Refrigerant charging discrepancies constitute the predominant malfunctions in air conditioning systems. Achieving the optimal charging level is crucial for system performance, underscoring the importance of precise refrigerant level prediction. This study introduces an algorithm designed for the quantitative detection of refrigerant charging errors by integrating the Markov Transition Field (MTF), Convolutional Neural Networks (CNN), and Multi-head Self-Attention (MSA) mechanisms. A high-precision enthalpy difference chamber was employed to establish a Variable Refrigerant Flow (VRF) refrigerant charging test bench. This setup facilitated the analysis of system parameter sensitivity to charging faults and aided in the creation of a training dataset for the algorithm. Comparative analysis was conducted against Support Vector Machines (SVM), Random Forests (RF), CNN with Self-Attention (AT), and MTF-CNN-MSA. The findings reveal that our method adeptly captures temporal dependencies and dynamic shifts in time series as visual representations, offering novel insights for discerning fault patterns within such data. Notably, the maximum pressure variations at high-pressure and low-pressure points were 0.25 MPa and 0.07 MPa, respectively, with temperature shifts of 12°C and 3.5°C at the high and low-temperature points. The high-pressure and high-temperature points are particularly sensitive to changes in refrigerant charging, and parameters from these sections were utilized to construct the dataset. The CNN-MSA algorithm demonstrates consistent performance across various fault types, effectively delineating fault characteristics. The accuracies achieved by SVM, RF, CNN-AT, and MTF-CNN-MSA were 84.38%, 73.75%, 88.13%, and 93.75%, respectively. In comparison, the CNN-MSA algorithm was able to more accurately detect refrigerant charge faults at different levels.