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

AI listens to every motor, drive or gearbox at the end-of-line test and flags noisy units instead of a human tester (FAZUA, Johnson Electric)

Added 2026-10-30 · Last updated 2026-10-30

Manufacturers in this case: FAZUA · Johnson Electric · all manufacturers using AI

Role

End-of-line testing and final inspection of motors, drives, gearboxes, pumps and other moving assemblies · Typical users: quality engineer, test engineer, end-of-line inspector or team lead, production manager

The problem

Many products that move (electric motors, e-bike drives, gearboxes, steering systems, doors, pumps) are checked at the end of the line by a person who listens or feels for abnormal noise or vibration. That check is subjective, depends on the tester's hearing, focus and the noise around them, and does not keep up when cycle times change. A unit that slips through becomes a warranty claim. At Johnson Electric, a global maker of electric parts for vehicles, the acoustic inspection of windshield wiper motors was done by trained people, an approach the vendor's case study calls 'very subjective and error prone'. At a Tier-1 driveline supplier, the end-of-line test of electric power steering systems recorded four vibration sensors over eight tests, 96 signals per unit, and a team of engineers evaluated the results by hand. At FAZUA, a maker of e-bike drive systems that is part of Porsche eBike Performance, parts were checked manually by skilled quality staff.

How AI helps

Sensors (microphones, structure-borne sound or vibration sensors) record each unit during its test run. A machine learning or deep learning model is trained on recordings labelled by the experienced testers, learns what a good unit sounds like, and flags units whose sound or vibration pattern deviates, in real time. The quality engineer reviews the flagged units with spectrograms (pictures of the sound) and feeds corrections back to improve the model. At FAZUA, MHP's Sounce software detects deviations in the structure-borne sound patterns of components; possible causes, checked in detail afterwards, include drive system, bearing, assembly and electronics defects and contamination. At Johnson Electric, a Netico edge device records and analyses the sound of each DC motor under test and raises an alarm when a bad motor is detected. At the Tier-1 supplier, Acerta built a model on features engineered from the 96 vibration signals.

Step by step

Measure today's test: units per shift, how many the testers reject, how many noise-related complaints or warranty returns get through, and how much tester time it takes.
Record the sound or vibration of each unit during the existing test for a few weeks, together with the tester's verdict and the serial number, without changing anything else.
Have your best testers label a set of good and bad recordings, including the fault types you care about (bearing, gear, assembly, contamination); collect real bad units, which are rare, from rework and returns.
Train or configure the model, then run it in parallel with the human test; compare every disagreement and agree a target, for example no faulty unit passed and a small, set share of good units sent for re-test.
Go live with a human check on every flagged unit, and keep recording so the model is retrained when products, suppliers or the line change.
Track escapes (complaints and returns for noise), false rejects and tester time before and after.

Tools

MHP Sounce (AI acoustic quality control, software as a service): used at FAZUA and piloted at a car manufacturer
Netico Edge Acoustic Sense 100 and Netico Hive (edge device and machine learning on motor sound): piloted at Johnson Electric
Acerta (machine learning on end-of-line vibration data): used at the Tier-1 driveline supplier
Other tools to evaluate: the acoustic or vibration analysis module of your test bench supplier, industrial acoustic monitoring systems, open-source audio anomaly detection for a do-it-yourself pilot
For a first step without new software: your end-of-line test results and warranty returns in a spreadsheet, plus any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below

In the AI tools index: Acerta · MHP Sounce · Netico Edge Acoustic Sense · all AI tools used by manufacturers

Result

FAZUA (e-bike drive systems, Germany, part of Porsche eBike Performance), MHP Sounce, in use since April 2024: more than 96% of anomalies detected and 60% working time saved in quality assurance (MHP press release, July 2024, a vendor claim). Car manufacturer (not named), MHP Sounce pilot on car doors: over 96% accuracy in correctly identifying defects, 10% less processing time in FTE, a quality result under five seconds after the first noise recording and over 10,000 data points a day (AWS partner blog, June 2023). Tier-1 driveline supplier (not named), Acerta: in testing, the model met the target of 0% false negatives (no faulty unit passed) and under 1% false positives; it was trained on about 700 good units and tested with 27 failed units (Acerta case study, a vendor claim). Johnson Electric (wiper motors), Netico: a pilot under intensive testing; the vendor says the test process needed changes to build better training data and that new models were being tested, with no result published.

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

Human review: cautions

The FAZUA, car-door and Tier-1 results are published by the vendors (MHP, Acerta), and two of the three customers are not named. Acerta's figures come from testing on past data with only 27 failed units, not from a long production run. Johnson Electric's project has no published result.
Bad units are rare, so a model can look accurate while missing the faults that matter; judge it on missed faults and false rejects, not on overall accuracy.
Background noise (air guns, hammers, nearby machines) can fool acoustic models; test in real line conditions, and consider structure-borne sound or vibration sensors where the area is noisy.
If the product, a supplier or the test bench changes, the model must be checked and often retrained; keep a human check on flagged units and on safety-critical products.

Try it yourself

You are a quality engineer responsible for the end-of-line test of [product, for example electric motors or gearboxes].

I will paste: (1) our end-of-line test results for the last [3 to 6] months [date, shift, serial number, tester, pass or fail, reason for failure]; (2) noise- or vibration-related complaints and warranty returns for the same products [date, serial number, description, finding after analysis]; (3) a short description of the current test (how the tester listens or measures, cycle time, test conditions).

Do this:
1. Summarise the test results by shift, tester and failure reason, and tell me which spreadsheet formulas or pivot tables to use for the rates; do not compute totals or percentages yourself.
2. Match returns to test results by serial number and list the units that passed the test but came back for noise or vibration, with what was found.
3. List signs that the current test is inconsistent (differences between testers or shifts, failure reasons that are vague), quoting the rows.
4. Propose a plan to pilot automatic sound or vibration testing in parallel with the human test: what to record, how many labelled good and bad units we need, how to compare the two, and what targets to set for missed faults and false rejects.
5. Write 10 questions to ask a vendor (sensors, training data needed, how false rejects are handled, retraining when the product changes, integration with our test bench, price).

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 13 Quality use cases →