified for its internal force response through classical examples,resulting in a ratio of 1.04 between t he calculated and measured shear capacities.(3)Based on the flexural-shear algorithm,response data such as the ultimate bearin g capacity,inclined crack angle,and compressive strain of the concrete in the compressi on zone of T-beams at the critical state were obtained.A prediction model was develope d using the Backpropagation(BP)neural network algorithm for learning and training.A two-factor analysis was conducted to determine the coupling relationships between shea r span ratio,stirrup ratio,concrete compressive strength,rebar yield strength,reinforcem ent ratio,critical inclined crack angle,and compressive strain in the compression zone.A n additional equation obtained through stepwise regression was introduced into the iterat ive MCFT process,allowing for explicit calculation of the iterative MCFT algorithm.Fi nally,the model errors of representative codes were discussed.The study found that the BP neural network prediction was outstanding,and the accuracy of the integrated learnin g model was higher than that of the single learning model.The two-factor analysis based on the BP network revealed the coupling effects between multiple input parameters,wh ich could influence the shear strength law of a single parameter.The strength characteris tic value ρv·fyv/fc′ was the main factor affecting the critical inclined crack angle θ and t he top edge compressive strain εtom at beam shear failure.The reinforcement ratio ρl als o had a significant impact on the critical inclined crack angle θ,while the shear span rati o was the key factor affecting the top edge compressive strain εtom at beam shear failur e.The predictions of shear capacity of T-beams with web reinforcement in all four codes were found to be conservative.(4)Based on Bayesian theory,a simplified probability model formula for the shear capacity of reinforced concrete T-beams was derived.The effectiveness of the model and the characteristic values at different confidence levels were verified.The prediction acc uracy of traditional codes and precise models was compared,followed by local paramete r sensitivity analysis based on the local database.The study found that different prior inf ormation of the model affected the updating accuracy of the probability model,and the d ispersion of various parameters decreased as the number of batches introduced increased.The posterior model corrected the parameter influence of the prior information accordin gly,effectively eliminating the bias caused by uncertainty.With the stirrup characteristic value ρvfv as the dividing line at 2.8,the accuracy calibration brought by the probabilit y model significantly increased.The calculation results of the precise formula were clos er to the experimental values compared to the code,and the prediction accuracy of the m odel based on the theoretical foundation was higher than that of the model based on the code,although the dispersion was high for both.Compared to other formulas,the probab ility model could correctly reflect the influence trend of all parameters.