Physics-supported GP surrogates: towards physics-informed probabilistic structural health monitoring

Damage quantification in active-sensing, guided-wave Structural Health Monitoring (SHM) under varying conditions is a challenging problem, for which advanced machine learning techniques have shown significant progress in the literature. However, one persistent drawback of accurate machine learning approaches is the need for training data, the acquisition of which can be very costly, or impossible altogether. This problem is addressed in this study where a probabilistic SHM framework integrating physics-based and data-based models is proposed here for active-sensing, guided-wave SHM under varying damage and loading states. A physics-based load-compensation model is used to reconstruct guided-wave signals under varying loads for the component being studied at the healthy state for supporting Gaussian Process regression models (GPRMs) in the training and state prediction processes. Two frameworks are proposed herein based on the scenario being treated. The first scenario is the unavailability of experimental training data at varying damage states under non-zero loading conditions for training GPRMs that can be used for damage quantification under load. In this case, physics model-reconstructed signals are used to assist one-dimensional GPRMs trained using experimental data from the unloaded component in expanding their applicability to loaded states by compensating load effects in the incoming test data for which damage size is being predicted. The second scenario addressed here is the sparsity of training data in the loading-state dimension when training multi-dimensional GPRMs that can simultaneously quantify damage size as well as the operational condition (i.e. load herein). The physics model-reconstructed signals are used in this case for synthesizing “healthy” and “damaged” signals at the loading conditions missing in the 2D training space. Results of applying both frameworks on an Aluminum coupon are shown, and prediction accuracy is compared with that of GPRMs trained using all-experimental data. It is shown that, especially in the second scenario, prediction accuracy is not affected too much when model-synthesized data is added into the prediction/training loop, which proves the importance of the approach presented in this study in reducing the dependence of machine learning techniques upon experimental training data. The proposed probabilistic SHM framework allows for efficient models that capture the advantages of physics- and data-based approaches for entertaining accurate and robust damage detection and quantification.