A digital twin is a reduced-order representation of a sophisticated real-world system or process, which leverages known physics and real-time data for the ultimate goal of status awareness, design enhancement, manufacturing optimization, condition-based maintenance and retrofit, operational risk minimization, etc. Yet any such metamodeling effort is inevitably constrained by the lack of high-fidelity physics knowledge, and corrupted by the poor observability of the sensing hardware and various kinds of uncertainties.
As the primary functionality of digital twins, research on accurate estimation of the system dynamics and prediction of its deterioration will be focused in the first half of this presentation. Specifically, deep neural network (convolutional long short-term memory) will be introduced to characterize the operation conditions of rotating machineries, and predict the remaining useful life. To obtain a better quality of the prediction and interpretability of deep learning models, physics is also involved through adaptive filtering to regulate the estimations. Run-to-failure tests on bearings are conducted to validate the algorithm.
The second half the presentation concerns noncontact sensing using video cameras, which is a fundamental departure from traditional measurement using instrumented physical sensors and drastically improve the potential of digital twins. In particular, phase-based motion extraction and amplification is adopted to recognize system status and complex structure dynamics. Modal identification is implemented on wind turbine blades, and subtle invisible dynamics is successfully extracted with full-field information maintained. Probabilistic uncertainty quantification model will also be introduced in the end, which intrinsically enhances the capability of digital twins in detecting system changes with statistical significance.