An effective Federated Learning system for Industrial IoT data streaming

Abstract

Due to its outstanding privacy-related characteristics, Federated Learning (FL) has recently become a popular solution for the IIoT’s data privacy and scalability issues. However, more research is needed to determine how unique streaming data in IIoT Settings affects FL-enabled IIoT architectures, with unique streaming data affecting accuracy and reducing convergence performance. To achieve this goal, this paper explains the streaming data learning problem in an IIoT framework enabled by FL. Afterward, it outlines two unique issues relevant to this situation: convergence and the catastrophic forgetting that occurs throughout training. This article presents FedStream, a practical FL framework for IIoT streaming data applications, considering these challenges. In particular, we develop a straightforward and effective pairwise similarity-based streaming data replacement training method that systematically replaces original data samples with ones that show high similarity during the iterative training process. This not only improves the accuracy but also reduces the convergence process and catastrophic forgetting problem. Comprehensive case studies support the effectiveness of the proposed method. Finally, the article recommends potential research areas, encouraging academics and industry professionals to explore these emerging topics further.

Keywords

• IIoT
• Federated Learning
• Streaming data
• Pairwise similarity