Definition
Degree to which a predictive maintenance model RUL estimates correspond to actual component life achieved, measured by comparing predicted RUL at time of alert to actual hours remaining at time of removal, binned by confidence level.
The Measurement Gap
The Measurement Gap
Predictive maintenance platforms generate remaining useful life estimates for heavy mobile equipment components. Maintenance teams use these RUL numbers to decide whether to pull a component now, schedule it for the next shutdown, or continue monitoring. The critical question is whether those numbers are accurate, and specifically whether the accuracy varies by the model's own stated confidence level. Legacy metrics like MTBF and PM Compliance measure whether maintenance happens on schedule. They cannot measure whether the predictions driving that schedule are trustworthy. Physical Availability tells you whether the asset is running, not whether the prediction that kept it running was correct. A model that consistently overestimates remaining life (RULCS below 1.0) creates unexpected failures that legacy metrics attribute to "random breakdowns" rather than to systematic prediction error. A model that consistently underestimates remaining life (RULCS above 1.0) drives premature replacements that inflate Cost per Ton without any metric identifying the prediction as the cause. The consequence compounds across fleets of 50 to 200 haul trucks where each component decision carries a $50K to $500K price tag.
Decision Supported: Can the maintenance team trust the RUL number when deciding whether to pull a component now, schedule it for next shutdown, or continue monitoring, and is the model more reliable for some component types or operating conditions than others?
Formula
Plain language: For each component removal, look back at the most recent RUL prediction issued before the removal event. Compare the predicted remaining hours to the actual remaining hours (time from prediction to removal). Group by the model's confidence level and compute the ratio. A perfectly calibrated model produces a ratio of 1.0 at every confidence level.
Notation:
RULCS(b) = Mean(Actual_RUL) / Mean(Predicted_RUL) for predictions in confidence band b
Overall: RULCS_weighted = Sum(n_b x RULCS(b)) / Sum(n_b)
| Variable | Definition | Unit |
|---|---|---|
| Predicted_RUL | Model's stated remaining useful life in hours at time of alert | Hours |
| Actual_RUL | Actual hours of operation between the prediction timestamp and the component removal event | Hours |
| b | Confidence band (e.g., [60-70%), [70-80%), [80-90%), [90-100%]) | Band |
| n_b | Number of predictions in band b | Count |
Proposed formula, pending validation pilots.