Skip to content
0727
InsightsAcademyCasesEventsSign in
中文/EN
Cases/Bosch
← Back to cases
← Previous6 / 127Next →

Bosch

Synthetic defect data for stator-weld visual inspection
First-party enterprise source Automotive component manufacturing Enterprise deployment ⚙ Technical reference Deep case | Key delivery chain is substantially documented Evidence level A Plant pilot / cross-site replication
A Evidence level
Generative AI in manufacturing
Publisher: Bosch · First-party technical case · Direct case-level source
Claim origin: Company project-team disclosure · Independent verification: No · Accessed: 2026-09-19
Evidence level measures whether a source can be located and reviewed; it does not mean vendor-reported claims were independently audited.
Deep case | Key delivery chain is substantially documented 12 / 12
Business context2 / 2
Transformation workflow2 / 2
Technical workflow2 / 2
Human roles & governance2 / 2
Measured outcomes2 / 2
Source traceability2 / 2
All six dimensions meet the current completeness threshold.
Bosch uses generative AI to create training data rather than final inspection decisions, then connects camera inspection, automated rework routing, and expert false-positive calibration into a production loop.
Business problem

Automated optical inspection needs images covering many defect types, but Bosch’s manufacturing quality is high, so a new line cannot naturally accumulate enough defective parts. Waiting for real defects delays model and line commissioning; deliberately damaging parts is expensive and still may not cover the full defect space.

Solution

Bosch Research used images from an earlier product and only a double-digit number of real examples per defect class to perform domain transfer and generate roughly 15,000 synthetic defect images. Twelve 2D/3D camera views capture stator welds. A vision model trained on real and synthetic data classifies each part and automatically routes repairable defects back into production, while imaging experts review false positives and retrain the model.

Technical architecture & production workflow
Step 01
Images from an earlier product plus a small set of real defects
→
Step 02
Domain transfer generates synthetic images for the new product
→
Step 03
Vision model trained on combined real and synthetic data
→
Step 04
Twelve 2D/3D camera views capture stator welds
→
Step 05
Online classification automatically routes repairable parts back to production
→
Step 06
Experts review false positives, adjust acceptance rules, and retrain
Key technology & infrastructure components
Generative vision modelSynthetic dataDomain transfer2D/3D industrial camerasAutomated optical inspectionRework routingExpert retraining
Human roles & accountability

The vision model performs high-speed screening and controls the rework route. Imaging engineers review false positives—for example, whether a weak weld is still acceptable—and update the OK/defect boundary through manual retraining. Non-repairable defects remain subject to the plant’s quality process.

FDE delivery actions
  • Define the six weld-defect classes and their repairability
  • Inventory reusable legacy images and the limited new-product samples
  • Use synthetic data to cover long-tail defects instead of waiting for failures
  • Connect model output directly to the rework flow
  • Have plant experts continuously recalibrate false positives and acceptance boundaries
Reusable delivery patterns
  • Synthetic data can address scarce defect examples in high-quality manufacturing
  • Validate synthetic data together with real samples
  • Map model decisions to concrete rework and scrap actions
  • Domain experts must continue to own the production quality boundary
Business outcomes & delivery results

Bosch reports that approximately 15,000 synthetic images were produced from a double-digit number of real images per defect type. The plant expects the approach to shorten the project by six months and generate annual productivity gains in the six-figure euro range. These are projected benefits while the new line is still being trained and expanded.

Evidence boundaries & verification notes
  • The six-month reduction and six-figure euro gain are Bosch projections; they are not stated as audited realized benefits.
  • The public article does not provide train/test splits, false-positive rates, or out-of-sample evaluation.
Primary source: Generative AI in manufacturing ↗
Traceable does not mean independently audited
Open primary public source
0727.ai · Trusted agents, built together.Case research: FDE-case-library ↗ · MIT