In the excavation of rock tunnels, rocks behind the face have to be known in advance, especially in the case that rocks like serpentinite, which would cause squeezing while tunneling, might exist. A long advancing boring from the working face gives such information with cuttings, but recognizing its rock type is difficult even for geologists. Accurate, simple rock-identification technique of cuttings helps to draw out a rational scheme beforehand and improve productivity of tunnel excavation. Recently, Convolutional Neural Network (CNN) using hyperspectral images of samples established a rock-identification technique where a rock is identified only with a picture taken by a special camera called “hyperspectral camera.” Here, we apply this technique to cuttings obtained by advancing boring for tunneling.
This study employed four kinds of cuttings samples obtained from a tunnel ground: greenstone 100% (G100), serpentinite 100% (S100), greenstone 90% + basalt 10%, and greenstone 90% + serpentinite 10%. Appearances of these samples are quite similar and are difficult to discriminate visually. Using hyperspectral images of the samples, two types of CNN models were trained; one is for G100 and S100 (Model A), and the other is for all of the four cuttings samples (Model B).
The validation accuracy of Model A was 99.40%, whereas that of Model B decreased to 73.10%. A mislabeling for a mixed cuttings sample by geologist would account for the low accuracy of Model B. The CNN model for mixed cutting samples should better be trained using hyperspectral images of artificially calibrated mixed samples. On the other hand, the high accuracy of Model A shows that this technique is already practical to cuttings purely composed of a single rock type. The CNN model developed in this study gives a correct, quick discrimination of serpentinite and greenstone with no professional technique, which will give a great help to avoid troubles and contribute to safety and profitability in tunnel excavation of the ground including serpentinite.
