Resource-Efficient Federated Learning and DAG Blockchain With Sharding in Digital-Twin-Driven Industrial IoT

The development of Industry 4.0 relies on emerging technologies of digital twin, machine learning, blockchain, and Internet of Things (IoT) to build autonomous self-configuring systems that maximize manufactory efficiency, precision, and accuracy. In this article, we propose a new distributed and secure digital twin-driven IIoT framework that integrates federated learning and directed acyclic graph (DAG) blockchain with sharding. The proposed framework includes three planes: 1) the data plane; 2) the blockchain plane; and 3) the digital twin plane. Specifically, the data plane performs federated learning through a set of cluster heads to train models at network edges for twin model construction. The blockchain plane, which supports sharding, utilizes a hierarchical consensus scheme based on DAG blockchain to verify both local model updates and global model updates. The digital twin plane is responsible for constructing and maintaining twin model. Then, an efficient resource scheduling scheme is designed by considering performance of both federated learning and DAG blockchain with sharding. Accordingly, an optimization problem is formulated to maximize long-term utility of the digital twin-driven IIoT. To cope with mapping error in the digital twin plane, a multiagent proximal policy optimization (MAPPO) approach is developed to solve the optimization problem. Numerical results illustrate that comparing with traditional approach, the proposed MAPPO improves utility by about 37 %, and reduces time latency by about 14%. Moreover, it also can well adapt to the mapping error.