基于多场信息数据驱动的滑坡演化多模式切换概率预测和控制研究

With the complexity of the geologic environments in the Three Gorges region, frequent landslide hazards are serious threats to people's life and property. Due to the complicated disaster-pregnant mechanisms of landslide, the various environmental factors and action processes, precise mechanism model of landslide can hardly be established by traditional engineering geological methods. With the fast development of landslide monitoring technologies, it's urgent to develop new landslide data mining and system modeling methods which can effectively analyze the landslide multi-field information big data. This project will focus on several typical kinds of landslides in the Three Gorges region. Advanced computational intelligence will be employed to establish multi-pattern switched probabilistic predictive control models of landslide evolution based on multi-field information data-driven methods. We will create more reasonable adaptive methods to divide the evolutional state of landslide using co-training style semi-supervised learning and imbalanced data classification approaches. To quantify the uncertainty of landslide prediction, new fast learning algorithms of neural networks with random weights will be developed to establish a probabilistic prediction model which will be switched based on the evolutional state of landslide. And then, we will generalize the landslide point probabilistic prediction model to a landslide surface probabilistic prediction model. Through the analysis of the different disaster-pregnant mechanisms and key control factors in the different evolutional states of landslide, we will lay a theoretical foundation for landslide process control based on different evolutional stages of landslide. In summary, the research of this project will play an important role in the fields of landslide multi-field information data mining, switched prediction, process control theory, etc.

我国三峡库区地质环境复杂，滑坡灾害多发且严重威胁库区人民生命财产安全。滑坡孕灾机理复杂，外界环境影响因素及作用过程多样，传统的工程地质方法难以建立精确的滑坡机理模型。随着滑坡监测技术的不断提高，急需发展新的滑坡数据挖掘和系统建模方法来有效地分析滑坡多场信息大数据。本项目以三峡库区典型滑坡为研究对象，发展先进的计算智能方法建立多场信息数据驱动的滑坡演化多模式切换概率预测控制模型。利用协同半监督学习和不平衡数据分类技术，试图创建更加合理的滑坡演化状态自适应划分方法。发展新的随机权值神经网络快速学习算法，建立依滑坡演化状态切换的概率预测模型，量化滑坡预测不确定性。进一步研究将滑坡点概率预测模型推广到滑坡面概率预测模型。分析滑坡不同演化状态下的不同孕灾模式和关键控制因子，初步建立滑坡演化分阶段过程控制体系。本课题的研究将对滑坡多场信息数据挖掘、切换预测、过程控制理论等研究产生一定的推动作用。

三峡库区滑坡数量多、规模大、密度高，严重威胁库区人民的生命和财产安全，因此，对滑坡演化趋势的预测和控制研究具有十分重要的意义。近年来，随着大量滑坡多场信息监测数据的获取，亟需发展基于先进机器学习方法的滑坡多场信息数据挖掘分析技术。本项目以三峡库区的多个滑坡为研究对象，从多场信息数据模式分析、分阶段切换的滑坡概率预测、滑坡演化分阶段过程控制三个方面进行了研究。具体地，重点研究了渗流场（降雨、三峡库区水位波动）与滑坡位移场之间的关系，设计了相应的多模式分析方法；提出了度量滑坡位移预测不确定性的概率预测模型，研究了基于神经网络的区间预测方法；针对滑坡突变点预测的难点，提出了一种基于不平衡数据分类器的切换预测方法；对数据驱动的滑坡预测控制进行了初步的探索。本项目的研究对多场信息数据融合方法、 滑坡演化阶段判别体系、 滑坡预测预报、 滑坡过程控制理论的发展起到一定的促进作用。