Efficient clustered federated learning for Industrial Internet of Things: enhancing predictive performance and training time

Abstract The Industrial Internet of Things (IIoT) brings together industrial devices in a network that gathers and analyzes data in real-time for making data-driven decisions. Federated learning is a popular approach for collaboratively training multiple industrial edge devices using an intermediate server in multiple rounds. This approach can be applied in various fields, including anomaly detection, asset management, energy efficiency, quality control, and predictive maintenance. However, predictive performance is affected by limited and non-independent, identically distributed (non-IID) data. Additionally, edge devices also face resource constraints for training large datasets. This paper proposes a cluster-assisted custom federated learning approach for improving the prediction performance and resources required for training. The edge server initializes the model by broadcasting initial parameters, and then the edge devices start training. After training on the current round’s data, edge devices transmit the updated parameters, performance, and data distribution back to the edge server. Then, the edge server clusters edge devices based on their data distribution and performance to minimize non-IID. Parameter aggregation is undertaken within the cluster to improve prediction performance and the aggregated parameter is sent back to the respective cluster members. Assuming a secure internal network, edge devices work together to share samples of the current round data within the cluster to increase the dataset size and diversity. Earlier portion of the datasets are excluded from the current round of training to reduce the resources required for training and to minimize data drift. Comprehensive experimental evaluation with testbed datasets proves the effectiveness of the proposed approach over the current state-of-the-art.