Supermarket refrigeration systems represent an important type of energy demanding appliances, which is in such widespread use that any development in the associated technology can have a huge impact on general health and global warming. Using automatic fault detection and diagnosis may for instance improve energy efficiency and reduce food waste as well as reduce expenses for the supermarket owners. In this paper, three modelfree classification algorithms are tested on faulty/non-faulty data obtained from an actual refrigeration system. It is found that support vector machines (SVM) are able to classify fan faults in a real refrigeration system with near-100% classification accuracy, independent of the number of input variables. The classification performance and robustness against an unseen operation mode, low-resolution data, noisy data, and data of different operating points is tested for three different classifier configurations. The results show Principle Component Analysis (PCA)-SVM is highly robust to different operating points, disturbances, and gives the best computational efficiency, as it is able to reduce the feature space to only two dimensions. It is concluded that while all of the examined methods are insensitive to noise, and effective in terms of detecting faults from relatively small amounts of data, overall, PCA-SVM is slightly more computationally efficient.