Gage Repeatability and Reproducibility (GR&R)

The variation in measurements obtained with one measurement instrument when used several times by the same operator while measuring the identical characteristic on the same part.

• Represents the inherent precision of the gage itself
• Also called within-system variation or EV (Equipment Variation)
• Assessed through repeated trials under identical conditions
• High repeatability issues indicate the sensor, camera, or fixture needs mechanical improvement

The variation in the average of measurements made by different operators using the same measuring instrument when measuring the identical characteristic on the same part.

• Captures operator-to-operator inconsistency
• Also called between-system variation or AV (Appraiser Variation)
• Critical when inspection relies on human positioning or subjective judgment
• In AI vision systems, reproducibility issues often stem from inconsistent part presentation or lighting setup across shifts

The actual differences between the parts being measured, representing the true process variation that the measurement system must be capable of detecting.

• Denoted as PV (Part Variation) in ANOVA calculations
• Must be significantly larger than the measurement system variation
• A measurement system with GR&R > 30% of total variation cannot distinguish between good and bad parts
• Selecting parts spanning the full tolerance range is essential for valid study results

The combined effect of repeatability and reproducibility expressed as a percentage of tolerance or percentage of total variation.

• %Tolerance = (GR&R / Tolerance Width) × 100
• %Study Variation = (GR&R / Total Variation) × 100
• < 10%: Measurement system is acceptable
• 10-30%: May be acceptable based on application criticality and cost
• > 30%: System requires improvement before use for process control

A metric derived from the GR&R study indicating how many statistically distinct groups the measurement system can reliably separate within the process variation.

• Represents the resolution of the measurement system
• ndc ≥ 5: System is adequate for process control
• ndc < 2: System cannot distinguish parts, effectively producing only noise
• Calculated as (Part Variation / GR&R) × √2, rounded down to the nearest integer

The Analysis of Variance approach is the preferred statistical method for GR&R studies because it quantifies the operator-by-part interaction in addition to repeatability and reproducibility.

• Decomposes total variation into: part, operator, operator×part interaction, and equipment
• The interaction term reveals whether certain operators measure specific parts differently
• More accurate than the simpler X-bar and R method which ignores interactions
• Essential for validating AI inspection systems where human operators define ground truth labels

The overarching statistical framework for evaluating the entire measurement process. While GR&R focuses specifically on precision (variation), MSA assesses both accuracy (bias) and precision. It uses designed experiments to decompose total observed variation into components from the part itself, the operator, and the gage, ensuring the system is fit for purpose before any process control decision is made.

The direct quality cost of poor measurement precision. In an AI inspection context validated by GR&R:

• Type I Error (False Reject/FRR): A good part is measured as bad, causing unnecessary scrap and rework costs.
• Type II Error (Escape Rate): A bad part is measured as good, allowing a defect to reach the customer. High GR&R variation inflates both error rates, undermining the confusion matrix metrics of the classifier.

A fundamental distinction in metrology. Precision (assessed by GR&R) refers to the tightness of repeated measurements—how much variation exists when measuring the same part. Accuracy (Bias) refers to how close the average of those measurements is to the true reference value. An AI camera can be highly precise (low GR&R) but inaccurate if camera calibration is off, requiring both to be independently validated.

The categorical counterpart to GR&R for visual inspection. When an AI system or human operator makes a pass/fail judgment, this analysis measures how often they agree with themselves (within-appraiser agreement), with each other (between-appraiser agreement), and with a known standard (ground truth). It uses Cohen's Kappa statistic to account for random chance, validating the reliability of subjective defect classification.

The preferred statistical technique for modern GR&R analysis. Unlike the simpler Average and Range method, Analysis of Variance (ANOVA) mathematically partitions total variation into:

• Part-to-Part: The desired signal.
• Operator: Reproducibility error.
• Operator-by-Part Interaction: Whether operators measure specific parts differently. This interaction term is critical for diagnosing complex AI training failures where a model behaves inconsistently on specific defect types.

A key GR&R output metric that answers: 'Can this measurement system tell parts apart?' It calculates how many statistically distinct groups the system can reliably separate from the total part variation. A value of 5 or greater is required for effective process control. An ndc of 1 or 2 indicates the system can only distinguish between 'good' and 'bad' but cannot detect subtle process shifts, limiting its use for continuous model learning.