Flutter Performance Optimization Of Long-span Suspension Bridges With Box Girder Based On Data-driven Models

Abstract:The number of long-span bridges constructed worldwide has increased drastically in recent decades.The increased number is followed by the rise in the main span length,making such structures very sensitive to wind-induced excitation due to their low structural damping.After the collapse of the Old Tacoma Narrow Bridge(TNB),flutter has been identified as the most dangerous wind-induced excitation phenomenon on the long-span Bridge.Moreover,many earlier studies proved that the aerodynamic performance of such structures was highly related to the shape of the deck adopted during the design process rather than the structural properties.To meet the increasing aerodynamic requirement of such structures,scholars are now turning to data-driven approaches(surrogate and machine learning models),which are very efficient compared to the traditional wind tunnel method.The primary aim of this dissertation is to develop innovative data-driven approaches to solve the aerodynamic shape optimization problem concerning flutter constraint.The main findings of this study are as follows:(1)The data collection step constitutes the bedrock of this study for the accuracy of all models involved in this thesis depends on the quality of the generated dataset.The data was generated using 2D computational fluid dynamics(CFD)simulation through ANSYS FLUENT software.An unsteady Reynolds-average Navier-Stokes(URANS)model was utilized during the CFD simulation to ensure the tradeoff between the accuracy of the results and the computational time.Rigorous verification and validation approach was adopted to ensure the dataset’s quality and accuracy.Grid independence test and time step sensitivity analysis were conducted.Meanwhile,wind tunnel tests were carried out in the CSU-I wind tunnel laboratory on five bridge decks for further validation.The experimental results presented a significant agreement with the CFD results.(2)Secondly,an artificial neural network(ANN)was built to predict the aerodynamic coefficients of a streamlined bridge deck based on the CFD data generated in step(1).The ANN included three optimization algorithms:Bayesian regularization,Levenberg-Marquardt,and scaled conjugate gradient algorithms.The ANN model was constructed using the most influential design parameters as inputs.After model training,the results indicated that the ANN accurately predicted the aerodynamic coefficients.The ANN model built with the Bayesian regularization algorithm presented the best predictive performance.The model has good generalization capability since the difference between the training accuracy and the testing was less than 2%.(3)After that,a polynomial surrogate model was combined with a genetic algorithm to determine the optimal cross-section of a bridge deck concerning flutter constraint.The accuracy of the polynomial surrogate was evaluated using statistical metrics such as R-squared coefficient of determination,root means squared error(RMSE),and sum squared error(SSE).Notably,the result of the R-squared is superior to 0.90 for the flutter velocity approximation.Meanwhile,the approach significantly improved the existing Bridge’s aerodynamic performance by increasing the flutter velocity by 25%to 35%.It is essential to mention that increasing the number of the sampling design from 57 to 73 points does not affect the optimization results since the sampling was uniformly distributed in the design domain to reduce bias.(4)A probabilistic machine learning model based on hierarchical Bayesian modeling(PML-HBM)was used to accurately predict the critical flutter velocity of long-span suspension bridge.The results indicated that the model accuracy and the computation time increased with the sample size.The statistically significant results showed that the PML-HBM model is accurate and can be an alternative approach to overcome the lack of big data problems in machine learning(deep learning)model construction.Moreover,the proposed PML-HBM model overperformed the conventional machine learning models such as extreme gradient boosting regression(XGBR),random forest(RF),and support vector regression(SVR).As opposed to conventional machine learning algorithms,the PML-HBM models include uncertainty in the data during the training process,which is an important feature since there is uncertainty in design variables as well as aerodynamic load.122 Figures,33 Tables,291 References...