Improved Hybrid Deep-Learning Framework for Pavement Performance Prediction Across Service Life

Abstract

Pavement performance prediction aims to forecast the future condition of pavement as accurately as possible, enabling proactive planning and optimized maintenance and rehabilitation (M&R) interventions. Pavement failure has traditionally been predetermined during the design stage when performance indicators such as the international roughness index (IRI), cracking, and other distresses exceed predefined thresholds. While such projections may be useful during material-selection phases, pavements often experience varying deterioration patterns over their service life, making original forecasts less reliable. To overcome these challenges, we propose to use advanced predictive models based on machine-learning techniques while leveraging text embeddings extracted from pretrained models. This study specifically presents two IRI prediction models: a long short-term memory model for forecasting IRI during long-term deterioration, and an artificial neural network model for predicting post-maintenance IRI. Our study further experiments with the semantic encoding power of different large language models, including transformer models, to provide numerical representations of raw maintenance logs. We show the importance of precise parameter selection for text-embedding models along with readily available pavement-design input parameters passed to a customized artificial neural network (ANN) model for post-maintenance IRI prediction. The proposed framework supports integration of a wide variety of maintenance actions, including complex scenarios involving multiple maintenance applications. We trained the models using data extracted from the long-term pavement performance (LTPP) database and obtained R 2 scores of 93% and 86% for IRI long-term deterioration and post-maintenance test datasets, respectively. These results underscore successful application of intelligent models across the pavement’s entire service life.