Research On Resource Adaption And Privacy Protection In Data-driven Industrial Internet

Data and its full-cycle processing are the core elements of industrial intelligence,they play vital roles in driving the application deployment of industrial Internet.However,the current industrial Internet are facing many challenges in data modeling,data-driven resource adaption,and data privacy protection.First of all,traditional data analysis methods exist low flexibility problems,which makes it difficult to meet the demand of diversified industrial applications for real-time and intelligent processing of massive data.Second,traditional static and coarse-grained resource management mode makes it difficult for the network to cooperatively adapt multi-dimensional resources(spectrum,cacing,and computing)in industrial scenarios.Finally,the centralized model training methods have the problems of data privacy leakage and low confidentiality,which is difficult to simultaneously satisfy the data sharing and privacy protection requirements.In recent years,Alliance of Industrial Internet put forward “Industrial Internet Architecture 2.0”,aimed at building a new type of Internet function structure with typical characteristics of “data”,“network”,and “security”,which provides a new technical route for effectively solving the above problems.Therefore,this dissertation,based on the Industrial Internet Architecture 2.0,focuses on the key issues in data-driven industrial scenarios and carries out research on data modeling,resource adaptation,and privacy protection.The main works and innovation points are as follows:(1)Aiming at the data modeling problem of industrial Internet,a deep learning based end-cloud collaborative data modeling approach was proposed.Firstly,based on the automatic washing equipment of high-speed railway in real scenarios,an “end-cloud”collaborative data acquisition system is designed to form a large-scale dataset supporting the sufficient training of service models.Secondly,an intelligent service model based on Deep Neural Network(DNN)was established to represent the global dependence of industrial operation and maintenance data through attention weights,so as to provide intelligent prediction service for equipment operation and maintenance.Then,through the operation mode of offline training and online inference,the service model can be equipped with the ability to make operation and maintenance decisions,to complete the instant and accurate perception of the operation state of industrial equipment.Finally,a prototype system platform was built,and performance evaluating for the proposed model was conducted over real-scene datasets.Experimental results show that the proposed approach achieves high accuracy of operation and maintenance decision,and thus the feasibility and effectiveness of the proposed approach are verified.(2)Aiming at the resource adaptation problem of industrial Internet,a multidimensional resource dynamic adaptation approach based on reinforcement learning is proposed.Firstly,an end-edge-cloud collaborative network architecture is designed to flexibly support service model deployment and inference task scheduling in resourceconstrained industrial scenarios.Secondly,in order to guarantee DNN inference service with low latency and high accuracy in industrial scenarios,the system modeling is carried out for communication,caching,and computing resources,and the multi-dimensional resource optimization problem is formulated with the goal of maximizing inference accuracy.Then,the formulated problem was transformed into a Markov decision process,and a resource adaption algorithm based on the Twin Delayed Deep Deterministic Policy Gradient was proposed,and thus the optimization problem can be solved efficiently.Finally,the convergence performance,task scheduling success rate,and caching resource utilization of the proposed algorithm are verified via simulation experiments.The results show that,compared with the deep deterministic policy gradient algorithm,the proposed algorithm has advantages in improving the success rate of task scheduling and the utilization of edge caching resource.(3)Aiming at the data privacy protection problem of industrial Internet,an efficient data privacy protection approach based on Federated Learning(FL)is proposed.Firstly,a multi-edge collaborative hierarchical FL mechanism is designed to complete parameter iteration of the DNN service model through the collaborative operation of end nodes,edge servers,and cloud servers,so as to meet the data privacy protection requirements of end-layer nodes effectively.Secondly,in order to ensure the efficient operation of FL mechanism,mathematical models of distributed training delay and energy consumption are established,and a stochastic optimization problem with the objective of minimizing the evaluation loss is formulated.Then,considering the opacity of information between multiple edge controllers,the formulated problem is transformed into a partiallyobservable Markov decision process,and a resource adaptation algorithm based on the Multi-Agent Soft Actor-Critic is proposed to solve the problem.Finally,simulation experiments are carried out over the built prototype system.Simulation results show that,the proposed algorithm can effectively reduce the energy consumption and save time cost while satisfying data privacy preserving requirements of end nodes.