Industrial models fail in boring ways: they overreact to ringy sensors, chase spikes, and burn budget on nonsense features. A common fix is “more model.” A cheaper fix is often cleaner inputs—with honesty about what the cleaner cannot do.
Where GSRF fits in the stack
Sensor → GSRF (calm / compress) → features / model → decisions
Detector / alarms → separate path (do not force GSRF to be the FA engine)
Locked strengths that matter before inference:
- Mid-band ring kill — ~+35% osc31 and ~−70% sweet-band spectral power vs EMA on locked actuator-style packs
- Peak compression toward a declared normal — ~38% mean peak-dev advantage; 49/52 wins
- Adaptive normal when operating point permanently moves (frozen x* fails; adaptive helps)
What not to sell the model team
- “This replaces your residual alarm logic.” Fixed-FA residual recall favored EMA (~0.83 vs ~0.44) on synthetic excursions.
- “Drop this into delayed plant feedback as the command path.” Delay packs show large MAD-to-measured regressions.
- “It always tracks better than EMA.” Tracking MAD: 0/52 wins—by design.
Recommended evaluation path
- Run the evaluation snippet on a calm-prefix x* and a ringy channel.
- Score oscillation / peak metrics that match cost (not only MAD).
- Shadow GSRF offline before live shaping.
- Keep detectors on their own path; use GSRF to reduce garbage into the model.
Who this note is for
PdM platforms, industrial AI feature pipelines, and OEMs embedding a validated “calm then infer” block—with a published safety cage. See customers for packaging tiers.
Sources: Evidence · customers industrial pitch · excellence hunt v2 · phase6.2 PSD · probes v5. /evidence · /customers