Physio-Semantic Compute Orchestration in Cyber-Physical Systems: A Mathematically Constrained Multi-Agent Edge-Cloud Architecture for Real-Time Manufacturing Intelligence
Main Article Content
Abstract
AI-enabled diagnostics in industrial cyber-physical systems (CPS) increasingly combine lightweight edge models with large cloud-hosted language models. Existing model-routing methods choose among models by estimating the difficulty of a query and trading response quality against monetary cost. In a machine that is degrading, however, the consequence of a late answer depends on the physical condition of the machine, which a query-only router cannot observe. This paper proposes a physiosemantic orchestration formulation in which the physical state of the machine enters the routing decision explicitly. Each diagnostic event carries a semantic descriptor and a physical state estimate; the physical state determines a response deadline and an effective age under a Weibull degradation model, and the penalty for delay is the conditional probability of failure during the response interval. Routing is posed as an expected-cost minimization over a heterogeneous model set spanning edge, fog and cloud tiers, with terms for cost, diagnostic quality, statedependent degradation, deadline risk and live queue state. We state the assumptions under which the formulation holds, show by a standard decisiontheoretic argument when physical state has strictly positive value, and describe a three-tier multi-agent architecture in which cloud outputs remain advisory. The formulation is evaluated by discrete-event simulation on three synthetic surrogate scenarios structured after high-speed CNC machining, industrial acoustic anomaly detection (MIMII-structured) and turbofan degradation (C-MAPSS-structured), using 30 replications of 4000 events each and six baselines. In these simulations the proposed router reduced the mean realized objective relative to a semantic-only router by approximately 21% in the CNC scenario and by approximately 4% and 2% in the acoustic and turbofan scenarios, respectively, and the difference vanished when physical state was made uninformative. Queue-blind routing produced unbounded backlog. The pre-specified targets of a 70% reduction in latency and cost with 98% accuracy retention relative to cloud-only routing were not jointly met in any scenario. The results illustrate the behaviour of the formulation under stated synthetic assumptions and are not evidence of performance on real machines; validation on the original datasets and on hardware is identified as future work.