Connecting physics based and data driven models: the best of two worlds

Tuesday, March 6, 2018 - 9:00am - 10:00am
Lind 305
The use of a priori digital models to build, evaluate and validate designs has become a standard practice in the product creation process. Furthermore, their use in production, operation and service has been increasing in the last years, for example in model predictive control, maintenance or assist systems. These models are often of gradually increasing complexity during the development cycle, evolving from a mere mathematical description or a conceptual system representation to a physics-based high-fidelity model accurately representing the actual physical asset as a “Digital Twin”. Once a physical system (prototype or product) is available, experimental data are used to validate, update or extend these models to maximize the value and applicability of the digital representation. Continuing and significant efforts focus on upgrading modeling capabilities in terms of speed, accuracy and system complexity. Challenges include dealing with multiscale methods reconciling micro-physical accuracy with macroscopic performances and with heterogeneous multi-representation models.
Over recent years, a second evolution has taken place starting from massively acquired experimental data. This ranges from the long-term observation of a single physical asset to monitoring huge fleets of similar products. Event-feature correlation, faulty and healthy functioning, occurrence of failure modes and operation under various operational and loading conditions generate a wealth of raw data which are reduced to explicit or implicit data-driven models by deep learning techniques and/or the adoption of machine learning approaches. These techniques prove very powerful in monitoring, diagnosing and troubleshooting the system in operation but their predictive capabilities are often subject to criticism.
While it is not uncommon to have both methods positioned as antagonizing to each other, the real challenge is to combine them, capitalizing on the strength and information present in each of them. Examples are in using data driven techniques to validate or improve simulation models with test data or to reduce the vast amount of simulation model data to formats feasible for integration in system of system models, in controllers or in virtual sensors. Alternatively, simulation models can be used to design optimal tests or in the training of neural networks for fault detection. Common to these applications is that the use of physics based models and data driven approaches is interwoven and explored without prejudice. The lecture will be illustrated with examples from the mechanical and mechatronic industries