Evaluating Resilience of Infrastructures Towards Endogenous Events by Non-Destructive High-Performance Techniques and Machine Learning Regression Algorithms

Nicholas Fiorentini,Pietro Leandri,Massimo Losa

doi:10.5194/egusphere-egu2020-21183

Abstract

&lt;p&gt;In order to plan infrastructure maintenance strategies, Non-Destructive Techniques (NDT) have been largely employed in recent years, achieving outstanding results in the identification of infrastructural deficiencies. Nevertheless, the extensive combination of different NDT that can cover various factors affecting infrastructure durability has not yet been thoroughly investigated.&lt;/p&gt;&lt;p&gt;This paper proposes a methodology for evaluating the resilience of infrastructures towards endogenous factors by combining different NDT outcomes. Machine Learning (ML) Regression algorithms have been used to predict the pavement surface roughness connected to a set of potential endogenous conditioning factors. The development, application, and comparison of two different regression algorithms, specifically Regression Tree (RT) and Random Forest (RF) have been carried out.&lt;/p&gt;&lt;p&gt;The study area involves 4 testing sites, both in the rural and urban context, for a total length of 11400 m. In addition to the International Roughness Index (IRI) calculated by profilometric measurements, a set of endogenous features of the infrastructure were collected by using NDT such as Falling Weight Deflectometer (FWD), and Ground Penetrating Radar (GPR). Moreover, a set of topographical data of roadside areas, information on properties of materials composing the subgrade and the pavement structure, traffic flow, rainfall, temperature, and age of infrastructure were gathered.&lt;/p&gt;&lt;p&gt;The database was randomly split into a Training (70%) and Test sets (30%). With the Training set, through a 10-Fold Cross-Validation (CV), the models have been trained and validated. A set of three performance metrics, namely Correlation Coefficient (R&lt;sup&gt;2&lt;/sup&gt;), Root Mean Square Error (RMSE), and Mean Absolute Error (MSE), has been used for the Goodness-of-Fit (GoF) assessment. Also, with the Test set, the Predictive Performance (PP) of the models has been evaluated.&lt;/p&gt;&lt;p&gt;Results indicate that the suggested methodology is satisfactory for supporting processes on planning road maintenance by National Road Authorities (NRA) and allows decision-makers to pursue better solutions.&lt;/p&gt;

Full Text