Multi-objective optimisation of in-service asphalt pavement maintenance schedule considering system reliability estimated via LSTM neural networks

Abstract

This article presents a novel system reliability-based framework for multi-objective optimisation of preventive maintenance (PM) management of in-service asphalt pavement. To accurately predict the international roughness index (IRI) sequence of pavement sections, a long short-term memory (LSTM) neural network that considers the spatiotemporal correlations between IRI sequences is trained with data retrieved from the long-term pavement performance (LTPP) program. Based on time-dependent limit-state functions (LSFs) incorporating the uncertainty associated with LSTM neural network prediction and the observational error involved in IRI measurement, Monte Carlo simulation (MCS) with importance sampling (IS) is adopted to calculate the reliability of pavement sections. Pavement sections located in New Mexico and Montana are selected as illustrative examples. Tri-objective optimisation processes are investigated by maximising user benefits (i.e. improved system reliability) and agency benefits (i.e. extended service life) while minimising the associated life-cycle cost (LCC) (i.e. user and agency costs) with multi-objective genetic algorithms (GAs). The obtained Pareto solution sets may assist decision-makers in the selection of well-balanced solutions to identify the optimal timing for applying PM treatments to in-service asphalt pavement.