EDGE AI for Heterogeneous and Massive IoT Networks

Abstract

By combining multiple sensing and wireless access technologies, the Internet of Things (IoT) shall exhibit features with large-scale, massive, and heterogeneous sensors and data. To integrate diverse radio access technologies, we present the architecture of heterogeneous IoT system for smart industrial parks and build an IoT experimental platform. Various sensors are installed on the IoT devices deployed on the experimental platform. To efficiently process the raw sensor data and realize edge artificial intelligence (AI), we describe four statistical features of the raw sensor data that can be effectively extracted and processed at the network edge in real time. The statistical features are calculated and fed into a back-propagation neural network (BPNN) for sensor data classification. By comparing to the k-nearest neighbor classification algorithm, we examine the BPNN-based classification method with a great amount of raw data gathered from various sensors. We evaluate the system performance according to the classification accuracy of BPNN and the performance indicators of the cloud server, which shows that the proposed approach can effectively enable the edge-AI-based heterogeneous IoT system to process the sensor data at the network edge in real time while reducing the demand for computing and network resources of the cloud.