A Virtual CO2 Sensor for Vehicle Cabins Using Deep Learning and Physics- Based Modelling

Abstract

The air quality inside vehicle cabins is an important factor affecting driver alertness, occupant comfort, and occupational health. Among the constituents of cabin air, carbon dioxide (CO2) is of particular interest because it is produced continuously by occupants and accumulates rapidly in the enclosed space of a modern vehicle. Sustained exposure above 1,000 ppm measurably degrades driver alertness and decision-making, creating a clear need to monitor cabin CO2 during operation. Physical CO2 sensors are the conventional approach and are already used in some vehicles to trigger ventilation switching, but they rely on infrared or photo-acoustic measurement principles that are susceptible to drift from temperature and humidity, require periodic recalibration as optics age, and degrade under the vibration and thermal cycling of vehicle operation. These limitations render them unreliable for continuous, fleet scale monitoring. This thesis addresses this gap by developing a virtual CO2 sensor that predicts cabin concentration from signals already available on the vehicle controller area network (CAN) bus, eliminating the dependence on dedicated hardware. Three modelling approaches are developed and compared: a purely data-driven deep learning model, a physics-based mass-balance model, and a hybrid approach in which the physics model output is incorporated as an input feature to the data-driven model. Seven architectures are evaluated, trained, and tested on 445,017 samples corresponding to 123 hours of real-world driving data collected across cold, mild, and warm ambient conditions. The best performing architecture is Hybrid TCN-LSTM, with a mean absolute percentage error of 8.35% using only CAN bus signals. Adding the physics-based equilibrium signal as an input feature reduces this to 7.93%, with consistent improvement seen across all seven architectures and all temperature conditions. The physics signal captures ventilation history in a way that raw CAN bus signals do not directly express, and its contribution shows that the two approaches carry different information. The results establish that CAN bus signals alone are sufficient for accurate cabin CO2 prediction, and that incorporating physical knowledge into the model pushes that accuracy further still.