BLOG
Virtual Sensors and AI Enabling "What If" Scenarios in Thermal Management
Advancements in sensor technologies, digital twins and artificial intelligence (AI) in recent years are now converging to provide dynamic, real-time temperature monitoring and adaptive thermal controls across various applications. The automotive sector has been an early adopter, but embedded predictive AI is now being applied to a broad range of applications – from data centers to building management and manufacturing plants.
Last year, the closing session at the Thermal Management Expo conference explored how real-time embedded AI can optimise on-board thermal management. One of the speakers representing the end-user perspective on the panel was Daniel Szepanski, Head of Software Development Electrification at SEG Automotive, who highlighted how critical thermal management is for their electrical motors. SEG's products span internal combustion, hybrid, and e-drive applications, each with different heat management and protection requirements operating within millisecond to microsecond time constraints.
Szepanski emphasized the distinction between the use of digital twins in the design phase versus real-time monitoring. While design relies on complex FEM analysis and CFD simulations that demand significant runtime, these cannot operate on embedded microcontrollers. This necessitates alternative approaches for real-time monitoring. The Italian startup Newtwen, represented on the panel by co-founder Francesco Toso, provides an engineering platform that generates and deploys virtual thermal sensors and fast-running embedded AI models.
The combination of sensors, digital twins and embedded AI can be employed to optimise thermal management for many applications, from data center cooling to building management, manufacturing plants, and industrial applications, but automotive is leading the field. Daniel Szepanski remarked in the session that although the automotive sector has longer experience with the need for thermal prediction, that experience can also be a hurdle to introducing new ways of doing things.
‘Especially with artificial intelligence trained on data,’ Szepanski said, ‘there's often huge skepticism because people are afraid something unpredictable might happen when they don't understand what's behind it.’ Alessandro Fauda, R&D Manager at VHIT, agreed:
‘The market is ready, but there is fear,’ he said, ‘because there is pressure to reduce cost. Engineers asked to eliminate physical sensors must trust the virtual ones.’
Alessandro's team generally proposes customers to do the validation together and compare the physical data coming from the physical assets with the virtual data. Virtual sensors can also replace physical sensors where two sensors are required for safety reasons.
Daniel Canchola, Portfolio Development Executive at Siemens Digital Industries Software, argued for ‘a hybrid approach’ to win over skeptical engineers, by using ‘a mixture of the physics the engineers trust combined with artificial intelligence.’ Francesco Toso agreed that ‘hybrid solutions are always the best.’ Finding the right people who understand AI implementation for production is a huge problem. Skilled engineers do great work, ‘but when asked to apply the same approach to different components like inverters instead of motors, manual work needs to restart because inverters differ from motors and require different knowledge.’
Software platforms like those from Newtwen and Siemens encapsulate all this know-how and technology with a user-friendly interface and make it available to engineers who are not computer science or AI experts. ‘The market is ready in the sense that there's a willingness to reduce costs,’ Newtwen CEO Francesco Toso said. ‘What's missing is know-how and knowledge inside companies. There's still work to be done with software tools. Many companies manually develop digital twins with smart people, but to scale that across all products requires significant investment.’
Application of virtual sensing should be more than just ‘a bill of material reduction’, according to Francesco Toso on the panel last year. Cost is an important consideration, but a system of virtual sensors can do much more. ‘It's a model that can make predictions, that can support “what if” scenarios that real sensors cannot. We have to consider virtual sensing not just as a substitute for real sensors, but as a complement.’
Anthony Vivek works as a CFD engineer for AFRY, a Swedish-Finnish engineering, design and advisory services company with over 17,000 employees and a global presence. His team has implemented machine learning for optimising thermal management in various applications including heat exchangers, batteries for electric vehicles and data centers. A recent project focused on predictive models and control systems for building heating, ventilation and air conditioning (HVAC) applications that can maintain optimal indoor climate conditions under varying environmental circumstances.
‘For example, for a café to maintain the perfect temperature on a sunny day with windows facing the sun, or if it's cloudy, or if there are more people,’ Anthony explains. ‘We have a test room where we train a predictive model on numerous CFD simulations under different conditions – a sunny summer, cold winter, dark, gloomy and rainy, Swedish weather.’ The CFD model, taking readings from sensors in the room, controls the HVAC to keep the climate inside the room at the perfect target.
The model is trained on close to 900 different points with 8 varying input parameters. The training incorporates the underlying physics, resulting in a model that is not just predicting values blindly but also tries to understand the correlation between various sensor inputs and fluid dynamics equations. The model can replace CFD simulations to a certain degree. If the mass flow changes by 10%, the HVAC controller can run the predictive model instead of a costly CFD simulation and make adjustments in less than half a second.
Anthony Jayanath Vivek, Development Leader in AI and CFD at AFRY, and Francesco Toso, co-founder and CEO at Newtwen, will provide more insight in AI-optimised real-time thermal management on a webinar on Wednesday, July 9, moderated by Farhad Nazar Pour, AI Strategy and Promotion Lead at Audi.
Register for free here.
