Federated multi scale vision transformer with adaptive client aggregation for industrial defect detection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Traditional deep learning for defect detection requires centralized data, posing privacy risks and limiting adaptability in distributed manufacturing.
• Existing federated learning methods may not adequately address data quality and domain shift issues across diverse industrial sites.

Purpose of the Study:

• To develop a privacy-preserving and adaptive federated learning framework for industrial defect detection.
• To enhance the accuracy and robustness of defect detection in distributed manufacturing environments.

Main Methods:

• Proposing Fed-MSVT, a Federated Multi-Scale Vision Transformer with Adaptive Client Aggregation.
• Utilizing multi-scale Vision Transformers (MSVTs) for comprehensive defect feature extraction.
• Implementing an Adaptive Client Aggregation (ACA) mechanism for dynamic model weighting.
• Employing a Contrastive Feature Alignment (CFA) module to reduce domain discrepancies.

Main Results:

• Fed-MSVT demonstrates superior accuracy and robustness in detecting diverse industrial defects.
• The Adaptive Client Aggregation (ACA) mechanism effectively handles variations in client data quality and domain shifts.
• Contrastive Feature Alignment (CFA) improves model generalization across different manufacturing sites.
• Evaluations show enhanced performance compared to existing federated learning approaches.

Conclusions:

• Fed-MSVT offers a scalable, privacy-preserving solution for real-time industrial defect detection.
• The proposed adaptive aggregation and feature alignment strategies are crucial for effective federated learning in manufacturing.
• This approach supports the development of adaptive and intelligent smart manufacturing systems.