The Challenge
A precision electronics manufacturer producing components for automotive, medical, and aerospace industries faced escalating quality control challenges. Their products required inspection for defects as small as 50 microns, including scratches, contamination, dimensional variations, and surface finish anomalies.
Human inspectors, working with microscopes and magnification equipment, could only inspect 120 units per hour while maintaining acceptable accuracy. At production volumes of 50,000 units per day, quality inspection had become the primary bottleneck. The company employed 45 full-time inspectors across three shifts, yet still experienced quality escapes that resulted in costly field returns and customer complaints.
The situation was unsustainable. Inspection fatigue led to inconsistent results, especially during night shifts. Training new inspectors took 6 months, and turnover was high due to the demanding nature of the work. The company needed a solution that could increase throughput while improving consistency and reducing escapes.
The Solution
Jasnova designed and deployed an automated visual inspection system using high-resolution imaging and deep learning to detect defects at production line speeds with superhuman accuracy.
Technical Architecture
- High-Resolution Imaging: Custom-designed imaging stations with 20-megapixel cameras, telecentric lenses, and structured lighting to capture consistent, high-quality images from multiple angles
- Multi-Stage Detection Pipeline: Coarse-to-fine inspection architecture that first identifies regions of interest, then applies high-resolution analysis to suspected defect areas
- Defect Classification: Deep convolutional neural networks trained to classify 47 distinct defect types with confidence scores and severity ratings
- Edge Inference: NVIDIA GPU-based edge computing for sub-second inference, enabling real-time sorting at production line speeds
- Anomaly Detection: Unsupervised models that detect previously unseen defect types, alerting quality engineers to emerging issues
Defect Categories
The system detects and classifies defects across multiple categories:
- Surface Defects: Scratches, pits, stains, corrosion, fingerprints, and contamination particles
- Dimensional Defects: Warpage, edge chips, incorrect dimensions, and alignment issues
- Coating Defects: Uneven coating thickness, bubbles, orange peel texture, and delamination
- Assembly Defects: Missing components, incorrect orientation, and solder defects
Implementation
Phase 1: Data Collection & Labeling (Weeks 1-8)
Captured 500,000+ images of both good and defective parts. Worked with quality engineers to label defects with precise boundaries and classifications. Developed a labeling tool with zoom, measurement, and annotation capabilities to ensure consistent labeling.
Phase 2: Model Development (Weeks 6-14)
Trained detection models using the labeled dataset with extensive data augmentation (lighting variations, rotations, synthetic defects). Achieved 99.2% detection rate with 0.5% false positive rate on held-out test sets. Validated performance on defects across all 47 categories.
Phase 3: Hardware Integration (Weeks 10-18)
Designed and installed imaging stations at 6 production lines. Integrated with existing conveyor systems and PLCs. Calibrated lighting and camera positions for each product type. Implemented automatic part handling for sorting accepted and rejected units.
Phase 4: Production Deployment (Weeks 16-24)
Gradual rollout with parallel human inspection for validation. Fine-tuned detection thresholds based on production data. Trained quality team on system operation and exception handling. Full handover with monitoring dashboards and alerting.
Results
After 12 months of production operation:
- 99.2% Detection Accuracy: Exceeding human inspector performance of 94% while maintaining consistency across all shifts
- 10x Throughput Increase: Each inspection station processes 1,200 units per hour vs. 120 for human inspectors
- 75% Reduction in Quality Escapes: Field returns due to quality issues dropped from 0.8% to 0.2%
- 0.5% False Positive Rate: Down from 3% with human inspection, reducing unnecessary rework
- $8M Annual Savings: From reduced labor costs, fewer returns, and lower warranty claims
- 24/7 Consistent Operation: No quality degradation during night shifts or high-volume periods
Continuous Learning
The system continues to improve through active learning:
- Uncertain Sample Review: Parts where the model has low confidence are routed to human experts for labeling, and these samples are added to training data
- New Defect Discovery: The anomaly detection module identifies potential new defect types, alerting quality engineers to investigate
- Model Updates: Monthly model retraining incorporates new labeled samples, maintaining accuracy as products and processes evolve
- Performance Monitoring: Automated tracking of detection rates, false positives, and correlation with downstream quality metrics
"Jasnova's vision system has revolutionized our quality control. We're catching defects we never could have seen consistently with human inspectors, and we've eliminated quality as a production bottleneck. It's paid for itself many times over."
- Director of Quality, Precision Electronics Manufacturer
Expansion
Following the success of the initial deployment, the manufacturer has expanded the system to:
- 12 additional production lines across 3 facilities
- New product families with different inspection requirements
- Incoming material inspection for supplier quality verification
- Integration with statistical process control for real-time quality trending
