A local industrial-agent middleware for resilient manufacturing execution systems under weak external connectivity

Abstract

Industrial agents can retrieve shop-floor knowledge and propose business actions, but a probabilistic proposal should not directly acquire authority over manufacturing execution system (MES) state. This paper develops a plant-side industrial-agent middleware for MES recovery under weak external connectivity and cloud-service disruption. The methodological contribution is a deterministic execution contract that combines current MES preconditions, role-checked user attribution, versioned local retrieval, manual-state protection, reversibility-aware compensation, and auditable invocation of registered MES services. The implementation uses C#/.NET and Microsoft semantic kernel with Redis, InfluxDB, MySQL, local Qdrant, and Docker/K3s. A single-site field evaluation on three computer, communication, and consumer-electronics power-supply lines used 302,418 MES records, 9.24 million industrial internet of things time-series points, 47 283 vector records, and 500 chronologically selected recovery cases. Failed, blocked, and downgraded outcomes remained in the cohort. The middleware restored consistency in 477 cases, corresponding to a consistency restoration rate of 0.954 (Wilson 95% interval: 0.932–0.969), a secondary damage rate of 0.082, and an invalid reversal rate of 0.014. Mean end-to-end recovery latency was 1.53 s and the 95th percentile was 2.38 s. Controlled service-degradation tests further examined local retrieval, runtime-state reconstruction, and fail-closed behavior. The evidence supports bounded local automation in which agent reasoning remains available while the MES retains sole authority for side-effecting execution.