講演情報

[AI-06]Long Short-Term Memory–Based Prediction of Flexible Pavement Surface Layer Performance Using Inspection Time-Series Data

*Bernardo Mota Lontra1, Leo Françoso Dal Piccol Sotto2, Jan-Hinrik Borchers2, Gina Torres-Alves1, Artem Yushchenko2 (1. TNO (Netherlands), 2. BASt (Germany))

キーワード:

Pavement performance prediction、Long-short term memory model、time-series inspection data、Pavement management system

Pavement inspection data is the primary input for the development of pavement performance prediction models, which play a key role in Pavement Management Systems by supporting monitoring schedules, maintenance planning, and budget allocation. As pavement condition data are collected at fixed time intervals, they form time-series that enable the identification of temporal trends, anomalies, and future conditions. This characteristic motivates the application of data-driven approaches capable of capturing temporal dependencies in pavement deterioration processes. This study investigates the application of Long Short-Term Memory (LSTM) networks for the prediction of pavement condition, with a focus on rutting and unevenness. The proposed framework includes data preprocessing and data augmentation strategies to address data sparsity and improve model robustness. Historical pavement condition data from the German highway system were used to train and evaluate the models. Two datasets were considered: (i) a national network dataset with measurements collected at four-year intervals and (ii) a permanent performance section dataset with measurements collected twice a year over an eleven-year period. Pavement condition predictions were generated for time horizons of 4, 8, and 12 years and compared against traditional mechanistic-empirical pavement performance models using metrics such as the Root Mean Square Error (RMSE) and the coefficient of determination (R²). The results demonstrate a significant improvement in prediction accuracy of the LSTM-based models compared to the mechanistic-empirical approaches across all prediction horizons and datasets. Results of the model trained on the permanent performance section to learn patterns across a longer horizon and then applied to the national network are also presented. The findings highlight the potential of LSTM-based pavement performance models to more accurately capture the temporal evolution of pavement condition. The proposed approach can provide asset managers with more reliable long-term predictions, thereby supporting improved planning of maintenance interventions and more efficient allocation of infrastructure budgets.