The different types of PCBs each have their own test sequence with different test steps that are run through. Pseudo faults are represented or pronounced with varying frequency on each type of PCB - as a result, we now have one model per PCB in operation. Due to the data situation, only model classes that can learn well with few examples come into question. We test different ensemble methods, SVMs and simple decision trees to find the best model for each PCB. However, all models learn the same classification task: test data of a PCB -> LP defective, retest, test station error?