Machine-Learning Based Virtual Sensor for Thermal Management

Abstract

Battery Electric Vehicles (BEVs) are increasingly prominent in the transition toward sustainable transportation. However, they face challenges compared to internal combustion engine vehicles, including range anxiety, the relatively shorter technical lifetime of their battery cells, and the risk of thermal runaway. A critical component in addressing these issues is the thermal management system (TMS), which relies on accurate real-time measurements or estimations of coolant pressure and flow. Due to cost and space constraints, the number of physical sensors in the TMS is limited, and alternative methods for providing these key performance indicators (KPIs) must be explored. This thesis investigates the development of a machine learning (ML)- based virtual sensor to predict the KPIs using data from steady-state simulations of a digital twin modeled in GT-SUITE.

Five traditional ML algorithms and three artificial neural network (ANN) architectures were developed and evaluated based on accuracy, storage size, and, ultimately, inference time, to meet the computational constraints of vehicle control units (VCUs). The dataset, derived from GT-SUITE simulations, was preprocessed and split into training, validation, and test sets. Hyperparameter tuning was conducted to optimize model performance, and multiple architectures were explored, including models specifically designed to handle pressure and flow targets separately. To facilitate deployment in embedded systems, the final selected model was converted into automotive hardware-optimized C code for integration and testing in a simulated software-in-the-loop environment.

Among the models evaluated, ANN-based approaches showed strong performance in predicting the KPIs of the TMS. The multilayer perceptron (MLP) model, in particular, offered the best trade-off between accuracy, simplicity, and integration feasibility, making it suitable for embedded implementation.

The results demonstrate that ML models can effectively function as virtual sensors in TMS, offering a viable alternative where physical sensors are not feasible. MLbased sensing shows strong potential for VCU implementation, with the selected MLP model exhibiting characteristics favourable for successful integration into a BEV. The thesis also highlights ethical considerations, emphasizing the importance of model transparency and accountability in safety-critical automotive systems.