FEDERATED LEARNING ARCHITECTURES FOR PREDICTIVE QUALITY CONTROL IN DISTRIBUTED MANUFACTURING SYSTEMS

Abstract

This study investigates how federated learning (FL) architectures influence predictive quality control (PQC) performance in distributed manufacturing environments characterized by heterogeneity, privacy constraints, and high data velocity. Predictive quality control leverages machine learning to forecast process deviations before defects occur; however, in globally distributed production networks, data-sharing restrictions and non-identically distributed (non-IID) data complicate centralized model training. To address these challenges, this research develops and empirically tests a quantitative, cross-sectional, multi–case framework linking FL architectural design, hub-and-spoke, hierarchical, and peer configurations—with PQC performance across IIoT-enabled plants. Constructs including infrastructure readiness, communication efficiency/update cadence, data heterogeneity, and privacy/trust governance were operationalized using validated Likert-scale instruments complemented by objective indicators such as AUC, F1, false-alarm rate, and time-to-detection. Regression and mixed-effects analyses reveal that infrastructure readiness and communication efficiency exhibit strong positive associations with PQC outcomes, whereas cross-site heterogeneity negatively affects performance. Crucially, hierarchical FL architectures moderate these relationships, attenuating the detrimental effects of heterogeneity and amplifying the gains from efficient communication. Privacy and trust governance correlate positively, though modestly, with PQC, underscoring that robust security and compliance practices enhance rather than hinder collaborative learning effectiveness. The findings establish that architecture is not merely an IT topology but a determinant of statistical and operational performance, transforming federated updates into a controllable mechanism for cross-plant quality intelligence. By integrating IIoT, edge computing, and privacy-preserving analytics within a measurable empirical model, this research advances both theoretical understanding and practical implementation of FL-enabled PQC. It offers an actionable blueprint for manufacturers: invest in edge readiness and orchestration, enforce cadence service levels, adopt hierarchical clustering for heterogeneous sites, and embed privacy governance into federation lifecycles. Collectively, the study demonstrates that disciplined technical, organizational, and architectural alignment enables distributed manufacturing systems to achieve predictive quality improvements without compromising data confidentiality.