Estimating the real-time value of state variables is crucial in various industrial embedded applications, such as automotive and aeronautical systems. The most straightforward way to obtain such information is through physical sensors. However, their implementation can be both costly and complex, prompting the use of models as substitutes, often referred to as virtual sensors. In this context, artificial intelligence (AI) has recently demonstrated its suitability for designing these advanced solutions. In this work, an embedded machine learning (ML) algorithm is explored as a virtual sensor to predict the actual temperature of an automotive power electronics component, thereby ensuring effective thermal management. This methodology is proposed as an alternative to conventional modeling approaches based on simplified physical laws, which are often used in the automotive industry, such as in internal combustion engine or electric machine controls. The study outlines each phase, from data preprocessing to prototype testing, including feature engineering, model design, and the embedding process. Several ML models, including linear regression and neural networks, are evaluated and demonstrated to be excellent and relevant alternatives to traditional modeling in terms of both offline and online performance.
Lamarque et al. (Tue,) studied this question.