Machine-Learning-Based Screening of Relative Eccentric-Wear Severity in Shield TBM Disc Cutters Using 3D-Scan Morphology Labels

Disc-cutter wear in abrasive strata is spatially non-uniform, yet mean wear depth cannot indicate where it concentrates. This retrospective feasibility study develops a machine-learning-based condition-monitoring framework that predicts the relative severity of eccentric wear from engineering data available before inspection, supervised by 3D-scan morphology labels. An Eccentric-Wear Morphology Index (EWI) is constructed from post-replacement 3D morphology and used solely as a relative-severity label; its tertile-based grades are cohort-relative rather than universal engineering thresholds. The analysis cohort comprised 244 quality-controlled 19-inch cutter rings, and an engineering-prioritized redundancy review condensed 58 candidate variables into a frozen 22-variable set. In five-fold out-of-fold evaluation, the Random Forest achieved 0.779 accuracy, 0.775 macro F1, and 0.939 high-severity recall. With nested threshold selection, in which the operating threshold was chosen only within the training folds, the pooled held-out screening result reached 0.988 recall and 0.946 F2 while including 41.0% of the samples in the review pool, and this operating point was insensitive to false-negative-to-false-positive cost ratios between 5:1 and 15:1. Grouping both rings of each twin cutter into the same fold left the screening operating points essentially unchanged. The framework shows potential to support within-project inspection prioritization; external validation and calibration remain necessary because the screening signal is strongly associated with service exposure and the project-specific cutter-change schedule.