Yard Detection Sensitivity Drift Index (YDSDI)

GMI-RAIL-035Rail TransportationIntegration-RequiredPattern O-2: Model Drift / Currency

Definition

Change in yard computer vision mechanical defect detection sensitivity over time, measured by comparing CV detection rates against periodic manual inspection audits on the same car populations.

The Measurement Gap

The Measurement Gap

Yard CV systems report their detection counts: how many defects found per shift, per week, per car. What they don't report is whether their detection rate has changed since deployment. A CV system that found 95% of defects in its calibration audit may be finding 80% now due to camera fouling, lighting changes, fleet composition shifts, or model staleness. The detection count stays high (more cars processed = more defects found), masking the declining per-car sensitivity.

This gap matters because operations teams treat CV as a reliable replacement for manual inspection, but the system's reliability is time-dependent. Without periodic sensitivity audits comparing CV to manual inspection on the same cars, nobody knows if the CV safety net has holes.

Decision Supported:

Is the yard CV system still detecting defects at its deployment-calibration sensitivity, or has environmental drift reduced its effective detection rate? YDSDI provides the longitudinal sensitivity tracking to trigger recalibration before the drift causes missed defects.

Formula

Plain Language: Periodically run manual carmen inspections on a sample of cars that have also passed through the CV detection zone. Compare the defects found by manual inspection to those detected by CV. Track the ratio over time.

Notation:

YDSDI = CV_detection_rate_current / CV_detection_rate_baseline

Where CV_detection_rate = defects_detected_by_CV / defects_detected_by_manual_audit (on the same car sample)

Variables:

Variable Definition Unit Source Classification
defects_detected_by_CV Mechanical defects the CV system identified on the audit sample cars Count CV system logs NATIVE
defects_detected_by_manual_audit Mechanical defects identified by qualified carmen on the same cars Count Manual inspection records MANUAL
CV_detection_rate_baseline Detection rate at initial deployment calibration Ratio Historical calibration data NATIVE

Sign Convention: 1.0 = no sensitivity change from baseline. <1.0 = sensitivity has degraded. >1.0 = sensitivity improved (unlikely without recalibration; may indicate fleet's defect profile changed).

Edge Cases:

  • Manual inspection has its own miss rate. YDSDI measures CV relative to manual, not absolute detection. If both miss the same defects, YDSDI won't reveal the shared blind spot.
  • Fleet composition changes: new car types may have defect presentations the CV model wasn't trained on. This isn't drift: it's coverage gap. Track separately.
  • Minimum sample: 200 cars per audit, quarterly audits.

Proposed formula, pending validation pilots.

At a Glance

Category
Yard Operations and Consist Integrity
← Hazmat Separation Margin Consumption Rate (HSMCR)

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