# Dynamic Identification Method for Highway Subgrade Soil Compaction Based on Embedded Attitude Sensors

## TL;DR

This paper introduces a new method using embedded sensors to dynamically assess soil compaction in road construction, enabling real-time quality control.

## Contribution

A novel dynamic identification method for subgrade soil compaction using embedded attitude sensors and machine learning models.

## Key findings

- XGBoost model achieved an R2 exceeding 0.995 for high compaction levels.
- Yaw angle is most sensitive to vertical settlement, while pitch and roll angles provide lateral and rotational behavior insights.
- Transient masking interpolation outperforms traditional wavelet thresholding in preserving baseline trends.

## Abstract

Compaction quality is a critical factor in ensuring the long-term performance of subgrade structures; however, traditional testing methods are limited by their destructive nature and delayed feedback. To address these shortcomings, this study proposes a dynamic identification method for subgrade compaction based on embedded attitude sensors. A customized sensor unit integrated with an inertial measurement module was embedded in soil samples to record triaxial acceleration and attitude angles during the compaction process. Signal processing techniques, including an improved wavelet-based denoising strategy, were employed to separate long-term compaction trends from transient impact disturbances. Attitude features such as cumulative angular change, angular velocity, root mean square values, and a comprehensive inclination index were extracted as predictive variables. Ridge regression, random forest, and XGBoost models were constructed to establish the mapping relationship between attitude features and compaction degree. Experimental results on clay, loam, and sand samples indicate that the yaw angle is most sensitive to vertical settlement, while pitch and roll angles provide complementary information on lateral and rotational behaviors. Comparative analysis of filtering methods shows that the transient masking interpolation (TMI) approach outperforms the traditional asymmetric wavelet thresholding (AWT) method in effectively preserving baseline trends. Among the regression models, XGBoost demonstrated the best predictive performance, achieving an R2 exceeding 0.995 at high compaction levels. The proposed method has been experimentally demonstrated as a laboratory-scale proof of concept, showing strong potential for future real-time field application, offering a novel technological pathway for intelligent quality control in road construction.