Federated Edge Learning for Predictive Maintenance in 6G Small Cell Networks
Abstract
The rollout of 6G networks introduces unprecedented demands for autonomy, reliability, and scalability. However, the transmission of sensitive telemetry data to central servers raises concerns about privacy and bandwidth. To address this, we propose a federated edge learning framework for predictive maintenance in 6G small cell networks. The system adopts a Knowledge Defined Networking (KDN) architecture in Data, Knowledge, and Control Planes to support decentralized intelligence, telemetry-driven training, and coordinated policy enforcement. In the proposed model, each base station independently trains a failure prediction model using local telemetry metrics, including SINR, jitter, delay, and transport block size, without sharing raw data. A threshold-based multi-label encoding scheme enables the detection of concurrent fault conditions. We then conduct a comparative analysis of centralized and federated training strategies to evaluate their performance in this context. A realistic simulation environment is implemented using the ns-3 mmWave module, incorporating hybrid user placement and base station fault injection across various deployment scenarios. The learning pipeline is orchestrated via the Flower framework, and model aggregation is performed using the Federated Averaging (FedAvg) algorithm. Experimental results demonstrate that the federated model achieves performance comparable to centralized training in terms of accuracy and per-label precision, while preserving privacy and reducing communication overhead.