Tag Failure Analysis: The Read Rate You Measure Is Not the Reliability You Need

Every tag project begins with a demonstration in which nothing fails. Every tag project that survives begins with a failure analysis process that assumes things will. The gap between those two statements is where most deployments quietly lose their business case.

Failure is a distribution, not an event

A tag does not simply work or not work.

• The same model fails at different rates in different environments, and the rate is the number that matters.
• A 2 percent dead-on-arrival rate is normal; a 2 percent in-field failure per year is not.
• Failure modes cluster: a bad lot, a bad adhesive, a bad printer setting, a bad application angle.
• The relevant metric is reads per pass at the required confidence, not reads per second in a lab.
• Reliability is a curve over temperature, humidity and time, and you need at least two points on it.

The failure modes that actually appear in the field

Know the taxonomy before you argue with a supplier.

• Dead on arrival: no response at any power level, usually a damaged chip or a broken strap.
• Weak but alive: reads only at close range, often a partial strap bond or a cracked trace.
• Detuned: works on a plastic shelf and fails on a metal one, because the environment shifted the resonance.
• Adhesive failure: the tag is intact and on the floor, which is a labelling process problem, not a tag problem.
• Printer damage: a crushed inlay that failed at the point of encoding and was never verified.
• Read collision: individually readable, unreadable in a crowd, which is a reader and layout problem.
• Environmental decay: works in the warehouse, fails after a summer on a loading dock.

Factory QC: who controls which defect

The supply chain has three parties and each owns a different failure.

• The chip vendor controls the die and the wafer-level test.
• The inlay converter controls the strap attach, the antenna print and the reel-level test.
• The label converter controls the face stock, the adhesive and the die cutting.
• The printer and the applicator control the encoding and the physical placement.
• Ask each party for the test coverage, and be suspicious of anyone who claims all of it.

Incoming inspection that means something

Sampling is a statistical claim, and it has to be designed.

• Define the acceptance quality level before the first delivery, not after the first complaint.
• Sample across the reel, not from the first metre, because defects cluster near splices.
• Test at the frequency and the power you will actually use, not at the vendor's optimum.
• Record the serial range of every reel, so a failure can be traced to a lot rather than to a brand.
• Keep a retained sample per lot; it is the only evidence you will have in six months.

The read test that matters, and the one that lies

Bench tests measure the wrong thing very accurately.

• A single tag on a wooden table at 30 centimetres tells you almost nothing about a pallet of 60.
• Test the population, at the density, in the geometry, with the reader you will deploy.
• Measure the percentage of tags read in a fixed dwell time, repeated enough times to get a distribution.
• Include the worst-case orientation, because the worst case is what generates the exception.
• Report the confidence interval, or the number is decoration.

Stress screening before deployment

Accelerated ageing is cheaper than a failed rollout.

• Thermal cycling catches the adhesive and the strap bond failures that arrive in month four.
• Humidity exposure catches the label stock and the printing degradation.
• Abrasion and impact testing matches the actual handling, not a generic standard.
• Chemical exposure should use the real substance, at the real concentration, for the real contact time.
• Run the screen on the production lot, not on a golden sample the vendor hand-picked.

Identity integrity is a quality attribute

A tag that reads but carries the wrong number is worse than a dead tag.

• Verify the encoded value against the intended value, at the point of encoding, every time.
• Check the EPC length, the filter value and the checksum, not just the serial.
• Test that the same tag cannot be read twice into two different identities.
• Keep the encoding log, so a duplicate can be traced to a printer or an operator.
• Treat a mis-encoded tag as a quality escape, with the same rigour as a physical defect.

Field returns and the failure analysis loop

The loop is what turns incidents into improvements.

• Require a return of every failed tag, with the location and the date it was applied.
• Cut open the returned tag and record which failure mode it is; do not accept "it stopped working".
• Aggregate by lot, by site, by applicator and by season, because the pattern is the finding.
• Feed the finding back into the incoming test plan, and change the sampling if the pattern moved.
• Publish the failure rate internally; a number nobody sees never improves.

Supplier quality agreements and lot traceability

Put the obligation in the contract, or it does not exist.

• Specify the test coverage, the acceptance level and the notification duty for lot changes.
• Require advance notice of any change in inlay, antenna, adhesive or chip, because silent substitutions are common.
• Require lot-level traceability from the reel to the shipment.
• Define the remedy for a lot that fails incoming inspection, including the cost of the recall.
• Audit the supplier's test process at least once, and look at the data rather than the certificate.

Deployment verification on site

The last test is the only one that counts.

• Read a known-good reference tag at every read point before the shift starts.
• Walk the actual route with the actual vehicle or the actual person, at the actual speed.
• Record the pass rate for a full shift, not for a five-minute demo.
• Check the read points that were added late, because they are the ones that were never modelled.
• Sign off with the operations team, not only with the project team.

When the failure rate is two percent

Two percent is either acceptable or fatal, and the difference is arithmetic.

• Calculate the cost per exception: labour, delay, rework, and the customer conversation.
• Compare it against the cost of a better tag, a better applicator or a better read point.
• Decide whether the fix is technical or procedural; a process fix is usually cheaper.
• Consider a hybrid: tag plus a manual verification step for the high-value items only.
• Do not accept a rising trend as normal variation; the trend is the early warning.