基于坡面位移的土质边坡潜在滑面智能预测

Intelligent prediction of potential sliding surface of soil slopes based on slope displacement

摘要

【目的】针对当前边坡潜在滑面预测分析难、精度低等现状，以实时坡面位移为基础，尝试借助神经网络实现土质边坡潜在滑面定量预测。【方法】基于随机潜在滑面生成原理，采用离散元程序3DEC模拟随机滑面下边坡失稳全过程；后结合数据样本和级联算法建立可预测潜在滑面的神经网络模型，并开展室内试验和工程案例双重分析和验证；最后建立边坡动态安全预警系统。【结果】室内试验结果显示：所建立的级联相关神经网络(简称“CC”神经网络)模型和反向传播前馈神经网络(简称“BP”神经网络)模型在一定程度上均能映射出坡面位移与潜在滑面之间的隐含关系；而相较于BP神经网络，由CC神经网络所预测的滑面实时曲线平均吻合度为0.973,同比增大0.07,且所预测滑面土体黏聚力的平均相对误差为14.99%,同比降低9.93%,内摩擦角的平均相对误差为10.12%,同比降低10.25%。采用CC神经网路模型开展实际工程应用表明：所预测的滑面实时曲线吻合度均大于0.99,而土体抗剪强度实时预测值的相对误差均小于14%,且预测精度整体呈上升趋势。【结论】结论表明：边坡潜在滑面的生成及稳定性状态可由坡面表观位移分布情况及发展趋势所进一步体现；分析室内边坡模型试验结果可知CC神经网络因其独有的自适应网络结构，在边坡滑面预测中体现出更高的精确度及更强的适应性；利用边坡坡面位移，搭配“人工智能-神经网络预测”技术，可短时间、高精度智能化预测边坡潜在滑面及其力学特性，进而实现便捷、高效且可靠的判断当前边坡可能发生的潜在失稳灾害。

Abstract

[Objective] In light of the current challenges in predicting potential sliding surfaces of slopes, such as high difficulty and low accuracy, this study attempts to quantitatively predict the potential sliding surface of soil slopes using neural networks, based on real-time slope displacement monitoring data. [Methods] Based on the principle of stochastic potential sliding surface generation, this study employs the discrete element program 3DEC to simulate the entire process of slope instability under randomly generated slip surfaces. Subsequently, a neural network model capable of predicting potential sliding surfaces is developed by integrating sample data and a cascading algorithm, which is then validated through both laboratory tests and real-world engineering case studies. Finally, a dynamic slope safety early-warning system is established. [Results] The indoor test result show that both the established cascade correlation neural network(referred to as the “CC” neural network) model and the backpropagation feedforward neural network(referred to as the “BP” neural network) model can, to some extent, map the implicit relationship between slope displacement and the potential sliding surface. Compared to the BP neural network, the average consistency of the real-time sliding surface curve predicted by the CC neural network is 0.973, by comparison increase of 0.07. Moreover, the average relative error of the predicted cohesive force of the sliding surface soil is 14.99%, by comparison decrease of 9.93%, and the average relative error of the internal friction angle is 10.12%, by comparison decrease of 10.25%. The application of the CC neural network model in practical engineering projects shows that the consistency of the predicted real-time sliding surface curves is greater than 0.99, while the relative error of the real-time predicted values of soil shear strength is less than 14%, with the overall prediction accuracy exhibiting an upward trend. [Conclusion] The result indicate that the formation and stability state of the potential sliding surface in a slope can be further reflected by the distribution and evolutionary trend of surface displacement. Analysis of laboratory slope model tests shows that the CC neural network, owing to its unique self-adaptive architecture, achieves higher accuracy and stronger adaptability in predicting slope sliding surfaces. By integrating slope surface displacement with AI-powered neural network prediction technology, it is possible to rapidly and accurately intelligently predict potential sliding surfaces and their mechanical properties. This approach enables convenient, efficient, and reliable assessment of potential instability hazards in slopes.

关键词

实时坡面位移 / 边坡滑面预测 / 离散元程序 / 神经网络模型 / 室内试验 / 安全预警 / 抗剪强度 / 影响因素

Key words

real-time slope displacement / slope slip surface prediction / discrete element method(DEM) program / neural network model / laboratory test / safety early warning / shear strength / influencing factors

基金资助

国家自然科学基金项目(51178402)

四川省自然科学基金(2025ZNSFSC0305)