Real-time Video Offload Strategy Based On Industrial Edge Computing

With the booming development of artificial intelligence and industrial Internet of Things,a large number of cameras have entered factories.Industrial quality inspection is gradually changing from labor-intensive manual quality inspection to automated and intelligent industrial artificial intelligence(AI)quality inspection.Due to the increasing complexity and refinement of industrialized products,there are still problems with industrial AI quality inspection such as lack of accuracy and timeliness.Edge computing offloads partial or all of pending tasks to the edge server for processing to reduce the processing latency of the takes,providing a solution to meet the low latency demand.In this paper,we propose a video computing offload system based on industrial Io T for industrial video quality inspection applications,model the problem for the system.We model the problem of this system and propose different offloading strategies for two different types of scenario requirements,respectively.The specific work in this paper is as follows:Firstly,in this paper,the problem is modeled for the multi-user video computing offload strategy problem for this scenario.Considering the system real-time requirements as well as the energy cost,we define the system cost as the weighted sum of system delay and energy consumption,and then convert the offloading problem into a system cost minimization problem.Since the traditional heuristic offloading strategy cannot adapt to the dynamic network fluctuations in the plant,and the complexity of the algorithm increases with the number of edge devices,it cannot meet the requirement of timeliness.Therefore,this paper addresses this problem by proposing a deep reinforcement learningbased offloading strategy for solving it.The offloading policy is based on the Twin Delayed Deep Deterministic Policy Gradient(TD3)with the addition of a dynamic behavioral clone module.We build a dual experience pool to store valid and useless experiences separately,and add a Variational Autoencoder to construct state and action distributions.This makes full use of previous experience and allows the intelligence to adapt to dynamic network fluctuations and multi-user offloading scenarios.Simulation experiments show that the proposed strategy has a stable convergence capability.Compared with other benchmark strategies,our strategy consumes less system cost and can effectively save the system cost of multi-user video computing offload systems.Second,based on the above problem,this paper considers the accuracy requirements in industrial quality inspection scenarios and adds accuracy as a problem constraint.Accuracy,as the evaluation index of the video target detection model,is also one of the important factors affecting the computational offloading strategy.The more parameters the video processing model has,the higher the accuracy is and the processing delay becomes relatively longer.To address this problem,this paper reconstructs the Markov decision process model by adding an accuracy reward function and an additional reward based on the benchmark value.This reward can guide the intelligence to explore better strategies.Then,we propose a TD3-based computational offloading strategy to solve the problem.In order to avoid the accuracy-seeking of the intelligence to offload tasks to the server in large batches,a long short-term memory recurrent neural network is added to the strategy to predict the congestion level of the server queue.Simulation results show that this scheme can perform well compared to other benchmark schemes with better system cost savings while ensuring accuracy.