Digital twin of UAV provides predictive maintenance
The Dynamic Data-Driven Application Systems (DDDAS) digital twin project includes researchers from ty of Texas at Austin (UT Austin), MIT, Akselos and drone maker Aurora Flight Sciences. The twin represents each component of the UAV, as well as its integrated whole, using physics-based models that capture the details of its behaviour. The digital twin also uses on-board sensor data from the UAV and integrates that information with the model to create real-time predictions of the health of the vehicle.
“It’s essential that UAVs monitor their structural health,” said Karen Willcox, director of the Oden Institute for Computational Engineering and Sciences at The University of Texas at Austin (UT Austin) and project lead. “Big decisions need more than just big data, they need big models, too. These big problems are governed by complex multiscale, multi-physics phenomena. If we change the conditions a little, we can see drastically different behaviour.”
The project combined computational modeling with machine learning to produce predictions that are reliable, and also explainable. In the case of the digital twin UAV, the system is able to capture and communicate the evolving changes in the health of the UAV. It can also explain what sensor readings are indicating declining health and driving the predictions.
The project also uses an approach called model reduction. This identifies approximate models that are smaller but still include he most important dynamics so that they can be used for predictions. “This method allows the possibility of creating low-cost, physics-based models that enable predictive digital twins,” she said.
Rather than simulate the entire vehicle as a whole, Akselos helped to break the model into pieces such as a section of a wing and compute the geometric parameters, material properties, and other important factors independently, while also accounting for interactions that occur when the whole plane is put together.
Each component is represented by partial differential equations and at high fidelity, finite element methods and a computational mesh are used to determine the impact of flight on each segment, generating physics-based training data that feeds into a machine learning classifier.
The combination of model reduction and decomposition made the digital twin 1000 times faster than other methods, cutting simulation times from hours or minutes to seconds while maintaining the accuracy needed for decision-making.
“The method is highly interpretable,” she said. “I can go back and see what sensor is contributing to being classified into a state.” The process naturally lends itself to sensor selection and to determining where sensors need to be placed to capture details critical to the health and safety of the UAV.