Currently, roads authorities apply deterministic (rule based) digital tools to recommend lengths that should be considered for maintenance (using condition data provided by automated and human assessors). However, these recommendations do not robustly reflect the decisions that engineers would themselves make. This work investigates how an outcome-based model could be developed to better identify lengths for treatment. The development of the model draws on network level condition data (from the Strategic Road Network) that includes visual condition, roughness and skid resistance, and contextualising information such as construction, traffic, material and age. These are collated and aligned with data on the actual treatments that were carried out on the network, in order to train and test a set of machine learning models. The best performing of these models deploys the Random Forest Classifier, which is referred to in this work as the Digital Engineer.
A comparison between the locations identified for treatment by the Digital Engineer, the locations identified by rules-based tools, and the locations that were actually treated, shows that the Digital Engineer provides a significantly higher level of overall accuracy in the identification of these lengths. The results suggest that additional contextualising information assists in achieving outcomes from digital tools that better agree with the decisions made by engineers, and that Machine Learning techniques may be used to apply this additional information. It is recommended that further development and testing of the Digital Engineer approach should: better understand the influence of features on tool decisions; more comprehensively verify the model; and determine the route to implementation.