基于高斯过程回归的变形智能预测模型及应用

Deformation Intelligent Prediction Model Based on Gaussian Process Regressionand Application

岩体或建构筑物的变形通常具有复杂性和非线性等特性,一般的回归模型难以精确地进行回归预测,应用高斯过程回归理论对变形监测数据呈现出的非线性特征进行时间序列分析。考虑到监测数据的不断更新和累积,以及超参数与样本集的适应性,首先研究了＂递进-截尾式＂超参数自动更新模式和训练样本集的选择方法;在此基础上构建了以时间作为输入项的高斯过程回归变形智能预测模型（GPR-TIPM）;将该模型应用于矿山边坡监测点非线性时间序列分析中,通过分析变形趋势,最终采用Matérn 32和平方指数协方差函数相加的方式进行核函数组合。实验结果表明,采用组合核函数的预测性能较单一核函数有所改善,该方法提高了模型的泛化能力,GPR-TIPM模型在短期内的预测效果较理想。

The deformation of building structures or rock mass usually has the features of complexity and nonlinearity that the general regression model cannot accurately predict.In this paper,gaussian process regression（GPR）theory is applied in time series analysis of nonlinear deformation monitoring data.Considering the unceasing updates and massive accumulation of monitoring data,the hyper-parameter and the adaptability of sample set,a ＂progressive^truncation type＂hyper-parameter automatic update mode and selection method for training sample set was developed.On this basis,a GPR time-driven deformation intelligent prediction model（GPR-TIPM）was constructed.This model was applied to the nonlinear time series analysis of monitoring points on a mine slope.By analyzing the deformation trend,a composite kernel optimization method including the＂Matérn32＂and square exponential covariance kernel function is proposed.The experimental results showed that the prediction performance of the combined kernel function is better than that of the single kernel function,and improved the generalization ability of the model The prediction effect of GPR-TIPM model is better in the short term.