Green HERO: Energy-Efficient Hierarchical Federated Learning with Client Association in Industrial IoT

The Industrial Internet of Things (IIoT) plays a pivotal role in advancing the automation and intelligence of industrial production processes. Its integration with various technologies is driving innovation in intelligent applications within industrial scenarios. Among these innovations, Federated Learning (FL) stands out by offering a distributed training service for AI-driven applications in IIoT settings. However, the complexity of industrial control systems and the demands for intensive computing capabilities have raised concerns regarding energy consumption and carbon emissions. These concerns have prompted the need for the existing FL framework to evolve towards more sustainable and efficient practices. In this work, we propose a green distributed intelligence framework: Multi-Agent Deep Reinforcement Learning (MA-DRL) based Hierarchical fEderated leaRning with client assOciation (HERO), which consists of three layers: the client layer, the edge layer (contains the Edge Server (ES) and the Edge Parameter Server (EPS)), and the cloud layer. This architecture enables two-layer aggregation, namely Edge-aggregation and Cloud-aggregation. The Edge-aggregation, in particular, can optimize data distribution and alleviate data heterogeneity difficulties, all of which contribute to an acceleration in model convergence speed and decrease training rounds. Eventually, it will lead to a reduction in energy consumption. Moreover, avoiding frequent data transmission between users and the cloud through adopting the computing resources at the edge can also lower communication energy consumption. In performance evaluation, we conduct extensive experiments on different datasets. The results demonstrate that the proposed HERO outperforms the baseline in energy consumption and training convergence rate.