Twin Fidelity Drift (TFD)

GMI-MRO-020MROInfrastructure-Dependent

Definition

TFD measures the divergence between twin-predicted states and physically observed states over a rolling window, combining distributional difference and time-series trajectory difference.

The Measurement Gap

Commissioning validation tests the twin under nominal conditions. The easiest case. Once the asset undergoes repairs, feedstock changes, or seasonal load shifts, the twin can drift silently from reality while every IT metric (uptime, API latency, model availability) shows green. OEE and MTBF describe the asset, not the decision tool. RMSE and R² describe historical fit, not present accuracy. The result is subtly wrong what-if analyses, RUL estimates, and maintenance recommendations driven by a model representing an asset that no longer exists. Commissioning validation (the most common twin quality gate) only measures fidelity under nominal conditions, the easiest and least informative test. The operationally relevant question is whether the twin still works after the asset has degraded, been repaired, changed feedstock, or shifted operating regimes. No existing reliability KPI defines "twin accuracy over time" operationally.

Decision Supported:

Whether to act on twin outputs as-is, schedule recalibration, or downgrade the twin's role for the affected asset class.

Formula

Plain language: Compare the distribution of observed sensor values to the distribution of twin-predicted values, and also compare their time trajectories. Combine into one score and track weekly.

Notation:

TFD = D_KL( P_observed ‖ P_twin ) + λ · DTW( x_observed(t), x_twin(t) )

Variables:

Variable Definition Unit
D_KL Kullback-Leibler divergence between observed and predicted distributions over the window Dimensionless
P_observed Probability distribution of physically observed state values from sensors Distribution
P_twin Probability distribution of twin-predicted state values Distribution
DTW Dynamic Time Warping distance between observed and predicted time series Dimensionless
λ Weighting factor tuned so both terms contribute equally at baseline Dimensionless

KL divergence catches "right neighborhood, wrong tail"; DTW catches "right average, wrong timing." A 15% month-over-month TFD growth is a reasonable initial flag.

Confidence Grading (required alongside every TFD value):

Grade Basis Interpretation
A: Empirical 5+ observed events in measurement window provide direct tail observation Measurement grounded in observed outcomes; highest confidence
B: Hybrid 1-4 observed events plus Monte Carlo simulation with Bayesian updating Blends observation and simulation; moderate confidence
C: Simulated Zero observed events; derived entirely from Monte Carlo simulation Modeled, not observed; treat as indicative, not validated

A TFD of any value at Grade C is a simulation that agrees with itself. It tells you the model is internally consistent but not that it matches reality. Grade C scores should trigger proactive stress-testing against historical near-miss events and cross-asset analogues rather than acceptance as calibrated.

Proposed formula, pending validation pilots.

At a Glance

Unit
Index
Trend Direction
Lower; and stable month-over-month
Reporting Frequency
Weekly rolling
Category
Digital Twin & System Resilience
← Topological Resilience Reserve (TPRR)Carbon Payback on Repair vs. Replace (CPRR) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment