An ML-based framework for predicting prestressing force reduction in reinforced concrete box-girder bridges with unbonded tendons

The paper presents a machine learning (ML) based framework to predict the prestressing force reduction in prestressed reinforced concrete (PSC) box-girder bridges with unbonded tendons. In the field of road network safety, the reliable assessment of some bridge typologies, such as PSC box-girder bridges, depends on different aspects, among which the inaccessibility of internal unbonded tendons, the difficulty in measuring the effective prestressing force reduction over time, the design of an efficient structural health monitoring (SHM) system. To address the above issues, the proposed approach exploits the results of experimental tests on a scaled PSC box-girder to validate a nonlinear modelling strategy and, in turn, to generate a sample dataset for training different ML algorithms. To ensure generalizability of the proposed ML model, the variability of several parameters, including geometrical and mechanical properties, was accounted for. The obtained results, evaluated in terms of statistical metrics and through an eXplainability approach, revealed that the proposed surrogate model is able to predict the prestressing force reduction for this bridge typology, knowing the current prestressing force, the elastic modulus of the concrete, and the strain variation in specific cross-sections of the structure. The application of the framework on a scaled PSC box-girder experimentally tested, demonstrated its suitability for: i) estimating the prestressing force reduction without employing periodic and expensive onsite tests; and ii) providing the best strategy for employing a sensor-based SHM system. • A ML surrogate model to predict prestressing force reduction in prestressed concrete (PSC) box-girder bridges is proposed. • Experimental test of a scaled PSC box-girder, simulating an effective prestressing loss by removing prestressing cables. • Calibration of FE modelling strategy suitable for simulating prestress losses and generating data to train the ML model. • Use of eXplainability to define the most relevant features influencing prestress losses and indicate the best SHM strategy.