Definition
Measures the strength of the condition-quality correlation under the current product mix compared to the correlation under the product mix the model was trained on.
The Measurement Gap
Model drift is a recognized problem in predictive analytics, and existing metrics track temporal drift where model accuracy degrades over time. CQMCI addresses a different drift mechanism that existing model currency metrics miss: product-mix drift. The CM-quality model may be current in calendar time but trained on a product portfolio that no longer matches what the plant is actually producing. A model trained when the plant ran 80% Product A and 20% Product B may deliver unreliable quality predictions when the mix shifts to 50% Product A, 30% Product B, and 20% Product C.
The decision CQMCI supports is whether the quality-linked CM system can be trusted for the current product mix or whether it needs revalidation. If the condition-quality correlation has weakened for specific product types, the CM triggers based on those correlations may be generating both false alarms (triggering maintenance when quality is fine for the current product) and misses (not triggering when quality is degrading because the degradation pattern differs from what the model expects). Without CQMCI, these trust failures are invisible until a quality escape occurs or a pattern of unnecessary maintenance is noticed through other channels.
Formula
CQMCI compares the coefficient of determination (r-squared) between condition monitoring parameters and quality metrics under the current product mix to the same coefficient computed on the training product mix.
CQMCI = r_squared_current_mix / r_squared_training_mix
Where r_squared_current_mix is the coefficient of determination between CM parameters (vibration, temperature, current draw) and quality metrics (Cpk, defect rate, dimensional deviation) computed on the current product mix (dimensionless, 0 to 1). r_squared_training_mix is the same coefficient computed on the training or validation product mix (dimensionless, 0 to 1).
A CQMCI of 1.0 means the correlation is unchanged and the model's quality predictions are as reliable as when it was trained. Below 1.0 means the correlation has weakened and the CM system is less predictive of quality for the current mix. Above 1.0 means the correlation has strengthened, which is uncommon but possible if a new product is more sensitive to the monitored condition parameters.
Edge cases require specific handling. If training mix data is no longer available, use the first six months of operation as a proxy baseline. When multiple CM parameters are monitored, compute CQMCI per parameter and then aggregate weighted by each parameter's importance in the quality prediction model. Because product mix changes gradually, use a rolling 90-day window for the current mix and compare to the original training-period window. Minimum sample requirements are 200 production records per product type per measurement period to ensure statistical reliability of the correlation computation.
Proposed formula, pending validation pilots.