Decentralized Manufacturing Management Based on Federated Learning with Stacking Ensemble

ICLR 2026 Conference Submission18286 Authors

Keywords: Manufacturing Management System, Federated Learning, Anomaly Classification, Stacking Ensemble, Distribution Shifts, Frozen Encoder

Abstract: We propose a new intelligent management system to overcome the limitations of privacy, security, communication efficiency, and real-time analysis of data generated in smart manufacturing environments. As the digital transformation of the manufacturing industry accelerates, the importance of data utilization has grown, but the existing centralized approach involves data leakage risk and network load issues. To overcome these limitations, we propose a three-layer federated learning architecture consisting of cloud–anchor–edge. In particular, the anchor layer applies a stacking ensemble technique that combines predictions from multiple models to accurately identify complex anomaly patterns that are difficult to detect with a single model and maximize the robustness of model predictions. Compared to the accuracy of 0.5585 achieved by a single 1D-CNN model, the model applying stacking to federated learning significantly improved performance to an accuracy of 0.7438. Furthermore, to address the continuously changing data distributions in manufacturing environments, we propose a data distribution change detection and edge reallocation mechanism to enhance system flexibility and adaptability. The proposed system demonstrates significantly faster inference times than centralized learning models, presenting it a powerful alternative that ensures data privacy.

Primary Area: learning on time series and dynamical systems

Submission Number: 18286