HomeSolutionsNewsCareersAboutLive DemoContact
🏭 CASE STUDY — AUTOMOTIVE

AI Vision Defect Detection
on Stamping Line

How MordeTech replaced manual visual QC with a real-time deep learning inspection system on a high-speed metal stamping line — achieving 94% reduction in defect escapes within 90 days.

ClientTier-1 Automotive Supplier
LocationPune, Maharashtra
IndustryAutomotive — Body Parts Stamping
Duration14 weeks (deployment + validation)
CompletedMarch 2025

The Challenge

The client's stamping line ran at 800 parts per minute producing structural body components. Manual visual inspection at this speed was impossible — operators were catching fewer than 40% of surface defects, micro-cracks, and dimensional deviations before parts reached downstream assembly.

The existing inline gauging system only checked a small sample (1-in-50) and couldn't detect surface anomalies. Warranty claims from assembly partners were rising, and an IATF 16949 audit had flagged the QC process as a high-risk non-conformance. The client needed a 100% inline inspection solution with zero impact on line speed.

Our Approach

MordeTech conducted a 3-day site assessment to map defect types, lighting conditions, and PLC integration requirements. We identified four primary defect categories to target:

  • Micro-cracks and hairline fractures on stamped surfaces
  • Dimensional deviations beyond ±0.3mm tolerance
  • Surface scratches, dents, and paint/coating voids
  • Edge burrs and material tears at cut points

We designed a custom inspection cell with four Basler acA4024 cameras (12.3MP) arranged around a 360° coverage fixture, triggered by the existing PLC conveyor encoder. An NVIDIA Jetson AGX Orin edge AI module runs the inference on-device, ensuring sub-15ms latency well within the part's 75ms conveyor window.

The deep learning model was trained on 28,000 labelled images captured from the line over a 2-week baseline period, augmented with synthetic defect generation. Training used a custom ResNet-based architecture fine-tuned with transfer learning from an automotive-specific defect dataset.

"We had tried off-the-shelf vision systems before — they couldn't handle the reflective surfaces and speed. MordeTech built something completely custom that just works. The ROI was visible within the first month."

— Quality Manager, Tier-1 Automotive Supplier, Pune (name withheld at client request)

Technical Implementation

  • 4× Basler acA4024 GigE cameras with structured LED ring lighting (eliminating reflections on metal)
  • NVIDIA Jetson AGX Orin — 32 TOPS inference, runs 4 camera streams simultaneously at 120 FPS
  • Custom PyTorch model (ResNet-50 backbone) quantised to INT8 with TensorRT for edge deployment
  • Siemens S7-1500 PLC integration via PROFINET — reject signal triggers pneumatic ejector within 2ms of detection
  • MordeTech OEE Pro dashboard shows real-time pass/fail rates, defect type breakdown, and shift reports
  • Operator HMI panel shows defect images with bounding boxes and confidence scores for every rejection

Results After 90 Days

  • Defect escape rate reduced by 94% (from ~3,200 ppm to ~190 ppm)
  • Zero false-negative detections for Class A defects over 6-month operational period
  • False-positive rate below 0.03% — line efficiency unaffected
  • Inspection throughput: 100% of parts at full line speed (800 parts/min)
  • Warranty claim costs reduced by ₹42 Lakhs in first 6 months post-deployment
  • IATF 16949 non-conformance resolved — audit passed without finding on QC process
  • System paid back in full within 11 months of go-live

More Case Studies

See how MordeTech has delivered measurable results across manufacturing, energy, and process industries.