In response to the problems of low efficiency in manual detection, poor reliability of sensor monitoring, and scarce label data in the wear monitoring of hobs of the full section tunnel boring machine, a prediction method based on graph Laplacian regularization (GLR) deep learning model was proposed. Based on a high-altitude tunnel excavation project, an efficient data preprocessing system was constructed: the excavation cycle dynamic extraction method was used to accurately identify and eliminate non excavation and empty push section data, the quartile method was used to eliminate outliers, the SG (Savitzky-Goloy) filtering for noise reduction was combined, then the data quality was improved. Furthermore, By integrating GLR and deep learning technologies, a data manifold structure was constructed using k-NN (k-nearest neighbor) graph. The Laplacian matrix was used to constrain the smoothness of adjacent sample prediction, generating pseudo-labels with high confidence to expand the training set, and combining with long short-term memory (LSTM), deep neural network (DNN) and convolutional neural networks (CNN) to develop three prediction models: GLR-LSTM, GLR-DNN, and GLR-CNN. The experimental results showed that the GLR-LSTM model had the best prediction performance. Compared with traditional small-sample machine learning methods such as ridge regression, support vector regression and gradient boosting regression tree, the prediction accuracy of the GLR-LSTM model improved significantly. This method can accurately predict the real-time wear rate of the hobs by only requiring the operating parameters such as the TBM cutterhead torque and the total propulsion force, providing technical support for reducing the opening inspection and optimizing maintenance decisions.