(General session) Mining and underground construction machineries / Mining technologies

[3201-08-02] Deep Learning and Hyperspectral Image Optimised Lithology Classification from Tunnel Boring Machine

○Brian Bino Sinaice¹, Kursat Kilic¹, Narihiro Owada¹, Jun Abe¹, Natsuo Okada², Hajime Ikeda³, Hisatoshi Toriya¹, Tsuyoshi Adachi¹, Youhei Kawamura² (1. Akita University, 2. Hokkaido University, 3. National Institute of Technology, Asahikawa College)

Chairperson：吉光 奈奈（京都大学）

Keywords:Hyperspectral Imaging, Deep Learning, Lithology Classification, Convolutional Neural Network, Tunnel Boring Machine

The tunnel boring machine (TBM) is a sophisticated engineering tool that automates tunnel excavation. It features a rotating cutter head equipped with cutting tools capable of penetrating a variety of geological formations, from soft soil and clay to hard rock and abrasive materials. Consequently, lithology classification is a crucial factor in optimizing tunnelling performance and managing time constraints during excavation. However, the mechanization of TBM tunnelling reduces human presence at the tunnel face, complicating the identification of lithology during the process. To address this challenge, this paper introduces the use of hyperspectral camera data, processed through a deep learning convolutional neural network (CNN), to identify and classify lithology from the excavated cuttings, thereby enhancing tunnelling performance. The study focuses on a 118-meter highway bypass tunnelling project in Japan, where a micro tunnelling method was employed to install umbrella pipe supports. Excavated material was collected on a screw conveyor and sampled with each advance. The samples were captured using a 204-band hyperspectral camera and analysed with a CNN, revealing the geological compositions as sandy clay (0 – 40 m), sand (40 – 58 m), brown clay (58 – 66 m), blue clay (66 – 84 m), and sandy clay (84 – 118 m). By training the CNN through a 5-fold cross-validation, the CNN categorised these lithologies into four main classes, thereby achieving accuracy and average per-class precision exceeding 90%. These results demonstrate the potential of integrating hyperspectral imaging and deep learning technologies in TBM operations, providing a robust method for real-time lithology classification. This approach not only improves the efficiency and accuracy of tunnelling projects, it ensures better management of geological variations encountered during excavation.