An iterative regression-based thermal response prediction methodology for instrumented civil infrastructure

Normal environmental and operational variability continues to be a challenge in the measurement interpretation of structural health monitoring (SHM) data. This variability masks important changes within the collected data that could be indicative of damage, and its removal has been shown to significantly improve structural anomaly detection. Temperature-based measurement interpretation approaches are employed to remove this variability, however they are typically knowledge-intensive processes and require significant data processing. Current thermal response prediction methodologies have been predominantly developed using simplistic methods of engineering informatics using data from laboratory experiments, with low-levels of available computing power and data storage capabilities, and have generally had limited application to real-world infrastructure. This study proposes the iterative regression-based thermal response prediction (IRBTRP) methodology that harnesses high-performance computing to optimise the thermal response regression models through: (i) advanced regression algorithms (e.g. support vector regression and artificial neural networks); and (ii) a grid-search that determines the optimal hyperparameters, defined as customisable variables that affect the prediction accuracy. The process of determining the optimal hyperparameters is automated and this enables the IRBTRP methodology to be easily applied across a wide range of applications. The measured thermal responses from a Laboratory Truss and the MX3D Bridge (the world's first structure to be produced using metal additive manufacturing) are accurately predicted, with average errors of 3.1% and 4.9% respectively, across unseen testing datasets. Accurate predictions are obtained for the testing period of the MX3D Bridge dataset despite significant differences in temperature conditions between the training and testing subperiods. The IRBTRP methodology can therefore facilitate more accurate and reliable temperature-based measurement interpretation for real-world infrastructure, such as to support damage detection.