Definition
The proportion of an AI maintenance model's predictions for a given asset class that rest on failure history from other asset classes rather than the class being predicted, weighted by how well those donor classes have historically transferred.
The Measurement Gap
There is no metric that tells you which of the model's predictions rest on failure history from the asset class being predicted versus history borrowed from other classes, which means predictions for sparse-history equipment carry hidden extrapolation risk that never surfaces on any dashboard.
Formula
Plain language: For each asset class, determine what fraction of the training signal behind the model's predictions came from that class itself versus other classes. Weight the borrowed fraction by each donor class's historical transfer accuracy (how well its patterns have predicted this class in backtests). A class scoring near 1.0 is predicted from its own history; a class near 0 is predicted almost entirely by extrapolation.
Notation:
ACEI_c = (S_own_c + Σ_d (S_d→c × T_d→c)) / S_total_c
Variables:
- S_own_c = training signal (failure and degradation events) from asset class c itself (count or weighted count, DERIVED from model training data audit)
- S_d→c = training signal borrowed from donor class d applied to class c (count, DERIVED)
- T_d→c = historical transfer accuracy from class d to class c, 0–1, from backtesting (dimensionless, DERIVED)
- S_total_c = total training signal behind predictions for class c (count, DERIVED)
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Can we trust the model's predictions for asset classes with thin failure history, or is the model quietly extrapolating from dissimilar equipment while presenting the same confidence it shows for well-covered classes?