An intelligent compaction-based framework for predicting and evaluating compaction quality of cement-treated base in airport pavements

Reliable compaction control of cement-treated base (CTB) layers is essential for airport pavement performance, yet conventional spot tests cannot resolve the continuous evolution and spatial heterogeneity of compaction quality. This paper establishes an intelligent-compaction measurement-to-control framework that converts roller vibration responses into interpretable and actionable construction criteria. Field measurements were used to compare four intelligent compaction measurement values, and the Vibrating Compaction Value (VCV) exhibited the strongest relationship with in situ compaction degree of R 2 =0.6758. A CNN–GRU–SE regression network was developed to capture nonlinear coupling among construction parameters and vibration feedback. On the present field dataset, the CNN-GRU-SE model achieved , MAE, and RMSE values of 0.909, 0.661, and 0.852, respectively, while simpler models such as MLR and RF also showed useful predictive capability ( =0.780 and 0.823). Compared with CNN-GRU, SE module further reduced MAE and RMSE by 20.2% and 18.4%, respectively. SHAP analysis identified compaction-pass number as the dominant variable, followed by VCV, and revealed physically consistent nonlinear effects associated with cumulative energy input and roller–layer interaction. Model inversion yielded a field-calibrated VCV threshold of 82.04 corresponding to the required 98% compaction degree. An entropy-weighted TOPSIS assessment integrating compaction adequacy, spatial uniformity and inter-pass stability further identified the sixth pass as the transition toward acceptable compaction with a stable-point rate of 91.33%, whereas the seventh and eighth passes reached 98.00% and 100.00% stability and all comprehensive scores exceeded 0.91. The framework advances compaction assessment from discrete post-construction inspection toward continuous, interpretable and spatially resolved quality control.