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

Defect detection in industrial applications is essential for maintaining product quality and operational efficiency. However, traditional deep learning methods require centralized data collection, raising privacy concerns and limiting adaptability in distributed manufacturing environments. To overcome these challenges, we propose Fed-MSVT, a Federated Multi-Scale Vision Transformer with Adaptive Client Aggregation for industrial defect detection. Our approach leverages multi-scale Vision Transformers (MSVTs) to capture both fine-grained local defects and global structural patterns, enhancing detection accuracy across diverse defect types. Unlike conventional federated learning models, we introduce an Adaptive Client Aggregation (ACA) mechanism that dynamically assigns weights to client models based on data quality, domain shift, and consistency. Additionally, a Contrastive Feature Alignment (CFA) module mitigates inter-client domain discrepancies, improving generalization. Evaluations on multiple defect datasets demonstrate superior accuracy, robustness, and scalability compared to existing approaches, enabling real-time, privacy-preserving, and adaptive defect detection for smart manufacturing systems.