深度學習用於網路切片分類之實作

Abstract

中文摘要 隨著物聯網 (IoT)的發展，網路的用戶量急劇增加，服務特徵複雜， 因此 5G網絡服務需要極多樣化的服務質量 (QoS)要求。 目前 5G標準根 據場景和用戶需求將網絡進行劃分，不同場景提供不同的延遲、可靠 度、用戶連接數量等等客製化網路，透過 SDN及 NFV技術實現網絡 切片。 在未來設計 5G、 6G的研究中，如何在有限的頻寬中合理分配 資源、提升服務質量，網路切片和 AI技術備受關注。 第一部分使用 Kaggle 資料集用於神經網絡的訓練，搭建 深度 神經 網路 (Deep Neural Networks, DNN)，根據不同的延遲時間、封包損失率 等等用戶使用特徵，進行切片類型的分類。切片類型之間關係各自獨 立，對類別的輸入與輸出使用 one-hot Encoding表 示，消除類別間的潛 在關係。針對模型訓練結果透過權重正規 化 (weights regularization)、減 少模型大小、加入 Dropout解決過擬合 (overfitting)問題。 第二部分使用模糊平均 (fuzzy-c-means)為主方法將沒有標記的訓練 資料進行聚類 (clustering)及降維 (dimensionality)任務。對於多用 戶 、 多業務共存的網絡，通過 FCM 對每個服務進 行 處理，最後將具有相 似特徵 的服務分配到同 ⼀ 個切 片 上。 針對切片分類的任務分成三階段 進行，首先會由切片粒度著手，將不同服務 做優先權的排序。其次使用 fuzzy-c-mean 對用戶進行深度聚類，優先將資源分配給目標用戶。 最後針對複雜網路環境中各用戶情況，靈活調整切片資源，動態適應 通訊需求變化。

3 逢甲大學學生報告 ePaper(2023 年) Abstract With the development of the Internet of Things (IoT), the number of network users has increased dramatically, and the service characteristics are complex. Therefore, 5G network services require extremely diverse Quality of Service (QoS) requirements. The current 5G standard divides the network according to scenarios and user needs. Different scenarios provide customized networks with different delays, reliability, and number of user connections. Network slicing is realized through SDN and NFV technologies. In the future design of 5G and 6G research, how to rationally allocate resources and improve service quality in limited bandwidth, network slicing and AI technology have attracted much attention. The first part uses the Kaggle data set for the training of the neural network and builds the neural network. Slice types are classified according to different user usage characteristics such as delay time and packet loss rate. The relationship between slice types is independent, and one-hot Encoding is used to represent the input and output of the category, eliminating the potential relationship between categories. For the model training results, through weights regularization, reduce the size of the model, and add Dropout to solve the problem of overfitting. The second part uses fuzzy-c-means as the main method to perform clustering and dimensionality tasks on unmarked training data. For a network where multiple users and multiple services coexist, each service is processed through FCM, and services with similar characteristics are finally assigned to the same slice. The task of slicing classification is divided into three phases. First, the slice granularity will be used to prioritize different services. Secondly, use fuzzy-c-mean to perform deep clustering on users, and allocate resources to target users preferentially. Finally, according to the situation of each user in a complex network environment, flexibly adjust slice resources and dynamically adapt to changes in communication requirements.