.png){kind=icon}
.png){kind=icon}
Return
Interpretable machine‑learning classification of TBM types: a comparative study of eleven algorithms using geological and project parameters
.png){kind=icon}
PRE
AI
E
D
DOI：10.1016/j.tust.2026.108016.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Selecting the correct Tunnel Boring Machine (TBM) type—Slurry, Earth Pressure Balance (EPB), Open‑face, Double Shield (DS), or Single Shield (SS)—is a critical decision that governs excavation performance, ground stability, and project cost in mechanized tunneling. This study presents a comparative, interpretable machine‑learning framework for TBM type classification using four design‑stage parameters: geological condition (nine classes), machine diameter, water head, and fine content. Eleven supervised classifiers—CART, C4.5, PART, OneR, Random Forest, Bagging, LightGBM, AdaBoost, XGBoost, K‑Nearest Neighbors, and Support Vector Machine—were evaluated using 10‑fold cross‑validation. Feature‑importance analysis identified the geological condition as the dominant predictor (100%), followed by diameter (33.9%), water head (19.4%), and fine content (14.6%). The CART model achieved the highest performance (accuracy 85.2%, precision 87.0%, recall 87.3%, F1‑score 86.8%, ROC‑AUC 0.964) and revealed interpretable diameter thresholds (≈6.4 m and ≈4.2 m) that distinguish Slurry from EPB TBMs under low water‑head and low fine‑content conditions—directly aligning with engineering practice. Misclassifications occurred primarily between Slurry and EPB types, reflecting their overlapping applicability in soft‑ground environments. By comparing eleven algorithms, this study demonstrates that rule‑based models (CART and PART) balance predictive accuracy with transparency, while the simple OneR model serves as a valid baseline. The proposed framework provides an auditable, data‑driven basis for TBM selection, directly linking geological taxonomy to machine configuration and supporting transparent, reproducible preliminary design. Practitioners can use the extracted decision rules and determinacy levels to rapidly identify when the baseline suffices, when one additional parameter is needed, or when full conditional models are required.
Keywords:
Tunnel boring machine
Machine learning
Decision tree
CART
Ground classification
TBM selection
Interpretability
Journal
IF:
7.4
Papers:
6.8K
Citations:
3.5W
Organization
Cited Papers