Efficient method based on recurrent neural networks for pavement evenness detection

Abstract

Surface evenness is an essential indicator for pavement evaluation; however, there is a lack of economical, efficient, and end-to-end assessment methods. This study proposes a pavement evenness estimation framework based on advanced recurrent neural networks. First, the method uses a 3D acceleration sensor to acquire the unsprung mass acceleration signal. The network is trained using a combination of acceleration and velocity inputs. Second, the trained model outputs the power spectral density (PSD) value. The R2 of the results reached 0.9880, and the classification accuracies were all above 91.7%, as evaluated by the simulation tests. Transfer learning is then used to transfer the pre-learned knowledge between similar vehicles. Finally, experiments were conducted in the field to estimate the PSD and to categorize six different pre-selected roads. The experimental results demonstrate that the method is robust to speed. The proposed framework achieves a cost-effective and efficient detection of pavement evenness.