A Hybrid Federated-Edge Learning Framework with Dynamic Model Pruning for Real-Time Anomaly Detection in Smart Manufacturing Networks

The rapid advancement of intelligent systems has accelerated the adoption of data-driven solutions across diverse industries, creating an increasing need for models that are both efficient and privacy-preserving. While traditional centralized machine learning approaches offer strong predictive capabilities, they often struggle with challenges related to data privacy, network latency, and computational inefficiency-especially in distributed environments with heterogeneous devices. To address these limitations, recent research has explored hybrid learning frameworks that integrate federated learning, edge computing, and dynamic model optimization techniques. These hybrid approaches enable models to process and learn from data closer to the source while maintaining stringent privacy requirements by keeping raw data localized. Additionally, the incorporation of pruning strategies, adaptive model compression, or multimodal data fusion contributes to improved speed, scalability, and accuracy in real-time inference tasks. Such frameworks have demonstrated notable promise in settings characterized by high data volume, operational complexity, and the necessity for fast anomaly detection or decision-making. However, despite these advancements, several challenges remain, including synchronization delays across edge nodes, variability in hardware capabilities, and the need for more efficient aggregation algorithms. Future developments may involve leveraging next-generation pruning techniques, energy-aware edge scheduling, decentralized orchestration protocols, or the integration of digital twin technologies to further enhance performance. Overall, hybrid distributed learning frameworks represent an important evolution toward more intelligent, secure, and autonomous computational ecosystems capable of supporting the next wave of smart applications.