A Compact Deterministic Multi-Regime Stochastic Configuration Network for Industrial Soft Sensing Across Held-Out Operating Regimes

Random row splits can overstate industrial soft-sensor performance when an operating regime is absent from training. We propose a quasi-Monte Carlo worst-regime monitored stochastic configuration network (QMC-WRM-SCN), which maps a deterministic Sobol stream to group-sparse candidates and accepts a node only when every observed training regime satisfies an SCN residual-reduction inequality. On 14 ZeMA hydraulic single-held-state tasks, mean RMSE was 13.552, compared with 19.199 for a five-seed ExtraTrees ensemble and 28.751 for an official-code-compatible SCN. The one-sided sign-flip comparison with ExtraTrees survived Holm adjustment under its exchangeability assumption, but the direction-only sign test did not (9/14 wins; p=0.212), and tasks across condition axes reuse cycles. On five held-out gas-turbine years, the proposed model had the second-lowest mean RMSE (7.441 versus 7.354 for ExtraTrees); no corrected external comparison was significant. The originally reported 0.916-microsecond value is amortized batch throughput, not one-sample latency. Withholding two hydraulic states together reversed the ExtraTrees ranking, while an exploratory linear-versus-ensemble screen identifies plausible application boundaries without validating a universal threshold. The results support further study of a compact constructive model for the specific single-regime shifts evaluated here.