Evaluation of Machine-Learning-Guided Adaptive Sampling for Industrial Monitoring Systems

Machine learning (ML) can guide measurement allocation, but predictive accuracy does not establish operational value. We evaluate whether a bounded ML allocator improves industrial monitoring when a frozen variance detector retains alarm authority. The evaluation combines five run-grouped test folds on the Skoltech Anomaly Benchmark (SKAB), an illustrative two-event MetroPT-3 replay, and 1800 controlled policy runs. On SKAB, ML rollout achieved physical-event recall of 0.636 versus 0.697 for the smallest fixed policy; the paired 95% interval for the difference included zero. It required 5.985 additional observations per system epoch, an acquisition proxy rather than measured energy or monetary expenditure. Precursor skill varied across runs, and controlled tests exposed sensitivity to distribution shift. Delay-only break-even analysis identified a high cost ratio needed to favor ML. MetroPT-3 illustrates cadence sensitivity but cannot establish field reliability. The contribution is a reproducible seven-gate evaluation procedure linking precursor validity, detector suitability, resource use, uncertainty, and deployment decisions. The findings support withholding autonomous deployment of this specific univariate variance-monitoring configuration; they do not establish the performance of multivariate or distributionally robust allocation methods.