Federated Deep Learning with Contrastive Representation for Node State Identification in Distributed Systems

Abstract

This study proposes a distributed node state identification framework that integrates contrastive learning and federated optimization to address the heterogeneity, latency, and dynamism of distributed systems. At the local node level, the framework introduces contrastive feature constraints by constructing positive and negative sample pairs to achieve self-supervised alignment in the feature space, thereby enhancing representational consistency and discriminative capability within nodes. At the global level, it employs a federated optimization mechanism for parameter aggregation to enable collaborative learning and global consistency across nodes. The framework consists of four core stages: feature encoding, contrastive representation learning, federated aggregation, and consistency regulation, allowing efficient global identification while preserving data privacy. To verify its effectiveness, multidimensional sensitivity experiments were conducted, including analyses of hyperparameters, environmental factors, and data perturbations. The results show that the framework maintains stable performance under varying weight decay coefficients, temperature parameters, communication delays, and noise intensities, achieving significant improvements over traditional centralized and single-node models in accuracy, precision, recall, and F1-score. Further analysis demonstrates that the contrastive learning module effectively suppresses noise interference and feature drift, while the federated optimization mechanism mitigates data distribution bias among heterogeneous nodes, ensuring good convergence under high-latency and unbalanced conditions. This study confirms that the integrated learning strategy can balance feature robustness, model consistency, and computational efficiency, providing a feasible solution for building secure, stable, and efficient distributed intelligent identification systems.