We partnered with Viston to deploy predictive maintenance and vision-based quality inspection to cut downtime and scrap in two automotive component plants.

Business challenge

• Unplanned downtime on CNC lines and variable defect rates on final assembly inspection.
  
• Siloed OT/IT data, strict works council requirements, and GDPR constraints.
  

Our approach and solution

• Connected PLC/SCADA data and camera feeds, built failure prediction models, and deployed an edge vision model for defect detection.
  
• Implemented unified monitoring and change control with human acceptance gates.
  

AI applications and benefits delivered

• Remaining Useful Life (RUL) prediction for spindles and bearings.
  
• Computer vision (YOLOv8) for surface defects, missing fasteners, and misalignment.
  
• Prescriptive maintenance scheduling integrated with SAP PM.
  

Cost of implementation

• EUR 670,00 end-to-end, including edge hardware and model ops.
  

Time to implement

• 6 months to steady-state production across two plants.
  

Tools and technologies used

• AWS IoT Greengrass, Kafka, Databricks, Delta Lake, XGBoost/CatBoost, YOLOv8, OpenCV, NVIDIA Jetson at the edge, Prometheus/Grafana, Airflow, MLflow, Great Expectations, SAP PM connector.
  

Quantitative outcomes

• 41% reduction in unplanned downtime on targeted assets.
  
• 32% reduction in scrap on inspected SKUs.
  
• OEE improvement of 9 percentage points.
  
• Payback in 11 months.
  

Key performance indicators (KPIs) tracked

• Mean time between failures, mean time to repair, defect rate PPM, OEE, model precision/recall, false reject rate, maintenance labor hours.
  

Pre- and post-implementation metrics

• Downtime (hrs/month targeted lines): 62 → 36.
  
• Scrap rate (PPM on target SKUs): 2,900 → 1,970.
  
• Inspection throughput (units/hr): 420 → 560.
  

Stakeholder quotes or testimonials

• “We finally predict failures before weekend shifts. The maintenance plan is proactive, not reactive.” — Plant Manager.
  
• “Vision AI caught defects our manual spot-checks missed.” — Head of Quality.
  

Regulatory or compliance considerations

• GDPR, data minimization, on-prem edge inference, no biometric data, works council approvals, TISAX-aligned controls, data residency in EU.
  

Lessons learned and next steps

• Edge-first deployment avoided network constraints and sovereignty issues.
  
• Next: expand vision to torque verification and add anomaly detection on assembly torque traces.
  

Attribution

• Name and designation: Confidential, Plant Manager.
  
• Company: Confidential Tier-1 Automotive Supplier, Germany.