| The Io T system consists of sensors,actuators,network nodes used for data transmission,and remote servers near the site.Each part of the system has the characteristics of wide geographical distribution,logical layers,and obvious affiliation.Thanks to the increased computing power of today’s edge devices,Io T systems can distribute computing and services to the edge of the network to extend the traditional centralized service paradigm,making it more suitable for large-scale,widely distributed,and data-intensive applications and services.However,the current mainstream distributed computing technologies cannot meet the requirements of Io T systems for real-time performance,high efficiency,and low communication overhead when facing the ever-increasing network scale.In addition,the security of the system is also a problem.As a solution,near-sensor data analytics advocates processing sensory data close to its source rather than collecting it to a server for centralized processing.Near-sensor data analytics technology can reduce communication costs,especially for geographically dispersed network sensor systems.Compared with centralized servers,sensors and edge devices have limited computing and storage capabilities,so a single point of workload needs to be distributed to other devices collaboratively.Finally,since information sharing is a necessary part of multi-node cooperative machine learning,how to optimize the communication overhead between nodes is the key to improving the efficiency of distributed machine learning.This thesis studies the edge collaborative computing architecture of the Internet of Things and the data interaction of edge devices,which proposes an online edge collaborative computing method based on decentralization,data analytics model compression,transmission,and error compensation methods.Aiming at the problems of large central node load and poor robustness existing in traditional strong central distributed systems,a decentralized edge collaborative computing architecture is proposed,which realizes the application of artificial intelligence with high concurrency,low latency,and strong computing characteristics in widely distributed,deployment on edge devices with high heterogeneity and limited computing power;Aiming at the problem of high resource consumption in the data interaction process of edge devices,the quantitative compression and sparse compression transmission methods of data analysis models are proposed to reduce the amount of communication data;for model compression transmission In this thesis,an error compensation method based on quantization compression is firstly proposed,and the convergence analysis of the accumulated error is given.Secondly,as an improvement,this thesis proposes an error compensation method based on extrapolation compression,which not only ensures the accuracy of model training but also realizes efficient training of machine learning models on resource-constrained edge devices;for collaborative computing,the data flow generation rate is high,with strong time-series features,a proximity sensor data analysis method based on multi-task online machine learning is proposed.The decentralized online machine learning algorithm with communication compression and error compensation capabilities proposed in this thesis can be used as the core of adjacent sensor data analysis tasks and is suitable for sensor network systems with energy and bandwidth constraints.The experimental results show that the accuracy of the method proposed in this thesis is comparable to that of the traditional centralized training model.Under the condition of ensuring the convergence of the cumulative error,the training efficiency and accuracy of the model are improved,and the communication overhead is controlled at a reasonable level,which improves the robustness and scalability of the system. |
