Development and implementation of medium-fidelity physics-based model for hybrid digital twin-based damage identification of piping structures

Current predictive maintenance technologies are mostly data-driven, where they identify complex relationships using statistics and machine learning (ML) models for damage prediction. The main disadvantage of data-driven or ML models is their high dependency on training data, making them poor in extrapolating and predicting untrained events. Hence, engineers prefer a physics-based model in most cases due to its strong interpretability that aids in supporting critical engineering decisions. However, high-fidelity physics-based models are computationally exhaustive. To preserve the merits and alleviate the inadequacy of both data-driven and physics-based models, recent years have shown an increase in works on hybrid digital twin (DT) models which integrate both methods. This work presents the development of a medium-fidelity physics-based model of a piping structure and its implementation in a hybrid DT for damage identification. Two modelling approaches for the piping support bolted connections were investigated, i.e., bonded contact and spring-based model. The developed physics-based models were correlated with the modal testing data. Results showed that with suitable spring stiffness, the spring-based model has better dynamical representation than the overly stiff bonded contact model with an average natural frequencies deviation below 10% and an average Modal Assurance Criterion (MAC) value of at least 0.75 for both undamaged and damaged conditions. The correlated medium-fidelity spring-based model was used to simulate damage cases for ML training. Results showed that the trained model achieved an accuracy of 95% in identifying the damage at the physical piping structure, thus validating the proposed hybrid DT in damage identification.