Enhancing Data-Driven Models with Knowledge from Engineering Models in Manufacturing

Abstract

Data-driven models of production plants used for anomaly recognition usually require long learning periods to obtain the normal production state of the equipment. Some evaluation methods are based on correlations which may be spurious correlation rather than a causality. In the meantime, during the system design of a plant a high amount of knowledge regarding the system behaviour and the interconnection of engineering objects is specified. Recent advances in the engineering process allow the usage of vendor-supplied behaviour models of mechatronic components during the process, adding detailed knowledge about components from their vendors in a standardized model format. This work proposes to use this a priori knowledge, which is a spin-off from the engineering phase, to reduce the training time and improve the meaningfulness of statistical models by adding causality information and providing a possibility to train the models.