Definition
Error between AI-predicted defect progression timelines and actual progression measured at subsequent inspections, segmented by operating load regime. Progression models are defect-level twins, and a timeline that assumes steady-state conditions becomes dangerously optimistic under high load.
The Measurement Gap
The Measurement Gap
Defect progression models predict when a known defect will reach the severity threshold requiring repair. These predictions drive major capital and logistics decisions: when to order crane mobilization for a blade repair, whether to defer a bearing replacement to the next scheduled outage, how to prioritize scarce repair resources across a portfolio of tracked defects. The models generate these predictions based on historical progression data, but most models are trained on aggregate data that does not segment by operating load regime. A blade leading-edge erosion defect progressing under normal wind conditions behaves differently than the same defect class under sustained high-wind regimes that accelerate erosion rates. The model reports a single timeline without disclosing that its accuracy degrades substantially when the asset operates outside the training data's dominant load profile.
Without load-segmented accuracy measurement, the organization cannot distinguish between a reliable 18-month progression estimate and an optimistic one that assumes average conditions the asset does not experience. The financial consequence is misaligned maintenance planning: crane mobilization scheduled for month 16 when the defect actually reaches threshold in month 9 under high-load conditions, or conversely, premature repair triggered by conservative assumptions when the actual progression under low-load conditions would have allowed safe deferral for another year.
The Decision It Enables
DPAUL provides load-regime-specific confidence intervals for every defect progression prediction. When DPAUL is low (accurate) for a given load regime, maintenance planners can trust the timeline and commit resources accordingly. When DPAUL is high for extreme-load regimes, planners apply explicit safety factors to progression predictions for assets in high-wind or high-irradiance zones. This transforms defect management from a single-estimate planning process into a risk-stratified approach where prediction confidence varies by operating context and resource commitments scale accordingly.
Formula
Primary Formula:
DPAUL(r) = (1 / N_r) x SUM( |Predicted_Severity_i minus Actual_Severity_i| / Actual_Severity_i )
Variable Definitions:
| Variable | Definition |
|---|---|
| r | the operating load regime between inspections, classified as low, normal, high, or extreme based on operating hour distribution during the inter-inspection period |
| Predicted_Severity_i | the AI-projected defect severity at the re-inspection date, as predicted at the time of the previous inspection |
| Actual_Severity_i | the measured defect severity at re-inspection using the same measurement methodology as the original inspection |
| N_r | the number of defect-to-reinspection measurement pairs in load regime r |
Calculation Example:
A wind fleet tracks 80 blade defects across two inspection cycles. Between inspections, 50 defects experienced normal wind conditions and 30 experienced high-wind conditions. For the normal-regime defects, the average absolute percentage error between predicted and actual severity is 15%. For the high-wind regime defects, the average error is 38%. DPAUL(normal) = 0.15. DPAUL(high) = 0.38. The 23-point gap between regimes is the load acceleration blind spot: the model loses substantial accuracy when wind conditions stress the defect beyond normal operating assumptions.
Edge Cases and Measurement Boundaries:
Defects repaired between inspections are censored data points because actual progression was not observed. Track these as "repaired before re-inspection" separately but do not include them in DPAUL calculations. Regime classification between inspections should use the dominant regime covering more than 50% of operating hours for cumulative-load-driven failure modes, or the peak regime experienced for peak-load-driven failure modes. New defect types not present in the progression model's training data should have DPAUL tracked separately from known defect types to avoid contaminating the accuracy assessment for established models.
When DPAUL for extreme regimes exceeds 0.40, apply a safety factor to all progression predictions for assets in high-load zones. A recommended starting approach is to multiply predicted time-to-threshold by 0.6, effectively accelerating the predicted timeline by 40% until the model is recalibrated with load-dependent training data.
Proposed formula, pending validation pilots.