基于不同核函数跨海大桥变形高斯过程预测精度分析

Accuracy Analysis of Sea-crossing Bridge Deformation Based Prediction on Gauss Process with Different Kernel Functions

跨海大桥处于海洋复杂运动环境中,受到海浪、飓风等因素的扰动,常规桥梁变形时间序列模型不能全面反映其变形的内在驱动性.选取了高斯过程回归模型,根据青岛胶州湾跨海大桥某处桥墩30期挠曲变形实测数据,构建训练样本,通过训练样本获得跨海大桥变形的先验参数,对测试样本进行预测.为了避开跨海大桥变形模型非线性映射函数形式表达式及"高维数"等技术难题,引入了SE、NN、RQ三种单一核函数及SE、RQ形成的组合核函数,对这四种核函数分别进行变形值的高斯过程回归,并计算各种核函数模型的预测值相对误差,获得了跨海大桥变形值最优核函数预测值.同时,基于核函数高斯过程回归,实现了变形值间的非线性映射,解决了复杂模式预测问题.

A sea-crossing bridge locates in the complex marine environment,and is disturbed by ocean waves,hurricanes and other factors.The conventional bridge deformation time series model cannot fully reflect the inherent driving force of the deformation.The Gaussian process regression model is selected to predict the sea-crossing bridge deformation.According to the 30 th-stage measured flexure deformation of a certain pier of Qingdao Jiaozhou Bay Sea-Crossing Bridge,the training samples are constructed to obtain the prior parameters of the deformation,and then the test samples are predicted.Three single kernel functions SE,NN and RQ and a combined kernel function formed by SE and RQ are adopted to avoid the technical problems such as the nonlinear characteristic and ＂high dimension＂ of the sea-crossing bridge deformation model.The Gaussian process regression of these four kernel functions are carried out respectively,and the optimal kernel function is obtained based on the relative errors of predicted deformation values of each kernel function model.Based on the kernel function Gaussian process regression,the nonlinear mapping is achieved between deformation values and the complex prediction problem is solved.