【作者】 张斌；

【导师】 佘兢克； 李景龙；

【作者基本信息】 湖南大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 工业互联网的背景下,随着制造和生产过程的集成数字化,工业环境伴随着越来越多的数据。而数据普遍具有价值,人工智能与深度学习的方法分别为机器学习数据开发提供了巨大的优化潜力,如提高效率、灵活性和生产过程的个性化。通过状态监测、异常检测以及预防性维护或改进,以数据驱动的过程建模,能够更快地识别任务状态,使工业生产模式自主性得到增长。这提供了生产设备的预测性维护,并改善了工业过程的监控模式。在边缘计算背景中,因为带宽资源的不足和网络自身的不稳定性,基于深度神经网络模型的任务调度可能会有较高的延迟,从而影响用户的体验。并且,深度神经网络模型由于自身的复杂度和规模,对设备的存储能力和计算能力有一定的要求,不能直接部署在资源不足的设备端。所以,基于深度神经网络模型的应用如何满足用户对低延迟、高精度和高性价比的需求,成为任务调度的核心问题。本文在故障分类任务上对基于注意力机制的SK模型进行了改良,提出了一种新模型SK-ECA,该模型不仅保留了SK模型多卷积核特征信息处理能力,还吸取了ECA快速一维卷积的特性,提升了原SK模型在故障分类任务的性能指标。基于SK-ECA模型的双池化特征融合实验,吸取了全局平均池化归纳全局信息的能力,也有全局最大池化突出重要特征的特点,进一步挖掘了模型的潜力,拓展成七种注意力机制模型。同时,针对故障分类任务,采用基于注意力机制的神经网络模型,将边缘设备的便捷性和边缘节点的计算能力相结合,在边缘计算任务调度问题上展开了系列研究,提出了一种面向边缘设备的故障分类任务调度策略。该策略以协调边缘计算节点和设备端的方式,充分考虑了边缘节点出色的计算性能和移动设备的便捷性,兼顾用户需求和模型优势,完成注意力机制模型在边缘端和设备端的动态部署,并对分类任务进行按需调度,从而发挥深度注意力机制模型的优势,提升任务调度的效率,满足用户的需求。实验结果表明,选择的最佳调度模型在分类精度上比原SK模型提高了28%,平均召回率比原SK模型提高了26%,F分数比原SK模型提高了26%,推理时间比原SK模型降低了36%。更多还原

【Abstract】 In the context of the Industrial Internet,with the integration and digitization of manufacturing and production processes,the industrial environment is accompanied by more and more data.While data is generally of value,artificial intelligence and deep learning methods respectively provide huge optimization potentials for machine learning data development,such as improving efficiency,flexibility,and personalization of production processes.Through condition monitoring,anomaly detection,preventive maintenance or improvement,and data-driven process modeling,task status can be identified faster and the autonomy of industrial production models can be increased.This provides predictive maintenance of production equipment and improves the monitoring model of industrial processes.In the context of edge computing,due to the lack of bandwidth resources and the instability of the network itself,task scheduling based on the deep neural network model may have a higher delay,thereby affecting the user experience.Moreover,due to its own complexity and scale,the deep neural network model has certain requirements on the storage and computing capabilities of the device,and cannot be directly deployed on the device with insufficient resources.Therefore,how to meet the needs of users for low latency,high precision and high cost performance based on deep neural network model applications has become the core issue of task scheduling.In this paper,the SK model based on the attention mechanism is improved on the fault classification task,and a new model SK-ECA is proposed.This model not only retains the multi-convolution kernel feature information processing capability of the SK model,but also draws on the fast one of ECA.The characteristic of dimensional convolution improves the performance index of the original SK model in the fault classification task.The dual-pool feature fusion experiment based on the SK-ECA model has absorbed the ability of global average pooling to summarize global information,and also has the feature of global maximum pooling to highlight important features,further tapping the potential of the model and expanding into seven attention mechanisms model.At the same time,for fault classification tasks,a neural network model based on the attention mechanism is used to combine the convenience of edge devices with the computing power of edge nodes.A series of studies have been carried out on the problem of edge computing task scheduling,and an edge-oriented approach is proposed.The equipment’s fault classification task scheduling strategy.This strategy takes into account the excellent computing performance of edge nodes and the convenience of mobile devices by coordinating edge computing nodes and devices,taking into account user needs and model advantages,and completing the dynamic deployment of the attention mechanism model on the edge and device ends,and The classification tasks are scheduled on demand,so as to take advantage of the deep attention mechanism model,improve the efficiency of task scheduling,and meet the needs of users.Experimental results show that the selected optimal scheduling model has 28% higher classification accuracy than the original SK model,an average recall rate 26% higher than the original SK model,F score 26% higher than the original SK model,and reasoning time is higher than the original SK model.The SK model is reduced by 36%.更多还原