Collaborative Improvement of Smart Manufacturing using Privacy-Preserving Federated Learning

• Autor(a) principal: Costa, Alexandre Miguel Manta da
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/153177
• topic: Federated Learning; Privacy-Preserving Federated Learning; Smart Manufacturing; Framework; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Nowadays, data sharing among different sources is is very challenging in the manufac- turing domain, mainly due to industry competition, complicated bureaucratic processes, and privacy and security concerns. Centralized Machine Learning (ML) poses an essential aspect in several industries, including smart manufacturing. However this approach may lead to several issues regarding security and performance. In response to these problems, Federated Learning (FL) was created. FL is an innova- tive and decentralized approach to ML, focused on collaboration and data privacy. In this approach, data is kept in each source where it is trained locally, and only model weights or gradients are shared to create a global model. Although several works have already been implemented towards this problem, there are still many unresolved issues concerning the application of FL frameworks in smart manufacturing scenarios. Among the several issues found in the analysed works it is important to emphasize the disregard facing industry 4.0 architectures, strategies and the unavailability to improve those frameworks further. This work aims to build a FL framework for smart manufacturing with specific con- cerns in privacy and applicability in industrial scenarios. The main focus of this frame- work is to facilitate a collaborative approach in the application of ML to manufacturing by enabling the knowledge sharing for this purpose and taking privacy as a special concern. In addition, the implementation and testing of privacy-preserving algorithms, while im- proving the framework for industrial scenarios are emphasized. A modular approach is chosen to create a framework adapted to various industrial cases by implementing several nodes that focus on specific aspects of data collection, data treatment, connection with the FL system, and ML model management. The results revealed a competitive model performance of the framework compared to the centralized approach while keeping data at each source, protecting its privacy. The implemented framework also proved to be compliant with the IEEE Std 3652.1-2020 standard guidelines, attaining the established requirement levels.