摘要
针对岩土力学参数正算反演方法存在的瓶颈问题,研究了将支持向量回归机(ε-SVR)与粒子群-差分演化杂交智能优化算法(BBDE)相结合的参数反演方法,采用ε-SVR的预测功能代替耗时的数值仿真计算,将其嵌入到改进了控制参数取值方法的BBDE算法中建立了相应的反演程序,应用该程序对一个实际工程进行了弹塑性力学参数反演,并对反演程序的可行性和有效性进行了研究.结果表明:利用ε-SVR具有的BP神经网络所不可比拟的泛化推广能力,在保证反演精度的同时提高了反演效率;引入BBDE算法使得在减少算法控制参数的同时提高了解的全局收敛性和收敛速度;将反演所得参数输入数值仿真模型获得的测点计算增量位移与实测增量位移比较吻合,相对误差不超过10%.
To solve the bottleneck problems in forward calculation-inverse analysis method for geotechnical mechanic parameter identification,an approach of combining support vector regression machine（hereinafter refer to ε-SVR） with particles swarm-differential evolution hybrid optimization algorithm（hereinafter refer to BBDE） was investigated,and a program was developed in which ε-SVR,functioning as predicting incremental displacement in measuring point（in such way,time-consuming finite difference numerical calculations could be avoided）,was imbedded into BBDE main routine as a subroutine.Meanwhile,the method of determining reasonable value of control parameter in BBDE was improved in order to promoting the performance of algorithm.The feasibility and effectiveness of inverse analysis program were testified by applying it to identify elasto-plastic mechanic parameters in an engineering site.The results show that generalization ability of ε-SVR is far beyond that of BP neural network,using ε-SVR as a substitute for numerical calculations can improve the efficiency of inverse analysis significantly and ensure solving precision.BBDE performs very well in finding global best solution and speeding up convergence but has less control parameters than PSO and DE algorithm.Finite difference numerical calculation was executed again after inverse analysis,it shows that the calculated incremental displacements approximately agree with those of field measurement in all measuring points,and the fractional errors are less than 10%.
出处
《中国矿业大学学报》
EI
CAS
CSCD
北大核心
2011年第1期95-102,共8页
Journal of China University of Mining & Technology
基金
国家自然科学基金项目(50911130366,50979030)
关键词
支持向量回归机
粒子群-差分演化杂交算法
弹塑性
反演
support vector regression
particles swarm-differential evolution hybrid algorithm
elastoplastic
inverse analysis
