Multi-Source Operational Feature-Driven Cutterhead Torque Prediction in Shield Tunnelling Using an IALA-Optimized Fuzzy Ensemble Deep RVFL Model

Abstract

Cutterhead driving torque is the primary load indicator of earth-pressure-balance shield machines, yet its dependence on strongly coupled multi-source operating parameters limits the reliability of empirical formulations. This study proposes IALA-edRVFL-FIS-Reg, a fuzzy ensemble deep random vector functional link regression model optimized by an improved artificial lemming algorithm (IALA). The base learner maps continuous operating parameters into fuzzy-state features through a Gaussian-membership Sugeno inference layer, propagates the concatenated raw and fuzzified inputs through stacked randomized hidden layers with direct input links, and obtains layer-wise output weights by regularized closed-form least squares before ensembling, thereby combining fuzzy-state representation with deep random feature mapping without gradient back-propagation. Distinct from the standard ALA, IALA introduces three explicitly defined mechanisms: an error-feedback exploration–exploitation transition factor normalized by the initial-population loss, which replaces the fixed energy factor; an adaptive step size coupling sigmoid error-gating with cosine annealing to preserve jumping capability while refining local search; and a stagnation-counter-triggered directional-disturbance jump for escaping local optima. Using 48,646 valid tunnelling records from 301 rings of Beijing Metro Line 22 and 65 raw and mechanism-based engineered features, the model attains R2 = 0.9555, RMSE = 382.52 kN·m, MAE = 302.95 kN·m and MAPE = 9.18%, outperforming eleven benchmarks on a ring-disjoint holdout, previously unseen rings of the same section, IALA yields an R2 gain of 0.0104 over ALA.