Probabilistic graphical models for performance diagnostics : Methods and applications to a print quality case
report
This report introduces a methodology for diagnosing performance issues in high-tech production systems. The approach uses probabilistic graphical models to combine machine data with expert and design knowledge. By reasoning across multiple representations of the data, multiple resolutions, and across time, the approach can interpret defectivity patterns and infer their actionable root causes. Validation on an industrial inkjet printing case, focused on diagnosing patterns of non-jetting nozzles resulting in print quality issues, shows that the method achieves high diagnostic accuracy (86%) and aligns well with experts’ expectations. The study presented in this report highlights the importance of scoped modeling for such complex diagnosic tasks, as well as the need for multi-representation reasoning, and expert involvement, and identifies opportunities for future improvements such as automatic learning of model parameters and handling continuous (random) variables.
TNO Identifier
1023296
Publisher
TNO
Collation
32 p.
Place of publication
Eindhoven