Rutting depth prediction model for asphalt pavements based on a dual branch spatiotemporal attention network

Figures (8) Tables (6)

Figure 1.
Longitudinal section of typical structure.
Figure 2.
DSAN model overall architecture diagram.
Figure 3.
DualTCN temporal block structure.
Figure 4.
Comparison of observed and predicted values.
Figure 5.
Effect of scatter diagram of each model. (a) DSAN. (b) LSTM. (c) LSTM-Attn. (d) TCN_LSTM.
Figure 6.
Comparison of the rutting depth of asphalt structures with different thicknesses over time.
Figure 7.
Absolute value distribution of rutting depth variation of asphalt structures with different thicknesses.
Figure 8.
Distribution characteristics of errors in microstructure and thickness.

Structure No.	AC layer thickness (cm)	Base course type
STR1	12	Thin asphalt semi-rigid structure
STR6	16	Ordinary semi-rigid structure
STR7	18	Ordinary semi-rigid structure
STR13	24	24~28 cm thick asphalt structure
STR11	28	24~28 cm thick asphalt structure
STR16	36	36 cm thick asphalt structure
STR19	48	Full-depth asphalt structure
STR18	52	Full-depth asphalt structure

Table 1.

Table of typical structural parameters.

Parameter name Value

Look_back 12

Hidden_dim 128

Layers of TCN 2

Layers of LSTM 1

Heads of MSA 8

Learning_rate 0.001

Table 2.
Hyperparameters for model training.
Model RMSE MAE R²

LSTM 4.581 2.696 0.93

TCN_LSTM 4.527 2.728 0.932

LSTM_Attn 4.562 2.665 0.931

DSAN 4.272 2.452 0.943

Table 3.
Comparison of the prediction performance of various models.

Thickness grouping	Model	Evaluation metric
Thickness grouping	Model	RMSE	MAE	R²
Thin layer structure group	DSAN	5.3	3.32	0.913
	TCN-only	5.721	3.632	0.898
	LSTM-only	5.573	3.541	0.903
	LSTM + MHA-only	5.424	3.44	0.908
	DualTCN-only	5.514	3.53	0.904
	DSAN (without TCN)	5.628	3.579	0.902
	DSAN (without MSA)	5.649	3.588	0.9
Middle structure group	DSAN	5.514	3.706	0.959
	TCN-only	6.626	4.612	0.942
	LSTM-only	5.991	4.17	0.951
	LSTM + MHA-only	5.629	3.917	0.956
	DualTCN-only	6.165	4.29	0.947
	DSAN (without TCN)	5.728	3.986	0.955
	DSAN (without MSA)	5.93	4.126	0.952
Thick layer structure group	DSAN	5.808	3.819	0.958
	TCN-only	6.482	4.131	0.947
	LSTM-only	6.29	4.01	0.951
	LSTM + MHA-only	5.905	3.763	0.956
	DualTCN-only	6.391	4.073	0.95
	DSAN (without TCN)	5.905	3.763	0.955
	DSAN (without MSA)	6.288	4.013	0.951

Table 4.

Ablation study of proposed models.

Thickness grouping	Model	Evaluation metric				Relative improvement (%)
Thickness grouping	Model	RMSE	MAE	R²		RMSE	MAE	R²
Thin layer structure group	DSAN	5.3	3.32	0.913		4.9%	6.2%	1%
Thin layer structure group	LSTM	5.573	3.541	0.903		4.9%	6.2%	1%
Middle structure group	DSAN	5.514	3.706	0.959		8.0%	11.1%	0.8%
Middle structure group	LSTM	5.991	4.17	0.951		8.0%	11.1%	0.8%
Thick layer structure group	DSAN	5.808	3.819	0.958		7.7%	4.8%	0.7%
Thick layer structure group	LSTM	6.29	4.01	0.951		7.7%	4.8%	0.7%

Table 5.

Comparison of model performance under different thickness groups.

Structure number	Evaluation metric
Structure number	RMSE	MAE	R²
STR1	5.3	3.32	0.913
STR6	4.599	2.657	0.933
STR7	5.118	3.127	0.966
STR13	4.993	3.239	0.917
STR11	5.514	3.706	0.959
STR16	4.549	2.527	0.949
STR19	4.976	3.299	0.9
STR18	5.808	3.819	0.958

Table 6.

Leave-one-structure-out (LOSO) tests of proposed models.