Definition
The alignment between a technician's belief distribution over an asset's possible states and the AI's belief distribution over the same states, calculated as one minus the Jensen-Shannon divergence between the two distributions.
The Measurement Gap
Classical situational awareness frameworks (SAGAT, SART) measure the human alone, treating the automation as a fixed reference. In predictive maintenance, the AI is another agent with its own beliefs that can be wrong, outdated, or miscalibrated. MTBF is a population-level statistic: it says nothing about real-time state of a specific asset. OEE measures output, not understanding. Even AI-side metrics (precision, recall, F1) compare the model to ground truth, not to the technician who must act on its output. Disagreement between human and AI beliefs is often the earliest leading indicator of failure, and nothing on the maintenance dashboard surfaces it.
Decision Supported:
Where to focus investigation, model recalibration, or technician engagement. Using divergence between human and AI beliefs as an early-warning trigger that no single source produces alone.
Formula
Plain language: Express what the AI "believes" about an asset as a probability across possible states (healthy, early-stage failure mode A, failure mode B, etc.) and capture the technician's belief in the same categories. Compute how different the two belief sets are using Jensen-Shannon divergence. Subtract from one. Perfect alignment scores 1.0; complete disagreement approaches 0.
Notation:
SSA = 1 − JSD(P_human ∥ P_AI)
Variables:
| Variable | Definition | Unit |
|---|---|---|
| P_AI | AI's probability distribution over asset states (derived from health score, anomaly scores, failure-mode probabilities) | Distribution |
| P_human | Technician's probability distribution over the same states (from structured inspection rating or revealed-preference action) | Distribution |
| JSD | Jensen-Shannon divergence: symmetric, bounded measure of difference between two distributions; computed as average KL divergence to the midpoint distribution | Score (0–1) |
A trend matters more than a single score: a drop from 0.95 to 0.70 over three inspection cycles is the leading-indicator signal.
Proposed formula, pending validation pilots.