Planning AI for missing tablet inspection in blister packaging starts with a practical question: which defect can the camera actually see, which physical pack will be rejected, and which record proves the result? A useful project scope includes missing, broken and wrong tablets, synchronized rejection and retrievable image evidence. This guide walks through a request for proposal (RFP), factory acceptance testing (FAT) and site acceptance testing (SAT) for a pharmaceutical packaging line in Thailand. The manufacturer’s quality team must determine which regulations and validation activities apply to its products and markets.
The decision unit is the pack, from image to confirmed rejection
A demonstration can show a red box around a missing tablet. Production acceptance needs more: the finding must be attached to the right pack, the downstream reject mechanism must remove that pack, and the system must record whether removal succeeded. Camera classification alone does not reveal an encoder slip, a missed trigger, low air pressure, a full reject bin or a communication failure. False rejects of good packs matter as well.
Map one pack through formed pocket, feeder, imaging station, decision engine, sealer, reject station and record system. A published WHO inspection report describes a camera positioned after the feeder and before sealing to detect missing or broken tablets; it also discusses a challenge test for the rejection sensor. This is a site-specific example, not a prescribed layout for every plant. WHO inspection report
Define the optical inspection boundary
Typical visible targets include an empty pocket, a broken or chipped tablet, an apparent colour or shape mismatch and a suspected wrong tablet. Positioning a camera before sealing can make the pocket contents visible, but the actual optics depend on transparent or reflective material, pocket depth, powder, multiple tablets and product geometry. Test representative physical packs under real lighting. Machine vision suppliers describe blister quality checks and inspections for missing, incorrect and damaged contents. Their application pages demonstrate possible uses, not guaranteed performance for a given product. Cognex pharmaceutical applications and inspection examples
A camera viewing tablets before sealing does not prove heat-seal integrity, microscopic leak tightness, barrier performance or material compatibility. Printed lot and expiry information also needs a suitable view and a separate verified task. Define these exclusions in the user requirements specification (URS). For related controls, see our articles on packaging heat-seal inspection and date-code print inspection.
Missing, broken and wrong tablets require different evidence
An empty pocket may differ clearly from a filled one, so a stable rule-based measurement can sometimes handle missing tablets. Breakage and chipping change appearance with tablet orientation, the exposed fracture surface, powder and shadows. A wrong product may be distinguishable by colour or shape, yet similar tablets and an imprint on the unseen face may be impossible to separate in that image. The camera cannot identify chemical composition. Evaluate each defect class for optical visibility and robustness instead of assigning one accuracy figure to all defects.
The WHO case lists missing, broken, chipped, colour-mismatched and other-product tablets in its URS and discusses DQ, OQ and PQ evidence. Its stated breakage and chipping thresholds are specifications of that inspected installation; they must not become universal acceptance limits. Set limits from product risk, visual detectability and controlled challenge samples. WHO inspection report
Show what happens before and after sealing
The pre-seal camera sees contents well but does not observe defects caused later. A post-seal camera may inspect appearance, printing or cut position; it may not see through an opaque foil. Where both imaging points are needed, state what each judges and how both outcomes link to the same physical pack. Define how the line handles unclassified packs during stoppage, restart, film splice, recipe change or camera reboot.
Choose rules, AI or a controlled combination
AI is not mandatory for every defect. A deterministic rule using position, area or contrast may be easier to explain and maintain where pocket geometry and lighting are stable. A trained classifier or anomaly model may help when breakage, powder, orientation or printed surfaces create greater visual variation. Compare candidate methods on the same physical samples and conditions. Supplier descriptions of AI-enabled vision do not establish that a particular model is best for every line. Cognex pharmaceutical applications
A practical architecture may use rules for presence, AI for variable surface appearance and a controlled safe response for low confidence. Quality must approve whether uncertainty causes rejection or line stop after assessing the risk of passing a defect and the cost of false rejects. A model score is not a product-quality guarantee. Model performance can shift when material, lighting or production conditions change.
Build a representative dataset, not a collection of attractive examples
Record product, lot, packaging material, pocket position, line speed, lighting condition, orientation, vibration, powder and changeover state for each sample. Include normal variation across good product, otherwise the classifier may reject large volumes of acceptable packs. Simulated defects can support testing when real ones are scarce, but document exactly what they simulate and where the appearance differs from an actual failure.
Separate training, threshold tuning and final evaluation data. Adjacent video frames and samples from the same lot appearing in both training and test sets can overstate real-world performance. Include unseen lots and material conditions in the final set. Report missed defects, false rejects, indeterminate decisions and pack-tracking failures by defect class, along with denominators and sample counts. A single overall accuracy hides important failure modes.
Design reproducibility and change control
Heat maps can help investigations, yet they do not replace validation. Retain model version, product recipe, camera and lighting settings, time, equipment ID, line speed and thresholds so a decision can be reconstructed. A model update needs a reason, impact analysis, approval, comparison with the previous version and a rollback path. EU GMP Annex 11 addresses risk-based control of computerised systems used in GMP activities; Annex 15 addresses qualification and the impact of changes. Applicability depends on the manufacturer’s regulatory context. EU GMP Annex 11 and Annex 15
Select a camera around physical line conditions
Pixel count alone does not determine the smallest detectable defect. Field of view, working distance, lens distortion, shutter and exposure, light angle and wavelength, transport speed, trigger timing and foil reflections interact. Ask for images from edge pockets, deeper pockets, different tablet orientations, creased film and dusty conditions, not only ideal centre pockets. The vendor should explain how pack pitch, camera-to-reject distance, encoder resolution and actuator response prevent rejection of the wrong pack.
For multi-lane equipment, require a coordinate system for lane and pocket. State allowable camera movement and re-alignment steps. Include window cleaning, light replacement, spares, heat, humidity, vibration, protection during stops and network fault behavior. For support in Thailand, establish remote-access authority, local replacement capability and retrieval of fault logs before contract signature.
Test inspection and rejection as separate physical events
An on-screen NG label is not proof that the correct pack left the line. Multiple packs travel between camera and reject station. Challenge stop and restart, backward movement if possible, low air pressure, solenoid failure, failed reject confirmation and a full bin. Specify whether a failure stops the line or takes another approved safe action. A confirmation sensor should support evidence of the intended pack’s removal, not merely that an actuator fired.
Use a known missing-tablet pack to trace detection, pack identity, reject command, physical rejection and record. Control insertion and recovery of challenge pieces, then reconcile quantities so none remain in a commercial batch. The WHO report’s rejection-sensor challenge test offers an example of why the whole chain needs evidence. WHO inspection report
Put acceptance evidence into the RFP
Start the RFP with products, tablet forms, pocket and lane counts, packaging materials, speed range, changeovers, PLC/SCADA/MES interfaces and installation constraints. Then specify each defect class, optical limits, unacceptable misses, allowable false rejects and the response to an uncertain result. Final thresholds should be fixed in an approved protocol after product-risk and sample review. Ask the supplier for challenge pieces, sample counts, original images, recipes and reject logs as deliverables.
| Requirement | Define in the RFP | Verify at FAT/SAT |
|---|---|---|
| Scope | Visible missing, broken, chipped and wrong tablets; seal integrity managed separately | Defect samples and exclusions |
| Imaging | Camera position, field, light, synchronization, speed and cleaning | Images from edge pockets and restarts |
| Decision | Rule/AI roles, versions, thresholds and uncertain-result response | Per-defect result table on fixed version |
| Rejection | Tracking, confirmation and fault response | Physical pack removal and simulated faults |
| Records | Image/event key, clock, retention, access and audit history | Trace one NG to image and rejection |
| Changes | Approval and re-evaluation of recipes, models, lighting and PLC logic | Version history and rollback procedure |
Assign responsibilities. Catalogue performance and acceptance with the actual product on the actual line are different claims. Name who prepares challenges, establishes ground truth, owns images, approves retraining and carries unresolved FAT items into SAT. Agree on remote-support security, software maintenance and recovery arrangements before installation.
Pair every numerical target with its denominator
“99% detection” alone does not say which defects, sample conditions or number of observations were used. Record tested packs, confirmed defective packs, detections, misses, good packs and false rejects by class. Include intended speed range, edge pockets, post-changeover operation and adverse-but-credible lighting. A statistical performance claim needs a justified sample plan. No universal acceptance percentage fits every product; quality must approve criteria based on defect severity, downstream controls and applicable requirements.
Evaluate throughput beyond algorithm runtime: decision latency at maximum speed, false-reject workload, retest effort, changeover time, recovery and storage capacity. Set image retention using applicable record rules, contracts and the company’s record policy rather than inventing a generic number of years.
Connect FAT, SAT and qualification
FAT can demonstrate compliance with URS and functional specifications before dispatch. SAT then tests the installed system with real lighting, vibration, PLC signals, reject equipment and networks. EU GMP Annex 15 describes URS, DQ, FAT/SAT and IQ/OQ/PQ, and allows a justified choice not to repeat some FAT work later when transport and installation do not affect the tested function. Map each test and its evidence to the approved risk-based plan. EU GMP Annex 15
At FAT, use a frozen recipe and version with known missing, broken and wrong-tablet samples. Challenge edge pockets, orientation, speed, missing triggers, power recovery, communication and reject sensor failures where feasible. Link original image, result, pack identity, command, confirmation, timestamp and settings in the record. At SAT, retest under installed plant conditions with the actual product, foil, conveyor, reject position and required system interfaces. A successful bench demo does not establish installed fail-safe behavior.
Challenge examples and failed-test handling
Candidate cases include sound packs, empty pockets, partial fractures, a similar-coloured wrong product, powder, film reflection, missing reject confirmation, an NG immediately before stop and another immediately after restart. Select cases by risk; the list is not a universal protocol. Identify and reconcile all challenge pieces. An uncertain result or lost communication must not silently turn into a good-pack verdict. Choose an approved line-stop or safe-reject response for the process.
When a test fails, changing the threshold and rerunning the same image is not sufficient to close it. Record the deviation and investigate optics, mechanics, model, challenge-piece ground truth and reject mechanism. Approve the revised version and retest scope, including any earlier passing cases affected by the change. FDA process-validation guidance frames process design, qualification and continued verification as a lifecycle; that supports monitoring as product and equipment conditions evolve. FDA Process Validation guidance
Make the image trail usable in an investigation
For one NG event, connect product, lot, line, timestamp, physical pack ID, pocket location, original image, decision, model/recipe version, reject command and reject confirmation. Synchronize clocks and use a robust join key across systems. The choice between saving all images and saving only NG images depends on quality risk, investigation needs, capacity and applicable record obligations. NG-only storage may limit trend analysis; full storage has capacity and search costs.
Separate permissions to view records, alter settings and approve models. Test backup restoration, offline queues, clock synchronization and export for audits. Annex 11 is an EU GMP reference for computerized-system controls where applicable; the manufacturer must establish its own regulatory scope. EU GMP Annex 11
Reassess after a model or process change
Tablet colour, foil, lighting, camera position, product recipes and software versions can change the classification problem. Document impacted defects and processes, then choose a proportionate retest. Compare new and old versions on a fixed evaluation set and inspect fresh plant images. Monitor unexpected false-reject trends or suspected misses, with an approved stop or rollback procedure. The European Commission’s proposed new AI Annex 22 has been published for consultation; do not describe that draft as a current binding requirement. Commission consultation
A staged implementation plan
First, collect process drawings and defect history. Measure camera positions, visible faces, lighting, reject distance and representative product variation. Establish what cannot be seen. Bring quality, production, maintenance, IT/OT and the vendor together to approve the URS and responsibility boundaries.
Second, compare rule-based and AI candidates on the same physical samples. Keep training and final evaluation separate. Measure per-defect misses, false rejects, indeterminate decisions, latency, pack tracking and confirmed rejection. For multiple products, test recipe selection and prevention of the wrong setting. Feed the remaining limitations back into the RFP and acceptance criteria.
Third, test function and evidence at FAT, installed conditions and physical rejection at SAT, and connect the results to the approved qualification and operating procedures. In service, monitor NG and false-reject trends, lighting, cleaning, model versions and reject-equipment maintenance. Regularly confirm that an investigation can retrieve the image and physical reject result. Optical inspection is one control within the broader manufacturing process.
RFP issues often overlooked
Who defines ground truth and controls challenge material?
Borderline chips or fractures may be disputed even by trained reviewers. Define the defect classification against product specifications before scoring the camera. Document how disputed samples are handled; quietly removing them can inflate apparent performance. Wrong-tablet challenges introduce a separate mix-up risk. Define authorization, identification, quantities, recovery and line clearance so challenge tablets never remain in commercial production. Contractually define whether the vendor may reuse customer images for general model training.
Write a labelling guide that separates missing tablets, displacement, chipping, powder and normal variation. If reviewers disagree, revisit the definition before tuning the algorithm. Image cropping, preprocessing and resolution are also part of the effective inspection system; changing them can change results even if the model file is unchanged. Retain data origin, permissions, exclusions and versions.
Include changeover, cleaning and maintenance work
The camera recipe and packaging-machine recipe must agree at changeover. Define product-code transfer from the PLC, access to manual changes and verification of the first packs after a switch. Include window cleaning, post-cleaning checks and lighting replacement in standard work. If maintenance bypasses the camera, prevent uninspected output and perform an approved restart challenge.
For shared systems, confirm approved recipe versions at each line. Local decision-making during network loss requires later reconciliation without missing or duplicate records. Remote maintenance should have access logs and change records. Operators need an understandable fault screen; quality investigators need the original image and settings history. Evaluate both interfaces.
Assess escape risk and excessive rejection separately
Investment analysis should include potential defect escape, investigation and isolation, false-reject material loss, maintenance, storage, validation and additional quality checks, not just saved inspection labour. Sparse historical defects are a weak estimator of future events; examine failure modes and weaknesses in existing controls. A high-impact wrong-tablet event cannot be reduced to average accuracy and simple payback alone.
Deployment can be phased, for example stabilizing missing-tablet detection and rejection before adding product-specific breakage detection. State what remains under human or other process controls during each phase. If AI may be added later, assess lighting, image acquisition, storage, computing and version-control capacity at initial selection. Manage unresolved uncertainty visibly rather than over-specifying an untested system.
Operating KPIs and triggers for reassessment
After acceptance, define daily indicators by product, recipe, line and lot: inspected pack count, indeterminate results, NG count by defect, confirmed false rejects, reject-confirmation failures and camera downtime. Put cleaning, maintenance and recipe-change timestamps on the same timeline so a rising NG rate can be investigated as a product issue, dirty light or shifted camera. A zero count of reject-confirmation faults is not proof that the sensor works; challenge it periodically. Averages across the whole line may conceal edge-pocket or product-specific failures, so review data at a useful level of detail.
Document triggers for reassessment before they occur: new foil, tablet colour change, lamp replacement, camera adjustment, speed change, AI update, reject repair or restart after a long stop. Map the changed component to affected defect classes and select the necessary images, physical pack challenges and reject tests. A full repeated protocol is not automatically needed for every change; a decision to omit a retest needs documented reasoning and approval. Before rolling back to an earlier version, confirm that it remains suitable for the current material and equipment conditions.
Quality reviews should examine suspected escapes investigated, quantity lost to false rejection, unexplained NG cases and time needed to retrieve evidence alongside overall accuracy. Use those findings to add challenge examples, service lighting, revise operating steps or reconsider acceptance criteria. Such ongoing evidence supports maintenance of process control beyond the initial demo and is consistent with the lifecycle approach in FDA’s process-validation guidance. FDA Process Validation
Frequently asked questions
Is an AI-enabled blister inspection machine required?
No. Rule-based vision may suffice for visually stable empty-pocket detection. Compare AI for complex variation such as breakage, powder and orientation. Select on actual per-defect performance, false rejects, maintainability and change control.
Can the missing-tablet camera guarantee seal integrity?
No. A pre-seal tablet camera observes visible contents. It does not directly prove heat-seal integrity, microscopic leaks or barrier performance. Define separate seal controls and tests with the quality team.
How do FAT and SAT differ for blister AI inspection?
FAT focuses on pre-delivery equipment and functional evidence. SAT verifies the installed line, including lighting, synchronization, correct-pack rejection, fault behavior and records. Link both to the URS and risk-based qualification plan.
Will image AI always detect a wrong tablet?
No. Similar-looking tablets, an unseen imprint or a difference in chemical composition may be invisible to the camera. Combine optical inspection with mix-up prevention, line clearance and material controls.
What detection percentage is acceptable?
There is no universal number. Quality should preapprove criteria using severity, representative samples, false rejects, downstream controls and applicable requirements. Always show the defect-specific denominator and misses.
Summary
Specify missing, broken and wrong tablets separately; image them where visible; reject the correct physical pack; and preserve evidence of what happened. Compare rules and AI on real samples. Put exclusions, conditions, reject confirmation, records and change control in the RFP. Demonstrate equipment functions at FAT and installed-line behavior at SAT. A tablet image classifier alone cannot guarantee seal integrity or regulatory approval.
If you are still defining a camera location or reject interface for an existing line, contact TOMAS TECH to discuss the process drawing and defect scope. We can help frame the questions for an RFP and acceptance test against the conditions of your plant.
Primary sources
- WHO Prequalification, Remington inspection report (2021) — site-specific camera, URS, qualification and reject-sensor example.
- European Commission, EU GMP Annex 15 (2015) — URS, FAT/SAT and qualification.
- European Commission, EU GMP Annex 11 (2011) — computerised systems.
- US FDA, Process Validation guidance (2011) — process lifecycle.
- Cognex, Pharmaceutical and Medical applications — machine-vision applications including blister inspection.
- Cognex, pharmaceutical inspection examples — missing, incorrect and damaged contents.
- European Commission, draft Annex 22 consultation — consultation, not a current binding rule.
