Articles

Mevronic Neural Networks for Explicit Reliability Regime Control. A Semantically Anchored Architecture with Analytical Properties and an Evaluation Protocol

/0 Comments/in /by

Autonomous systems face a difficult problem when the information they rely on becomes unreliable: they must decide not only what to predict, but also how much confidence to place in those predictions. This paper introduces Mevronic Neural Networks (MNNs), a design approach that connects a neural representation network with a persistent state describing its operating reliability. Rather than proposing a new computational architecture, the focus is on making the relationship between incoming evidence, stored information, network behavior, and operational decisions more explicit. The proposed model uses three main state variables representing estimated prediction correctness, accumulated anomaly evidence, and sensor quality. These variables influence the system’s alertness and determine how much adaptation is allowed. Their meanings are defined in advance to make the mechanism easier to examine, although the estimates used to update them may still contain errors. The paper explains how the system collects evidence, maintains its internal state, and modifies network features within specified limits. Then it switches to alternative responses when conditions require it. The mathematical results confirm that the state variables stay within their defined boundaries and show how earlier observations influence subsequent updates. They also provide an exact expression for the time needed to reach a given threshold when the incoming evidence remains constant. A numerical example helps illustrate a practical compromise: smoothing reduces rapid fluctuations in the system’s response, but it can also delay reactions to changes in the environment. These results describe the behavior of the regulatory component and do not, by themselves, demonstrate that the complete system is safer or more accurate. To examine whether the proposed approach is useful in practice, the paper presents an experimental plan based on synthetic multisensor tracking. The plan includes comparisons with models given equivalent information, separate datasets for training and testing, and measures that assess both prediction quality and operational behavior. The discussion also considers reference with established techniques, such as recurrent gating, feature-wise modulation, hypernetworks, neuromodulated recurrence, differentiable plasticity, and runtime assurance. Whether MNNs provide sufficient practical benefits to justify their additional complexity remains an open question that requires experiments with trained models and validation in real-world settings.

Data-Driven Robust MPC for an Industrial Evaporator under Model Uncertainty and Sensor Degradation

/in /by

Model predictive control is attractive for industrial processes because constraints and multivariable interactions can be handled explicitly, but its performance depends on the credibility of the prediction model and the measurements supplied to the controller. This article develops a data-driven robust MPC framework for an industrial evaporator in which model uncertainty and sensor degradation are treated jointly. The open DaISy industrial evaporator benchmark (code 96-010) supplies the data basis: 6305 samples, three manipulated inputs and three measured outputs. Publicly reproducible system-identification results for this benchmark report a first experiment of 3300 samples, an order-eight prediction-error model, and five-step output mean-square errors of 0.0576, 0.1564 and 0.0193. These error scales are used to define relative residual uncertainty for a closed-loop benchmark-constrained surrogate. The controller combines a trust-weighted state estimate with scenario-based robust prediction across a finite uncertainty ensemble. Bias, drift, precision loss and dropout are injected into the sensor channels, and at the same time plant matrices are perturbed independently across Monte Carlo trials. In 200 composite-degradation scenarios with 12% plant-model uncertainty, the proposed robust MPC reduces mean normalized tracking RMSE from 0.2750 for nominal MPC to 0.2421, and reduces normalized performance-envelope exceedance from 14.12% to 6.02%. The improvement is obtained at a 6.75% increase in average control effort. The results show that a practical data-driven robust MPC architecture should not treat identification error and sensor integrity as separate problems; both must enter the online state estimate and the predictive optimization layer.