Abstract
An automated detection and diagnosis of machine and process faults is a core element of modern maintenance strategies. Data-driven methods, often summarized by the term machine learning, potentially provide manifold advantages in this area compared to conventional approaches. However, the necessity of a comprehensive database, which covers a variety of operation and fault cases, poses a remaining challenge for the application of data-driven methods. One way to mitigate the issue of costly data acquisition is to merge datasets from different, but related systems and applications to a larger dataset. In this contribution, the mixing of data obtained from a hydraulic press and a corresponding simulation model is investigated. For this, datasets comprising regular and faulty machine operation are generated in both domains. Using different mixing ratios of real and simulated data-instances, models for fault classification are derived and evaluated. The results show that an instance-based transfer from the simulated domain to the real domain can help reducing the amount of training data required from the physical machine. The mixing ratio appears to be decisive for a beneficial effect of mixing data from the two domains.
