Predicting tunnel boring machine penetration rates in rock masses using knowledge distillation with limited samples

Accurate prediction of Penetration Rate (PR) in Tunnel Boring Machine (TBM) construction helps optimize tunneling parameters. In this paper, a limited-sample prediction model was proposed to predict PR based on Knowledge Distillation (KD). A dataset was utilized consisting of 151 samples and six rock parameters. The Pearson Correlation Coefficient (PCC) Method is used to select the features of the input network, and the highly correlated features were removed. The prediction results demonstrate that for the machine learning (ML) models, Long Short-Term Memory (LSTM) and Bi-directional LSTM-Attention (BiLSTMA), the prediction accuracy is not satisfactory due to the inability to overcome overfitting caused by limited samples. However, when LSTM is employed as the student network within the KD model, the Mean Absolute Percentage Error (MAPE) of LSTM is reduced by 1.05%, and the determination coefficient ( R2 ) value is increased from 0.7884 to 0.8191. This improvement is because the KD technique enables the student network to not only learn the feature information of the original data but also obtain the implicit information from the teacher network. Therefore, the proposed model provides a reliable technical solution for practical engineering applications under limited samples.