Automatic pavement maintenance predictions using language and machine learning models.

Abstract: Maintenance and rehabilitation (M&R) activities are routine programs designed to mitigate pavement deterioration and ensure a sustainable, high level of service throughout the pavement’s remaining design lifespan. A critical component of planning future M&R activities. involves evaluating the effectiveness of past M&R efforts. Well-organized, timestamped maintenance records not only assist in determining future schedules but also facilitate seamless preventive and routine maintenance practices. However, it is common for pavement management systems or public databases, such as the Long-Term Pavement Performance (LTPP) database, to have incomplete or missing M&R activity records. While missing numerical data, such as rutting, can be addressed using statistical imputation methods, it is significantly more challenging to fill in missing M&R activities due to their textual nature. To address this challenge, this paper proposes a framework that automatically extracts complex relationships between recorded pavement features within a specified time window to predict the category of M&R activities conducted. Using a few-shot approach, the framework utilizes the OpenAI ChatGPT-4 language model to automatically annotate extracted maintenance records as major or minor activities. These classified records and numerical pavement section features, such as the international roughness index, rutting, distresses, and pavement layers, are organized into a sequential time series format. Various machine learning models, including Long Short-Term Memory (LSTM), Support Vector Machine (SVM), XGBoost, and Convolutional Neural Networks (CNN), are trained to predict whether major, minor, or no maintenance occurred during the selected time frame. This paper presents the procedures and key findings from the development of both the language and machine learning models, further expanding previous works with applications for different pavement M&R activities.

Keywords: maintenance and rehabilitation, large language models, machine learning models, pavement management systems, time series, LTPP datasets.