基于深度学习的摇摆自复位桥墩设计参数优化方法

Design Parameter Optimization of Self-centering Pier Based on Deep Learning

在墩底与承台间设置摇摆提离面,并将无黏结预应力钢筋及耗能钢筋内置于桥墩中,是实现桥墩自复位的主要方式之一。桥墩的几何参数及材料力学参数直接影响自复位桥墩的抗震性能,但各设计参数之间相互制约,其对摇摆自复位桥墩的耗能能力、残余变形及承载能力3项重要抗震性能指标的总体影响规律尚不明确,摇摆自复位桥墩的设计参数优化问题有待解决。本文提出了一种基于深度学习的摇摆自复位桥墩设计参数优化方法,通过NSGA和MOEA/D算法实现了函数的多目标优化求解,利用AHP–熵权法使得各目标间权重系数能够考虑主客观因素,最后利用TOPSIS方法对Pareto最优解进行排序决策。以1座内置耗能钢筋的摇摆自复位桥墩构件为例,建立了有限元分析模型,兼顾残余位移、等效黏滞阻尼系数及承载力3项抗震性能指标,优化摇摆自复位桥墩的关键设计参数。结果表明:该方法可同时考虑桥墩的几何及材料力学性能参数的随机性,对多个设计目标进行统筹设计,快速筛选出综合性能最佳的设计参数;通过深度学习模型的计算,获得了各决策参数对抗震性能指标的影响规律;较长的耗能钢筋无黏结段能有效降低残余变形,较大的预应力筋初始预张力对结构承载能力与残余变形有利,使用高等级耗能钢筋提升摇摆桥墩耗能和承载能力具有较为明显的效果。利用深度学习方法创建有限元结构代理模型,能够引入结构几何和材料力学性能随机参数提升模型鲁棒性能,使得摇摆自复位桥墩多目标优化效率明显提升。

Separating the pier and platform at the base of the pier and embedding the prestressed steel bars and reinforcements into the pier body facilitates the creation of a self-centering pier. The pier’s geometric and material mechanical parameters affect its seismic performance;however, the impact on critical seismic performance indices (energy-consuming capacity, residual deformation, and load-bearing capacity) remains unclear due to the coupling of design parameters. Optimizing these parameters for the rocking self-centering pier is crucial. This study proposes a deep learning-based method for optimizing the design parameters of self-centering piers. The NSGA and MOEA/D algorithms are employed to achieve a multi-objective optimization solution. The AHP-entropy weighting method considers subjective and objective factors when determining the weight coefficients among targets. The TOPSIS method is applied to rank the optimal solutions for Pareto. A finite element analysis model is de-veloped to optimize the critical design parameters of the self-centering pier with embedded prestressed steel bars and reinforcements, considering the three seismic performance indices (residual displacement, equivalent viscous damping coefficient, and load-bearing capacity). The results demonstrated that the method can account for the randomness in the pier’s geometric and material mechanical performance parameters, harmon-ize multiple design objectives, and rapidly identify design parameters with the best comprehensive performance. In addition, longer unbonded re-inforcements can effectively reduce residual deformation, higher initial pretension of prestressing tendons enhances structural load-carrying capa-city and mitigates residual deformation, and high-grade reinforcements are advantageous for improving load-carrying capacity and reducing resid-ual deformation. A finite element structural agent model created through the deep learning method can incorporate random parameters related to geometric and material mechanical properties to improve the model’s robustness and the self-centering pier’s multi-objective optimization effi-ciency.