Semi-supervised ensemble model for TBM rock mass classification

Abstract

Rock mass classification is crucial for tunnel construction with tunnel boring machines (TBMs). Due to the limited measured rock data, existing machine learning-based classification methods often use empirically derived data to supplement the rock mass datasets for model training, which compromises the classification reliability. To address this issue, this study proposes a semi-supervised soft-voting ensemble (semi-supervised SVE) model for TBM rock mass classification, utilizing limited labeled data (i.e., data including measured rock mass grades) and extensive unlabeled data (i.e., data lacking measured rock mass grades). The model is initially pre-trained with limited labeled data, which include rock mass grades derived from a national code based on rock mass integrity and physical properties. The pre-trained model is then used to identify the rock mass grades of unlabeled data to produce pseudo labels. Multiple training iterations are conducted to incorporate the pseudo-labeled data, continuously expanding the training dataset of the proposed model. By integrating the boosting and bagging ensemble strategies through the soft voting method, the robustness of the proposed model in rock mass classification is enhanced. A high-quality dataset from the Yinchuo-Jiliao Project in China is used for illustration. The proposed model outperforms alternative supervised and unsupervised models on both labeled and unlabeled data. The superior performance on labeled data is directly evident from evaluation indices including precision, recall, and F1 score, whereas its effectiveness on unlabeled data is demonstrated indirectly through rock-breaking indices. A 5-fold random cross-validation shows that the soft-voting ensemble classifier is more robust than individual classifiers. In addition, the confidence level of the proposed model matches actual classification accuracy, providing useful insights into model uncertainty for decision-makers.