Evaluation And Adaption Of Maintenance Prediction Methods In Mixed Production Line Setups Based On Anomaly Detection

Mixed production line setups are common in real world scenarios but are sparsely covered by current research in anomaly detection. A mixed production line setup poses several challenges and properties that make using anomaly detection more complex compared to monotonous production processes. We consider a high variation in the produced products, diluted temporal dependencies, imbalance in the frequency of product types and sensor measurements differing based on the produced product type. We gather contextual information using the OPC UA standard and extract information such as running program name.In this work we adapt and evaluate common anomaly detection methods to such a scenario. By building ensembles of anomaly detectors, we account for different setups. Each single work-piece gets evaluated for anomalous behaviour. We evaluate the anomaly detection by using data from a mixed production setup in the automotive domain. The context based approach is compared to a sliding window approach, which we use as a baseline.The results highlight that our product type-based approach shows a higher precision and recall for all applied detection techniques, by utilizing contextual information. Our experiments additionally show that in the selected industrial case study our approach achieves good results even when only limited data per product type is available.