Adaptive neuro-fuzzy ensemble architecture with performance-driven weight allocation for geotechnical settlement prediction in transport infrastructure

Abstract

Accurate prognosis of ground settlement constitutes a critical and ongoing challenge in intelligent asset condition evaluation, directly influencing the operational safety and service lifespan of transport assets within digitally enabled urban ecosystems. Conventional prognostic methodologies—spanning deterministic continuum-mechanics simulations to simplified empirical curve-fitting—routinely fall short when addressing streaming geotechnical observations characterized by high-dimensional nonlinearity, marked spatiotemporal heterogeneity, and nonuniform sampling frequencies. This study presents a novel computational intelligence framework that integrates the transparent, rule-based interpretability of fuzzy logic with the autonomous nonlinear learning capability of artificial neural networks. The central innovation lies in an Adaptive Neuro-Fuzzy Inference System (ANFIS) reinforced by a performance-adaptive model amalgamation protocol, which continuously readjusts the contribution of each constituent predictor through rollinghorizon residual monitoring and reward-guided self-calibration heuristics. This autonomous weight-regulation scheme automatically regulates the influence of base models—such as sigmoidal saturation curves and exponential attenuation profiles—according to the progressive stages of soil consolidation, thereby achieving fully automated model selection without any manual tuning. The framework was deployed and rigorously validated on prolonged observational records gathered along a strategic expressway route penetrating rugged alpine topography. Comparative evaluations against contemporary machine-learning baselines—including recurrent neural networks with long-term memory gates, integrated autoregressive moving-average schemes, and a non-adaptive ANFIS variant—demonstrate that the proposed self-optimizing ensemble topology delivers substantially superior predictive accuracy and robustness, yielding a coefficient of determination (R²) of 0.95, a root-mean-square error (RMSE) of 15.35 mm, a mean absolute error (MAE) of 12.22 mm, and a mean absolute percentage error (MAPE) of 8.38%. By offering a powerful, entirely data-driven tool for proactive geotechnical integrity oversight and automated hazard mitigation, this methodology supplies reliable algorithmic support for automated decision-making during both the construction stage and the subsequent postconstruction monitoring phase.