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

AI predicts which boards really need the slow test, so X-ray and end-of-line test stations stop being the bottleneck (Siemens Amberg, ZF, Flex)

Added 2026-11-13 · Last updated 2026-11-13

Manufacturers in this case: ZF · Flex · Siemens · all manufacturers using AI

Role

Quality and test engineering (end-of-line, X-ray and functional testing) · Typical users: test engineer, quality engineer, process engineer, electronics or mechatronics plant manager, data analyst working with quality

The problem

In electronics and mechatronics plants, every unit often goes through the same long test sequence, even when the process is stable and almost every unit passes. The slow test station then limits the whole line, and the usual answer is to buy another one. At Siemens Electronics Works Amberg (Germany), output of the circuit board line was limited by time-consuming X-ray tests of solder joints, and every further X-ray machine would cost about EUR 500,000. At ZF, electronic boards go through three 100% electrical checks, including a run-in test at -40, +20 and +120 degrees Celsius, with about 1,000 measurements on every board: costly, slow and energy hungry. At a Flex customer in Guadalajara (Mexico), one section of the circuit board test had more than 50 steps and lasted over two hours, and a failure late in the sequence sent the board back for debugging and a full retest.

How AI helps

The model learns from the data the line already produces to decide, unit by unit, whether the expensive test is needed. At Siemens Amberg, data from the solder paste printer, solder paste inspection, pick-and-place machines and automatic optical inspection (more than 40,000 production parameters captured in real time) trains a model in the cloud; the model then runs on an edge device beside the line and tells the manufacturing execution system, board by board, whether an X-ray test is needed or can be skipped. At ZF, a predictive model learns from the in-circuit test results, flags the boards likely to have anomalies and sends only those to the physical stress test; quality engineers check the model's decisions in a browser app and add comments. At Flex, a machine learning application learned from test data to move the most frequently failing steps to the front of the sequence and to point out redundant tests. Engineers decide which tests may be removed or skipped, and the full test stays available.

Step by step

  1. Pick one bottleneck test station and measure its cycle time, its share of line capacity and its failure rate per test step over several months.
  2. Collect the data upstream of it for the same units (inspection results, process parameters, earlier test results), linked by serial number.
  3. Start with the easy win: reorder steps so the ones that fail most run first, and list steps that never fail for engineering review.
  4. Train a model that predicts pass or fail at the slow test, and run it in shadow mode: every unit is still tested, and you compare.
  5. Agree with quality and your customers on the rules: which tests may be skipped, the minimum sample rate, and when the line goes back to 100% testing (process change, new lot, drift alarm).
  6. Once in production, audit skipped units with periodic full tests and track escapes and field returns.

Tools

  • Siemens edge computing with a cloud-trained quality prediction model linked to the Simatic manufacturing execution system: used at Siemens Electronics Works Amberg
  • KNIME Analytics Platform with Python and a browser-based data app: used at ZF
  • In-house AI/ML test optimisation application: used at Flex (Guadalajara and a second North American site)
  • Other tools to evaluate: the analytics module of your test executive or MES, statistical test-time reduction tools from your test equipment supplier, Python or KNIME on your own test logs
  • To start without new software: a test results export and any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below

In the AI tools index: KNIME Analytics Platform · all AI tools used by manufacturers

Result

Siemens Electronics Works Amberg (Siemens publications): X-ray tests reduced by up to 30% with a quality rate of 100%, and a EUR 500,000 investment in a further X-ray machine avoided. ZF (KNIME success story, the software vendor): physical end-of-line testing reduced by 95%, seven-digit savings on stress-test equipment investments and expected seven-digit yearly energy savings. Flex (company article): at the Guadalajara site, overall test time down 30% and down 50% when a unit fails, built in two months; at a second North American site, 10 to 25 hour test times cut by 20 to 25% and 35% more electronics produced and tested per day. Caution from research: a 2026 peer-reviewed study on 28,000 circuit board runs (Frontiers in Mechanical Engineering) found reduced test plans that saved 18.78% and 91.57% of test time with no escaped defects on past data, but a fixed reduced plan let 110 and 8 defects escape when the process drifted later, while an adaptive policy that returned to the full plan when the process became unstable brought escapes back to zero.

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

Human review: cautions

  • Figures come from the companies themselves (Siemens, Flex) or the software vendor (KNIME for ZF); they are not independently audited.
  • Skipping tests is a quality and sometimes a contractual or regulatory decision: check customer requirements and standards before removing any test.
  • A reduced test plan that works on past data can let defects escape when the process changes; keep sampling, drift alarms and a way back to 100% testing.
  • The model is only as good as the traceability: every unit must be linked to its upstream data by serial number.

Try it yourself

You are a test engineer in an electronics or mechatronics plant. I want to know where our end-of-line test time can be cut safely.

I will paste: (1) a test results export for the last few months [serial number, date, test step ID, step name, measured value, limits, pass or fail, step duration]; (2) our line takt time and the cycle time of the test station.

Do this:
1. For each test step, give its failure rate, its average duration and how many failures it caught that no other step caught.
2. List the steps that never failed in the period, and the steps whose failures always come with a failure of another step (possible redundancy).
3. Propose a new order of steps that finds failures earlier, and estimate the time saved per good unit and per failed unit.
4. List what would have to be true before any step could be skipped (sample size, process stability, customer requirements, standards), and the data we are missing.
5. Draft a shadow-mode trial plan: what to compare, for how long, and what result would stop the trial.

Do not recommend removing a test only because it rarely fails: mark each proposal as a hypothesis for engineering review. 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.

Want more like this? Join the newsletter

Related cases

All 19 Quality use cases →

Sources

Weekly: 1–2 practical AI use cases for your plant

Newsletter in

Free, one email a week. By subscribing you agree to receive the weekly newsletter (and its sponsor message) from AI Agents for Manufacturing. Every email includes an unsubscribe link, and you can unsubscribe anytime. See our Privacy Policy.