Federated Learning Architectures for Distributed Engineering Project Knowledge Management A Simulation-Based Analysis

Abstract

The increasing complexity of distributed engineering projects has created significant challenges in knowledge management, particularly regarding knowledge sharing, data privacy, communication efficiency, and collaborative decision-making across geographically dispersed stakeholders. Federated learning has emerged as a promising decentralized machine learning approach that enables collaborative intelligence generation without requiring direct sharing of sensitive organizational data. This study investigated the effectiveness of federated learning architectures for distributed engineering project knowledge management through a quantitative simulation-based analysis. Specifically, the study evaluated and compared centralized, decentralized, hierarchical, and hybrid federated learning architectures across key performance dimensions, including learning accuracy, convergence efficiency, communication overhead, scalability performance, computational efficiency, and knowledge-sharing effectiveness. A simulation environment consisting of 100 distributed engineering stakeholder nodes was developed to represent project owners, contractors, consultants, design teams, maintenance units, and operational management entities. The experimental design utilized synthetic engineering project datasets containing design records, construction data, operational logs, maintenance information, and project knowledge repositories distributed across heterogeneous nodes. Statistical analyses included descriptive statistics, one-way analysis of variance, correlation analysis, and multiple regression analysis at a significance level of p < 0.05. The findings revealed statistically significant differences among the evaluated architectures. The hybrid federated learning architecture achieved the highest learning accuracy (95.1%), knowledge-sharing effectiveness (93.6%), communication efficiency (92.6%), and computational efficiency (94.5%). Hierarchical federated learning demonstrated the strongest scalability performance (94.2%) and the fastest convergence rate, requiring only 31 communication rounds to achieve stable learning performance. Centralized federated learning produced reliable learning outcomes but generated the highest communication overhead (18.6 GB), while decentralized federated learning reduced dependence on centralized coordination but exhibited lower overall learning performance (87.9%). Regression analysis indicated that knowledge-sharing effectiveness (β = 0.53), communication efficiency (β = 0.48), and convergence speed (β = 0.42) were significant predictors of overall learning performance. The study concluded that hybrid and hierarchical federated learning architectures provided the most effective balance between learning quality, scalability, communication management, and computational efficiency. These findings contribute to the growing body of knowledge on privacy-preserving artificial intelligence and demonstrate the potential of federated learning as an effective framework for distributed engineering project knowledge management.