Definition
Degree to which the health model's training data is biased by non-uniform inspection/sampling cadence: measuring whether assets sampled less frequently appear artificially healthier than assets sampled more frequently.
The Measurement Gap
Utility condition monitoring programs are not uniformly applied across the asset fleet. High-criticality assets (large power transformers serving major substations) typically receive annual DGA sampling, frequent thermal scans, and regular partial discharge testing. Lower-criticality assets (smaller distribution transformers, rural line equipment, older secondary switchgear) may receive biennial or triennial sampling, or irregular sampling triggered by nearby events rather than a fixed schedule. This variability is an operational and economic reality, not a data quality failure.
The tracking gap is that the AI health model receives all of this data as equivalent inputs. A transformer with annual DGA history going back 10 years provides a rich feature set. A transformer with three DGA samples in 12 years provides a sparse one. Current analytics platforms do not automatically flag the confidence level difference: both assets receive a point-estimate health score, and neither score carries an explicit uncertainty band reflecting data density.
Legacy systems have this same limitation, but the problem compounds in AI models because the sparsely-sampled population's apparent health patterns get encoded into model coefficients that then apply to all similar assets, not just the lightly-inspected ones.
The decision SBI enables: determining which sub-populations of the asset fleet need either accelerated sampling (to reduce uncertainty before the next capital cycle) or explicit uncertainty premiums applied to their health scores (acknowledging that "no news is good news" is not a sound basis for capital deferral).
Without SBI, the utility may defer assets in the infrequently-sampled population with the same confidence it defers well-monitored assets: treating data sparsity as health evidence. With SBI, the capital planning team knows which deferred assets are sitting on confident scores versus which are sitting on scores that reflect limited data more than limited degradation.
AI can correct this by flagging data recency and density as explicit model features, but only if the organization first measures the bias that sparse data creates. SBI is that measurement.
Formula
SBI = (Score_infrequent − Score_frequent) / Score_frequent − (FailureRate_infrequent − FailureRate_frequent) / FailureRate_frequent
Variables:
- Score_infrequent: average health score for assets sampled less frequently than the median sampling cadence for that asset class (dimensionless score on the model's native scale). Source: Asset analytics platform, segmented by sampling frequency from CMMS
- Score_frequent: average health score for assets sampled at or above the median cadence (dimensionless). Source: Asset analytics platform
- FailureRate_infrequent: observed failure rate for the infrequently-sampled population (failures per asset-year). Source: OMS + CMMS failure records, segmented by sampling frequency group
- FailureRate_frequent: observed failure rate for the frequently-sampled population (failures per asset-year). Source: OMS + CMMS
Normalize for asset age and criticality before computing: the goal is to isolate the sampling frequency effect from age-driven health differences. Compute the residual SBI after controlling for age cohort and asset class.
Calculation Frequency:
Annual; recomputed when sampling protocols change or the health model is retrained
Output: Dimensionless. Positive SBI = infrequently-sampled assets score higher than their actual failure rates justify (bias present, action required above 0.15). Zero = no detectable bias. Negative SBI = infrequently-sampled assets score worse than failure rates warrant (conservative bias, may cause excess capital spend). Minimum 100 assets per frequency group for statistical validity.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Is the AI health model treating infrequently-sampled assets as "no news is good news," or is it correctly accounting for the uncertainty introduced by sparse data?