Optimized state estimation by application of machine learning

The requirements concerning the technical availability as part of the overall equipment effectiveness increase constantly in production nowadays. Unplanned downtimes have to be prevented via efficient methods. Predictive, condition-based maintenance represents a valuable approach for fulfilling these demands, but precise models for state estimation are missing. From the manufacturers’ point of view the challenge consists in wear models with the capability of specifying the correct component’s state as well as providing reliable failure forecasts. Unfortunately, nowadays creation of wear models is based on specific stress tests or design of experiments from the manufacturer. The integration of the production phase or even data feedback and user knowledge does not take place. New potential is promised by cross-cutting technologies from ICT like cloud technologies—in general virtual platform concepts—or approaches of machine learning as enabling technologies. The objective of this paper is to adopt existing algorithms to the new application of condition monitoring in order to evaluate the applicability for automated training of robust wear models. In that context the most commonly used algorithms are described and the reader gets an impression what challenges have to be met when dealing with machine learning. A selection of about ten algorithms with 45 variants is evaluated for four different features within a packaging machine. In the outlook the embedding of the trained model in a cloud architecture is presented.