Identification of Pavement Issues Using Latent Dirichlet Allocation Machine Learning

The Federal Aviation Administration (FAA) is developing new design procedures to extend airport pavement life beyond 20 years. One element of this research is a new distress “mega-index” whose components are intended to represent independent aspects of airport pavement serviceability: low foreign object damage (FOD) potential, low skid potential, and smoothness. The proposed components of the mega-index appear intuitively correct but require validation. This paper is part of the research to validate the assumption that these three components completely describe airport pavement serviceability. Machine learning methods were used to review maintenance and rehabilitation records to determine the types of issues that the airport owners and operators were willing to expend funds and effort to resolve, which is a strong indicator of factors that cause a pavement to fail to meet expectations. Researchers developed topic models using Latent Dirichlet Allocation (LDA) on a corpus of funding request documents and then examined the topics for issues related to pavements. The records review included the detailed maintenance records collected for the FAA’s Extended Airfield Pavement Life project and Department of Defense Standard Form 1391 requests for funding. Researchers expected the LDA result topics to contain FOD, friction, or roughness. Early results indicated that using LDA to identify the topics in funding request and work order documents was viable. Topics identified using work-order sources only included items seemingly related to cracks and roughness. The results did not scale well once the Forms SF1391 were integrated into the data set. The Forms SF1391 included many non-pavement projects, so while the analysis was able to identify that pavements needed repair, it did not identify specific pavement issues. The analysis was able to identify specific issues for other facilities, including asbestos and lead paint abatement for buildings and leak repairs for roofs. This indicates the validity of the approach given an appropriate corpus of funding documents.