Causality-Guided Explainable Deep-Learning Model for Managing Tunnel-Induced Ground Settlement Risks

The use of deep learning (DL) has been growing in tunnel-induced ground settlement risk modeling, eliminating the necessity for extensive prior risk management knowledge. Despite the success of deploying a DL model to predict risk, challenges prevail. (1) DL requires high-quality data, which is expensive and time-consuming to prepare, and (2) DL is a “black-box,” which is difficult to understand and interpret. In this instance, we address the following question in this paper: How can we accurately predict ground settlement with limited monitoring data using DL and concurrently provide effective explanations for the generated results? We propose a new DL approach combining explainable techniques to improve ground settlement risk modeling accuracy and explainability. Our approach comprises the following: (1) an interval type-2 fuzzy system to process limited data and improve its usability; (2) a novel causal-based feature selection to determine input parameters that have strong causal effects with ground settlement risk; and (3) a soft-type attention module to evaluate and allocate feature importance to inputs, guiding neural networks to concentrate on learning features in a more targeted manner. A case study is used to validate the feasibility and effectiveness of our approach, demonstrating its superiority in predicting ground settlement risk with a high degree of robustness. We suggest that our approach can help decision makers better understand the “how” and “what” of DL-produced outputs, improving decision making associated with managing safety in construction.