An efficient determination of the geological characteristics and soil-rock type ahead of a tunnel face is critical for adjusting construction parameters during shield tunnelling. In general, operational engineers rely on visual observations of mucky soil types from belt conveyors. This results in shield halting and involves both time and cost implications. This paper proposes a deep learning approach designed to identify mucky soil by monitoring a video installed on the strut of a belt conveyer. The proposed approach comprises four steps: (1) image acquisition, (2) enhanced you-only-look-once (YOLO) modelling, (3) model performance evaluation, and (4) soil identification based on an optimal analysis. The enhanced YOLO model is a deep image detection algorithm. It was introduced by integrating two innovative strategies: data augmentation and imbalance learning. This enhancement accelerates the speed of image identification and improves the overall classification performance. A case study of shield tunnelling in the soil-rock mixed strata of the Guangzhou–Foshan intercity railway line was conducted to validate the proposed approach. The results indicate that the enhanced YOLO model achieves a classification performance comparable to that of the highly optimised AlexNet and GoogleNet. Additionally, the proposed approach more effectively detects the muck soil content than manual observation. This demonstrates its potential for real-time applications in shield tunnelling operations.
Read full abstract