Case Study: Defect Detection in a Production Line

This case study presents the implementation of an automatic defect detection system for a manufacturing company using Techsolut's computer vision tools. The system identifies various types of anomalies on industrial products in real-time, thereby improving quality and reducing costs related to undetected defects.

Context and Challenges

Company Profile

A consumer electronics manufacturing company producing approximately 5,000 units per day across multiple assembly lines.

Initial Issues

• 1.2% defect rate escaping manual quality control
• High costs of customer returns and warranty repairs
• Slow visual inspection process subject to operator fatigue
• Wide variety of potential defects: defective soldering, missing components, scratches, etc.

Project Objectives

1. Reduce the rate of defective products shipped to less than 0.1%
2. Increase inspection speed to keep pace with production
3. Collect data on defect types to improve upstream processes
4. System capable of operating 24/7 without performance degradation

Implemented Solution

Hardware Configuration

• 8 high-resolution industrial cameras (12MP) positioned at strategic points
• Controlled LED lighting to ensure consistent light conditions
• 2 GPU servers for real-time image processing
• Automated rejection system to remove defective products from the flow

Software Architecture

• Techsolut interface for centralized management
• Specific anomaly detection models for each type of defect
• Real-time processing pipeline with latency < 300ms
• Feedback system for continuous model improvement

Development Approach

Phase 1: Data Collection and Preparation

1. Collection of over 50,000 product images (defective and compliant)
2. Meticulous annotation of 8 different types of defects
3. Data augmentation for rare defects (rotation, brightness, etc.)
4. Division into training (70%), validation (15%), and test (15%) sets

Phase 2: Model Development

1. Use of a hybrid approach combining:
2. Object detection for localized defects (YOLOv8)
3. Semantic segmentation for surface defects (U-Net)
4. Unsupervised anomaly detection for unknown defects
5. Training on GPU infrastructure for 2 weeks
6. Model optimization for real-time inference

Phase 3: Deployment and Integration

1. Gradual implementation on a pilot line
2. Fine calibration of the system with operator feedback
3. Integration with the factory's MES (Manufacturing Execution System)
4. Training of maintenance staff and quality operators

Results and Impact

Technical Performance

• Precision: 98.7% correct defect detection
• Recall: 99.3% of defects actually detected
• Speed: Processing of 6 units per minute (higher than production rate)
• False positive rate: 0.4% (compliant products incorrectly rejected)

Business Benefits

• 87% reduction in customer returns for defective products
• Estimated annual savings of 1.2 million euros
• ROI achieved in 7 months
• Reassignment of 60% of inspection staff to higher value-added tasks

Process Improvements

• Identification of root causes through analysis of collected data
• 23% reduction in defects at source through corrective actions
• Complete defect traceability and improved quality documentation

Challenges and Solutions

Challenges Encountered

Lighting Variability

• Problem: Brightness changes affecting detection
• Solution: Installation of a controlled lighting system and image normalization

New Types of Defects

• Problem: Appearance of defects not present in training data
• Solution: Implementation of a generic anomaly detection system and continuous feedback

Integration with Existing Systems

• Problem: Legacy proprietary systems difficult to interface with
• Solution: Development of an adaptive API layer and joint work sessions with the IT team

Resistance to Change

• Problem: Operator concerns about automation
• Solution: Early team involvement, training, and repositioning towards system supervision roles

Lessons Learned and Best Practices

Key Success Factors

1. Data Quality: Investment in meticulous collection and annotation was decisive
2. Hybrid Approach: The combination of several techniques allowed coverage of all defect types
3. Progressive Deployment: The pilot phase allowed system adjustment without disrupting production
4. User Involvement: Collaboration with operators improved acceptance and efficiency

Recommendations

1. Allow time for fine calibration in real conditions
2. Build a continuous feedback system from the beginning
3. Precisely document image acquisition conditions
4. Train an internal team capable of maintaining and evolving the system

Future Developments

The next phase of the project plans for:

1. Extension of the system to all production lines
2. Addition of predictive capabilities to anticipate quality drifts
3. Implementation of continuous learning to adapt models to product evolutions
4. Integration with cobots for automatic correction of certain defects

Conclusion

This project illustrates how a well-designed computer vision solution can transform a critical industrial process. Beyond simple detection, the system has become a continuous improvement tool generating valuable data for manufacturing process optimization.

The key to success was the holistic approach combining adapted hardware, sophisticated AI models, and careful integration with existing processes, all accompanied by effective change management.