Real-Time Thermal Data Assimilation for Power Electronics at the Edge

Abstract Physical systems are often far too complex for a virtual model to fully encapsulate the behavior of a system. Requiring the user to fine tune parameters of the virtual model over the course of a system’s life cycle. Data assimilation addresses this problem by continuously updating the model with sensor data to construct a personalized model of the physical system. This personalized model is known as a digital twin. Digital twins of ship systems can provide insight into the future state of the ship to enable operators to make informed decisions to increase health of the ship. However, there are a few key challenges that need to be overcome while updating a model. The first problem is reducing latency between the physical system and the digital twin. While ensuring that the digital twin has enough time and data to update. The second problem is verifying the model is an accurate representation of the physical system. This paper proposes a methodology that uses real-time sensor data and a particle swarm optimization algorithm to update model parameters for an instrumented thermal loop developed as a stand-in for liquid-cooled power electronics. The swarm of particles represent different configurations of a multi-domain model that constitutes the digital twin of the thermal loop. All computations are done on the edge to emulate a real world system. Results demonstrate that the particle swarm algorithm can reliably update a digital twin of the thermal loop as external changes are made to the system (radiator turned on and off) with a root mean squared error of under 0.35 °C over the whole system. All models are updated in real-time with a maximum compute time of 38.4 s; demonstrating the proposed methodologies applicability for real-time data assimilation within a digital-twin framework.