Prototype-guided learning for geological condition inference from TBM operational data

Accurate classification of surrounding rock masses is crucial for the safe and efficient operation of tunnel boring machines (TBMs), particularly in complex geological conditions. While traditional classification methods are widely used, their reliability in heterogeneous formations is limited by reliance on subjective field judgments and empirical indices. Although machine learning approaches utilizing TBM operational data offer a promising alternative, their performance is frequently compromised by label noise originating from inconsistent geological logs and engineering assessments. To address this fundamental challenge, this study proposes a novel weakly supervised learning framework that effectively mitigates label noise through domain-informed soft supervision, which converts uncertain labels into probabilistic targets using engineering knowledge, and class prototype learning, where representative feature vectors guide semantic clustering for each rock mass category. However, applying soft supervision directly to noisy, sequential TBM data may lead to unstable training. To further enhance robustness, we introduce a teacher-student self-distillation mechanism, where self-supervised representation learning captures structural regularities in TBM data, and the teacher model propagates high-confidence predictions to stabilize training and suppress noise amplification. Experimental results on real-world TBM tunneling data demonstrate a 31.6% reduction in label inconsistency compared to original annotations, along with significantly improved alignment with operator judgments. The framework demonstrates strong noise robustness, maintaining over 80% prediction consistency under 30% uniform or Gaussian label corruption, thereby providing a reliable solution for rock mass classification that enhances both safety and efficiency in TBM tunneling operations.