TB-9619 · REV A · Technical newsheet
Metrology & CalibrationDevice profile
Pharmaceutical Station Bias Distorts Process Capability Data
A 70% station contribution in an illustrative fill-weight dataset exposed how automated checking can distort capability, hide drift and misdirect process corrections.
By Nathan Brooks4 min read825 words
Features
- Approximately 70% of observed variation in the illustrative dataset traced to the checking station.
- Figure 1 compares the same 400 measurements pooled and colored by checking station.
- Station 5 had measured 1,248 units since drift began, including 156 borderline readings.
- The 156 borderline readings spanned four lots: two on site and two shipped.
- Recalibration and standardization reportedly improved agreement and reduced false rejections and false acceptances.

By Nathan Brooks
In an illustrative pharmaceutical fill-weight dataset, approximately 70% of observed variation traced to the checking station rather than production variation.
What did the production reports miss?
Several machines filled pharmaceutical units in parallel. Automated stations downstream verified net fill weight, and each unit went to whichever station was available. Pass rates were healthy, while control charts remained in control, so the standard reports did not test whether the stations agreed.
Each station kept a separate measurement log. The logs were consolidated in SQL and joined to production records while retaining the production-machine identity, checking-station identity, measured value, pass/fail result, unit ID and timestamp. That chain allowed each reading to be followed through the line.
Plotting every reading produced a cloud around the target. Coloring the points by station separated that cloud into bands. Figure 1 shows the same 400 measurements pooled and then colored, with shared specification limits exposing differences that the combined view concealed.
Near a specification limit, a station that reads consistently high or low can change whether a borderline unit passes. The combined distribution also enters process-capability assessment, which raises the central measurement question: how much apparent production variation comes from the act of checking the units?
How much variation came from checking?
Station averages did not explain the pattern. Comparing each production machine across stations produced nonparallel lines in Figure 2, meaning station disagreement changed with the machine that made the unit. The data therefore showed both an overall station effect and a production-machine-by-checking-station interaction.
The case used ANOVA-based variance-components analysis in Python within the measurement system analysis framework of a Gage R&R study. The model separated product variation, equipment-to-equipment reproducibility and the machine-by-station interaction rather than treating every departure from target as process variation.
That distinction matters because a shared target does not make two measurement systems equivalent. In variance-components terms, the station and interaction terms represent measurement-system contributions. Adjusting the filling process without removing those terms would leave the team judging its correction through the same biased readings.
In the illustrative dataset, approximately 70% of observed variation traced to the checking station. Process work alone could not remove that contribution. The equipment identifiers instead pointed corrective work toward recalibrating and standardizing the stations.
After that correction, the team reported better agreement, fewer false rejections and false acceptances, and more useful Cpk and Ppk capability measures. The original case therefore changed both the interpretation of the control chart and the direction of quality work.
Can agreement survive after recalibration?
Agreement still had to survive service, replacement and continued use. A station could drift again after the initial correction, so the team built a measurement-system health view in Power BI. Figure 3 tracks station agreement over time and current comparisons, with a declining trend exposing a station separating from its peers.
The underlying records continued to connect every reading to its unit and equipment. Once the team saw drift, it had to establish when it began, identify all units measured since then and trace those units to lots. Material on site and material already shipped required different follow-up.
An AI agent supplied that traceability from the consolidated records. It identifies drift onset from the trend, retrieves units measured during the interval and tests readings near a specification boundary. For each borderline case, the agent estimates whether the station offset could reverse the original pass/fail assessment.
The output gave the quality engineer the full affected population and a smaller priority group. The agent then followed those units into lot records, separated on-site material from shipped material and drafted a recommendation with its supporting evidence.
Figure 4 shows the illustrative result for Station 5. Since drift began, the station had measured 1,248 units; 156 were borderline across four lots. Two lots remained on site and two had shipped. The draft proposed investigating the station, increasing monitoring of affected lots and reviewing disposition of borderline units.
How far should the AI agent go?
The quality owner could inspect the records and proposed response together, then approve or reject the recommendation with the decision logged. Recalibration, material holds and deviations stayed outside the agent's authority. The software assembled an investigation that otherwise required work across separate systems; it did not make the disposition.
By the end of the work, the team could detect drift earlier and connect it to material needing review. The station identity had been present in the logs from the start, but joining it to production and lot records changed the chart interpretation, the corrective action and the questions available when drift returned.
Before adjusting a process that appears to vary, teams should ask what the measuring equipment contributes. The adoption question is whether pharmaceutical operations can require station-level identity in routine records, monitor agreement continuously and preserve human approval for lot-level actions.
via bnpmedia.com (Original)
Filed under
- gage-r-r
- measurement-system-analysis
- station-bias
- process-capability
- pharmaceutical-manufacturing
More from Nathan Brooks
Show full bio
Staff writer covering people, launches and day-to-day practice in Instruments, test and measurement at Testbench Report.
105 articles