Identifying the most suitable machine learning approach for a road digital twin

Abstract

Road infrastructure systems have been suffering from ineffective maintenance strategies, exaggerated by budget restrictions. A more holistic road asset management approach enhanced by data-informed decision making through effective condition assessment, distress detection, future condition predictions can significantly enhance maintenance planning, prolonging asset life. Recent technology innovations such as Digital Twins have great potentials to enable the needed approach for road condition predictions and a proactive asset management. To this end, machine learning techniques have also demonstrated convincing capabilities in solving engineering problems. However, none of them has been considered specifically within digital twins context. There is therefore a need to review and identify appropriate approaches for the usage of machine learning techniques within road digital twins. This paper provides a systematic literature review of machine learning algorithms used for road condition predictions and discusses findings within the road digital twin framework. The results show that existing machine learning approaches are to some extent, suitable and mature to stipulate successful road digital twin development. Moreover, the review whilst identifying gaps in the literature, indicates several considerations and recommendations required on the journey to road digital twins, and suggests multiple future research directions based on the review summaries of machine learning capabilities.

Research on the application of digital twin technology in smart highway real-time simulation

“…China Unicom summarized the current situation and pain points of the smart highways digital construction [5] , proposing a digital base architecture. It elaborated on the important role of big data platform to achieve comprehensive perception capability and improved management and control capability, etc.…”

A GIS based approach for predicting pavement deterioration on the UK road network.

“…Random forest approaches can be used for either regression or classification without fundamentally changing the underlying algorithm; the key difference is at the final aggregation stage, where for example either majority vote (for classification) or mean value (for regression) could be used to predict a target variable (Liaw and Weiner, 2002). (Gong et al, 2018) successfully used a random forest approach to predict IRI from the construction and condition data, and in their systematic review of ML approaches in civil engineering Chen (Chen et al,2022) notes the potential for random forests for pavement prediction, and suggests key inputs to the predictive model.…”

“…Following the work of (Gong et al 2018) and (Chen et al 2022) this paper looks to investigate the use of Random Forest approaches for prediction of pavement deterioration on the UK MRN and identify key datasets that can be used to support the prediction process within the context of the UK LRN. Novel use of authoritative spatial and contextual data alongside flexible predictors such as the Random Forest algorithm could support the predictive process in lieu of hard-to obtain information such as construction, materials, and historic maintenance.…”