Machine Learning for Predictive Maintenance in Production Environments

System condition-based predictive maintenance is an important factor for the availability and performance of complex machines and offers cost-saving potential due to shorter and better planned outages. For this purpose, complex models are commonly used to detect possible faults and degradations in an early stage. These models need information about the faults and all operating points as a training set. Often these data sets are not available or sufficient, so that alternatively training data can be generated artificially. For well-researched rotating machines, scientific literature, which provides information about the modeling of different kinds of damage conditions, can be used to achieve this. We develop a demonstrator grid and a system for detecting degrading and sudden faults on rotating systems. This requires data to be recorded from the sensors (acceleration, noises, etc.) under variation of the speed on the demonstrator. Afterwards the training set for the diagnosis models has to be generated using a simulation of a set of possible faults. For this, the identified system and modeling of the possible faults is needed. The models have to be implemented, trained on this data set, and evaluated on real but manually caused damage data at the demonstrator. The aim is to investigate the extent to which models trained with simulated data can be used for diagnosis in reality. Also, different kind of models will be compared regarding their performance.