Applied System Innovation

High-mix manufacturing systems are vulnerable to stochastic equipment failures, changeovers and bottlenecks that deterministic capacity planning cannot adequately represent. This study develops a hybrid Markov discrete-event methodology for selecting bottleneck capacities in an unreliable serial–parallel multi-product foundry producing aircraft-engine components. A five-state discrete-time Markov chain of a bottleneck machine, derived from model inputs, estimates the stationary shares of processing, changeover, idle time, and two-phase repair and bounds the feasible configurations, whereas an AnyLogic model represents eleven operations, eight finite buffers, three product types, changeovers, and equipment failures. Replacing green-sand molding with cold-hardening mixtures and increasing the bottleneck from one to two parallel machines raised the bottleneck capacity from 15.1 to 26.6 workpieces per hour and reduced the shift-order makespan from 487 to 414 min, whereas a third machine added 5% capacity. Two machines maximized the constrained criterion of productive utilization, makespan, and resource intensity at every loading time, and the monthly plan was met for a cold-hardening yield of at least 0.63; a deterministic calculation overestimated the capacity by 7–49% and accepted a single machine, which, in simulation, completed only 20% of the shift orders within the shift. The chain reproduced the simulated processing and idle shares within 1.5% and, when excluding four of the seven machine numbers, reduced the simulation runs by 57% without changing the selection. Practitioners should validate the unit times, criterion weights, and yield against independent workshop data.

Economic model predictive control (MPC) of induction motor drives selects the flux from a loss model that usually omits saturation, additional load loss, the core-loss increment under pulse-width modulation, and the junction-temperature dependence of switching loss. A thermally adaptive economic MPC is extended with learned components, and each effect is isolated. The exact DC-link power balance shows that losses supplied through the air gap are unobservable without a torque sensor, whereas the additional load loss is observable; its estimate is passed to the stator node of the thermal observer. The magnetizing curve is learned from the reactive-power balance by a set-membership estimator contracted by monotonicity and concavity, and the resulting flux band tightens the current, reserve, and voltage constraints of the economic layer. Fifteen controller variants were simulated on 2.2, 7.5, and 37 kW drives on pump and conveyor cycles, including 600 Monte-Carlo samples. In nominal runs, the winding-temperature error fell from 5.5–10.2 K to 1.1–1.6 K and the torque error from 0.7–1.6% to 0.2–0.4% of rated torque. Across the samples, the stator voltage exceeded its planning limit in one case, against 13–36 per 100 without learning, and conveyor-cycle energy fell by 0.26–0.42 percentage points; pump-cycle differences were not resolved.

Generative-AI systems are increasingly trained on datasets containing AI-generated content, creating recursive feedback that may cause progressive losses of tail mass, diversity, and effective variance in learned distributions–a phenomenon commonly associated with model collapse. Existing analyses are primarily empirical or based on discrete-generation mixture models, motivating the development of a continuous-time dynamical framework. We develop a SIRS-type compartmental model, in which a normalized training-data pool is divided into clean, AI-polluted, and flagged/recovered fractions, with synthetic seeding ($\sigma$) and real-data preservation ($\rho$). We establish well-posedness, positivity, boundedness, and positive invariance; derive the equilibria and an explicit contamination threshold governing the clean equilibrium in the unseeded system; and analyze local stability and global convergence using planar dynamical-systems methods. Numerical simulations and parameter sweeps illustrate the resulting dynamics. In addition, a proof-of-concept fit of the reduced contamination–seeding subsystem to independently reported web-content prevalence data reproduces the observed post-2022 growth trend ($R^2=0.977$). Contamination ($\beta$) and synthetic seeding ($\sigma$) increase equilibrium pollution, whereas detection ($\gamma$) and real-data preservation ($\rho$) reduce it. Within the parameter ranges investigated, increasing the cleaning rate ($\alpha$) can indirectly increase equilibrium pollution by returning more data to the clean compartment, where it is again exposed to contamination. These results identify continuous real-data preservation and effective detection as key mechanisms for limiting synthetic-data pollution in generative-AI training ecosystems.