Roughness Modeling for Composite Pavements using Machine Learning

A large number of paved highway surfaces comprises composite pavements as a result of concrete pavement rehabilitation that uses an asphalt overlay on top of the concrete surface. Annually, billions of dollars are spent on the maintenance and rehabilitation of road networks. Roughness is one of the several indicators of road conditions used to make objective decisions related to road network management. The irregularities in the pavement surface affecting the ride quality of road users can be described by a standard roughness index defined as the International Roughness Index (IRI). Roughness prediction models can identify rehabilitation needs, analyze rehabilitation effects, and estimate future pavement conditions to implement different Maintenance and Rehabilitation (M&R) activities to extend the pavement life cycle and provide a smooth surface for road users. This study intended to develop pavement performance models to predict roughness for asphalt overlay on concrete pavement sections using the Long-Term Performance Pavement (LTPP) program database. Artificial Neural Networks (ANNs) approach was used to develop roughness prediction models. A total of 52 pavement sections with 592 data points were analyzed. Five models were developed, and the best performing model, Model 5 was found with an average square error (ASE) of 0.0023, mean absolute relative error (MARE) of 12.936, and coefficient of determination (R²) of 0.88. Model 5 utilized one output variable (IRIMean) and 14 input variables (i.e., Initial IRIMean, Age, Wet-Freeze, Wet Non-Freeze, Dry-Freeze, Dry Non-Freeze, Asphalt Thickness, Concrete Thickness, CN Code, ESAL, Annual Air Temperature, Freeze Index, Freeze-Thaw, and Precipitation). The ANN model structure utilized for Model 5 was 14-9-1 (14 inputs, 9 hidden nodes, and 1 output). Environmental impacts and traffic repetitions can cause severe damage to the pavement if timely maintenance and rehabilitation are not performed. By considering the effects of the M&R history of the pavement, it is possible to obtain realistic prediction models for future planning. Therefore, the developed ANN roughness performance models in this paper can be used as a prediction tool for IRI values and guide decision-makers to develop a better M&R plan. Local and state agencies can use available historical traffic and climatological data in the developed models to estimate the change in IRI values. Utilizing these prediction models eliminates time-consuming data collection and post-processing, and consequently, a cost reduction. This low-cost tool will improve the condition assessment and effective M&R scheduling.