Communication-Efficient Federated Learning for IIoT Bearing Fault Diagnosis Using Multi-Head Deep Q-Network-Based Intelligent Client Selection
International Journal of Emerging Research in Engineering, Science, and Management · 07.08.2026 · 5 (3) · 67-80
DOI: 10.58482/ijeresm.v5i3.5 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Emerging Research in Engineering, Science, and Management, online erschienen am 07.08.2026. Verfasst von Bhupathirao Lakinana und 1 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben.
Abstract (Originalsprache)
Federated learning (FL), which allows distributed training across several industrial machines without exchanging raw vibration data, has emerged as a promising paradigm for privacy-preserving bearing fault diagnosis in Industrial Internet of Things (IIoT) scenarios. Nevertheless, current FL approaches for fault diagnosis face two significant limitations. First, uniform client participation in each communication round results in needless and prohibitive communication overhead. Second, static or random client selection strategies are indifferent to the informativeness of individual client updates under highly non-IID data distributions. This paper proposes a novel Multi-Head Deep Q-Network (Multi-Head DQN)-based client selection framework for communication-efficient federated learning applied to bearing fault diagnosis to overcome these limitations. In the proposed framework, the Multi-Head DQN agent jointly determines the optimal number of clients, k, to select in each round from a predefined range, and the identities of the most informative clients. To accurately replicate heterogeneous IIoT deployments in the real world, a stringent non-IID partitioning approach is implemented across 12 federated clients, each of which is allocated precisely two of four failure classes (Normal, Inner Race, Outer Race, and Ball). The Flower federated learning framework is used to compare six FL configurations: FedAvg, FedProx, FedAvg+FedRandom, FedProx+FedRandom, FedAvg+FedDQN, and FedProx+FedDQN. All experiments were conducted on the CWRU bearing fault dataset. Experimental results show that the proposed FedProx+FedDQN approach reduces cumulative communication costs by 19.15% compared with the all-client FedAvg baseline while achieving a global fault classification accuracy of 94.05%, which is comparable to the full-participation baselines (FedAvg and FedProx). The DQN-based selection prioritizes informative client updates, leading to faster convergence and lower communication overhead. These findings confirm that intelligent client selection is an effective strategy for improving communication efficiency in federated learning for industrial fault diagnosis applications.
Zitieren
APA
Lakinana, B., & Chilukuri, K. C. (2026). Communication-Efficient Federated Learning for IIoT Bearing Fault Diagnosis Using Multi-Head Deep Q-Network-Based Intelligent Client Selection. International Journal of Emerging Research in Engineering, Science, and Management, 5(3), 67-80. https://doi.org/10.58482/ijeresm.v5i3.5
Vancouver
Lakinana B, Chilukuri KC. Communication-Efficient Federated Learning for IIoT Bearing Fault Diagnosis Using Multi-Head Deep Q-Network-Based Intelligent Client Selection. International Journal of Emerging Research in Engineering, Science, and Management. 2026;5(3):67-80. doi:10.58482/ijeresm.v5i3.5
BibTeX
@article{Lakinana2026,
author = {Lakinana, Bhupathirao and Chilukuri, Kalyana Chakravarthy},
title = {Communication-Efficient Federated Learning for IIoT Bearing Fault Diagnosis Using Multi-Head Deep Q-Network-Based Intelligent Client Selection},
journal = {International Journal of Emerging Research in Engineering, Science, and Management},
year = {2026},
volume = {5},
number = {3},
pages = {67--80},
doi = {10.58482/ijeresm.v5i3.5}
}
EUVIMED