Comparative Evaluation of Federated Learning and Centralized Learning Architectures for Cross-Factory Productivity Optimization in Distributed Manufacturing Networks

Abstract

The present study aimed to comparatively evaluate the effectiveness of federated learning and centralized learning architectures in improving productivity optimization, predictive maintenance performance, operational responsiveness, and computational efficiency across distributed manufacturing networks. This applied quantitative study was conducted using a semi-experimental comparative design in 24 manufacturing factories located in Tehran, Iran. A total of 312 production units from automotive, electronics, industrial machinery, and polymer manufacturing sectors participated in the research. The factories were divided into two analytical environments including federated learning architecture and centralized learning architecture. Operational and productivity data were collected over a twelve-month period through Industrial Internet of Things sensors, Manufacturing Execution Systems, Supervisory Control and Data Acquisition platforms, and enterprise operational databases. Key variables included overall equipment effectiveness, production throughput efficiency, predictive maintenance accuracy, energy optimization, operational response speed, and defect reduction rates. Data analysis was performed using TensorFlow Federated, Python machine learning libraries, SPSS version 27, and R software. Descriptive statistics, multivariate analysis of variance, hierarchical regression analysis, and predictive machine learning performance evaluations were used to compare the two learning architectures. The findings demonstrated statistically significant differences between federated learning and centralized learning architectures across all productivity indicators (p<0.001). Federated learning achieved significantly higher overall equipment effectiveness, throughput efficiency, predictive maintenance accuracy, energy optimization, and operational response speed compared to centralized learning systems. The federated architecture also demonstrated superior machine learning performance with higher prediction accuracy, precision, recall, and F1-score values alongside lower mean absolute error and root mean square error. Moreover, federated learning substantially reduced communication overhead, computational latency, and model convergence time. Hierarchical regression analysis further revealed that federated learning architecture was the strongest predictor of productivity optimization in distributed manufacturing networks. The findings indicate that federated learning architectures provide a highly effective framework for productivity optimization within distributed manufacturing ecosystems by simultaneously enhancing predictive intelligence, operational responsiveness, scalability, cybersecurity resilience, and data privacy preservation. Compared to centralized learning systems, federated learning demonstrated superior adaptability to heterogeneous industrial environments and reduced computational inefficiencies associated with centralized data aggregation. The study highlights the strategic importance of decentralized collaborative intelligence systems for the future development of Industry 4.0 and smart manufacturing infrastructures.