The RACKET project addresses the problem of detecting rare and unknown faults by combining model-based and machine learning methods. The approach is based on the assumption that a physical or procedural model of a manufacturing plant is available, which is not fully specified and has uncertainties in structure, parameters and variables. Gaps and errors in this model are detected by machine learning and corrected, resulting in a more realistic process model (nominal model). This model can be used to simulate system behavior and estimate the future characteristics of a product.
Actual product defects can thus be attributed to anomalies in the output signal and to inconsistencies in the process variables, without the need for a known failure event or an accurate failure model. Errors have a wide range, i.e., geometric errors such as scratches, out-of-tolerance dimensional variables, or dynamic errors such as deviations between estimated and actual product position on a conveyor belt, process steps or incorrect path assignment in the production flow, etc., and can occur at the product and process level.