Beyond the detection limit: an event-anchored, fixed-reference detector for slow-onset energy faults in MES-embedded carbon-intensity monitoring

Real-time, MES-embedded carbon-intensity monitoring is increasingly deployed to surface energy anomalies in discrete manufacturing. A companion study showed that its rule-based form has a structural blind spot — the adaptive-baseline inertia trade-off: because “normal” is judged against a rolling recent-history baseline, a slowly developing fault is tracked by the reference and never alarms. This paper closes that blind spot and isolates why the fix works. The remedy is to replace the tracking reference with a fixed one — an event-anchored baseline pinned to a known-healthy MES event — optionally with a residual CUSUM. Our central result is a controlled ablation holding the detector statistic identical and varying only the reference: a CUSUM on a rolling reference still collapses, whereas two fixed references (an event-anchored baseline and a global calibration constant) both recover detection. The mechanism is therefore the fixed reference, not the statistic. Because a persistent above-floor fault accumulates without bound against a fixed reference, recovery is near-deterministic above the floor, so we foreground the detection floor, warning latency, and a false-alarm-constrained operating point. We identify and remedy the method’s own failure mode (a health gate prevents re-anchoring onto an active fault), show the result generalises across severity and fault morphology, and validate it on a semi-synthetic substrate built from real CNC spindle power (real-substrate, not real-fault). A robustness programme maps the result’s boundaries — legitimate regime changes, contaminated anchors, negative deviations, production interruptions, regenerative-braking treatment, and sensor faults — and positions the detector as one monitoring layer , an anomaly trigger rather than a diagnosis. Detection rates carry binomial (Wilson) intervals and false-alarm rates exact Poisson intervals throughout. All code and results are openly archived, and a live pilot on a naturally occurring fault is identified as the primary next step.