A digital twin-driven machine learning framework for structural condition monitoring using multi-datasets

Real-time condition monitoring (CM) utilizing vibration measurements offers a proactive approach to detect faults and enable predictive maintenance. The robustness and accuracy of CM applications significantly rely on the quality of training datasets. While numerical model-generated data is commonly used in current solutions, the efficiency of CM frameworks is greatly influenced by dataset quality. This research addresses the challenge by employing data-driven modeling to enhance the efficiency and accuracy of CM frameworks. A Hierarchical Bayesian Modeling (HBM) framework is introduced to estimate uncertainties in finite element (FE) model parameters at the healthy state. The HBM is particularly adept at addressing uncertainties in model parameters caused by variations in experimental data, material characteristics, assembly processes, and nonlinear mechanisms under diverse loading scenarios. These uncertainties in the parameters of the FE model obtained at the healthy state of the structure are maintained for the model representing the damaged state of the structure. The FE model with the associated uncertainties is used to generate data for training a convolution neural network model for the different health states. Such training based only on observed/inferred uncertainties at the healthy state significantly enhances the robustness and overall accuracy of the CM framework against previous approaches based on arbitrarily imposed uncertainties at the healthy and damaged state of the structure. To validate the effectiveness of the proposed method, multiple experimental structural dynamics tests on a small-scale laboratory truss are conducted at the healthy and damaged states. The results demonstrate the applicability and effectiveness of the developed approach in improving the structural health monitoring process.