Our group (Renjie Wu and Yuji Fujita) published a paper Integrating domain knowledge with deep learning models: An interpretable AI system for automatic work progress identification of NATM tunnels on Tunnelling and Underground Space Technology.
Abstract:
Finding a reliable and cost-effective approach to monitor the activities of the New Austrian Tunneling Method (NATM) tunnel construction automatically is a challenging yet important task. This study presents an interpretable artificial intelligence (AI) framework that automatically identifies NATM construction works using low-cost site surveillance images. The framework adopts the Bayesian statistics to combine the prior NATM construction knowledge with the visual evidence extracted by deep learning (DL) based computer vision models. The analysis results of Site CCTV surveillance videos of four NATM tunneling projects are presented to demonstrate its ability (i) to label NATM work cycles from the work timeline, (ii) to identify NATM work categories inside each work cycle, and (iii) to estimate the degree of plan-work deviation at the construction cycle level. The proposed framework yields promising results on a real NATM tunneling project.
Figure 1: Framework of the proposed system.
This work is also presented on Workshop: Recent Advances in Tunneling (Tongji- SFB837-UCC- ACTUE-UC Berkeley). See video below for our presentation.