Deep learning-based soil compaction monitoring: A proof-of-concept study.

• A deep learning-based compaction monitoring technique for estimating soil stiffness is developed. • Labeled training data was obtained by numerical simulation of soil-vibratory roller interaction model. • The proposed method is validated by conducting a field experiment at an embankment construction site. • The experiment confirmed a strong positive correlation between the estimated and measured soil stiffness. The dynamic behavior of the vibratory drum of a soil compactor for earthworks is known to be affected by soil stiffness. Real-time monitoring techniques measuring the acceleration of vibratory drums have been widely used for soil compaction quality control; however, their accuracy can be affected by soil type and conditions. To resolve this problem, a novel deep learning-based technique is developed. The method allows the regression estimation of soil stiffness from vibration drum acceleration responses. By expanding the range of applicability and improving accuracy, the proposed method provides a more reliable and robust approach to evaluate soil compaction quality. To train the estimation model, numerous datasets of noise-free waveform data are numerically generated by solving the equations of motion of the mass–spring–damper system of a vibratory roller. To validate the effectiveness of the proposed technique, a field experiment is conducted. A good correlation between the estimated and measured values is demonstrated by the experimental results. The correlation coefficient is 0.790, indicating the high potential of the proposed method as a new real-time monitoring technique for soil compaction quality. [ABSTRACT FROM AUTHOR]