Industrial Internet of Things (IIoT) systems require dynamical, safe, and competent control to address the dynamic industrial processes. Centralized Deep Reinforcement Learning (DRL) methods are however vulnerable to privacy, high communication overheads, and model tampering. In order to address these shortcomings, the present paper suggests a Blockchain-Secured Federated Deep Reinforcement Learning (BS-FDRL) system to privacy-guaranteed and resilient IIoT control. The suggested system has distributed edge agents, which locally train DRL models and exchange only encrypted policy updates through federated learning; this guarantees the privacy of data. A smart contract-based blockchain layer allows aggregation and safe validation of model parameters to be tampered with. Also, differential privacy is integrated to safeguard sensitive industrial data in the process of policy exchange. The framework is coded with Python 3.10, DDPG/PPO algorithms with PyTorch, Flower to coordinate federated, and Hyperledger Fabric to integrate with blockchains. Industrial control and anomaly detection experimental analyses show that performance improvements are significant. The increase in the cumulative reward of the proposed model is 18-22% and the convergence rate is about 28 times higher than the baseline approaches. Stability of control is improved to 91.3% and communication overhead is cut down by up to 25%. Moreover, the overall energy consumption is reduced by approximately 65%, which enhances the efficiency of the system. Security analysis indicates a 96% poisoning attack detection and 94% inference attack resistance. The system is robust and only 35% accuracy deteriorates in adverse conditions. The results prove that BS-FDRL offers scalable, secure, and efficient communication-based intelligent control in Industry 4.0 IIoT setups.
A.Mallika, Anitha Christy, P. Krishnamoorthy et al.· 2026 6th International Confe...· 0 citations
Due to the rapid proliferation of Industrial Internet of Things (IIoT) systems within smart industries, scalable, energy-efficient, and sustainable edge intelligence solutions are in demand. Despite the fact that Federated Learning (FL) allows collaborative training without raw data sharing, traditional FL methods have high communication overhead, huge model sizes, and require a lot of energy, which restricts their implementation in resource-constrained industrial settings. To address these issues, the present paper introduces a Green Federated Learning framework, GFed-AMP, a sustainable IIoT edge intelligent framework that includes Adaptive Model Pruning.The suggested framework incorporates the dynamic magnitude-based pruning of local updates to eliminate redundant parameters and lessen the computational complexity. A strategy of energy-conscious client selection assigns priority to the devices with larger residual energy and with lower carbon intensity, to make sure that participation is environmentally conscious. It has real-time energy and carbon monitoring modules that can be used to gauge power consumption and environmental effects during federated training. Moreover, a sparse-awares aggregation mechanism is an effective method which optimally manages pruned model updates and ensures stable convergence. Experimental assessment of industrial anomaly detection data sets shows that GFed-AMP can achieve up to 40% model reduction, 35% communication overhead reduction, and 30% overall energy reduction in comparison with traditional FedAvg. Notably, the decline in the accuracy of predictions is less than 1.5%, which proves robustness and stability of learning. Better scalability, bandwidth, and lower carbon emission are proven by statistical comparisons. All in all, the overall experience with GFed-AMP’s trade-off between accuracy, communication efficiency, and environmental sustainability is a viable green AI solution to Industry 4.0 deployments.
A. Priya, Vungarala Satya Kishore, A. Christy et al.· 2026 6th International Confe...· 0 citations
