A probabilistic framework for forward model-driven SHM

A challenge for many structural health monitoring (SHM) technologies is the lack of available damage state data. This problem arises due to cost implications of damaging a structure in addition to issues associated with the feasibility and safety of testing a structure in multiple damage scenarios. Many data-driven approaches to SHM are successful when the appropriate damage state data is available, however the problem of obtaining data for various damage states of interest restricts their use in industry. Forward model-driven approaches to SHM seek to aid this challenge. This methodology uses validated physical models to generate predictions of the system at different damage states, providing machine learning strategies with training data, to infer decision bounds. An ideal forward model-driven SHM framework is one in which one or more physical models are able to produce predictions that are statistically representative of data obtained from the physical structure. Validation of these physical models requires observational data. As a result, validation is performed on a component or sub-system level where damage state data can be more easily obtained. This methodology requires the combination of several low-level physical models via a multi-level uncertainty integration technique. This paper outlines such a framework using uncertainty quantification technologies and statistical methods for combining low-level probabilistic models whilst accounting of discrepancies that may occur in interactions with other low-level models. The method contains several statistical techniques for accounting for model discrepancies that may occur at any point throughout the modelling process. Model discrepancies arise due to missing physics or simplifications and result in the model deviating from the observed physics even when the ‘true’ parameters of the model are known. By accounting for model discrepancies throughout the framework the approach allows for further insight into model form errors whilst also improving the techniques ability to produce statistically representative predictions across damage states. The paper presents the key stages highlighting the relevant technologies and application considerations. Additionally, a discussion of integration with current data-driven approaches and the appropriate machine learning tools is given for a forward model-driven SHM approach.