Air conditioners have played an important role while providing comfortable air quality in buildings. However, the performance of the system may significantly deteriorate due to improper refrigerant charge or refrigerant leakage during operation. Conventional refrigerant charge diagnosis often relies on invasive measurements, complex thermodynamic modelling, or costly sensor configurations, limiting their applicability in large-scale field deployment. This study proposes a simplified, data-driven refrigerant leakage detection framework using only easily accessible measurements from the outdoor unit. An air conditioning unit is tested in a dedicated facility to characterize the system behavior under various refrigerant charge conditions. Power consumption and refrigerant pipe temperatures measured at the outdoor unit are employed as input parameters to ensure accessibility and cost-effectiveness for extensive field implementation. The most sensitive inputs are analyzed using neighborhood component analysis (NCA), while artificial neural network (ANN) is applied for defining classification models of different refrigerant charge conditions. The ANN model was optimized using a grid-search procedure with evaluating multiple input parameter combinations on independent training and testing datasets. The results demonstrate that refrigerant charge can be accurately estimated using the optimized ANN model, achieving a confidence level exceeding 96 %. © 2026 Elsevier Ltd