GSRF vs EMA vs Kalman
Everyone can smooth. Almost nobody will publish a domain map of where their smoother loses. We do — and we’re a soft thermostat, not a tracker.
Competitive posture (claim-safe)
- vs EMA: EMA is free and correct when tracking is the job. GSRF is a spring to a declared normal \(x^*\). Different tool.
- vs Kalman: Kalman needs a process model. GSRF doesn’t. Different job (state estimation vs soft thermostat).
- vs your PLC: Siemens / Beckhoff-class stacks already have “good enough” blocks inside the drive. We don’t replace them. We sit upstream — pre-model, pre-path — where mid-band ring poisons inference and shadow paths.
- vs stiction compensators: Kano-style models and friction feedforward live in the controller. We live in the signal path before the model sees it. Complementary, not competitive.
Moat is not “secret math.” It’s characterization + published failures + a clear identity. Actuators flagship → · Safety cage →
If you have this problem…
Salesy without inventing dollars. Pain → action. You measure the money.
| If you have… | Then… |
|---|---|
| Mid-band hunt / limit-cycle-class ring you can see on OP–PV or joint traces | Shadow GSRF vs best EMA on that series. Eval is free. If osc / travel proxies move in your favor, Pro is usually the cheap next step vs more engineering time. |
| Pre-model junk (PdM / residual models thrashing on chatter) | Treat input hygiene as product quality. We publish mid-band kill metrics; you score your model with/without the pre-stage. |
| 10–50+ channels (cell, multi-joint, multi-loop site) | Enterprise conversation — not Pro (≤~10 channels). Density multiplies cost of noise; still pre-stage, not a controller claim. |
| You need minimum MAD tracking of raw, delayed command feedback, or fixed-FA residual alarms | Buy free. Keep EMA / Kalman / vendor blocks. Our red lights say so. |
Job matrix (locked evidence posture)
| Job | Prefer | Why (public evidence posture) |
|---|---|---|
| Mid-band ring / chatter calm | GSRF Practical | ~+35% osc31 vs best EMA; ~−70% sweet-band spectral power on locked actuator-class packs |
| Peak chase reduction toward a normal | GSRF Practical | Peak-dev advantages in multi-family hunt; step settles intermediate with frozen x* |
| Minimum MAD tracking of raw | EMA / Kalman / k_return≈0 | GSRF 0/52 MAD wins vs best EMA when spring is on |
| Delayed plant measurement as command feedback | Not balanced GSRF | MAD-to-measured can be ~+59% to +106% worse vs EMA at 3–10 min delays |
| Fixed-FA residual event detection | EMA residual (or dedicated detector) | Matched-FA residual recall favored EMA (~0.83 vs ~0.44) on synthetic excursions |
| Probabilistic state estimation with noise model | Kalman (or similar) | Different objective; GSRF is not a Kalman substitute |
| Permanent level shift following | Adaptive x* GSRF or EMA | Frozen x* sticks; adaptive reduces mean regime error ~9.7→1.5 in probe |
One-line chooser
If the costly failure is ring and spike chase around a known normal → try GSRF. If the costly failure is lagging the truth → start with EMA/Kalman.
Questions people actually ask
When should I choose GSRF over EMA?
When mid-band ring or peak chase around a known normal is the costly failure (~+35% osc31 / ~38% peak on locked packs).
When should I keep EMA or Kalman?
Minimum MAD tracking, delayed plant command feedback, fixed-FA residual detection, or probabilistic state estimation.
Is GSRF a drop-in EMA replacement?
No — different identity (log-space soft thermostat). Read the safety cage before production.
Which filter for industrial PdM sensors?
GSRF candidate for pre-model mid-band chatter + peak chase; keep EMA/Kalman for tight MAD tracking or fixed-FA residual alarms. Pre-AI hygiene →