Adaptive Error Prediction for Production Lines with Unknown Dependencies

Abstract

Forecasting or predicting errors can dramatically reduce the downtime of machines in industrial settings and even allow to take counteractions long before the error affects the production system. A forecast system to predict upcoming critical values for identical production lines under different environmental circumstances is proposed. We focus on errors that result in multiple erroneous work pieces. These error patterns need manual corrections by a machine controller. An analysis of the system observed gathered the information about the types of errors that are observable. 30% of errors are measurement errors or single faulty work-pieces which are not influenced by previous work-pieces and do not show any indication to preceding work-pieces. These errors do not need any type of action by the machine controller. 70% of the observed errors are continuous system deviations which lead to multiple erroneous work-pieces in order or a high percentage of erroneous work-pieces in an observed time frame. We observe multiple production lines which consist of identical machines and produce the same product type. For the forecast of errors, we use the ARIMA, Holt and Holt-Winter method. Each production line and product type combination showed different results for the different forecast methods. We implemented a dynamic system that automatically detects the seasonality and trend of the specific combination to assign a correct forecast method and model. For 40 combinations of production line and product type the holt-winter algorithm performed best for 14, the holt-winter without seasonal or trend component performed best for 13 combinations and the holt-winter with only a trend component performed best for 10 setups. 3 combinations did not have a distinct best method for all observed results. By selecting the correct forecast methods, we were able to boost the forecast accuracy for the overall system over each single forecast method.