For many industrial machines and components, it is difficult or expensive to directly measure temperatures at critical points of interest, making accurate monitoring of local temperatures challenging. Often, only a few measurement points are available close to the point of interest (e.g. the junction temperature in PE modules) or even impossible to obtain (e.g. temperature at the gear flanks or in the bearing contact of gearboxes). These temperatures however strongly correlate to the performance and degradation rate of the component under consideration.

Most of these products exhibit a combination of complex thermal phenomena (conduction, convection, fluid flow, …) that can be described by 3D partial differential equations (PDE). For these applications there are a multitude of models available (finite volume, finite elements, lumped parameter models, …). These models can have reasonable accuracy of the order 1-2 °C but come with tremendous computational cost, often making real time monitoring impossible.

In DTF-PINN, we aim to develop PINN (Physics Inspired Neural Networks) based virtual sensors for thermal applications that are computationally lightweight compared to their physics based counterparts. Physics inspired neural networks have the ability to accurately capture thermal phenomena that are governed by PDEs, while having the potential to run significantly faster than their physics based (FE, FV, LP, ...) counterparts. By incorporating physical laws and boundary conditions, the PINN architecture has the ability to learn from fewer and sparser data compared to regular “data hungry” black box Machine Learning techniques.

If successful, this project will enable the integration of lightweight PINN based virtual sensors into the framework of the industrial machines and processes and it will lead to the development of better design and monitoring tools for a multitude of applications. The PINN framework leverages the accuracy of available design phase models while running significantly faster and reduces the effort needed to build and fit computationally light temperature estimators. If we succeed in developing a robust PINN architecture selection guideline or automated search, this will significantly reduce the deployment time for fast thermal monitoring from the order of months to the order of days or weeks. The virtual sensing framework may also reduce the cost of sensor solutions as nearby sensors can help to increase the accuracy of temperature estimations.