QualitySpecialized caseVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI checks every spot weld from the welding controller's data, instead of sampling one car a day with ultrasound (Audi, Volkswagen)

Added 2026-10-26 · Last updated 2026-10-26

Manufacturers in this case: Audi · Volkswagen · all manufacturers using AI

Role

Quality (in-process quality control: resistance spot welding in the body shop) · Typical users: quality engineer, welding or body-shop process engineer, quality inspector, production supervisor, automation or data engineer

The problem

A car body holds thousands of resistance spot welds (metal sheets joined by pressing electrodes together and passing a strong current), and the usual way to check them is destructive or manual sampling. At Audi's Neckarsulm plant, which assembles up to about 1,000 vehicles a day with about 5,000 welds each (5,300 in the body of an Audi A6), more than 5 million welds are made every day. The standard method was to pull one car off the line each day and have 18 engineers test its welds with ultrasound probes, recording the results on clipboards, which left no in-line check of the welds on the other 999 cars. 'Our big goal for this solution is to make it possible for us to inspect 100% of our welds with a very high degree of accuracy,' said Mathias Mayer of Audi.

How AI helps

A machine-learning model reads the data that the welding-gun controllers already produce for every weld: the electric voltage and current curves, the weld configuration, the metal types and the condition of the electrodes, about 200 variables per weld, each weld lasting 400 to 1,000 milliseconds. The model was trained by comparing its predictions with Audi's real ultrasound inspection results. It flags anomalies in real time, and a lineside screen tells the inspector which welds to check, so people test the suspicious welds instead of a random sample. The proof of concept ran with Intel's Industrial Edge Insights software at the edge (on a computer beside the line rather than in a remote data center); the production version runs in the Volkswagen Industrial Cloud on AWS. To make the method audit- and certification-proof, Audi developed it with the German Association for Quality (DGQ) and the Fraunhofer IAO and IPA institutes.

Step by step

1. Check what your welding controllers already record for each weld (current, voltage, resistance or energy curves, weld program, electrode tip-dress count) and whether it can be exported with a weld ID and a timestamp.
2. Collect the ground truth: past ultrasound, chisel or teardown results matched to the same weld IDs. Without this the model cannot be trained or checked, as Audi did with its own inspection data.
3. Start in advisory mode on one cell or one body-in-white (the bare welded car body, before painting) model: the model ranks welds by risk and inspectors test the high-risk ones plus a random sample.
4. Keep the random sample running during the trial and measure how many bad welds the AI found and how many it missed.
5. Agree with quality management and your customers how the AI check fits the control plan and audits; Audi worked with DGQ and Fraunhofer so the result of each AI-based test could be explained.
6. Retrain the model for each plant and each change in weld settings, sheet stack or electrode type, as Audi does before each roll-out.
7. Only then reduce the sample size, step by step, and review weekly with the welding team.

Tools

• Audi WPS-Analytics (in-house machine learning on welding-controller data): proof of concept with Intel Industrial Edge Insights, production on the Volkswagen Industrial Cloud (AWS, with Siemens for machine integration); used at Audi Neckarsulm; Audi announced in June 2023 a roll-out to Audi Brussels, Volkswagen Emden and Audi Ingolstadt (we have not checked the current status)
• Other tools to evaluate: the weld-monitoring and data-export options of your welding-controller supplier, and anomaly-detection platforms that read controller or PLC data
• For a first offline study: any LLM with a data-analysis or code tool (ChatGPT, Claude, Copilot, Gemini) on an export of weld data matched with past ultrasound or teardown results, with the prompt below

In the AI tools index: Intel Industrial Edge Insights · all AI tools used by manufacturers

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

• The AI points inspectors to suspicious welds; it does not release parts on its own. Keep a random sample running until the model's misses are measured on your own welds.
• A weld-quality model is specific to weld programs, materials and settings: Audi retrains it for each plant before roll-out.
• Changing the inspection method can affect your control plan and customer requirements (for example IATF 16949): involve quality management and, if needed, the customer before reducing sampling.
• The figures come from Audi's own press release and an Intel customer story; Audi has not published a defect-rate or cost result for the welds.

You are an experienced welding and quality engineer for a body shop.

I will paste: (1) an export of weld data, one row per weld [weld ID, timestamp, robot or gun, weld program, current, voltage or resistance values or curve summary, weld time, electrode tip-dress count, sheet stack or material]; (2) ultrasound, chisel or teardown results for some of those welds [weld ID, result, defect type]; (3) our current sampling plan.

Do this:
1. Describe the data: missing values, how many welds have an inspection result, and how many of those are bad.
2. Compare good and bad welds: which variables differ most, per weld program and per gun, quoting row counts. Mark which differences could be causes and which could be effects.
3. Propose simple first rules or a model I could test (for example a limit on a curve value, or a score per weld), and how I would check it against the inspection results without fooling myself.
4. Suggest how inspectors could test the highest-risk welds first while keeping a random sample, and what to measure during the trial.
5. List the questions for our welding-controller supplier and our quality manager before using this in the control plan.

Write every calculation as a spreadsheet formula or code I can run myself; do not compute statistics in your head. Mark every rule as unproven until tested.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Related cases

• Anomaly detection on machine and process data to cut scrap
• Vision defect inspection on the edge
• AI balloons the part drawing and fills the first article inspection report, instead of retyping every dimension (AdvancedPCB, Lake County Tool Works, B&B Airparts)
• Automation and AI check supplier certificates of analysis against your specifications before material is released (Georgia-Pacific, C.H. Guenther, two anonymous manufacturers)

All 12 Quality use cases →

Sources

• https://www.audi.com/en/press-releases/audi-begins-roll-out-of-artificial-intelligence-for-quality-control-of-spot-welds-15443
• https://www.intel.com/content/www/us/en/customer-spotlight/stories/audi-automated-factory.html
• https://www.automotivemanufacturingsolutions.com/oems/the-application-of-ai/534291