Integrated Sensing, Computing, Communication, and Control for Time-Sequence-Based Semantic Communications

TL;DR

This paper introduces a novel semantic communication framework for real-time wireless control in IIoT, integrating sensing, computing, communication, and control to reduce overhead and improve adaptive control through deep reinforcement learning.

Contribution

It develops an integrated ISC3 architecture with semantic inference, mutual information-based transmission control, and adaptive control gain adjustment, trained via a hybrid multi-agent deep reinforcement learning approach.

Findings

Significant reduction in communication overhead.

Enhanced control accuracy and responsiveness.

Outperforms baseline schemes in real-world experiments.

Abstract

In the upcoming industrial internet of things (IIoT) era, a surge of task-oriented applications will rely on real-time wireless control systems (WCSs). For these systems, ultra-reliable and low-latency wireless communication will be crucial to ensure the timely transmission of control information. To achieve this purpose, we propose a novel time-sequence-based semantic communication paradigm, where an integrated sensing, computing, communication, and control (ISC3) architecture is developed to make sensible semantic inference (SI) for the control information over time sequences, enabling adaptive control of the robot. However, due to the causal correlations in the time sequence, the control information does not present the Markov property. To address this challenge, we compute the mutual information of the control information sensed at the transmitter (Tx) over different time and identify their temporal semantic correlation via a semantic feature extractor (SFE) module. By this means, highly correlated information transmission can be avoided, thus greatly reducing the communication overhead. Meanwhile, a semantic feature reconstructor (SFR) module is employed at the receiver (Rx) to reconstruct the control information based on the previously received one if the information transmission is not activated at the Tx. Furthermore, a control gain policy is also employed at the Rx to adaptively adjust the control gain for the controlled target based on several practical aspects such as the quality of the information transmission from the Tx to the Rx. We design the neural network structures of the above modules/policies and train their parameters by a novel hybrid reward multi-agent deep reinforcement learning framework. On-site experiments are conducted to evaluate the performance of our proposed method in practice, which shows significant gains over other baseline schemes.