Pavement Roughness Prediction Based on Encoder-decoder Structure

Abstract:

A pavement roughness prediction model based on encoder-decoder structure was proposed， and a comparative analysis of different layers was conducted. Then， the effect of the layer number， the number of hidden units and the time window length on the accuracy of the model was discussed. To train and evaluate the model， an international roughness index （IRI） dataset was constructed based on long-term pavement performance （LTPP） database published by the US Department of Transportation. The results show that the encoder-decoder structure with gated recurrent unit （GRU） layer has the highest accuracy， its performance is better than that of the machine learning model XGBoost and single long short-term memory （LSTM） network. The importance of different input features was evaluated by randomly shuffling features， and the results indicate that the road structure and temperature are important for pavement roughness prediction. Therefore， the road structure and temperature data should be attached great importance during the construction of pavement database.