Federated Multi-Agent AI Framework for Privacy-Preserving Predictive Analytics in Smart Manufacturing Systems
Main Article Content
Abstract
Modern smart manufacturing environments produce diverse streams of sensors signals in real time and at high speed (HVS), often across widely distributed factory floors, resulting in significant conflicts between the value of sharing machine learning data among machines, and the need for the confidentiality of proprietary data. This book presents the new Federated Multi-Agent AI (FedMA-AI) framework, which combines federated learning with autonomous multi-agent coordination, differential privacy, secure multi-party computation and homomorphic encryption, to facilitate privacy-preserving predictive analytics without relying on sharing raw operational data. It introduces a Bayesian meta-learner for cross-factory model aggregation and a FedProx-enhanced optimization objective which explicitly penalizes agent heterogeneity drift. With 67.5% of the inter-node communication bandwidth saved at 64 nodes, it achieves classification accuracy of 96.5% (F1 = 0.961, AUC-ROC = 0.983) at the differential privacy budget of ε = 0.5, compared to the FedAvg baseline that achieves 92.1%. One-way ANOVA (F = 284.7, p < 0.001, η² = 0.89) and Cohen's d effect sizes greater than 3.84 compared to all competing baselines demonstrate statistically powerful superiority. The proposed framework is a well-founded approach for Industry 4.0 compliance in data-sovereign manufacturing networks.