Uncertainty-informed large vision model for automated recognition of excavated rock chips in tunnel boring machines

Real-time identification of rock chip size and shape parameters from muck images plays a critical role in intelligently optimizing cutterhead thrust and torque parameters for tunnel boring machines (TBM). However, complex light environments in field images are difficult to recognize via traditional methods. This paper proposes a U-Net-SAM framework integrating semantic segmentation and the vision foundation model-Segment Anything Model (SAM), combined with dropout-based uncertainty analysis, achieving efficient rock chip segmentation and parameter quantification. First, a U-Net is trained to identify the rock mass centroid as an automatic SAM prompt. Next, an overlap region optimization strategy based on Intersection over Union (IoU) and a noise filtering method is employed to tackle boundary blurring and particle adhesion. Finally, a Dropout layer is added to implement the committee-based uncertainty analysis model and quantify predictive uncertainty. Results show that: (1) U-Net-SAM improves mean F1-score and PA by 9.1% and 7.8% over U-Net; (2) A strong correlation between prediction SD and error rate validates the proposed uncertainty quantification strategy. This framework provides reliable rock chip perception for intelligent TBM tunneling, with potential applications in other engineering scenarios.