Pavement Performance Prediction for Continuous Road Sections Considering Spatial Correlation: A Deep Learning–Based Framework

Abstract

The accuracy of pavement performance prediction models directly affects the development of pavement maintenance and rehabilitation (M&R) decision plans. However, existing pavement performance prediction models ignore the interactions of adjacent road sections or do not combine pavement spatial correlation with artificial intelligence algorithms such as deep learning. Therefore, this study proposed a deep learning–based framework for pavement performance prediction considering spatial correlation. Two artificial neural network (ANN) models were first established to clean pavement performance data. The raw pavement section data could be transformed into spatial series data using the sliding window (SW) method. The bi-directional long short-term memory (BiLSTM) model could capture the dependencies of adjacent spatial series data. Therefore, the SW-BiLSTM pavement performance prediction model was established. The data used in this study were obtained from the Shanxi Province pavement management system (PMS). Three indicators commonly used in machine learning, mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2), were used to assess the accuracy between the predicted and true values. The results show that the data cleaning method used in this study is more accurate than the traditional adjacent data interpolation method. The SW-BiLSTM prediction model is compared with an SW-LSTM model that only considers spatial single propagation and an ANN model that does not consider spatial correlation. The SW-BiLSTM prediction model has the highest accuracy with MAE, RMSE, and R2 of 1.5090, 2.0331, and 0.9098, respectively. This shows that this framework for pavement performance prediction can provide more reliable prediction results, which can help engineers make more reasonable M&R decision making.

Topics & Concepts