Characterizing the as-encountered ground condition with tunnel boring machine data using semi-supervised learning

Knowing the as-encountered geology during mechanized TBM tunneling is important for machine operation and contractual dispute resolution. However, ground condition estimated before excavation can be locally highly inaccurate. One solution is to characterize the ground directly from tunnel boring machine (TBM) data with data-driven models, assuming there exists unique relationships between TBM behavior and the changing ground. Only supervised-learning models trained exclusively with TBM data obtained near boreholes, where the true geology is known, have been used. Due to borehole sparsity, such training data is limited in quantity, especially at the beginning of excavation, undermining the model performance and delaying its deployment during tunneling. This paper proposes a semi-supervised learning model capable of using TBM data both with and without borehole information, relying on the principle of "guilt by association". Using the Seattle Northgate Link Extension tunneling project for case study, the proposed model is shown to outperform popular SL models - multilayer perceptron, multinomial logistic regression, random forest and sparse autoencoder neural networks - both in fitting the borehole data, as well as in inferring the heterogeneous ground profile with many irregular geological interbeddings and shifts of geological transitions identified. The proposed SSL model performs significantly better than SL models when given limited boreholes for training, making it more effective for use during tunneling.