A Lightweight System for High-efficiency Edge-Cloud Collaborative Applications
Zhipeng Zhang, Wenting Ma, Hao Li, Hongyi Tang, Xiaohang Yuan, Hao Yuan, Jun Xiao, Hongshun He, Wei Liu, Zhiheng Zhou · Research Square · 2022
Abstract With the development of IoT (Internet of Things) technology, ubiquitous IoT devices and the generated data increase the computational consumption dramatically. In industry, more and more manufacturers attempted to upgrade their manufacturing/assembly lines with sensors to achieve automated and intelligent workshops. Yet, to handle the massive and complicated data generated by end-point devices, technical companies started to utilize edge computing and artificial intelligence (AI). However, the shortcomings of AI in computing efficiency have become an obstacle when applying AI to edge devices. The majority of production systems need to run multiple applications for different detection tasks simultaneously. These tasks require edge nodes capable of managing threads/processes at the least cost to save computational consumption as much as possible. Thus, it is of great significance to improve the edge computing system whilst the presence of performance requirements and the limited computing resources. We intended to tackle the issues by focusing on optimizing the system structure and the complexity of the AI models. First and foremost, we proposed a distinctive cloud-edge collaborative framework composed of one universal block (UB) deployed on the edge devices and multiple parallel task-specific blocks (TSBs) on the cloud servers. This framework makes it possible to only deploy low-cost edge devices on the spots. To cooperate with the edge-cloud framework, we also introduced a pruning method that can reduce the model parameters without lowering the performance, making the system more efficient.