Deep neural network partitioning method for unmanned aerial vehicle and edge computing server

Abstract

The invention discloses a deep neural network partitioning method for an unmanned aerial vehicle and an edge computing server. The method comprises the steps of: firstly building a system cost model, then converting a partitioning decision problem in the process in which an unmanned aerial vehicle and a edge computing server execute DNN into a particle optimization problem, and solving the binary particle optimization problem based on a chaotic variation binary particle swarm algorithm to obtain the optimal partition point between the unmanned aerial vehicle and the edge computing serve. Computing resources of the unmanned aerial vehicle can be utilized to the maximum extent, the system energy consumption and total time delay cost are minimized, and the task execution efficiency of the unmanned aerial vehicle is effectively improved. According to the invention, the DNN can be reasonably divided between a mobile device and the edge computing server according to the hierarchical structure and the output data size of the DNN and the energy consumption required by each layer of network computing.

Description

technical field [0001] The present invention relates to the technical field of data processing, in particular to a deep neural network partitioning method for drones and edge computing servers. Background technique [0002] In recent years, Deep Neural Networks (DNN) have achieved great success in fields such as computer vision, natural language recognition, and medical diagnosis. With the development of science and technology, the performance of smart mobile devices has been greatly improved, but there is still a gap between the ability of mobile devices to handle DNN inference tasks and the increasing processing requirements. Since the high-precision DNN model consumes a lot of computing power and storage space, the DNN network is usually deployed on the cloud, but the data collected by the mobile device is transmitted to the cloud for offline processing, which will still cause high end-to-end delay and cannot To meet the real-time requirements for completing tasks, this ...

AI Technical Summary

Problems solved by technology

However, completing DNN inference at the edge still faces challenges

Images