Joint Scheduling for Federated Learning in Battery-Powered IIoT with Multiple Services

Federated learning has been envisioned as a promising technique to enable the intelligence of Industrial Internet of Things (IIoT). An efficient resource management algorithm is critical yet hard to design in IIoT due to limited power devices and the variety of co-existent services. In this paper, we propose a multi-service joint scheduling algorithm for federated learning on IIoT devices to maximize model accuracy by forming a loss function minimization problem under long-term device energy constraints. The problem is then reformulated into a single-round optimization, which can be solved through a binary search and a greedy algorithm. The transmit power of devices, the CPU frequency for model training, and the selection of devices' services are jointly optimized. Simulations demonstrate that our proposed algorithm outperforms the benchmarks in model accuracy, especially when the energy of the device battery is constrained.

I. Introduction

INDUSTRIAL Internet of Things (IIoT is a shifting paradigm that is foreseen to reform the production of factories. Given the distributed nature of IIoT devices, it is hard for an IIoT device alone to train a model that could be widely applicable. Therefore, we need federated learning to enable the intelligence of IIoT [1]. A large number of IIoT devices can collect and train their own models and then share those models with a central service for model aggregation. However, the implementation of federated learning is difficult as IIoT devices are often battery-limited and face heterogeneity.