Definition
Cohort Homogeneity Score (CHS) measures the statistical similarity of assets within an analytics cohort across the variables that actually drive degradation: validating whether fleet-level comparisons are legitimate before anyone acts on them.
The Measurement Gap
Your fleet analytics dashboard is telling you that Pump P-0847 is an outlier. Its vibration signature is degrading 40% faster than the cohort average. The system flags it for early intervention. The planner creates a work order. The bearing gets replaced. Three weeks of useful life, wasted.
Because P-0847 wasn't an outlier. It was the only pump in the cohort running at 92% of rated capacity on a slurry application. The other fourteen pumps in the "same" cohort were running clean water at 60–75% capacity. P-0847's degradation rate was exactly what you'd expect for its actual operating conditions. The fleet analytics saw a deviation from the group average and flagged it. The group average was meaningless because the group wasn't actually comparable.
This happens constantly and almost nobody catches it. Fleet analytics platforms group assets by make, model, and sometimes age. They compare performance across the group and flag deviations. The underlying assumption (that these assets are similar enough for comparison) is rarely validated. And when it's wrong, the consequences run in both directions: assets that look like outliers may be normal for their actual conditions (false positives), and assets that look normal may be masking abnormal degradation because the cohort average is dragged in their direction by dissimilar peers (false negatives).
No standard reliability metric catches this.
MTBF
by asset class measures failure frequency across a population, but it doesn't validate whether the population is valid.
OEE
compares against rated performance, not against peer performance, so cohort validity isn't relevant. Even within fleet analytics platforms, the grouping is treated as a configuration choice, not a measured characteristic. You pick your cohort definition, and the system trusts it.
The problem gets worse as fleet analytics becomes more sophisticated. AI-driven degradation curves, population-level RUL estimates, cohort-based anomaly detection: all of these produce more precise outputs than traditional dashboards. But precision on a bad cohort is worse than imprecision on a good one. A precise answer to the wrong comparison is more dangerous than a rough answer to the right comparison, because the precision creates false confidence.
There is no metric that tells you whether asset cohorts used for fleet-level analytics are statistically valid given differences in operating context, maintenance history, and environmental exposure, which means fleet-level findings may be misapplied to individual assets.
Formula
Cohort Homogeneity Score (CHS) measures the statistical similarity of assets within an analytics cohort across the variables that actually drive degradation: validating whether fleet-level comparisons are legitimate before anyone acts on them.
In plain language:
For a group of assets that your fleet analytics treats as comparable, measure how different they actually are on the things that matter. Operating intensity, environmental exposure, maintenance history, load profile, age since last major overhaul. For each of these "degradation drivers," calculate how spread out the values are across the group. Average the spread. A low spread means the group is genuinely similar; a high spread means you're comparing apples to transmission housings.
The formula:
For cohort K with n assets and m degradation drivers:
CV_j = σ_j / μ_j for each driver j
CHS_K = 1 − (1/m) × Σ min(CV_j, 1)
Where:
- σ_j = standard deviation of degradation driver j across all n assets in cohort K
- μ_j = mean of degradation driver j across all n assets in cohort K
- CV_j = coefficient of variation for driver j
- min(CV_j, 1) = CV capped at 1 to prevent a single wildly heterogeneous driver from dominating
CHS ranges from 0 to 1. 1.0 = perfectly homogeneous (all assets identical on all drivers). 0.0 = maximally heterogeneous.
The metric uses coefficient of variation because it's unitless: you can compare the spread of operating hours (measured in thousands) against the spread of ambient temperature (measured in degrees) without normalization.
Worked example:
A fleet analytics platform groups 12 centrifugal pumps as "Cohort: Goulds 3196 LTX, installed 2018–2020." You evaluate CHS across five degradation drivers:
Operating hours since last overhaul: mean 14,200 hrs, std dev 3,100 hrs, CV = 0.218
Average load as % of rated capacity: mean 71%, std dev 18%, CV = 0.254
Process fluid abrasiveness index: mean 3.2, std dev 2.7, CV = 0.844
Number of seal replacements (lifetime): mean 2.4, std dev 1.8, CV = 0.750
Ambient temperature range (annual °C): mean 28°C, std dev 12°C, CV = 0.429
Average of capped CVs: (0.218 + 0.254 + 0.844 + 0.750 + 0.429) / 5 = 0.499
CHS = 1 − 0.499 = 0.501
This cohort scores 0.50: below the proposed 0.70 threshold. Look at what's driving it: operating hours and load are reasonably tight (CV < 0.30), but process fluid abrasiveness (CV = 0.84) and seal replacement history (CV = 0.75) are wildly dispersed. Some of these pumps are handling clean water; others are handling abrasive slurry. They wear out differently, degrade through different mechanisms, and fail in different modes. Comparing them as a single cohort will produce misleading outlier flags and invalid population-level degradation curves.
The fix:
Split the cohort by process fluid abrasiveness. Create a "clean service" sub-cohort and an "abrasive service" sub-cohort. Recalculate CHS for each. The clean service group will likely score above 0.70.
Now consider what happens if you don't calculate CHS and use the original cohort: An asset running clean water at 65% capacity with 16,000 hours will show normal degradation against the cohort average (which is dragged up by the abrasive-service pumps). The fleet analytics says "within normal range." But compared against ONLY clean-service peers, it might actually be degrading faster than expected, a genuine anomaly masked by the heterogeneous cohort. That's a false negative, and it's the dangerous one.
Edge cases:
If a driver's mean is zero (e.g., all assets have zero major repairs), CV is undefined. Exclude that driver, a variable with zero variance is perfectly homogeneous and shouldn't affect the score. If fewer than 5 assets are in the cohort, flag as statistically unreliable. For categorical drivers (indoor vs. outdoor, continuous vs. batch service), use Blau's diversity index instead of CV, normalized to the same 0–1 range. If more than 20% of values are missing for a driver, exclude the driver rather than computing CV on a partial sample.
Proposed formula, pending validation pilots.