Machine learning with monotonic constraint for geotechnical engineering applications: an example of slope stability prediction
Machine learning with monotonic constraint for geotechnical engineering applications: an example of slope stability prediction
复制标题
用于岩土工程应用的具有单调约束的机器学习：边坡稳定性预测的一个实例
DOI：
10.1007/s11440-023-02117-7
复制
发表时间：
2023-11
期刊：
影响因子：
5.7
通讯作者：
Te Pei;Tong Qiu
中科院分区：
文献类型：
--
作者：
Te Pei;Tong Qiu
由于数据科学的最新进展，机器学习（ML）算法已被广泛应用于分析岩土工程问题。然而，使用有限数据训练的灵活ML模型可能会表现出意想不到的行为，导致低可解释性和物理不一致性，从而降低了ML模型在风险预测和工程应用中的可靠性和鲁棒性。由于岩土工程应用的输入特征通常表示遵循内在且通常单调关系的物理参数，因此将单调性纳入ML模型有助于确保模型输出的物理真实性。在这项研究中，单调性作为一个软约束引入人工神经网络（ANN）模型，并与几个基准ML模型的结果进行了比较。在训练过程中，使用数据增强和逐点梯度来评估模型预测的单调性，并通过修改的损失函数来最小化单调性违规。从文献中的边坡稳定性的情况下的历史汇编用于模型开发，基准测试其性能，并评估单调性约束的影响。交叉验证程序用于所有模型性能评价，以减少样本选择中的偏倚。结果表明，无约束ML模型在输入空间的许多部分产生违反单调性的预测。然而，通过向ANN模型中添加单调性约束，在保持相对较高性能的同时有效地减少了单调性违反，从而提供了更鲁棒和可解释的预测。使用边坡稳定性预测作为代理，在这项研究中开发的方法，将单调性约束ML模型可以应用到许多岩土工程应用。所提出的方法增强了ML模型的可靠性和可解释性，从而为现实世界的应用程序带来更准确和一致的结果。
Machine learning (ML) algorithms have been widely applied to analyze geotechnical engineering problems due to recent advances in data science. However, flexible ML models trained with limited data can exhibit unexpected behaviors, leading to low interpretability and physical inconsistency, thus, reducing the reliability and robustness of ML models for risk forecasting and engineering applications. As input features for geotechnical engineering applications often represent physical parameters following intrinsic and often monotonic relationships, incorporating monotonicity into ML models can help ensure the physical realism of model outputs. In this study, monotonicity was introduced as a soft constraint into artificial neural network (ANN) models, and their results were compared with several benchmark ML models. During the training process, data augmentation and point-wise gradient were used to evaluate the monotonicity of model predictions, and monotonicity violations were minimized through a modified loss function. A compilation of slope stability case histories from the literature was used for model development, benchmarking their performance, and evaluating the effects of monotonicity constraints. Cross-validation procedures were used for all model performance evaluations to reduce bias in sample selections. Results showed that unconstrained ML models produced predictions that violate monotonicity in many parts of the input space. However, by adding monotonicity constraints into ANN models, monotonicity violations were effectively reduced while maintaining relatively high performance, thus providing a more robust and interpretable prediction. Using slope stability prediction as a proxy, the methods developed in this study to incorporate monotonicity constraints into ML models can be applied to many geotechnical engineering applications. The proposed approach enhances the reliability and interpretability of ML models, resulting in more accurate and consistent outcomes for real-world applications.