人工智能物联网中面向智能任务的语义通信方法  

Intelligent task-oriented semantic communication method in artificial intelligence of things

作　　者：刘传宏 郭彩丽[1,2] 杨洋[2] 冯春燕[1] 孙启政 陈九九 LIU Chuanhong;GUO Caili;YANG Yang;FENG Chunyan;SUN Qizheng;CHEN Jiujiu(Beijing Laboratory of Advanced Information Networks,Beijing University of Posts and Telecommunications,Beijing 100876,China;Beijing Key Laboratory of Network System Construction and Integration,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学先进信息网络北京实验室,北京100876 [2]北京邮电大学网络体系构建与融合北京市重点实验室,北京100876

出　　处：《通信学报》2021年第11期97-108,共12页Journal on Communications

基　　金：国家自然科学基金资助项目(No.92067202);国家重点研发计划基金资助项目(No.2018YFB1800805)。

摘　　要：随着物联网(IoT)和人工智能(AI)技术的融合发展,传统的数据集中式云计算处理方式难以有效去除数据中大量的冗余信息,给人工智能物联网(AIoT)中智能任务低时延、高精度的需求带来挑战。针对这一挑战,基于深度学习方法提出了AIoT中面向智能任务的语义通信方法。针对图像分类任务,在IoT设备上利用卷积神经网络(CNN)提取图像的特征图;从语义概念出发,将语义概念和特征图进行关联,提取语义关系;基于语义关系实现语义压缩,减小网络传输的压力,降低智能任务的处理时延。实验和仿真结果表明,对比传统通信方案,所提方法的复杂度仅约为传统方案的0.8%,同时具有更高的分类任务性能;对比特征图全部传输的方案,所提方法传输时延降低了80%,大大提升了有效分类准确率。With the integration and development of Internet of things(IoT)and artificial intelligence(AI)technologies,traditional data centralized cloud computing processing methods are difficult to effectively remove a large amount of redundant information in data,which brings challenges to the low-latency and high-precision requirements of intelligent tasks in the artificial intelligence of things(AIoT).In response to this challenge,a semantic communication method oriented to intelligent tasks in AIoT was proposed based on the deep learning method.For image classification tasks,convolutional neural networks(CNN)were used on IoT devices to extract image feature maps.Starting from semantic concepts,semantic concepts and feature maps were associated to extract semantic relationships.Based on the semantic relationships,semantic compression was implemented to reduce the pressure of network transmission and the processing delay of intelligent tasks.Experimental and simulation results show that,compared with traditional communication scheme,the proposed method is only about 0.8%of the traditional scheme,and at the same time it has higher classification task performance.Compared with the scheme that all feature maps are transmitted,the transmission delay of the proposed method is reduced by 80%and the effective accuracy of image classification task is greatly improved.

关 键 词：物联网 语义通信 图像分类 人工智能 语义压缩 

分 类 号：TN929.5[电子电信—通信与信息系统]