Fault-Tolerant Verification Architecture for Safety-Critical Learning-Based Control Systems

The adaptability of cyber-physical systems (CPS) can be enhanced through learning-based control, while prediction errors, distribution shifts, and faults make safety assurance challenging. This study evaluated a fault-tolerant verification architecture that separates learned-model trust, fault-relevant evidence, and physical-safety outcomes. The architecture combined nominal next-step dynamics with a trust gate based on state-distribution support, actuator-pattern support, historical sampling-interval support, and forward Δt support, together with physics-informed specialist evidence, redundant-sensor consistency, temporal constraints, and an observation-based Simplex supervisor. The benchmark comprised 55 runs, with nominal data separated into training, validation, and test sets, while out-of-distribution conditions and 13 real-fault runs were reserved for evaluation. Ridge achieved the lowest validation macro-scaled RMSE among the evaluated predictors (0.118233). The frozen real-benchmark verification track retained fault-relevant evidence in 11 of 13 fault runs (84.6%), while none of the seven held-out nominal runs generated fault-relevant evidence or fallback requests. In controlled simulation, Simplex supervision prevented 7 of 8 unprotected Modelica-domain boundary crossings in the repository-grounded reduced-order simulator (87.5%; exact two-sided McNemar p = 0.015625). Performance decreased at higher measurement-noise levels, and severe leakage remained unprevented because the one-step predictor provided no advance lead. The learned real-benchmark predictor and simulator supervisor were evaluated on separate tracks because real-to-Modelica scaling remained unresolved; therefore, the study does not provide end-to-end closed-loop validation of the complete architecture.