Skip to content

Raw-Signal Fallback Switch

Escape-path control switch — instantiates Predictive Residual Processing

Suspends residual-only processing and reverts to full observations the moment the model's validity conditions fail, trading bandwidth for completeness on demand.

A Raw-Signal Fallback Switch is the escape hatch that makes aggressive residual processing safe to run: it watches the conditions under which the predictive shortcut is trustworthy, and the instant they fail, it suspends residual-only mode and routes full, unabridged observations instead. Its defining role — distinct from every mechanism that keeps operating in residual mode — is to stop operating in it on purpose: when confidence collapses, a residual leaves the modeled range, the model goes stale, or a safety-critical condition appears, the switch decides the model can no longer be trusted to summarise the situation and reverts to ground truth. It is a circuit breaker for prediction, and its whole value is being wired to fail toward raw reality.[1]

Example

Offshore wind turbines report condition data to shore over an expensive satellite link, so each unit normally sends only residuals — deviations of vibration and temperature from its predicted healthy baseline. A Raw-Signal Fallback Switch on each turbine continuously checks four validity conditions: model confidence, residual magnitude against the modeled range, staleness since the last baseline resync, and a safety-critical vibration band. One unit begins throwing residuals far outside anything its model was trained on — an unmodeled bearing fault. The switch trips: it suspends residual-only reporting for that turbine and streams the full raw sensor feed to shore, accepting the bandwidth hit because a model that never learned this fault cannot be trusted to compress it. Engineers now see the complete signature, not a summary. Once the fault is characterised, a fresh baseline confirmed, and residuals return to the modeled range, the switch — after a deliberate dwell — reverts to residual mode.

How it works

  • Evaluate validity continuously. Confidence, residual magnitude versus modeled range, staleness, and a safety flag are checked every cycle; the switch is always deciding whether the shortcut still holds.
  • On failure, suspend and route raw. Residual-only processing is halted and full observations are requested or routed for the affected scope — the on-demand path is opened.
  • Bypass deliberation for safety. Designated safety-critical conditions skip all weighing and force raw immediately, on a path independent of the model that may itself be failing.
  • Revert with hysteresis. When validity is restored, the switch waits out a dwell period before returning to residual mode, so it doesn't flap on a marginal signal.

Tuning parameters

  • Trip thresholds — how bad conditions must get before failover. Sensitive thresholds are safe but expensive and prone to flapping; insensitive ones save bandwidth but risk not tripping when they must.
  • Trigger set — which conditions can fire the switch (low confidence, out-of-range residual, staleness, safety flag). More triggers catch more failure modes but raise the false-trip rate.
  • Hysteresis / dwell — how long to remain in raw mode before reverting. Longer dwell prevents oscillation around a threshold at the cost of extra bandwidth.
  • Fallback scope — whether a trip reverts one channel, one unit, or the whole system. Narrow scope preserves savings elsewhere; broad scope is safer when failures correlate.

When it helps, and when it misleads

Its strength is that it bounds the blast radius of model failure: a residual scheme can be run hard and lean precisely because there is a guaranteed, fast path back to complete observation when the model stops being trustworthy. It converts "the model might be silently wrong" — the archetype's central danger — into "the model is wrong and we've already switched to raw."

It misleads when the trip logic is mis-set. Thresholds too sensitive mean constant fallback, the bandwidth savings evaporate, and — worse — operators learn to raise the thresholds or ignore the trips, defeating the safety net through fatigue. Thresholds too insensitive mean the switch never fires when it should, and the model quietly filters out the very event it can't handle. The classic misuse is desensitising the switch after too many "nuisance" fallbacks, disabling the escape hatch right before the failure it exists to catch. The discipline is to set the safety-critical bypass from harm analysis rather than bandwidth cost, to route that bypass around the possibly-broken model, and to regularly test that the switch actually trips — an untested fallback is not a fallback.

How it implements the components

This tool owns the archetype's escape-to-raw slice:

  • fallback_to_on_demand_path — it is the on-demand full-observation path and the control that switches onto it.
  • decompression_trigger — the validity-condition logic that fires the reversion from compressed residuals to the full signal.
  • safety_critical_bypass_rule — the fast, non-negotiable route that forces raw immediately for safety-critical conditions, independent of normal weighing.

It does not detect the gradual drift that often trips it — that signal comes from Model Drift Monitoring — nor does it audit what residual mode was dropping while all looked well (Shadow Raw-Channel Sampling) or prioritise residuals within normal operation (Precision-Weighted Error Gate).

Notes

The safety-critical bypass must not depend on the model whose failure it guards against. If the switch decides whether to escape to raw using the same predictor that might be broken, a model failure can suppress its own alarm. Wiring the bypass to independent, model-free evidence — a raw threshold, a hardware interlock — is what keeps the escape hatch open in exactly the situation it is meant for.

References

[1] Graceful degradation (and the fail-safe / circuit-breaker pattern) is the standard engineering principle that a system losing a capability should fall back to a simpler, safer mode rather than fail outright. Applied here, the "simpler mode" is unabridged raw observation — deliberately more expensive but not dependent on a model that has stopped being valid. ↩