You already photograph every weld. Nobody checks them one by one: from today, something does

Do you recognise at least two of these?

• Every unit is photographed at assembly, but the photos end up in an archive nobody reopens.
• Weld defects are found at end of line, or worse, by the customer.
• Quality control checks a sample, not everything, because there is no time.
• A vision-system vendor proposed cameras, licences and a yearly fee before seeing a single one of your photos.
• A large customer asks for evidence of inspection on 100% of the joints.

If you said yes at least twice, you already own the most expensive ingredient: the photos. What is missing is something that looks at all of them, every day, without getting tired.

What a defect that gets out costs

Do the sums on your own product:

• A pack or assembly reworked at end of line costs the time to strip it, plus the part, plus the idle station.
• A defect found by the customer costs the return, the 8D report, and often the whole batch under suspicion.
• A defect in the field on a battery is not a cost: it is an incident, with recall and liability.
• Every photo taken and never looked at is an inspection you paid for and did not use.

With these numbers in front of you, a system that flags suspect welds to quality control within a minute is not compared with "nothing": it is compared with the next return.

What if every weld got a score, and you only saw the ones worth a look?

The system learns what a good weld looks like from the thousands of good welds already in your archive: it learns the normal, it does not need a catalogue of defects. Then it looks at every new photo, finds the welds, compares them and scores each one. If a photo is blurred or unusable it asks the operator to retake it within two seconds.

Quality control receives only the "to verify" queue, sorted by severity, with the crop and a heat map on the suspect spot, and three buttons: defect, false alarm, invalid photo. A person decides; the system never stops production and never judges operators. Every verdict becomes a lesson: month after month, useless alerts drop.

The unit's passport gets the line "visual inspection: zones checked N of M". Coverage is always stated, never implied: it is what a large customer wants to read.

What came out of 23,194 real photos (and why it matters to you)

At a battery-pack assembler in North-East Italy the whole station archive was analysed before proposing anything: 42,097 files, 23,194 unique photos, 3,524 packs, nine months. About 8,500 close-ups of welds, three joint types, five stations, and zero labelled photos.

• On one product family the hand-held camera frames about 1 weld in 5-8 and the back is never photographed; on the other, 3 photos per pack already cover 100% of the spot welds in 795 packs out of 859.
• Over 94% of photos have strong reflections on nickel: unmasked, they become the first cause of false alarms.
• 66 black or empty-bench photos, 1,197 duplicates, 3% at unusable resolution: to discard before training anything.

This is the part no vision-system vendor shows you: the limit is not the software, it is the photos. Your project is designed on yours, and the first thing you receive is this kind of analysis, not a quote.

How it works on the floor, without stopping anything

1. Quality filter at the station: black, blurred or sub-1080p photo → retake request within 2 seconds.
2. Sorter and locator: close-up or whole pack, then every spot-weld group and every solder joint, with what covers them marked (labels, cables, fingers).
3. Anomaly models, one per joint type, trained only on good joints: score and heat map.
4. Verdict per unit: OK, to verify, observed, not assessable; you choose the threshold on the number of false alarms per day quality control accepts.

Your management system stays in charge of the flow: it receives the photo as today, every minute passes the unjudged ones to the vision service, and the result lands in the panel quality control already uses. Everything runs on a computer inside the company: photos never leave. Only Apache 2.0 licensed models: no patents, no fee for the vision software, no proprietary hardware.

The rules that protect you: people, privacy, regulation

• Results are never linked to operator evaluation (EU AI Act, Annex III); reports show only station and batch.
• Cameras also capture hands and work rhythms: before a fixed camera, the local rules on workplace monitoring apply (in Italy, a union agreement or labour-inspectorate authorisation). The technical documentation is prepared for you; the filing stays with the employer.
• Impact assessment and privacy notice, data-processor appointment, worker information and training (EU AI Act, Article 4).
• Model version recorded in every result; a model change is a release with parity tests on the reference set. Retention per IATF 16949: verdicts outlive the photo.

It starts with a trial on your photos, and the decision is made with numbers

1. Study and reference set: cleaned history, 600 welds judged together with your quality manager, some thirty defects made on scrap material, first model trial. Closes with signed acceptance criteria: if the numbers do not convince, it stops there.
2. Shadow production: the chain runs on real photos, the system judges, quality control verifies, production does not change one bit.
3. Calibration and commissioning: thresholds tuned on the real line, automatic drift monitoring, retraining with parity tests.
4. Full coverage: fixed camera on an arm with diffused light and guided shooting row by row, from a sample to every weld, front and back.

What you need: a server with a GPU, 1080p or better webcams at the stations, half a day of the quality manager at the start and 10-15 minutes a day on the queue. The price of each phase is set after the study, with a written quote.

The next step commits you to nothing: send ten photos from your stations and in 30 minutes you know whether it can be done and with what coverage. Book the 30 minutes.