Data-Driven Inversion Model of Composite Dielectric Asphalt Pavements for Ground Penetrating Radar Scans

Abstract

The key to non-destructive density detection of newly constructed asphalt pavements using ground penetrating radar lies in measuring the dielectric constant. Deep learning-based dielectric constant inversion can significantly enhance the real-time performance and efficiency of density monitoring. This study presents a novel new pavement dielectric constant inversion network (NPDCI-Net), which utilizes an encoder-decoder architecture to perform real-time inversion of A-scan signals. The NPDCI-Net integrates an enhanced gated attention mechanism and a multi-scale fusion module to facilitate feature extraction. Forward simulations, accounting for material composition and structural characteristics, were applied to generate observation signals and construct a comprehensive training dataset. An optimized theoretical theory was utilized to calculate the true dielectric constant curves of composite material numerical models corresponding to the signals. Ablation experiments and test results revealed that the NPDCI-Net outperformed other classic and state-of-the-art models due to the effective introductions and modifications of the gated attention layers and multi-scale fusion module. Compared with other models, NPDCI-Net demonstrated significantly resistance to random noise and amplitude oscillations caused by the composite characteristics of asphalt mixtures. Finally, specially designed continuous equipment was employed to collect the observed signals from two different survey lines. Based on the electromagnetic mixing theory, NPDCI-Net's dielectric constant outcomes enabled accurate density prediction by comparing with the density of pavement core samples.

Keywords: Pavement construction, Density detection, Ground penetrating radar, Deep learning, Attention mechanism, Forward simulation.