Thesis title: La gestione delle infrastrutture stradali al fine del prolungamento della loro vita residua: applicazioni dell’Automated Machine Learning per la valutazione del rischio di ponti esistenti
Road infrastructure, crucial for mobility and economic development in modern societies, faces increasing challenges due to progressive deterioration. Bridges and viaducts, in particular, suffer degradation due to environmental factors, increasing loads, and aging materials. In Italy, this issue is exacerbated by the heterogeneity and age of infrastructure.
The recent “Guidelines for the Classification and Risk Management, Safety Assessment, and Monitoring of Existing Bridges” (LLGG) [1], [2] introduced a methodological approach for safety assessment and management of road bridges on a national scale. However, while they represent a step toward more structured infrastructure management, they entail a high application burden and the need for complex assessment processes. In this context, the development of computational tools to support the management and maintenance of existing bridges in line with the LLGG becomes particularly important.
The adoption of innovative technologies, such as Machine Learning (ML) algorithms, is emerging as a strategic solution to improve infrastructure safety, resilience and sustainability.
During the thesis, several useful algorithms were implemented and validated as support for the application of the LLGG-based management procedure.
First, an algorithm was developed for the analysis of normative parameters to generate and evaluate structural configurations with respect to risk, quantified in Classes of Attention (CdA). The use of Automated Machine Learning (AutoML) techniques, on the generated configurations, allowed the training of models with a reduced set of parameters readily available from census data and expeditious assessments, for predictive analyses of CdA with respect to structural and seismic risks. The models were experimentally validated through a sample of real bridges inspected during the thesis work.
In parallel, activities aimed at validating deep learning algorithms for automated identification of structural defects were also performed, with a focus on image analysis and degradation assessment on bridge piers. A study was, in addition, conducted on risk assessment for works with spans of less than six meters that are not included in LLGG.
The results showed that it is possible to accurately predict the CdA using a reduced set of parameters, avoiding the need to determine the Defectiveness Level, which is the most challenging parameter to assess in real cases. Moreover, the use of software for automatic recognition of structural defects by analyzing orthophotos obtained from drones has shown enormous potential in being able to be used as a tool to support inspection activities by reducing the rate of subjectivity in the process.
The development of a simplified method for risk assessment risk of small-span structures highlighted how a more streamlined procedure, both in terms of process burdensomeness and resources used, is capable of effectively capturing the aspects of the complete procedure.