Abstract
The Industrial Internet of Things (IIoT) provides real-time sensing and computing services and has been widely deployed and used in a wide range of applications today. However, the participation of IIoT devices in federated learning has led to network congestion and high energy consumption for intelligent services. A hierarchical federated learning approach for intelligent IIoT applications to reduce energy consumption is proposed. We perform a theoretical analysis of the convergence of the proposed approach, which guides the selection of local aggregation steps in the upper layer of the system. Moreover, the model quantization and bandwidth allocation scheme are integrated into the lower layer of the system, allowing IIoT devices to upload the quantized gradients to save communication resources. The proposed method has been compared with different quantization-based FedAvg algorithm. The results show that our approach achieves the same convergence accuracy while reducing energy consumption by 20%-34%. This demonstrates that our approach attains an improved balance between computation and communication, leading to reduced energy consumption and ensuring convergence throughout the process.