Blog

2026.11.10

AI Vision for Inspection-Fixture Wear: Deployment and Acceptance in Thailand

AI Vision for Inspection-Fixture Wear: Deployment and Acceptance in Thailand

If you want to find wear on manufacturing inspection jigs using an AI image judgment system, you must not decide the pass/fail of a product based solely on the camera’s detection results. Design a process that captures the state of locating pins and seating surfaces, verifies suspicions with dimensional measurements, and allows the quality manager to approve the suspension and resumption of jig use. This article provides a practical proposal for production technology, quality, and purchasing personnel at Thai factories to judge introduction specifications, PoC, FAT, and SAT with the same evidence.

1. Reasons for Monitoring Inspection Jig Wear

The deterioration of inspection jigs is different from the wear of tools that make products or the visual inspection that captures scratches on product surfaces. Inspection jigs are tools that physically create reference points for pass/fail. If the locating pin wears and the workpiece tilts slightly, it may be judged as non-conforming even if it is actually a good product, and conversely, it may show non-standard products as conforming. What the camera first detects is the change in jig condition, not the product’s compliance with standards.

If conventional inspections are only visual at the start of a shift, the timing of replacement can easily vary by operator, and it may be difficult to trace which jig, which part, or which product model was “normal” last time. Fixed cameras and reference images serve as a means to record conditions in the same field of view. However, if you immediately interpret differences in images as wear amounts, dirt and reflections may be mixed into the lifespan judgment. The investment purpose is not the introduction of AI itself, but to quickly find jigs that need to be stopped and ensure repairs before they affect quality.

The design standards for jigs, the concept of replacement units and tolerances are organized in Practical Jig Design and Manufacturing, and the shape and mounting methods of pins are organized in Locating Pin Selection. This article focuses on determining the condition of jigs in operation and managing their stoppage and return.

2. Separate Failure Modes and Verification Methods by Part

Jig PartSuspicions Detectable by ImageVerification Beyond ImageInitial Action
Locating PinChipped tip, abnormal gloss, bent appearanceDiameter, roundness, position from reference surface, workpiece seatingRe-image after cleaning, gauge measurement if necessary
Seating SurfaceChips, oil film, dents, wear marksFlatness, height, contact conditionRemove foreign matter and confirm seating with reference workpiece
StopperSuspected position change due to chipping or looseningStopper position and fixation stateCheck fastening and position
BushChipped mouth, one-sided scratchesInner diameter and play, relation to mating pinInspect paired parts as well
ClampSurface scratches, signs of return failureHolding force, stroke, safety mechanismPass to mechanism inspection

This table is a starting point for determining inspection items and does not mean that all abnormalities can be detected with a single top-view photo. Separate perspectives or probes/gauges are needed for shielded backsides, internal wear, and play that only appears under load. Oil films may disappear with cleaning, but if they return to the same spot in a short period, the cause remains. Teaching AI only “normal/abnormal” can confuse states that can be restored by cleaning with those requiring replacement.

3. Boundary Between Image Screening and Dimensional Judgment

The datums on design drawings, machining standards when making jigs, inspection standards during use, and product inspection standards are related but not necessarily identical. First, extract the critical characteristics of the product from the drawings and correlate the jig parts that affect the inspection results of those characteristics through causal relationships. For example, determine how changes in pin diameter affect workpiece position and measurement values, or how insufficient clamping force appears in reproducibility. Detecting unrelated scratches with the same priority will inflate the maintenance queue.

The system separates three outputs. Image analysis suggests “normal,” “requires cleaning and re-imaging,” or “candidate for suspension.” Quality and maintenance personnel conduct independent measurements or physical checks and approve whether to return the jig to use or stop it. Product pass/fail is judged by existing approved inspection methods and standards. Measurement uncertainty is important for determining quantitative tolerance boundaries, and ISO 14253-1:2017 deals with decision rules for conformity/non-conformity to GPS specifications of numerical values. This rule cannot be directly applied to image AI category classification.

4. Connecting Fixed Cameras, Masters, Gauges, and Work Orders

AI Vision for Inspection-Fixture Wear: Deployment and Acceptance in Thailand - figure 1

The design of the imaging point changes depending on whether the jig is removed and placed on a workbench or taken while fixed to the equipment during stoppage. The former makes it easier to align lighting and posture, but requires confirmation of repositioning during attachment/detachment. The latter may have the field of view obstructed by conveyors, safety fences, oil/chip splatter, or operator pathways. In either case, images are linked with jig ID, replaceable part revisions, camera/lens/lighting recipe versions, time, equipment, and product model.

Reference images are not just the first shot taken on the introduction day. They are created from jigs approved for use through physical measurement and masters or reference workpieces with managed conditions. Without deciding on the storage, inspection, replacement, and revision of masters, new jigs will continue to be matched to old standards. The camera shows image differences and positions, gauges or CMM confirm necessary numerical values, and HMI shows reasons for hold to people. Work orders record detection images, physical checks, cleaning, repairs, replacements, re-verification, approvers, and time.

5. Determine if Rules Are Sufficient or AI Is Needed

For the presence or absence of pins, clear chipping, misalignment, or contours outside specifications, rule-based image processing may suffice if lighting and imaging are stable. Cognex “Locate Fixture” defines a reference position within the image; its software term “fixture” is distinct from the physical inspection fixture discussed here. Image-position correction like Locate Fixture by Cognex and feature inspection like Inspect Feature can serve as references to clarify the position and features to be observed. Do not equate the manufacturer’s feature names directly with your fixture’s detection performance.

Consider AI when permissible manufacturing variations, continuous changes in wear, and overlapping lighting reflections make it difficult to manage the number of reviews or misses with simple thresholds. However, if a model considers a dangerous state of a chipped pin as “normal,” prioritize stopping rules. Compare implementation by evaluating the same test samples with and without AI, recording not only detection rates but also false stops, pending judgments, review workload, and time required for re-imaging. Advanced models are not always better.

6. Manage Imaging Recipes as Part of the Machine

Verify in practice the light source angle that suppresses mirror reflections on metal pins, the lighting direction that reveals surface irregularities, the necessary pixel resolution, depth of field, lens fixation, shading, and safe field of view during operation. Conditions for seeing oil films or chips may not be the same as those for seeing wear contours. If one image cannot satisfy both, treat two images with switched lighting as a set, fixing the order and settings. Include fixed reference points in the frame to determine if the camera has shifted or the fixture has changed.

Document pre-imaging cleaning conditions, conditions for placing and removing workpieces, equipment temperature range, lighting duration, and shooting triggers in the work standard. If focus or light intensity is out of tolerance, do not let AI judge it; label it as “poor image quality.” After lens replacement, lighting equipment replacement, or fixture revision, confirm whether the change is comparable to past reference images. Simply replacing images with new standards risks nullifying actual wear.

7. Link Training Data Labels to Repair Decisions

If learning is done with only binary “normal/abnormal” labels, dirt that can be washed off and wear that should be replaced become the same abnormality. At a minimum, define “usable,” “clean and re-image,” “measurement confirmation,” “candidate for use suspension,” and “poor image quality” as operational states. The model’s training labels and HMI display states do not necessarily need to be one-to-one. Decide how humans will verify before choosing classification granularity.

The true state cannot be automatically determined from images. Combine reference workpieces, actual measurements, repair history, replacement part conditions, and quality department approvals to label. Record which parts deteriorated in which mode and what response was required. Do not force examples with split judgments into “normal” by majority vote; record the measurement or design department’s decision. Taking many images of the same fixture from different angles to increase rare wear or chipping images does not increase independent defect cases.

8. Inspection Frequency and Escalation

Do not determine inspection cycles solely by usage count. Consider the product’s critical characteristics, past wear and replacement history, chip and oil adhesion during the process, fixture attachment and detachment frequency, and the production volume before the next inspection. Initially, placing imaging at existing control points like shift starts or part number changes and reviewing missed detections and review volume to revise the cycle is practical. Do not confuse fixtures suitable for constant monitoring by fixed cameras with those that should be safely imaged during stops.

If a “candidate for use suspension” arises, the site should verify the image and fixture ID, and even if cleaning dirt restores it, re-image without deleting the original image. Notify the quality manager of any suspicion affecting safety or product quality, and design not to automatically return to production before approval. When removing or replacing on-site, follow the energy isolation procedures for each facility at the Thai plant. U.S. OSHA documents can serve as references for hazardous energy management but do not substitute for Thai regulatory requirements.

9. Dataset and Model Validation

Simply dividing images randomly into training and evaluation sets can lead to lenient evaluations, as similar images of the same worn pin taken under the same lighting conditions may end up in both sets. Record jig ID, the series of periods for part replacement, shooting date, equipment, and shift, and decide in advance the unit to be used for evaluation. While it is useful to associate images before and after replacement, do not inflate the number of independent evaluation samples with the number of images.

Evaluation samples should include clean and conforming fixtures, oil films removable by cleaning, metal swarf, usage marks within acceptable limits, true chipping, wear close to dimensional limits, and images where the camera shook. Quality and maintenance confirm each condition on the physical fixture with independent measurement evidence. When new product models or part revisions are added after operation, do not force the model to output “normal,” but hold them as out of scope. The NIST AI RMF Measure Playbook provides insights into monitoring during operation and human review.

AI Vision for Inspection-Fixture Wear: Deployment and Acceptance in Thailand - figure 2

10. Evaluate False Acceptance and False Rejection Separately

What the quality department most wants to avoid is determining that a jig that should be stopped is usable and continuing to trust the inspection results. On the other hand, if there are many false rejections of normal jigs, the field will ignore the alarms. In the PoC, independently determine the true state of the jig and count false acceptance, false rejection, judgment hold, and poor image quality separately. Overall accuracy alone can hide the oversight of rare significant chipping.

Recognize the state where “AI is unsure” and list the jig ID, image, candidate area, previous state, confirmation deadline, and person in charge in the review waiting list. Separate human approval authority from product shipment pass/fail authority and designate in advance the approver for jig reuse. Do not move the threshold to a favorable value after seeing the test data, but leave the predetermined evaluation set and failure judgment in the acceptance document.

11. Dimensional Measurement and MSA are Different from Image Classification

Quantitative values such as positioning pin diameter and seating surface height are measured with calibrated gauges, CMM, etc., and verify whether the measurement method is sufficient for the required tolerance. ASQ’s Gage R&R deals with measurement system variability, including repeatability and reproducibility. For image AI that outputs categorical good/bad, do not directly apply continuous value Gage R&R, but separately confirm attribute judgment agreement among evaluators, false acceptance, false rejection, and hold rate.

ISO 10012:2026 specifies requirements for measurement management systems. When calculating dimensions from camera images, it is necessary to verify scale, geometric correction, environment, target surface, reference instrument, and measurement uncertainty. NIST’s Validation Upgrades for the NIST M48 Vision Measurement System provides examples of non-contact edge detection and device-specific uncertainty evaluation, but these figures cannot be transferred to the performance of jig cameras in Thai factories. Clearly state in the RFP whether the camera is used merely for condition screening or as a dimensional measuring instrument.

12. In PoC, Use Cleaning, Wear, and Poor Image Quality as Control Tests

Select one target jig for the PoC and prepare a reference workpiece, parts before and after replacement, and conditions of cleanliness, oil film, and chips, as well as actual lighting conditions on site. There is no need to intentionally create dangerous conditions on the equipment. Agree in advance on the range that can be reproduced with removed parts or safe test jigs. Test the distinction between parts with known chipping and normal products, re-imaging after cleaning, and the transition to poor image quality when lighting deteriorates.

Create an acceptance table with columns for “Input Conditions,” “Method of Confirming True State,” “Expected System State,” “Human Work,” and “Record.” Agree not to silently exclude images that could not be captured from the denominator, not to count consecutive images of the same part as independent cases, and to record threshold adjustments after testing as a separate version. Determine accuracy targets based on the quality impact when missed and the number of samples measured on site, rather than borrowing general values from vendors.

13. Fix the Scope Comparable in the RFP

In the request for proposal, specify the target jigs/parts and part numbers, installation diagrams, required field of view, whether it is during stoppage or operation, conditions of chips, oil, temperature, lighting, connection with existing MES and maintenance systems, network boundaries, storage period, and viewing permissions. Bidders should provide separate estimates for cameras and lenses, lighting, stands, jig modifications, gauges, model development, on-site training, maintenance, and retraining. Even if the model is cheap, if the maintenance cost of imaging conditions is high, comparisons will be incorrect.

Output specifications should not be limited to images and judgment values. Make it possible to export jig ID, part revision, image quality, model version, threshold version, reviewer, measurement results, and stop/release history, and specify a format that can be read after the contract ends. If using an external cloud, confirm the destination for image transmission, storage location, availability for retraining, and continuity of operation in case of failure. ISO/IEC 42001:2023 indicates the requirements for AI management systems, but having certification alone does not mean the model will pass in this field.

14. Separate Evidence in FAT and SAT

In FAT, before delivery, confirm the imaging recipe, result display, hold, permissions, history, export, and behavior during communication interruptions with the agreed test jigs and specimens. Passing FAT does not prove performance under the actual lighting, chips, temperature, and operator flow in the Thai factory. Save not only videos and screenshots of test results but also input image IDs, methods for verifying true values, model versions, and each judgment.

In SAT, fix the local stand and test normal operations including shift and part number changes, before and after cleaning, and equipment temperature changes. When a candidate for suspension is issued, confirm to whose screen it is notified, that the actual item and work ticket can be tracked, and that the jig does not return without approval. Attach deadlines and provisional operations to unmet acceptance items. If only AI detection succeeds and re-verification after replacement is not connected, the purpose of system introduction has not yet been achieved.

15. From Suspension to Replacement, Re-verification, and Release

AI Vision for Inspection-Fixture Wear: Deployment and Acceptance in Thailand - figure 3

Distinguish states such as “In Use,” “Needs Confirmation,” “Cleaning/Re-imaging,” “Suspended,” “Under Repair/Replacement,” “Awaiting Re-verification,” and “Approved for Resumption.” When a person confirms “Suspension,” not only attach a physical tag to the jig but also prevent misuse on the equipment or operational side. If there is suspicion of delayed suspension, track the inspected items from the last confirmed normal point by jig ID and lot. The quality department evaluates the risk to decide which products to re-inspect.

During replacement, identify removed pins and bushes, record part numbers and lots of replacements, fastening/installation, jig cleaning, seating of reference workpieces, and necessary dimensional measurements. Even if re-imaging is normal, do not release until the reference position and gauge readings are confirmed. Record the releaser and time, changed image recipe/model version, and next inspection. Follow safety procedures and energy isolation for each equipment during replacement work, and do not substitute AI screen approval for safety procedures. The reproducibility of attachment/detachment, as handled in Quick Change Fixture Selection, is also a confirmation target when using alternative jigs.

16. Calculate Cost-Effectiveness Within a Verifiable Range

Include not only cameras but also lighting, fixed stands, jig modifications, measuring instruments, software, MES/maintenance integration, training, maintenance, part replacement, and model updates in the introduction cost. Accumulate effects from inspection man-hours, excessive jig replacement, erroneous stoppages, rework, and quality leakage, comparing current and post-introduction under the same conditions. Do not confirm the amount of “quality accidents that did not occur” as actual reduction. Include increased review man-hours due to model improvement and re-verification costs after lighting replacement in operating costs.

First, in a PoC narrowed down to target jigs, measure whether images can be stably captured, whether cleaning and wear can be distinguished, and how many human judgments are needed. Accept product functions in FAT and local closed-loop in SAT, and decide on horizontal deployment by looking at actual stoppage/replacement history after operation. Numerical targets should be predetermined from important characteristics of the factory and past accident/maintenance data, and this article cannot set uniform accuracy or recovery years.

17. Use Jig Ledger as a Common Key for Images and Measurements

The first thing to create during implementation is not a pile of learning images, but a jig ledger. If there are multiple jigs with the same drawing number, assign each a unique ID. Assign different part numbers and revisions to interchangeable parts such as base plates, positioning pins, bushes, and receiving surfaces. Simply labeling as “JIG-01” would lead to comparing images before and after pin replacement as the same state. Design the image key by combining the jig unique ID, replacement episode ID, shooting session ID, and image ID, while keeping the product number and equipment number as related attributes, making it easier to track before and after replacements.

In operation, what needs to be confirmed is “which part of the jig currently attached to the line this image was taken of.” Link the shooting point ID indicating the fixed camera position, lighting recipe, reference image version, model version, and judgment threshold version to the image ID. If the method involves removing the jig and taking it to another station, not only read the jig ID but also confirm that the jig matches the product type and is correctly seated on the shooting stand guide before shooting. Passing seating abnormalities to AI’s wear judgment would lead to misdiagnosing “the jig is worn out.” If the method involves shooting while fixed to the equipment, include equipment stop status, safety fence position, and work presence in the imaging conditions.

The ledger should also include inspection methods for each part, replacement conditions, verification methods after replacement, and affected product characteristics. For example, if the receiving surface is dirty, perform re-imaging after cleaning first, if the pin diameter is suspicious, proceed to gauge measurement, and if a stopper has loosened, proceed to check the position and tightening state. Sending all photos to the same work ticket prevents maintenance personnel from estimating necessary tools and downtime. Confirm the treatment table by part before the model’s classification label, and later test if the HMI can correctly call that treatment.

Leave approval reasons for master image revisions. Unconditionally promoting new images to normal standards just because the jig was replaced risks standardizing machining scratches or misassemblies on new items. Attach approved drawings, dimensional records of the actual item, and functional confirmation with a reference work, and do not erase the old version but add an effective start date. When deploying to multiple factories, do not just copy image files, but re-confirm validity by unit of camera, lighting, stand, and jig revision combinations.

18. Do Not Pass Imaging Failures as “Normal Jig”

What occurs first on-site is not AI’s lack of learning but issues with images being incomparable. Loose screws on fixed cameras, oil on lighting diffusers, changes in pin metal luster due to seasons or temperature, and operators placing the master in reverse. These create differences similar to wear images while also hiding actual defects. Place an imaging quality gate before model inference to check if there are fixed reference points in the field of view, focus, exposure, light amount, shielding, and target posture within conditions. Images that fail the imaging-quality gate are not given a pass/fail score but sent for re-imaging or physical inspection.

Test specimens are not just “chipped pins.” Separate clean and normal jigs, receiving surfaces with only chips, the same receiving surface after cleaning, usable wear marks, parts that actually need replacement, images with changed lighting, and images with focus shifts. Judging something that can be restored by cleaning as a replacement increases labor, and judging a chip as mere dirt leaves quality risks. For any specimen, leave evidence that quality and maintenance confirmed the actual condition. Composite images with defects drawn on photos can aid training but do not replace acceptance tests with actual items.

Input ConditionExpected System BehaviorHuman Confirmation and Record
Clean, approved jigDisplay normal candidatesSave jig ID and reference image version
Removable chips on receiving surfaceSuggest cleaning and re-imagingSave images before and after cleaning as a pair
Chipped pin needing replacementSuggest usage stop and indicate partPhysical confirmation, stop approval, product range confirmation
Focus shift or shieldingHold judgment due to poor image qualityRecord cause correction and re-imaging
Jig after part replacementHold comparison with old imagesConfirm dimensions and reference work, approve new standard

The above table is an example from a specification document, not the acceptance threshold itself. In acceptance, leave a line for the true state of the specimen, shooting conditions, expected state, result, operator, and model version. If the same specimen is re-shot after changing lighting, confirm the difference between old and new images, and revise the reference image and model if necessary. When incorporating changed images into learning, do not mix the test set already used for acceptance into learning. Otherwise, the next evaluation loses independence.

19. Report Attribute Judgment and Dimensional Measurement Separately

If a PoC report is reduced to a single line of “AI accuracy rate,” important failures become invisible. First, divide the true value of the jig state into those that return with cleaning, usable wear, wear requiring replacement, judgment hold, and poor image quality. Next, create a combination table with system output, showing the number of cases where jigs needing replacement were treated as normal, the number of cases where normal jigs were stopped, and the number of cases held or re-imaged. Indicate both the number of images taken and the number of jig units as the denominator, and do not inflate independent cases by the number of consecutive shots of the same pin.

Pre- and post-replacement of the same jig, close-up images on the same day, and images taken by the same operator with the same lighting are correlated. During model development, consider separating learning, adjustment, and final evaluation by jig unit or part replacement episode. If there are plans to use under different part numbers or shifts, include those conditions in the final evaluation. For rare chips with few specimens, report the observed number of cases and unverified parts rather than concluding “high detection rate.” Leaving human inspection until additional physical items are gathered is a legitimate conclusion of the PoC.

Separate the report of image classification and the report of pin diameter and receiving surface height measurements. For the latter, look at calibration records, measurement range, repeatability, differences by measurer, temperature, and relation to product tolerances. While Gage R&R considerations are useful for systems measuring continuous dimensional values, do not directly apply the same calculations to image AI attribute classification. Confirm inter-evaluator agreement, false acceptance, false rejection, and holds for image attribute judgment. For quantitative values near tolerance boundaries, agree on judgment rules considering uncertainty in advance. This is to avoid confusing regulatory handling with AI’s suspicion threshold.

Leave images of failure examples and physical confirmation records in evaluation materials. Break down why something was missed into lighting, shape, label, model application range, and operation procedures. Simply writing “address with retraining” leaves ambiguity about who will approve specimens next and which version will be re-evaluated. After retraining, re-test both specimens that passed with the old version and new failure specimens to confirm no deterioration. The ability to revert model versions is also part of operational acceptance.

20. Hand Over to Operations Where Responsible Parties Can Sign

The state of inspection jigs involves production technology knowing the mechanism, quality judging the impact on inspection results, and maintenance handling replacement and recovery. IT/OT supports camera and history operation. Placing “AI manager” in just one department without the authority to stop actual items does not close operations. In the RFP, write roles for the person receiving the first notification, the person confirming the stop, the person determining the range of affected products, and the person approving the restart. If Thai and Japanese displays are needed, match the translation of judgment reasons, holds, usage stops, and restart approvals with on-site procedure manuals.

In operational handover, not only train on normal screen operation but also rehearse re-imaging for dirt, cases where the model does not provide an answer, communication interruptions, lighting replacement, jig part replacement, and incorrect recipe selection for part numbers. Decide per equipment whether jigs can be used during communication interruptions or if existing manual inspections should be reverted to. Also, test how to handle lost images or duplicate work tickets upon recovery. Simply continuing to save images makes them difficult to use for quality audits, so enable search by jig ID, period, parts, stop reason, and approver, and trace back to original images and measurement records.

After implementation, periodically review false acceptance candidates, false stops, holds, poor image quality, and incomplete work tickets. When product types, jig designs, lighting, camera positions, or suppliers of replacement parts change, review the model’s application range. Incorporate not only “AI retraining days” into maintenance plans but also camera cleaning, lighting inspection, reference instrument management, jig part replacement, and re-approval of judgment samples by the quality department. Maintain a state where the reason for stopping and evidence for resuming can be explained.

21. Create Estimate Comparison and Acceptance Tables with the Same Items

If you only ask for prices in the RFP, only look at screens in FAT, and only test maintenance integration in SAT, additional costs and blame-shifting occur. In the proposal stage, line up rows for “target jigs and parts,” “true value creation method,” “imaging conditions,” “judgment state,” “human approval,” “work tickets,” “data export,” and “procedures during failures,” and have each company specify standard support, additional development, and exclusions. Placing the estimate price column and acceptance evidence column in the same row reveals what is omitted in seemingly cheaper proposals.

For example, the phrase “detect pin chipping” does not clarify which diameter, material, surface treatment, lighting, angle, or chipping is targeted. Illustrate excluded backsides and shielded parts, leaving the responsibility for manual inspection there to the site. Even the phrase “includes model updates” should be divided into image acquisition, physical label approval, retraining, regression testing of existing specimens, switching to the operational version, and reverting to the old version. This not only clarifies post-contract change requests but also prevents the quality department from overestimating the system’s inspection capabilities.

Design the acceptance table to allow entry of preconditions, test specimen ID, expected output, actual output, image and measurement evidence, evaluator, and actions for unmet items for each test case. In FAT, confirm software and control integration before bringing it into the factory, and in SAT, focus on local light conditions and human operations. If temporarily operating with unmet items, clearly document which manual inspections to retain and who will make stop decisions during that period. Do not treat signatures meeting all items and conditional use approvals as the same.

22. Frequently Asked Questions

Can gauge measurement be omitted if AI passes?

If the design is for screening the fixture condition via images, it cannot be omitted. To determine dimensional compliance from images, the performance and uncertainty as a measuring instrument must be separately demonstrated. The frequency of independent measurements is determined by critical characteristics and site variability.

Can you distinguish between dirty and worn fixtures?

By combining images before and after cleaning with physical measurements, the range of distinction can be verified. However, conditions such as a chipped locating pin concealed by an oil film cannot be definitively determined by images alone. If indistinguishable, it is put on hold and not automatically resumed.

What do you stop when a locating pin is worn?

First, stop using the affected fixture. The range of inspected items that may have been impacted is determined by the quality department, considering the last normal check, fixture ID, product lot, and critical characteristics. It does not mean mechanically discarding all products.

How are PoC pass criteria and costs determined?

Agree individually on dangerous misses, false stops, holds, imaging failures, and review workload, and compare candidate methods with the same sample. Costs are estimated within the scope including site stands, fixture modifications, integration, and maintenance. Manufacturer’s nominal accuracy or unit price is not directly applied to the factory’s profitability.

23. Conclusion

In the introduction of AI image judgment for fixture wear, stable imaging, independent physical measurement, human-controlled stop and release, and exchange history must all connect to confirm effectiveness. AI judgment is the entry point for fixture management. Define target areas and responsibility boundaries in the RFP, investigate misjudgments and holds in the PoC, and confirm recovery after repairs in FAT and SAT.

If considering visual inspection of fixtures at the Thai plant or how to connect with existing gauge management, you can consult TOMAS TECH. It is fine to consult even before drawings, inspection records, and exchange history of the target fixtures are complete.

Reference Information

  1. ISO 10012:2026 — Requirements for measurement management systems.
  2. ISO 14253-1:2017 — Decision rules for proving conformance or non-conformance with specifications considering measurement uncertainty.
  3. ASQ Gage R&R — Repeatability and reproducibility of continuous measurement.
  4. NIST: Validation of upgrades to the M48 vision measurement system — Non-contact edge detection and equipment-specific validation.
  5. NIST: Selecting dimensional measurement equipment — Selection of measurement instruments according to inspection tasks.
  6. Cognex: Locate Fixture — Setting reference points on images.
  7. Cognex: Inspect Feature — Inspection functions for image features.
  8. NIST AI RMF Measure playbook — Evaluation, monitoring, and human review during operation.
  9. ISO/IEC 42001:2023 — AI management systems.