| With the development of wireless communication and artificial intelligence technology,the Industrial Internet of Things(IIoT)constructs an intelligent decision-making system integrating perception,communication,information processing and operation management to realize comprehensive intelligent interconnection of people,intelligent devices,industrial processes and data.Under the existing IIoT architecture,how to meet the timeliness of massive multi-source and multi-mode data collection is the key to quickly capture equipment status and improve the accuracy of industrial control.However,throughout the whole life cycle of data sampling,storage,transmission and calculation,which are mutually restricted and jointly affect the timeliness of data collection,thus reducing the value of fusion information.To solve this problem,facing the freshness of data,this paper researches data collection,node scheduling and resource allocation schemes by using 5G non-slot based transmission technology and reinforcement learning strategy.The main research contents include:(1)To solve the problem of data freshness in delay-sensitive IIoT network,this article designs a joint optimization mechanism of data sampling,caching and scheduling.Considering the constraints of node energy and queue capacity,dynamic channel variation and wireless resource shortage,etc.The transmission technology of 5G non-slot based scheduling is adopted to further reduce communication delay by setting data transmission time intervals(TTI)of different lengths.The age of information(AoI)at the destination and the age of information at the queue are used to represent the data freshness degree of the node and the backlog degree of queue respectively.Then,the maximum age of information(MAoI)among nodes minimization problem is established to improve the tolerance of system for the worst data freshness.The optimization problem is modeled as a Constrained Markov decision process(CMDP).The steady-state distribution probability of Markov chain is used to represent the closed-form solution of the objective function,and then a suboptimal sampling and semi-distributed scheduling policy is proposed.Compared with the centralized scheduling scheme,the communication overhead brought by state information exchange is greatly reduced.Simulation results show that the proposed strategy achieves the balance between queue sampling rate and service rate,and significantly reduces the MAoI.(2)Age of multimodal data fusion(AoMF)is proposed for the application scenario of multimodal data fusion of industrial Internet and the freshness of fusion task.In order to capture the characteristics of the actual collaborative nodes that jointly serve a real-time application,different upper bounds of Aol are used to distinguish the validity period of each modal data.The time-average AoMF minimization problem of multimodal data fusion is modeled as Markov decision Process(MDP).Considering the limitation of communication and computing resources,based on the non-orthogonal Multiple Access(NOMA)transmission technology,Soft Actor Critic with Discrete Actions(SACD)reinforcement learning algorithm is used to jointly design node scheduling and power allocation strategies.By remolding the reward of SACD algorithm,the convergence speed and performance are accelerated.Experiments show that the proposed scheme effectively reduce the average AoMF of the system.At the same time,the average task freshness does not increase monotonously with the increase of the number of superimposed transmitted signals,and there is an optimal number of node scheduling. |
