As the paradigm of Industry 4.0 accelerates the digital transformation of smart manufacturing, the utilization of massive, high-dimensional sensor data generated from Cyber-Physical Systems (CPS) has become paramount for predictive maintenance. Howeve...
As the paradigm of Industry 4.0 accelerates the digital transformation of smart manufacturing, the utilization of massive, high-dimensional sensor data generated from Cyber-Physical Systems (CPS) has become paramount for predictive maintenance. However, traditional centralized processing approaches face critical structural limitations, including data privacy violations, network bandwidth saturation, and prohibitive latency in real-time analysis. Moreover, the inherent heterogeneity of industrial data often leads to data silos, hindering the development of robust global models. To address these multi-faceted challenges, this study proposes a novel three-tier hierarchical Federated Learning (FL) architecture consisting of Cloud, Anchor, and Edge layers. This topology is designed to mimic the physical hierarchy of manufacturing systems, maximizing analytical efficiency while securely isolating sensitive raw data within edge devices. Specifically, to overcome the performance limitations of conventional weight-averaging methods in non-IID environments, a Stacking Ensemble technique is introduced at the anchor layer. In this scheme, a meta-model iteratively re-learns the non-linear relationships based on time-series features extracted by local 1D-CNN models. This approach effectively corrects the bias inherent in single models and ensures predictive robustness against noise. Furthermore, to actively respond to Concept Drift caused by frequent changes in production processes and equipment wear, a Dynamic Edge Reallocation mechanism based on the Wasserstein distance was implemented. This adaptive protocol allows edge nodes to autonomously reconfigure their learning paths by identifying and associating with anchors that share the most similar statistical characteristics, thereby minimizing performance degradation due to distribution shifts and ensuring long-term operational stability. Experimental results demonstrate that the proposed architecture achieves an anomaly classification accuracy of 0.7438, representing a significant improvement of approximately 33% compared to standard 1D-CNN-based federated learning (accuracy 0.5585). This substantial gain verifies the system's efficacy in identifying complex, non-linear anomaly patterns that single-tier models fail to capture. Additionally, the architecture recorded an inference speed approximately 6.3 times faster than centralized Random Forest models, confirming it as a powerful, lightweight alternative that meets the strict real-time requirements of manufacturing sites. In conclusion, this study presents a next-generation integrated framework for intelligent manufacturing management that simultaneously satisfies the critical requirements of data security, analytical precision, and adaptive capability in non-stationary environments.