Enhancing obstacle detectability in tunnel look-ahead systems through human-centered explainable AI and inverse inference

Yanlong Niu

doi:10.1016/j.tust.2025.107319

Enhancing obstacle detectability in tunnel look-ahead systems through human-centered explainable AI and inverse inference

2025-11-27

0

PRE

AI

Y

Yanlong Niu

H

Huaigu Tang

E

Enhedelihai Nilot *

Y

Yuhan Zhou

Y

Yunyue Elita Li

S

Siau Chen Chian

DOI：10.1016/j.tust.2025.107319

Original

Original request for help

Save

Abstract

En 中文

Tunnel Boring Machines (TBMs) increasingly use tunnel look-ahead technology to predict geological conditions and safeguard excavation. Seismic look-ahead methods are especially promising because they infer ahead-of-face obstacles from reflected wavefields recorded on tunnel sidewalls, reducing the need for pre-drilling. The central practical question, however, is: under what conditions are obstacles detectable, and how can detectability be enhanced? To address this, we develop a human-centered AI framework that quantifies detectability and provides actionable guidance for adjusting acquisition parameters. The workflow consists of: (1) physics-based simulation of synthetic shot gathers with semblance analysis and expert-validated, rule-based labeling into four detectability classes; (2) training an Extreme Gradient Boosting (XGBoost) classifier to predict the class; (3) explainable AI via SHAP (Shapley Additive Explanations) to quantify global and per-case feature contributions; and (4) an inverse-inference procedure that proposes constraint-compliant adjustments to controllable parameters to enhance detectability. Across 1250 cases, the classifier attains high accuracy on previously unseen combinations. SHAP consistently ranks shear-wave impedance contrast, detection distance, and obstacle geometry as the dominant factors, with density and P-wave impedance playing secondary roles. Joint SHAP analysis recommends a primary operating range of detection distance < 70 m with source frequency 380–420 Hz, and a secondary range of 70–82 m with frequency < 400 Hz, balancing resolution for typical tunnel environments. By integrating explainable machine learning with numerical simulation, the framework makes detectability assessment transparent, supports pre-excavation acquisition design, and enables rapid, defensible adjustments during tunneling—advancing safer and more reliable AI-assisted TBM operations.

Enhancing obstacle detectability in tunnel look-ahead systems through human-centered explainable AI and inverse inference

Abstract

Journal

Organization