Towards an inline quick reaction system for actuator manufacturing using data mining

Large-scale production lines aim to realize 0 ppm defects. This is getting more and more complicated, due to all the so far achieved process optimizations. However, our research showed that a huge amount of unpredictable disturbance variables influences production systems, which promote defects. Here, the modelling of every single influence like temperature, machine condition, tool wear and quality of supplied parts is almost impossible, regarding a fully automated assembly line for actuators. Yet conventional methods for process optimization like Six Sigma, Kaizen, etc. usually focus on single processes and are not suited for quick reactions when disturbances occur during manufacture. Therefore, we created and evaluated a novel method based on data mining. To speed up failure detection, process data and testing results as well as batch information and new methods are required. This paper introduces an inline anomaly detection system to automatically highlight critical conditions with very low delay. Here, three independent systems analyze the data in order to detect jumps and outliers of process values and to find an anomalous distribution of defective parts within processes. For further investigations of detected malicious conditions an efficient root cause analysis for a whole production line including assembly and quality processes is introduced, which uses clustering and decision trees. Based on the detected anomalies of the system, we propose cluster algorithms to discover complex combinations of malicious process influences on the quality of the final product.