Deep learning-based inversion of soil and rock compaction density utilizing Falling Weight Deflectometer (FWD) and Surface Waves

Abstract

Listen

Translate article

Methods relying on a single dynamic parameter to estimate the material density for the compaction quality testing of earth-rock dams and roadbed projects, present certain limitations in terms of applicability, testing accuracy, and work efficiency. The theory of elastic wave propagation reveals that the compaction density of the medium results from the combined interaction of multiple material properties. This study proposes a method that utilizes the transient surface wave method and a vehicle-mounted Falling Weight Deflectometer (FWD) to jointly enhance the accuracy of soil and rock material compaction detection. The transient surface wave method is employed to extract surface wave velocity and compressional wave velocity. The analysis of the FWD signal involves extracting dynamic parameters closely related to material density, including stiffness coefficients, central frequencies, and stiffness impedances, from time-domain, frequency-domain, and mechanical impedance analyses. A fully connected deep neural network is introduced to intelligently estimate the compaction density. The deep learning model is optimized by selecting the optimal activation function and optimization algorithm, as well as additional tuning. Extensive testing shows that deep learning model produce results closely approximating the true compaction density values. The average relative errors for the wet density and the dry density are 2.9% and 2.85%, respectively, meeting the quality control requirements for density testing. The proposed approach enables rapid detection and control of the compaction quality of soil and rock materials, providing a new method for the in-situ rapid detection of compaction quality.