Industrial defect detection for data privacy and domain differences: Based on federated learning and unsupervised domain adaptation

Industrial defect detection faces challenges from domain shifts in unsupervised domain adaptation (UDA) and data privacy in federated learning (FL). This study proposes a novel industrial defect detection method based on lightweight UDA and FL participant selection optimization (IDD-LUDA-FL), a lightweight UDA-FL framework for robust cross-domain defect detection. This method leverages MobileNet V4 for efficient feature extraction, dynamic feature pyramid network (DYFPN) for multi-scale processing, and quantized convolutional neural networks (Quantized-CNN) for model compression, while optimizing FL participant selection using Shapley values and second-order Markov chains to enhance privacy and mitigate attacks like label shuffling. Extensive studies on four industrial datasets show improved performance across a variety of source-target adaptation scenarios. Compared to other state-of-the-art methods, the proposed method achieves an average accuracy of 91%. Inference speed improves by over 50%, with detection speed reaching 58.6f/s. The holistic integration of lightweight UDA and secure FL mechanisms enables effective handling of data heterogeneity and adversarial threats, validating its utility for practical industrial applications. This work advances multi-domain defect detection, providing a foundation for scalable, privacy-preserving AI in manufacturing.