In recent years, 3D printing has seen a stellar rise despite its inability to deliver constant quality goods. This article presents a machine learning experiment that results in a model performing fault prediction, in the sense of forecasting the fault, on the printed parts so that printer parameters can be corrected before the faults appear. This model is able to predict faults in real-time during printing, even in the case of multiple faults. It relies on multiple sensors gathering time-series data during printing, a pre-processing of these data to extract the most relevant features and several machine learning algorithms, each suited and tuned to predict at best each fault. A benchmark for testing and tuning the different algorithms is presented. The resulting model has been implemented on a plastic delta 3D printer and tested for the prediction of eight different faults. The best performing model is a random forest, but decision trees are almost as good while explaining what causes the fault.