ABSTRACT
The road layer modulus backcalculation based on the road deflection basin obtained by the Falling Weight Deflectometer is a key issue in road engineering. Traditional Falling Weight Deflectometer backcalculation method based on Artificial Neural Network has the disadvantages of poor generalisation ability and low convergence accuracy in terms of the dynamic modulus. In this paper, a hybrid neural network structure, combined with Residual Neural Network, Recurrent Neural Network and Wide & Deep (Abbreviated as ResRNN–W&D) structure, was proposed for Falling Weight Deflectometer deflection basin backcalculation. A case study using the United States Long-Term Pavement Performance database verified that the ResRNN–W&D structure can train Falling Weight Deflectometer data on multiple roads together and achieve fast and high-precision convergence, thereby greatly improving the availability of the multi-source heterogeneous data. Moreover, two transfer learning methods for the ResRNN–W&D structure were proposed to improve the divergence issue. It was found that the ResRNN–W&D structure has stronger generalisation ability than traditional Artificial Neural Network.
Acknowledgements
The authors would also like to acknowledge the financial support provided by Bureau of Transportation Construction of Jiangsu.
Disclosure statement
No potential conflict of interest was reported by the author(s).