Advances in information and communication technologies for manufacturing applications enable the acquisition of huge amounts of various data from manufacturing processes. This opens the way for improving efficiency of manufacturing processes by data-driven analysis with machine learning. In most real-world applications, however, only a small portion of available process data has sufficient quality to be useful for data-driven analysis.
In this project, hybrid methods for fault detection and fault diagnosis (FDD) are proposed and investigated. Hybrid FDD methods combine data-driven analysis with further models derived using domain knowledge. This includes, e. g., physics-based models or models derived from typical knowledge sources in manufacturing such as FMEA documents. The proposed hybrid approaches to FDD aim to overcome the lack of sufficient data and to increase the FDD accuracy.
The developed approaches are validated with two manufacturing use cases provided by the industry partner of this project. The objective of the first use case is the data-driven quality prediction of a highly automated assembly line. The second use case implements FDD for servo-pneumatic welding guns via deep learning models as well as via two physical models as FDD approaches. Here, the data-driven and physical models are combined via a decision fusion framework. In both use cases, workers are also supported in correcting faults via a decision support system that is based on Bayesian networks.
Collaboration Partner and Funding
This project was funded by the Festo SE & Co. KG as part of the Graduate School of Excellence advanced Manufacturing Engineering (GSaME).