Abstract
The highway subgrade is the foundation for ensuring the safe operation of highways, and subgrade compaction degree is a crucial parameter in subgrade construction. This study employs field tests to investigate subgrade compaction degree, analyzing the effects of rolling passes, rolling speed, and moisture content of subgrade fill material on subgrade compaction degree. Subsequently, four machine learning models are used to establish prediction models for subgrade compaction degree: random forest model, sparrow search algorithm optimized random forest model (SSA-RF), AdaBoost model, and BP-AdaBoost model. The results indicate that: (1) Subgrade compaction degree increases with the number of rolling passes, decreases with rolling speed, and is greater when the moisture content of the subgrade fill material is closer to the optimum moisture content. (2) The AdaBoost model shows the poorest prediction performance and is not suitable for predicting subgrade compaction degree, whereas the SSA-RF model achieves the best training results, with correlation coefficients ($R^2$) of 0.98 and 0.94 for the training and testing sets, respectively, closely matching the actual results, validating the feasibility of this method for predicting highway subgrade compaction degree.
Introduction
The "subgrade"—the soil foundation beneath a highway—is the most critical element for road longevity and safety. If it isn't compacted correctly, the road will eventually settle, crack, or fail under heavy loads. Currently, engineers rely on "in situ" tests like the sand cone or ring knife methods to check compaction. While accurate, these methods are time-consuming, labor-intensive, and only provide a snapshot of a specific spot after the work is done.
This research moves highway construction away from slow, manual tests toward real-time predictive modeling. By combining field experimental data with advanced machine learning, the study aims to create a "digital twin" approach: a predictive model that allows engineers to understand how variables like roller speed and soil moisture will affect the final density before or during the construction process.
Experimental Design and Methodology
The researchers conducted 27 controlled field experiments on a 270-meter segment of a highway (K8 + 931 to K9 + 201). The study focused on three primary input variables:
- Rolling passes: 6, 8, and 10 passes.
- Rolling speeds: 3, 4.5, and 6 km/h.
- Moisture content: 10.4%, 12.4%, and 14.4%.
The output variable was the compaction degree (%) measured via traditional sand cone tests. To find the most accurate predictor, four machine learning architectures were compared:
- Random Forest (RF): An ensemble of decision trees.
- SSA-RF: A Random Forest model optimized using the Sparrow Search Algorithm, a bio-inspired technique that finds the "global optimum" settings for model hyperparameters.
- AdaBoost: A sequential ensemble model that corrects errors from previous "weak" learners.
- BP-AdaBoost: A hybrid approach using a Backpropagation (BP) Neural Network as the base learner within the AdaBoost framework.
Key Drivers of Compaction
The analysis revealed distinct relationships between construction parameters and the resulting soil density.
As shown in the correlation analysis, moisture content is the dominant factor influencing compaction. If the soil is too dry, friction between particles prevents them from settling; if it's too wet, water displaces the soil particles. The "sweet spot" identified in this study was 12.4%.
Furthermore, there is a clear negative correlation between rolling speed and compaction. Moving the roller too fast reduces the time available for vibrational energy to rearrange soil particles effectively. Conversely, increasing the number of passes consistently improves compaction, though with diminishing returns as the soil reaches its maximum density.
Model Performance and Results
The SSA-RF model emerged as the superior predictive tool. By using the Sparrow Search Algorithm to fine-tune the Random Forest hyperparameters, the researchers significantly reduced error rates compared to standard models.
- Accuracy: The SSA-RF model achieved a correlation coefficient ($R^2$) of 0.98 on the training set and 0.94 on the testing set.
- Error Metrics: It maintained the lowest Mean Absolute Error (0.11) and Mean Absolute Percentage Error (0.12).
- Peak Performance: The highest compaction recorded was 98.89%, achieved with 10 passes at 3 km/h and 12.4% moisture.
In contrast, the standard AdaBoost model performed poorly, suggesting it may be too sensitive to the noise or small sample sizes typical of field construction data.
Conclusion
This study demonstrates that the SSA-RF model is a highly effective tool for predicting subgrade compaction degree. While traditional methods remain the gold standard for final verification, predictive modeling offers a path toward "intelligent compaction," allowing for real-time adjustments during construction.
Future work should focus on expanding the dataset to include diverse soil types and environmental conditions, as geotechnical properties vary significantly across different regions. Developing a more universal model would enable widespread deployment of this technology across various highway projects.
Summary updated