Machine learning model for asphalt pavements performance prediction.

Abstract

The accurate forecast of pavement performance is crucial for Pavement Management Systems, as they guide maintenance decisions and budget allocation. With improvements in data collection, storage and processing, machine learning (ML) is gaining visibility as a behavior prediction method in the field of engineering. Several studies evaluated these algorithms' potential to predict pavement serviceability, however some challenges limit its use. The pavement performance history, structural information, and traffic load characteristics are not always available on data-oriented manner. The training dataset preprocessing has great impact on the model's predictive performance, is highly dependent on the modeler experience, and are not typically reported on the engineering related literature. Also, the long-term prediction using ML algorithms usually demand long historical time-series, which are not always available on a large scale. Therefore, the objective of this dissertation is to develop a methodology for the use of machine learning algorithms on the Asphalt Pavement Performance Prediction, comprehending: data collection and organization; training dataset definition; algorithm selection and configuration; and long-term performance model definition. The pavement's performance was based on the Surface Maximum Deflection (D0) and International Roughness Index (IRI). To achieve this goal, the three most used ML algorithms -Support Vector Machine (SVM); Random Forest (RF); and Artificial Neural Network (ANN) -in D0 and IRI short-term prediction were tested using 10 training datasets, composed of the data collected from 21,568 traffic lane kilometers. The long-term prediction model was based: on the short-term ML model; the Markov chain principle; and the recursive method. The results indicated that ANN is the most accurate technique with a RMSE of 16x10 -3 mm on the D0 prediction; and a RMSE of 0.19m/km on IRI prediction. The models' evaluation of the long-term prediction was obtained by the comparison of 20 pavement segments field data with simulated data. The best results were also obtained with ANN: they presented an average RMSE of 23x10 -3 mm on the D0 prediction; and a RMSE of 0.17m/km on IRI prediction. RF was also identified as an effective technique, generating similar results with less data preprocessing.