Federated Learning Strategies for Enhanced Quality Inspection Processes in Manufacturing Industry

The increasing reliance on vision-based deep learning for automated quality inspection in manufacturing emphasizes the need for collaborative model development across distributed production sites. Also, legal, privacy, and proprietary constraints often prevent centralized aggregation of such industrial data. This thesis investigates the use of Federated Learning (FL) as a decentralized framework for training high performance computer vision models without sharing raw data. A modular and scalable FL architecture is proposed, supporting both image classification and object detection under non-Identically and Independent Distributed (non-IID) client distributions, and leveraging lightweight, high-accuracy deep learning models suitable for industrial deployment. The experimental validation covers diverse industrial scenarios, including USB port and cabin windshield classification, as well as YOLO-based object detection for localizing quality-relevant components. The framework is further extended to hybrid setups that incorporate synthetic and real clients to address data scarcity, and validated through live inference across heterogeneous domains. In addition, a novel Federated Ensemble (FedEnsemble) strategy is introduced, in which a centralized dataset is partitioned into clients for federated training, yielding models that outperform conventional centralized baselines under domain shift. The main contributions of this thesis are: (i) a validated FL framework tailored for privacy-preserving quality inspection in industrial environments, (ii) empirical insights into training dynamics with a small number of heterogeneous clients, (iii) evidence of robust cross-domain generalization through hybrid federated setups combining real and synthetic data, and (iv) the introduction of FedEnsemble as a practical extension of FL that improves detection robustness in deployment. Together, these findings establish FL as a viable foundation for building distributed, scalable, and secure AI systems in smart manufacturing, enabling the development of robust and generalizable quality inspection models.