Publication
Collaborative Learning in Shared Production Environment Using Federated Image Classification
Vinit Hegiste; Tatjana Legler; Martin Ruskowski
In: Kosmas Alexopoulos; Sotiris Makris; Panagiotis Stavropoulos (Hrsg.). Advances in Artificial Intelligence in Manufacturing II. European Symposium on Artificial Intelligence in Manufacturing (ESAIM-2024), Cham, Pages 98-106, ISBN 978-3-031-86489-6, Springer Nature Switzerland, 2025.
Abstract
The application of federated learning (FL) in industrial settings offers promising advancements in maintaining data privacy while collaboratively training machine learning models. This study focuses on the comparative analysis of federated image classification versus locally trained models within a shared production environment. Specifically, we explore the classification of windshields in truck cabins, which is a crucial task for quality inspection in manufacturing of trucks. Our research involves four clients, each producing different types of truck cabins and research based on FL process between them. Various deep learning architectures, including VGG19, ResNet50, InceptionNetv3, DenseNet-121, and EfficientNetv2-s, were evaluated under a FL framework implemented using the FLOWER framework. A custom plain averaging strategy was used for weight aggregation. The global model's performance was assessed using a combined test set from all clients and compared against models trained locally by individual clients. The results highlight the effectiveness of FL in enhancing model generalization and adaptability to new product variations in industrial applications, promoting its adoption for collaborative quality inspection tasks.