Where does GSRF sit in a PdM ML pipeline?

Between raw sensors and features/models: Sensor → GSRF (calm/compress) → model. Keep residual/alarm detectors on a separate path.

Can GSRF replace my false-alarm detector?

No. Locked fixed-FA residual probes favor EMA residual recall. GSRF is not sold as a fixed false-alarm event detector.

Does GSRF improve every industrial valve or tag?

No. Synthetic Phase 5 batteries are regime-bounded. DAMADICS real-plant first look published 2/3 partial and 1/3 loss. Validate on your tags.

How do I try GSRF on my SCADA export?

Export a positive CSV, run gsrf-bench or the Try Python snippet, compare osc31/peak/MAD vs EMA, then read the safety cage before production.

GSRF for Predictive Maintenance Pipelines

Predictive maintenance and industrial AI models often fail on boring input problems: ringy sensors, spike chase, and features that encode noise the model should never see. More model is not always the fix. Cleaner inputs with an honest safety cage can be.

Stack placement

Sensor / SCADA tag → GSRF Practical (calm / compress) → features / model → decisions
Alarms / residual detectors → separate path (do not force GSRF to be the FA engine)

GSRF is a deterministic log-space soft thermostat: spring k_return toward declared normal x*, tanh-bounded observation pull. It is not a linear low-pass rebrand.

What locked packs support for PdM

Mid-band ring kill — ~+35% osc31 vs best EMA on phase6.2 locked packs; ~−70% sweet-band spectral power
Peak compression toward a normal — ~38% mean peak-dev; 49/52 dataset wins in excellence hunt
Adaptive x* when operating point permanently moves (frozen x* sticks; adaptive helps multi-regime industrial-style traces)
Industrial batteries — synthetic Phase 5 SCADA-style actuators under frozen rules; DAMADICS real-plant first look with partial/loss published

Value without fantasy ROI

Pre-model hygiene is a cost class (cleaner features, less thrash on ring) — not a promise of “3× training” or cloud GPU $. Measure train time yourself on a fixed recipe. For valve/joint mechanical motion, see the flagship actuators guide.

Red lights for PdM teams

Not a residual event detector at fixed false-alarm rate (EMA residual recall higher on locked probes) — do not budget man-hours saved on residual FA from GSRF
Not a MAD tracker of raw (0/52 wins vs best EMA) — by design of the spring
Not a multi-minute delayed command smoother if you close the loop on filtered commands with large plant delay
Not “wins all plant valves” — DAMADICS multi-day hardness + independent overall limits; pathology map on actuators

How to evaluate on your tags

Pick strictly positive sensor series (or offset to positive) with a known “normal” or calm prefix for x*
Run Try snippet or local gsrf-bench on CSV (osc31 · peak_dev · MAD · spectral)
Compare against EMA baselines you already trust
Keep alarm logic on a separate residual design — do not treat GSRF residual as the product claim
Read safety cage before production license

Verdict

Use GSRF in PdM when mid-band chatter and peak chase pollute features. Keep detectors and tight tracking on tools built for those jobs.

Deep dives: Pre-AI hygiene note · Pre-model stabilizer · Evidence industrial · Citations

GSRF for PdM pipelines — calm sensors before models