Abstract: Accurate and real-time classification of surrounding rock during TBM tunneling is essential for construction safety and efficiency.TBM operational parameters offer feasible input for surrounding rock classification,but the accuracy of traditional supervised machine learning models declines when training data contain noisy labels.To address this limitation,an Adaptive Confidence Threshold Mean Teacher model (ACT-MT) is proposed by integrating a confidence-based noise-label filtering strategy with a semi-supervised Mean Teacher framework.This method identifies mislabeled data using adaptive confidence thresholds,ensuring high-confidence pseudo-labels across all classes.The approach improves classification accuracy and computational efficiency when only a small portion of labeled data is available.Applied to a real-world TBM tunneling project,ACT-MT maintains higher accuracy than the standard Mean Teacher model as the proportion of labeled data decreases below 50%.When 1%~30% label noise is artificially introduced into the dataset,ACT-MT consistently outperforms conventional classifiers,preserving accuracy across noise levels.The model effectively mitigates the impact of label noise and reduces the need for extensive manual annotation.