Abstract
To address the technical limitations of traditional pavement performance prediction models in capturing temporal features and analyzing multi-factor coupling, this study proposes a Bayesian Optimization Dual-Layer Feature Fusion Model (BO-DLFF). The framework integrates heterogeneous data streams from embedded strain sensors, temperature/humidity monitoring nodes, and weigh-in-motion (WIM) systems, combined with pavement distress detection and historical maintenance records. A dual-stage feature selection mechanism (BP-MIV/RF-RFECV) is developed to identify 12 critical predictors from multi-modal sensor measurements, effectively resolving dimensional conflicts between static structural parameters and dynamic operational data. The model architecture adopts a dual-layer fusion design: the lower layer captures statistical patterns and temporal-spatial dependencies from asynchronous sensor time-series through Local Cascade Ensemble (LCE) ensemble learning and improved TCN-Transformer networks; the upper layer implements feature fusion using a Stacking framework with logistic regression as the meta-learner. BO is introduced to simultaneously optimize network hyperparameters and feature fusion coefficients. The experimental results demonstrate that the model achieves a prediction accuracy of R(2) = 0.9292 on an 8-year observation dataset, effectively revealing the non-linear mapping relationship between the Pavement Condition Index (PCI) and multi-source heterogeneous features. The framework demonstrates particular efficacy in correlating high-frequency strain gauge responses with long-term performance degradation, providing mechanistic insights into pavement deterioration processes. This methodology advances infrastructure monitoring through the intelligent synthesis of IoT-enabled sensing systems and empirical inspection data.