Definition
Degradation Projection Validity Rate (DPVR) measures what percentage of your degradation trend projections (remaining useful life estimates, time-to-failure predictions) proved accurate within a stated tolerance when you could actually check the outcome. Higher is better. If your DPVR is 80%, eight out of ten projections you made were within acceptable range of what actually happened. If your DPVR is 45%, you're wrong more than half the time, and you should stop scheduling maintenance around those projections until you understand why. The
The Measurement Gap
The number on the screen said 120 days remaining. The bearing had 63.
Turbine 11 sat in a coastal wind farm alongside 17 identical machines, all running AI-based prognostics on their gearbox bearings. The model had been trained on steady-state operational data, a reasonable calibration, because steady-state is what these turbines mostly ran. In early August, the grid operator issued curtailment orders. The farm went from running continuously to cycling through three or four start-stops per day. The turbines kept spinning. The model kept projecting. Nobody flagged that the projections were now based on physics the model had never seen.
When Turbine 11's main bearing failed 63 days into a 120-day projection, the failure mode was not subtle. The gearbox needed full replacement: $420,000. Emergency crane mobilization added $180,000. Lost generation during the repair window cost another $310,000. Call it $910,000 in damage from a projection that was confidently, precisely wrong.
The operational team had done what they were supposed to do. They had AI prognostics. They had a RUL display. They deferred the maintenance campaign to spring, just as the tool recommended. What they didn't have was any signal that the projection underneath that number had become invalid the moment operating conditions shifted.
This is the failure mode that legacy metrics cannot see.
MTBF (mean time between failures) tells you the historical average interval for that failure mode on that asset class. It's useful for budgeting and baseline planning. It tells you what happened in the past across a population of similar events. It says nothing about whether a specific forward projection for a specific asset is still calibrated to current operating conditions. An asset with an MTBF of 200 days can fail at day 63 if it's been pushed outside the operating envelope the model was trained on. MTBF will not warn you. It will only update after the failure is recorded.
P-F interval tracking has the same blind spot, pointed in a different direction. P-F interval measures the historical gap between a detectable precursor condition and functional failure. That gap is real and useful: it's the window inside which you can act. But P-F interval tracking measures what the interval has been. It doesn't validate whether a projection of future failure timing is still accurate. You can have a well-established P-F interval for a bearing under normal load and miss entirely when accelerated cycling collapses that interval by half.
The RUL dashboard display compounds the problem by showing a number with false precision. A digital readout that says "120 days remaining" looks authoritative. It invites you to schedule around it. What it doesn't display is the confidence interval, the operating context it was calibrated against, or any signal that those conditions have since changed. Operators shouldn't have to know to ask that question: the measurement framework should answer it automatically.
There is no metric that tells you whether degradation trend projections are still valid when operating conditions have changed.
That gap is what DPVR (Degradation Projection Validity Rate) is designed to close. Not by predicting failures more accurately, but by measuring how often the projections you're already trusting actually hold up when outcome data comes in. You can't manage what you don't measure. And right now, most operations are not measuring this.
Formula
Degradation Projection Validity Rate (DPVR) measures what percentage of your degradation trend projections (remaining useful life estimates, time-to-failure predictions) proved accurate within a stated tolerance when you could actually check the outcome. Higher is better. If your DPVR is 80%, eight out of ten projections you made were within acceptable range of what actually happened. If your DPVR is 45%, you're wrong more than half the time, and you should stop scheduling maintenance around those projections until you understand why.
The Formula
DPVR = P_v / P_t × 100
Where:
- P_v = count of projections where the actual outcome fell within tolerance
- P_t = total count of projections that reached an observable outcome
- Tolerance: |RUL_actual − RUL_projected| / RUL_projected ≤ 0.20
The default tolerance band is 20%. A projection of 100 days is "valid" if the actual failure or end-of-life event occurred between 80 and 120 days. That band isn't arbitrary: it reflects a practical scheduling window. If your maintenance campaigns are planned quarterly, a 20% error on a 90-day projection might still land in the right planning window. Tighten it to ±10% if you're making short-horizon decisions where precision matters more.
Worked Example
A mining operation runs AI-based RUL projections on conveyor drive components. Over a 14-month review period, 84 projections reached an observable outcome: the equipment either failed, was proactively replaced, or reached the end of its service life with full condition documentation. Of those 84:
- 49 projections fell within the ±20% tolerance band
- 35 projections fell outside it
DPVR = 49 / 84 × 100 = 58.3%Below the 60% threshold where you'd treat projections as advisory only. Worth investigating, but the aggregate number doesn't tell you enough. When the team segmented by operating condition stability:
- Stable operating conditions (no significant load or duty-cycle changes during projection period): 39 of 52 projections valid. 74.2% DPVR- Changed operating conditions (load shifts, process changes, throughput increases during projection period): 10 of 32 projections valid. 31.3% DPVRThat's the finding that changes behavior. Under stable conditions, the projection system is borderline acceptable: close to the 75% operational reliance threshold. When operating conditions change after the projection is issued, the system is wrong two out of three times. The models were calibrated on historical data that no longer described the equipment's actual operating environment. The aggregate 58.3% masked a system that was almost functional in one context and nearly useless in another.
Edge Cases
Four situations require special treatment or exclusion from the DPVR denominator. Assets still running when you compute DPVR: exclude them entirely and track them as "in-flight projections." An asset that hasn't failed or been replaced hasn't produced an outcome yet. Including it in the denominator would undercount valid projections and distort the rate. Assets decommissioned before failure for reasons unrelated to the projection: exclude but flag as censored data. If an asset was taken out of service because the process line shut down, not because the projection said to act, you don't have a valid comparison point. Accumulate these flags, a pattern of censored data can indicate process instability that's also affecting your model calibration. Operating condition changes after the projection was issued: don't exclude, but segment. This is the most important edge case. Projections issued under one operating context and evaluated under a different one should be computed separately, a "stable context" DPVR and a "changed context" DPVR. The gap between them is where your model risk lives. Very short RUL projections, under 24 hours: apply tighter tolerance, ±10% rather than ±20%. At short horizons, precision matters more and the consequences of being wrong are immediate. A 20% error on a 12-hour projection means you could be off by more than two hours: enough to miss a maintenance window or allow a failure to propagate.
Cross-References
DPVR doesn't exist in isolation. Two upstream KPIs produce the inputs that determine whether your projections are valid in the first place. DTFI (Digital Twin Fidelity Index (GMI-AR-019)) measures how accurately your digital twin or predictive model represents the physical asset's current state. When DTFI degrades, the model is producing RUL estimates from an increasingly inaccurate representation of reality. Low DTFI is a leading indicator that DPVR is about to drop. If your twin hasn't been updated to reflect a significant process change, the projections it generates are already wrong: DPVR will confirm that when outcomes arrive. MDDR (Model Drift Detection Rate (GMI-AR-018)) measures how reliably your system detects when a model's predictions are drifting from actual behavior. Drift happens gradually: a model calibrated on one operating regime slowly loses accuracy as conditions shift. If MDDR is low, drift goes undetected, and DPVR erodes over time without anyone noticing until the projection error accumulates into a real incident. Two downstream KPIs depend on DPVR being healthy. PTVR (Prediction Threshold Value Ratio (GMI-AR-004)) tracks the balance between true and false positives in your predictive maintenance alerts. Invalid projections distort that balance. A model that's projecting failures at the wrong time will generate false alarms that erode technician confidence and missed detections that turn into unplanned failures. PTVR absorbs the signal of DPVR degradation in alert quality. SOSI (Strategy Override Safety Index (GMI-AR-008)) measures what percentage of documented failure modes remain covered when an algorithm recommends changing a maintenance strategy. If DPVR is low, condition-based coverage claims in those strategy changes may be built on invalid RUL projections. You may believe you have condition-based coverage on a critical asset when the underlying projection is invalid. SOSI will surface instability in strategy execution; DPVR tells you one likely cause.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Can I trust the "remaining useful life" number on my screen to make run/repair/replace decisions, or are these projections unreliable for assets operating outside their historical norms?