Inference of Refrigerant Quantity in Car HVAC Systems: A Predictive Model for Leak Detection

This thesis addresses a critical issue in the automotive sector: customer dissatisfaction arising from complications with Heating, Ventilation, and Air Conditioning (HVAC) systems, notably Air Conditioning (A/C) performance reduction due to gas leakage. The primary objective of this study was to comprehend the behaviour of the air conditioning system under variable refrigerant levels, which facilitates detection of potential leakage. The initial study focused on understanding the HVAC system’s behaviour under various temperatures and identifying the minimal refrigerant charge that can guarantee performance equivalent to the nominal charge. This was done to establish the suitable range of charges for conducting experiments. The aim is to alert the client about potential HVAC system dysfunction before it occurs, based on this identified charge range. To compose the training set for a machine learning model, a comprehensive data collection strategy was employed, featuring a Latin Hypercube Design. The application of this method ensured a structured exploration of the multi-dimensional space of external temperatures, HVAC settings, and refrigerant levels that allowed for an optimized and systematic data collection of physical sensors. The collected data were segmented into smaller time series when there was a change in the HVAC configuration, with transitional effects and noise removed to focus on the stable behaviour of the HVAC system. The dataset was refined to minimize noise and emphasize the relevant information, excluding insignificant signals and consolidating highly correlated features into new composite features. The effectiveness of this feature consolidation was evaluated through a comparative analysis between the model using the original set of features and the model utilizing the combined features. The baseline model utilized an XGBoost regressor to infer the refrigerant quantity in the HVAC system under varying conditions, without the feature combination. This model achieved a Root Mean Square Error (RMSE) lower than 12 % of the considered charge range, reflecting a high level of prediction accuracy, serving as a performance benchmark for the combined feature model.