Safe MES-driven execution control for advanced manufacturing systems using constrained matrix zonotopes

Manufacturing execution systems (MESs) coordinate release, transfer, inspection feedback, maintenance lockout and device commands. A command that is locally feasible, however, may still violate production rules or physical limits at the system level. We study this execution-safety problem using constrained matrix zonotopes (CMZs). Audited MES rules are converted into finite Boolean combinations of bounded linear predicates over states, actions and production modes, and the resulting cells are checked with robust set-containment constraints. A projection-based filter selects the action finally sent to the equipment; semantic or language-enabled modules only supply candidates. All methods are evaluated on the same anonymized field replay from three electronic-production lines covering electronic modules and 48\,V power adapters over 42 production days, with 6,427 station-level decisions in the reported split. The filter obtained an empirical replay-level physical-safety rate of 0.985, a semantic violation rate of 0.058, an erroneous-release rate of 0.012 and a recovery time of 23.6± 6.1 s. Relative to pure CMZ filtering, it reduced semantic violations, manual intervention and recovery time, whereas pure CMZ retained slightly higher empirical physical-safety and erroneous-release performance. The proposed method is therefore interpreted as a more balanced semantic–physical execution trade-off rather than as uniformly superior on every metric.