Rock mass grade recognition of TBM tunnel based on multi-feature extraction, compression, and LightGBM

Currently, the research on the TBM tunnel rock mass grade recognition by using tunneling parameters rarely mines the geological information contained in the detailed distributions and relative changes of tunneling parameters. This research proposes a novel rock mass grade recognition method based on multi-feature extraction, compression, and LightGBM. Firstly, five key tunneling parameters are selected. Then, the collected data is truncated into data segments using the sliding window. Thirdly, multiple features are extracted from the data segments, including statistical, distribution, and correlation features, and these high-dimensional features are compressed using locally linear embedding. Finally, a mapping model between the compressed features and rock mass grades is established using LightGBM. The results show that compared to the commonly adopted method of directly using tunneling parameters as the model’s inputs, the proposed multi-feature extraction and compression method can improve the model’s performance by 12.4 %. Moreover, the LightGBM model performs the best with an F1 score of 0.851, which is 16.7 %, 5.4 %, 4.4 %, 1.2 %, 3.1 %, and 2.0 % higher than the CART, SVM, KNN, random forest, AdaBoost, and XGBoost models, respectively. Therefore, the proposed method can provide guidance for the safe and efficient construction of TBM.