基于多尺度卷积传递模型拓展的水库型滑坡智能同化方法研究

Systematic analysis of landslide has been seriously hampered by the great disparity between landslide monitoring data and test data, especially between displacement data and creep data, due to data magnitude, types and scales, and difficulties in combined utilization. Taking the monitoring data and test data of several typical reservoir landslides in the Three Gorges Reservoir Area as the study object, a landslide engineering geological model is primarily constructed to analyze the time-space coupling relationship between seepage field, stress field and strain field. A multi-scale convolution-transfer model is designed to link strain to deformation and displacement, and to extend the analysis of landslide response mechanism from mesoscopic analysis to macro-meso joint analysis. Subsequently a landslide dynamic process model is established to describe the whole process of geotechnical properties, stress, deformation, and displacement. An intelligent asynchronous assimilation algorithm is designed to exploit the observation data collected by different acquisition frequencies, to estimate the key parameters combined with the results of process model, and to optimize the process model dynamically through a decision-making and feedback correction according to geological regularities. The landslide intelligent assimilation method extended by the multi-scale convolution-transfer model extends the application of displacement monitoring data collected at macroscopic scale to the response mechanism analysis of landslide to water fluctuation and rainfall at mesoscopic scale. This method provides ideas and methods for comprehensive evaluation, analysis and prediction of landslide with high precision and informatization.

滑坡的监测数据和试验数据，特别是位移数据与蠕变数据，具有数据量悬殊、类型尺度各异、难以联合使用等特征，严重制约了滑坡的系统性分析。本项目以三峡库区多个典型水库型滑坡的监测数据与试验数据为研究对象，构建时空维度的滑坡工程地质模型，分析渗流场-应力场-应变场的耦合关系。设计多尺度卷积传递模型将应变和形变、位移进行联结，将滑坡响应机制的分析从细观拓展为宏细观跨尺度联合分析。以此两模型为内核构建起岩土性质-受力-形变-位移全过程的滑坡动态过程模型。设计智能异步同化算法对不同采集周期的监测数据进行反演学习，结合过程模型背景场对关键参量进行估计，并依据滑坡地质规律进行决策校验和反馈修正，动态优化过程模型。多尺度卷积传递模型拓展的滑坡智能同化方法将宏观尺度下的位移数据拓展应用到细观尺度下的滑坡对降雨库水的响应机理分析中，为构建高精度量化和信息化的综合性滑坡评价、分析、预测等提供新的思路和方法。

滑坡监测数据和试验数据的数据量级多样，类型尺度各异、相互关系复杂，严重制约了滑坡系统性分析。此外，如何将多源数据进行联合使用，也是国内外学者研究这类问题时的难点与挑战。项目组针对以上问题进行了深入研究。以三峡库区多个典型水库型滑坡的监测数据与试验数据为研究对象，在地质规律的约束下，构建时空维度的滑坡水动力模型、物理力学模型和本构模型，分析并揭示渗流场-应力场的时空耦合关系；设计多尺度卷积传递模型根据不同深度位移设计位 移量与形变量的传递并构造系列卷积模式综合蠕变特性的渐变叠加效果实现蠕变量-形变量的关联，继而构建起岩土性质-受力-形变-位移全过程的滑坡动态过程模型；首次提出并设计了以粒子滤波和智能算法为核心的智能多频同化算法对多频异步观测数据进行反演学习，结合过程模型背景场 对关键参量进行估计，并依据滑坡地质规律进行决策校验和反馈修正，动态优化过程模型，完成滑坡信息同化系统。项目组根据研究成果发表了多篇SCI、EI等国际高水准论文和发明专利，被国内外同行大量引用。