The Federal Aviation Administration (FAA) has initiated research to develop models of long-term airport pavement performance that can be incorporated into a methodology for extended-life pavement design. This report includes a review of machine learning (ML) techniques for prediction of long-term pavement performance. The study examined applicability of three ML techniques, namely artificial neural network, support vector machine, and random forest, based on FAA Extended Airport Pavement Life (EAPL) data. Two modeling approaches, namely static and autoregressive, were implemented for each ML method. Initial models were developed for pavement condition index (PCI) and for a subsidiary index based on PCI but considering only non-load-related components (designated anti-SCI) by considering pavement age as the primary predictor. The study also implemented various feature selection methods based on supervised learning techniques to determine the key environmental and traffic variables affecting the pavement performance. These methods were implemented to reduce the dimensionality of input space by identifying and removing a subset of irrelevant and redundant environmental and traffic variables. Feature construction methods based on unsupervised learning techniques, including k-means clustering and Principal Component Analysis, were also applied to reduce the dimensionality of input space by constructing new inputs. Findings from the feature selection and construction analyses were used to develop ML predictive models for the performance index anti-SCI, by implementing an autoregressive approach and using random forest as the learning algorithm.    

DOT/FAA/TC-21/44 Author: Ali Z. Ashtiani