If impending failures, malfunctions or defects in technical systems such as aircrafts, power plants or production facilities are not detected at an early stage, they can lead to significant safety risks for people and the environment as well as financial losses. By analysing sensor data and control commands in combination with expert knowledge, it is possible to determine the current condition of a system and predict its remaining useful life. This is referred to as predictive maintenance and helps to perform maintenance and reparation in a timely and effective manner. To investigate approaches of artificial intelligence for this purpose, I use the following Fischertechnik factory model to simulate a cyber-physical production system (CPPS) and its wear processes as well as complex failure modes (Paper). A comprehensive description of the factory simulation (e.g. type affiliation, component hierarchies, relationships betw. actuators and sensors, ...) was modelled by experts using Web Ontology Language (OWL) (Paper, Repo). My main research interest lies in the integration of such expert knowledge into machine learning models for improving the prediction performance. Therefore, I apply a special siamese neural network architecture, which allows an expert to select relevant data streams for each failure mode (Paper, Code, Data). At the moment, I am working on an anomaly detection based on an autoencoder, which can determine the data streams of the abnormality origin.
- Deep Learning
- Case-based Reasoning
- Predictive Maintenance
Supervised Theses & Student Projects:
Professur für Wirtschaftsinformatik II