Resource Management Research For Industrial Internet Of Things

With the rapid development of global communication and Internet technology,industrial manufacturing has now achieved the digitization,informatization and intelligentization of the entire production process.Technologies such as Federated Learning(FL),edge computing offloading,and Device-to-Device(D2D)offloading have been adopted to solve the problem of limited resources for industrial intelligent devices.FL achieves the utilization of data resources from other devices while protecting data privacy.Edge computing offloading and D2 D offloading can provide relatively abundant computing resources in locations close to devices.Due to the non-i.i.d.data and the complex task offloading decision-making environment in industrial scenarios,how to manage data and computing resources in response to these situations is still an important research direction.This article studies the management of data and computing resources in the Industrial Internet of Things(IIo T)scenario.To deal with the situation where device data is non-i.i.d.,a personalized Federated Learning algorithm is proposed to improve the accuracy of the final task model.To deal with the complex decision-making environment caused by the specificity of industrial devices and different task arrival rates,a task offloading algorithm based on deep reinforcement learning is proposed to reduce the overall system task latency.The innovative work of this article mainly includes the following aspects:(1)For the data resource management problem caused by multiple data distributions in the IIo T scenarios,label information is added to enable pre-aggregation of data which is known with same data distribution on the edge server can be preliminarily aggregated to improve clustering accuracy.Based on the difference among the model update quantities from different data distributions,a binary clustering method based on cosine similarity of the model update quantities is proposed for iterative clustering on the cloud.Whether the loss function descent directions of the participants are orthogonal is used as the basis for fast data clustering.The length of the model update magnitudes are used as the criterion for ending clustering,and a fast Federated Learning clustering algorithm based on cosine similarity(FFLCCS)is proposed.Simulation results demonstrate that the proposed algorithm is able to cluster data with different data distributions quickly,and is capable of producing task models with high accuracy even when training on small amounts of device data through data clustering.(2)To address the multi-task computing resource management problem where intelligent device resources are limited in the IIo T scenario,we propose three offloading decisions: compute tasks locally,edge offloading and Device-to-Device(D2D)offloading to efficiently utilize the system’s computing resources.Constraints on offloading strategies are proposed to meet the specificity of industrial devices in the IIo T scenario.Devices with time-varying task arrival rates are considered,as well as possible additional queueing delay caused by unfinished tasks from the previous time slot.Based on this environment,a time-varying task arrival offloading algorithm based on deep deterministic policy gradient(TVTAO)is proposed.Simulation results demonstrate that this algorithm can make excellent task offloading and resource allocation strategies under complex industrial scenarios,effectively reducing the average processing latency of the overall system compared to other task offloading methods,and it can better meet the low latency requirements of tasks.This paper includes 21 figures,4 tables and 62 references.