CN113992753B - Intelligent caching strategy for heaven-earth integrated satellite network node - Google Patents

The invention discloses a heaven and earth integrated information center network caching strategy based on an SDN, which comprises the following steps: establishing a satellite caching strategy model based on an SDN-ICN framework, collecting satellite node caching log files, and establishing a caching characteristic data set; using the feature data set to establish a cache decision model offline through a support vector machine for on-board online cache decision; and (4) performing on-satellite cache replacement by using a cache decision model and the maximum survival time to adapt to satellite dynamics. The strategy can effectively utilize the storage space of the satellite nodes, improve the cache hit rate of the satellite nodes and reduce the data transmission delay of the space-ground integrated network.

一种天地一体化卫星网络节点的智能缓存策略An Intelligent Cache Strategy for Satellite Network Nodes of Space-Earth Integration

技术领域technical field

本发明涉及缓存策略领域，具体涉及一种天地一体化卫星网络节点的智能缓存策略。The invention relates to the field of caching strategies, in particular to an intelligent caching strategy of a satellite network node integrating space and earth.

背景技术Background technique

随着卫星通信技术的发展以及用户接入网络需求的增加，天地一体化通信网络协作传输数据将是未来网络的趋势。一体化网络目前的拓扑动态时变性以及计算和存储资源的有限性，使得传统网络端到端通信服务在空间信息网络中存在高时延，高丢包率，转发负担重等问题。随着卫星节点缓存能力的增加，为保障天地一体化网络中各节点的数据获取速度，减少重复内容的检索，提高网络效率，需要制定一个合理有效的缓存方案。With the development of satellite communication technology and the increasing demand of users to access the network, the cooperative transmission of data in the space-ground integrated communication network will be the trend of the future network. The current topology dynamic time-varying of the integrated network and the limited computing and storage resources make the traditional network end-to-end communication services in the spatial information network have problems such as high delay, high packet loss rate, and heavy forwarding burden. With the increase of the caching capacity of satellite nodes, in order to ensure the data acquisition speed of each node in the integrated space-earth network, reduce the retrieval of duplicate content, and improve the network efficiency, it is necessary to develop a reasonable and effective caching scheme.

缓存策略主要分为缓存决策策略和缓存替换策略，缓存决策策略解决内容是否被存储，缓存替换策略决定缓存的内容是否被替换。卫星节点的存储和计算资源有限，而天地一体化网络中传输的内容是海量的。现有的通用缓存策略不能有效适应卫星节点间频繁的拓扑切换，虽然提升了缓存命中率，但也带来了很多缓存冗余。这些缓存冗余在地面节点上可以暂时忽略，但在存储资源有限的卫星网络中是不能容忍的。缺乏简洁有效的缓存替换策略会导致缓存资源的冗余和大量缓存空间的浪费，因此，设计一种智能缓存和替换策略以有效提高卫星网络中的整体分发性能具有重要的实际意义。The cache strategy is mainly divided into a cache decision strategy and a cache replacement strategy. The cache decision strategy determines whether the content is stored, and the cache replacement strategy determines whether the cached content is replaced. The storage and computing resources of satellite nodes are limited, and the content transmitted in the integrated space-earth network is massive. The existing general caching strategy cannot effectively adapt to frequent topology switching between satellite nodes. Although the cache hit rate is improved, it also brings a lot of cache redundancy. These cache redundancies can be temporarily ignored on ground nodes, but cannot be tolerated in satellite networks with limited storage resources. The lack of a concise and effective cache replacement strategy will lead to redundancy of cache resources and waste of a large amount of cache space. Therefore, it is of great practical significance to design an intelligent cache and replacement strategy to effectively improve the overall distribution performance in satellite networks.

发明内容SUMMARY OF THE INVENTION

针对天地一体化网络中文件数量繁多，种类复杂的特点，通过SDN的全局视角收集待缓存内容的多维特征，提出了一种基于SDN-ICN架构的卫星缓存策略模型，避免了由各个节点独立做出缓存决策所出现的局部最优现象。In view of the large number and complex types of files in the integrated space-earth network, the multi-dimensional characteristics of the content to be cached are collected from the global perspective of SDN, and a satellite caching strategy model based on the SDN-ICN architecture is proposed, which avoids the need for each node to independently do it. The phenomenon of local optima that occurs when the cache decision is made.

为实现上述目的，本发明的技术方案为：一种天地一体化卫星网络节点的智能缓存策略，包括：In order to achieve the above purpose, the technical solution of the present invention is: an intelligent caching strategy for a satellite network node integrating space and earth, including:

构建基于SDN-ICN架构的卫星缓存策略模型，收集卫星节点缓存日志文件，建立缓存特征数据集；Build a satellite cache strategy model based on SDN-ICN architecture, collect satellite node cache log files, and build a cache feature data set;

基于缓存特征数据集，通过支持向量机离线建立缓存决策模型，用于星上在线缓存决策；Based on the cache feature data set, the cache decision model is established offline through the support vector machine, which is used for online cache decision on the satellite;

使用缓存决策模型和最大存活时间进行星上缓存替换，适应卫星动态性。On-board cache replacement is performed using a cache decision model and maximum survival time to adapt to satellite dynamics.

进一步的，所述的SDN-ICN架构的卫星缓存策略模型将SDN控制器部署在GEO卫星上，SDN控制器的设计包括缓存决策策略和转发策略，用于管理低轨卫星的ICN网络。每颗低轨卫星都部署ICN架构，具备网络内缓存能力。LEO卫星从地面内容服务器中获取资源，用户直接向地面转发节点发送兴趣包进行资源请求。卫星网络缓存策略算法的设计分为四个模块，分别是低轨卫星转发模块，SDN控制器路由模块，SDN控制器缓存管理模块和地面离线训练模块。由于在其间传输的是计算结果，判别模型的建立不需要高额的通信和计算开销，大大降低了资源占用。卫星缓存策略模型可整体描述为：Further, the satellite caching strategy model of the SDN-ICN architecture deploys the SDN controller on the GEO satellite, and the design of the SDN controller includes a caching decision strategy and a forwarding strategy for managing the ICN network of the low-orbit satellite. Each low-orbit satellite deploys an ICN architecture with in-network caching capabilities. LEO satellites obtain resources from ground content servers, and users directly send interest packets to ground forwarding nodes for resource requests. The design of the satellite network caching strategy algorithm is divided into four modules, which are the low-orbit satellite forwarding module, the SDN controller routing module, the SDN controller cache management module and the ground offline training module. Since the calculation results are transmitted between, the establishment of the discriminant model does not require high communication and calculation overhead, which greatly reduces the resource occupation. The satellite cache policy model can be described as a whole as:

step1.低轨卫星收到兴趣包后，首先在自身缓存空间内寻找有无匹配项，如有，则直接从自身缓存空间内获取数据包并进入step4，如无，则向上转发至SDN控制器；step1. After the low-orbit satellite receives the interest packet, it first looks for a match in its own buffer space. If there is, it will directly obtain the data packet from its own buffer space and enter step4. If not, it will be forwarded to the SDN controller. ;

step2.SDN控制器收到兴趣包后，进入缓存管理模块检索所有卫星的缓存信息，如无匹配项，则直接从PIT表中找到原始服务器并获取路径，获取后进行step3；如有匹配项，则从缓存列表中获取最近的数据包地址，随后进入step3；step2. After the SDN controller receives the interest packet, it enters the cache management module to retrieve the cache information of all satellites. If there is no match, it will directly find the original server from the PIT table and obtain the path, and then go to step 3; if there is a match, Then get the latest packet address from the cache list, and then enter step3;

step3.控制器获取数据包地址后进行路由计算。然后将传输路径输入缓存决策模型，寻找相关卫星，并进行缓存策略生成，最后下发相关节点的控制命令。step3. The controller performs routing calculation after obtaining the address of the data packet. Then input the transmission path into the cache decision model, find the relevant satellites, generate the cache strategy, and finally issue the control commands of the relevant nodes.

Step4.沿途卫星执行SDN下发的数据传输方案，发送数据包后，执行缓存决策及缓存替换操作。同时将自身缓存状态上传到SDN的缓存状态表中。Step4. The satellites along the way execute the data transmission scheme issued by the SDN, and after sending the data packets, execute the cache decision and cache replacement operations. At the same time, upload its own cache state to the cache state table of SDN.

Step5.地面缓存训练模块读取卫星缓存状态表，收集全局的缓存日志文件，对数据进行归一化处理，构建训练数据集并进行SVM缓存分类器训练，将训练好的缓存决策模型上传到SDN控制器并分发到低轨卫星节点中，进行之后的缓存决策。Step5. The ground cache training module reads the satellite cache status table, collects the global cache log files, normalizes the data, builds a training data set and trains the SVM cache classifier, and uploads the trained cache decision model to SDN The controller is distributed to low-orbit satellite nodes for subsequent caching decisions.

进一步的，step5中收集到的卫星缓存日志文件需要用通过预处理使数据形式规范化，预处理将分为两个步骤进行，日志过滤和训练数据集的构建，在日志过滤中，不相关或是无效的请求将被从日志文件中删除，此后将进行数据集构建，从日志代理文件中提取所需要的信息，使用有预测价值的数据特征进行数据集的构建，这些特征如表1所示：Further, the satellite cache log files collected in step 5 need to be normalized by preprocessing. The preprocessing will be divided into two steps, log filtering and training data set construction. In log filtering, irrelevant or Invalid requests will be deleted from the log file, after which the dataset will be constructed, the required information will be extracted from the log proxy file, and the dataset will be constructed using the data features with predictive value, as shown in Table 1:

表1:缓存日志文件中提取的数据特征Table 1: Extracted data features from cached log files

其中，内容数据包类型m₂分为4类，分别为文本，图像，音频，视频。m₃前一跳节点距离由SDN控制器中记录的当前时间片的卫星节点拓扑计算得出。m₄内容流行度的计算公式为单位时间内目标内容被请求的次数，m₅为内容的全部历史请求次数。m₆,m₇的值直接读取目标内容的上次请求时间差和兴趣包检索时间。y_i为内容数据包的标签，y_i数据的标记主要由SDN控制器进行，y_i的初状态为0，若缓存的内容文件在一个缓存周期T内有第二次访问则将y_i记为1。如在较短时间t内被再次访问，则将y_i记为2。据以上数据构建多维向量x_i＝{m₁,m₂,…,m₈}后，对数据进行归一化处理，使用公式

对原始数据进行线性变换，使其结果值映射到[0,1],其中c_i为经过归一化处理后的m_i值，m_max和m_min分别为此类数据的最大值和最小值。此时多维向量x_i可表示为x_i＝{c₁,c₂,…,c₈}。D＝{(x₁,y₁),(x₂,y₂),(x₃,y₃),…,(x_n,y_n)}为经过归一化处理后的数据集。Among them, the content data packet type m ₂ is divided into four categories, namely, text, image, audio, and video. m ₃ The previous hop node distance is calculated from the satellite node topology of the current time slice recorded in the SDN controller. The formula for calculating the content popularity of m ₄ is the number of times the target content is requested per unit time, and m ₅ is the total number of historical requests for the content. The value of m ₆ , m ₇ directly reads the last request time difference of the target content and the retrieval time of the Interest packet. y _i is the label of the content data packet. The marking of y _i data is mainly performed by the SDN controller. The initial state of y _i is 0. If the cached content file is accessed for the second time within a cache period T, y _i will be recorded. is 1. If it is revisited within a short time t, _yi is recorded as 2. After constructing a multi-dimensional vector x _i ={m ₁ ,m ₂ ,...,m ₈ } from the above data, normalize the data and use the formula

Linearly transform the original data to map the resulting value to [0,1], where c _i is the normalized value of m _i , and m _max and m _min are the maximum and minimum values of such data, respectively . At this time, the multi-dimensional vector _xi can be expressed as _xi ={c ₁ ,c ₂ ,...,c ₈ }. D={(x ₁ , y ₁ ), (x ₂ , y ₂ ), (x ₃ , y ₃ ), . . . , (x _n , y _n )} is the normalized dataset.

步骤二中所述的缓存决策模型是基于支持向量机构建的，SVM算法的基本思路是将一个具有线性的样本映射到另一个更高维度的空间，然后用最大边际法求出一个独立的超平面，在高维空间中，支持向量机的目的是使超平面到最接近的训练样本的最小距离最大化。具体实现方法为：The cache decision model described in step 2 is constructed based on support vector machines. The basic idea of the SVM algorithm is to map a linear sample to another higher-dimensional space, and then use the maximum margin method to find an independent super Plane, in high-dimensional space, the purpose of SVM is to maximize the minimum distance of the hyperplane to the closest training sample. The specific implementation method is:

Step1：将清洗后的数据的70％作为训练集，30％作为测试集，输入SVM分类模型；Step1: Take 70% of the cleaned data as the training set and 30% as the test set, and input it into the SVM classification model;

Step2：引入径向基核函数

用于解决样本集线性不可分问题，其中σ为核函数参数；Step2: Introduce radial basis kernel function

It is used to solve the linear inseparability problem of the sample set, where σ is the kernel function parameter;

Step3：使用AOPSO算法进行支持向量机模型的参数寻优；Step3: Use the AOPSO algorithm to optimize the parameters of the support vector machine model;

Step4：使用决策函数

进行缓存决策，其中α_i与b为拉格朗日算子，用于进行辅助计算；Step4: Use decision function

Make a cache decision, where α _i and b are Lagrangian operators for auxiliary calculation;

步骤二中的支持向量机模型是根据粒子群优化算法构建的具体实现方法是：将粒子设计为一个二进制位串。其数据主要包含三部分，分别为参数C，参数σ和特征掩码，粒子的前n_C位代表参数C，中间n_σ位代表参数σ，最后n_F位代表特征掩码，其中n_C和n_σ根据精度要求来决定，n_F根据数据集的特征数决定。基于特征排布优化粒子群算法的SVM参数寻优流程如下：The support vector machine model in the second step is constructed according to the particle swarm optimization algorithm. The specific implementation method is: design the particle as a binary bit string. Its data mainly includes three parts, namely parameter C, parameter σ and feature mask. The first n _C bits of the particle represent parameter C, the middle n _σ bits represent parameter σ, and the last n _F bits represent feature mask, where n _C and n _σ is determined according to the accuracy requirements, and n _F is determined according to the number of features of the data set. The SVM parameter optimization process based on the feature arrangement optimization particle swarm algorithm is as follows:

Step1初始化粒子群；Step1 Initialize the particle swarm;

Step2根据粒子编码方式，将每个粒子转化为二进制表达，转化依据为SVM的参数C和σ及所选择的特征子集，然后调用SVM算法进行学习训练，测试并记录分类精度，根据式

计算粒子适应度，其中ω_α表示分类精度的权重，ω_f表示特征数目倒数的权重；svm_accuracy是当前分类精度，f_i是特征的适应度，n_F是特征掩码的数量；Step2 According to the particle encoding method, convert each particle into a binary expression, and the conversion is based on the parameters C and σ of the SVM and the selected feature subset, and then call the SVM algorithm for learning and training, test and record the classification accuracy, according to the formula

Calculate particle fitness, where ω _α represents the weight of classification accuracy, ω _f represents the weight of the reciprocal of the number of features; svm _accuracy is the current classification accuracy, f _i is the fitness of the feature, and n _F is the number of feature masks;

Step3根据粒子适应度，更新粒子的局部最优值和全局最优值；Step3 According to the particle fitness, update the local optimal value and the global optimal value of the particle;

Step4根据式v_id＝v_id+c₁×rand×(p_id-x_id)+c₂×raid×(p_gd-x_id),Step4 According to the formula v _id = v _id +c ₁ ×rand×(p _id -x _id )+c ₂ ×raid×(p _gd -x _id ),

更新粒子的速度；其中c₁、c₂是坐标参数，v_id是粒子速度，p_id、p_gd是最优粒子坐标，x_id、x_gd是当前粒子坐标；Update the speed of the particle; where c ₁ , c ₂ are coordinate parameters, v _id is the particle speed, p _id , p _gd are the optimal particle coordinates, and x _id , x _gd are the current particle coordinates;

Step5根据粒子的初始位置及粒子速度，更新粒子当前位置，判断是否达到最大迭代次数N，若达到，则输出当前最优的特征子集，参数C、σ及分类精度，否则返回Step2，继续迭代。Step5 According to the initial position of the particle and the particle velocity, update the current position of the particle, and determine whether the maximum number of iterations N is reached. If so, output the current optimal feature subset, parameters C, σ and classification accuracy, otherwise return to Step2 and continue the iteration .

步骤三中所述星上缓存替换基于步骤二中离线训练得到的SVM缓存分类模型，当在线组件收到用户请求的数据包对象g时，使用SVM智能替换组件判断是否将所请求的对象的副本复制到卫星网络缓存中。SVM分类器根据对象是否会被再次访问，将g的标签y_i标注为0,1,2。与此同时，使用SDN计算拓扑切换时间，将SDN发送的拓扑切换时间t作为缓存权重。每当SVM将对象标注为y_i≥0，则使其存活时间增加一个t值，内容在缓存堆栈内按照存活时间排序，存活时间为0时，当新的值得缓存文件到达时可以直接丢弃。The on-board cache replacement described in step 3 is based on the SVM cache classification model obtained by offline training in step 2. When the online component receives the data packet object g requested by the user, the SVM intelligent replacement component is used to judge whether the copy of the requested object is to be replaced. Copy to satellite network cache. The SVM classifier labels the labels _yi of g as 0, 1, 2 according to whether the object will be revisited. At the same time, the topology switching time is calculated using SDN, and the topology switching time t sent by the SDN is used as the cache weight. Whenever the SVM marks an object as y _i ≥ 0, its survival time is increased by a value of t, and the contents are sorted according to the survival time in the cache stack. When the survival time is 0, it can be directly discarded when a new worthy cache file arrives.

本发明由于采用以上技术方案，能够取得如下的技术效果：天地一体化卫星网络节点智能缓存策略利用SDN的优势，为ICN提供全局缓存决策方案，通过集中式SDN控制器，对卫星网络中资源文件的多维特征进行记录，使用支持向量机建立分类预测模型，从全局角度做出缓存决策。同时基于最大存活时间的缓存替换策略适应了卫星网络的动态变化，整体上改进了缓存空间的利用效率。Due to the adoption of the above technical solutions, the present invention can achieve the following technical effects: the intelligent caching strategy of the satellite network nodes of the space-earth integration utilizes the advantages of SDN to provide a global caching decision-making solution for the ICN, and through the centralized SDN controller, the resource files in the satellite network are monitored and controlled. The multi-dimensional features are recorded, and the support vector machine is used to build a classification prediction model to make caching decisions from a global perspective. At the same time, the cache replacement strategy based on the maximum survival time adapts to the dynamic changes of the satellite network and improves the utilization efficiency of the cache space as a whole.

附图说明Description of drawings

图1是基于SDN-ICN的卫星网络缓存系统缓存决策算法流程图。Figure 1 is a flowchart of the caching decision algorithm of the satellite network caching system based on SDN-ICN.

图2是粒子二进制结构示意图。Figure 2 is a schematic diagram of the particle binary structure.

图3是AOPSO优化算法收敛示意图。Figure 3 is a schematic diagram of the convergence of the AOPSO optimization algorithm.

图4是SVM缓存决策分类示意图。Figure 4 is a schematic diagram of SVM cache decision classification.

图5是单个仿真周期内数据传输的时延图。Figure 5 is a time-delay diagram of data transfer within a single simulation cycle.

具体实施方式Detailed ways

针对天地一体化网络中文件数量繁多，种类复杂的特点，通过SDN的全局视角收集待缓存内容的多维特征，提出了一种基于SDN-ICN架构的卫星缓存策略模型，避免了由各个节点独立做出缓存决策所出现的局部最优现象。提出了AOPSO-SVM(Arrangementoptimization particle swarm optimization algorithm)缓存决策算法，此算法使用SVM建立缓存决策模型，考虑内容的流行度，上一跳节点位置，缓存时间，文件大小等多维特征，改进了传统缓存决策策略只考虑单一维度特征，无法兼顾全局的问题，同时使用特征结构排布优化后的粒子群优化算法进行SVM参数寻优。同时针对卫星网络动态性高，拓扑时变的特点导致的缓存替换滞后性，设计了一种卫星网络自适应缓存替换策略。设计最大存活时间，保证卫星在切换时也具备较高的缓存命中率。In view of the large number and complex types of files in the integrated space-earth network, the multi-dimensional characteristics of the content to be cached are collected from the global perspective of SDN, and a satellite caching strategy model based on the SDN-ICN architecture is proposed, which avoids the need for each node to independently do it. The phenomenon of local optima that occurs when the cache decision is made. The AOPSO-SVM (Arrangementoptimization particle swarm optimization algorithm) cache decision algorithm is proposed. This algorithm uses SVM to establish a cache decision model, considering the popularity of the content, the location of the last hop node, cache time, file size and other multi-dimensional characteristics, which improves the traditional cache. The decision-making strategy only considers the features of a single dimension and cannot take into account the global problem. At the same time, the particle swarm optimization algorithm after the optimization of the feature structure is used to optimize the SVM parameters. At the same time, aiming at the hysteresis of cache replacement caused by the characteristics of high dynamics and time-varying topology of satellite network, an adaptive cache replacement strategy of satellite network is designed. The maximum survival time is designed to ensure that the satellite also has a high cache hit rate during handover.

下面结合附图和具体实施例对本发明作进一步详细的描述：以此为例对本申请做进一步的描述说明。The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: the present application will be further described and illustrated by taking this as an example.

如图1所示，本实施例提供一种天地一体化卫星网络节点智能缓存策略，具体包括：As shown in FIG. 1 , this embodiment provides an intelligent caching strategy for satellite network nodes that integrates space and ground, which specifically includes:

一、建立基于SDN-ICN架构的卫星缓存策略模型，收集卫星节点缓存日志文件，建立缓存特征数据集；1. Establish a satellite cache strategy model based on the SDN-ICN architecture, collect satellite node cache log files, and establish a cache feature data set;

本发明中所述的SDN-ICN架构的卫星缓存策略模型将SDN控制器部署在GEO卫星上，SDN控制器的设计主要基于RYU控制器进行，包括缓存管理模块和路由模块，用于管理低轨卫星的ICN网络。具体为：首先经由路由模块对天地一体化智能网络的数据包转发日志文件进行记录，然后使用缓存管理模块对数据进行数据特征提取和数据集构建。The satellite cache strategy model of the SDN-ICN architecture described in the present invention deploys the SDN controller on the GEO satellite. The design of the SDN controller is mainly based on the RYU controller, including a cache management module and a routing module, which are used to manage low-orbit orbits. Satellite ICN network. Specifically: firstly, the data packet forwarding log file of the integrated intelligent network of heaven and earth is recorded through the routing module, and then the data feature extraction and data set construction are performed on the data by the cache management module.

卫星网络缓存策略的整体设计分为四个模块，分别是低轨卫星转发模块，SDN控制器路由模块，SDN控制器缓存管理模块和地面离线训练模块。由于在其间传输的是计算结果，判别模型的建立不需要高额的通信和计算开销，大大降低了资源占用。本实例中，一个完整的兴趣包初次请求过程为：The overall design of the satellite network caching strategy is divided into four modules, namely the low-orbit satellite forwarding module, the SDN controller routing module, the SDN controller cache management module and the ground offline training module. Since the calculation results are transmitted between, the establishment of the discriminant model does not require high communication and calculation overhead, which greatly reduces the resource occupation. In this example, a complete initial request process for an Interest packet is:

step1.低轨卫星收到兴趣包后，首先在自身缓存空间内寻找有无匹配项，如有，则直接从自身缓存空间内获取数据包并进入step4，如无，则向上转发至SDN控制器；step1. After the low-orbit satellite receives the interest packet, it first looks for a match in its own buffer space. If there is, it will directly obtain the data packet from its own buffer space and enter step4. If not, it will be forwarded to the SDN controller. ;

step2.SDN控制器收到兴趣包后，进入缓存管理模块检索所有卫星的缓存信息，如无匹配项，则直接从PIT表中找到原始服务器并获取路径，获取后进行step3；如有匹配项，则从缓存列表中获取最近的数据包地址，随后进入step3；step2. After the SDN controller receives the interest packet, it enters the cache management module to retrieve the cache information of all satellites. If there is no match, it will directly find the original server from the PIT table and obtain the path, and then go to step 3; if there is a match, Then get the latest packet address from the cache list, and then enter step3;

step3.控制器获取数据包地址后进行路由计算。然后将传输路径输入缓存决策模型，寻找相关卫星，并进行缓存策略生成，最后下发相关节点的控制命令。step3. The controller performs routing calculation after obtaining the address of the data packet. Then input the transmission path into the cache decision model, find the relevant satellites, generate the cache strategy, and finally issue the control commands of the relevant nodes.

Step4.沿途卫星执行SDN下发的的数据传输方案，发送数据包后，执行缓存决策及缓存替换操作。同时将自身缓存状态上传到SDN的缓存状态表中。Step4. The satellites along the way implement the data transmission scheme issued by the SDN. After sending the data packets, the cache decision and cache replacement operations are performed. At the same time, upload its own cache state to the cache state table of SDN.

Step5.地面缓存训练模块读取卫星缓存状态表，收集全局的缓存日志文件，对数据进行归一化处理，构建训练数据集并进行SVM缓存分类器训练，将训练好的缓存决策模型上传到SDN控制器并分发到低轨卫星节点中，进行之后的缓存决策。Step5. The ground cache training module reads the satellite cache status table, collects the global cache log files, normalizes the data, builds a training data set and trains the SVM cache classifier, and uploads the trained cache decision model to SDN The controller is distributed to low-orbit satellite nodes for subsequent caching decisions.

获取卫星缓存状态日志文件后，首先将卫星缓存日志文件通过预处理使数据形式规范化，预处理将分为两个步骤进行，日志过滤和训练数据集的构建，在日志过滤时，从日志文件中删除不相关或是无效的请求，此后，从日志代理文件中提取所需要的信息，提取有预测价值的数据特征进行数据集的构建，由m₁：内容数据包大小；m₂：内容数据包类型；m₃：上一跳节点位置；m₄：内容流行度；m₅：历史请求次数；m₆：上次请求间隔；m₇：兴趣包检索时间；m₈：时延敏感度。其中，内容数据包类型m₂分为4类，分别为文本，图像，音频，视频。m₃前一跳节点距离由SDN控制器中记录的当前时间片的卫星节点拓扑计算得出。m₄内容流行度的计算公式为单位时间内目标内容被请求的次数，m₅为内容的全部历史请求次数。m₆,m₇的值直接读取目标内容的上次请求时间差和兴趣包检索时间。y_i数据的标记主要由SDN控制器进行，y_i的初状态为0，若缓存的内容文件在一个缓存周期T内有第二次访问则将y_i记为1。如在较短时间t内被再次访问，则将y_i记为2。After obtaining the satellite cache status log file, first normalize the data form of the satellite cache log file through preprocessing. The preprocessing will be divided into two steps, log filtering and training data set construction. During log filtering, from the log file Delete irrelevant or invalid requests, after that, extract the required information from the log proxy file, extract data features with predictive value to construct a data set, m ₁ : content packet size; m ₂ : content data packet type; m ₃ : last hop node position; m ₄ : content popularity; m ₅ : number of historical requests; m ₆ : last request interval; m ₇ : Interest packet retrieval time; m ₈ : delay sensitivity. Among them, the content data packet type m ₂ is divided into four categories, namely, text, image, audio, and video. m ₃ The previous hop node distance is calculated from the satellite node topology of the current time slice recorded in the SDN controller. The formula for calculating the content popularity of m ₄ is the number of times the target content is requested per unit time, and m ₅ is the total number of historical requests for the content. The value of m ₆ , m ₇ directly reads the last request time difference of the target content and the retrieval time of the Interest packet. The marking of _yi data is mainly performed by the SDN controller. The initial state of _yi is 0. If the cached content file is accessed for the second time within a cache period T, _yi is recorded as 1. If it is revisited within a short time t, _yi is recorded as 2.

根据以上数据构建多维向量x_i＝{m₁,m₂,…,m₈}后，对数据进行归一化处理，使用公式

对原始数据进行线性变换，使其结果值映射到[0,1],其中c_i为经过归一化处理后的m_i值，m_max和m_min分别为此类数据的最大值和最小值。此时多维向量x_i可表示为x_i＝{c₁,c₂,…,c₈}。是经过归一化处理后的数据集为D＝{(x₁,y₁),(x₂,y₂),(x₃,y₃),…,(x_n,y_n)}。After constructing a multi-dimensional vector x _i ={m ₁ ,m ₂ ,...,m ₈ } based on the above data, normalize the data and use the formula

Linearly transform the original data to map the resulting value to [0,1], where c _i is the normalized value of m _i , and m _max and m _min are the maximum and minimum values of such data, respectively . At this time, the multi-dimensional vector _xi can be expressed as _xi ={c ₁ ,c ₂ ,...,c ₈ }. is the normalized dataset D={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),(x ₃ ,y ₃ ),…,(x _n ,y _n )}.

二、使用特征数据集通过支持向量机离线建立缓存决策模型，用于星上在线缓存决策；2. Use the feature data set to establish a cache decision model offline through the support vector machine, which is used for online cache decision on the satellite;

首先将归一化处理后的数据集用于离线训练SVM缓存分类模型。本发明通过使用特征选择排布优化的粒子群优化算法进行SVM缓存分类模型训练，以得到最终的缓存分类模型。该方法改进了粒子群优化算法和SVM分类算法，提高了算法的准确率，减少了计算量，降低了模型的复杂度，提高了模型分类的速度。First, the normalized dataset is used to train the SVM cache classification model offline. In the present invention, the SVM cache classification model is trained by using the particle swarm optimization algorithm optimized for feature selection and arrangement, so as to obtain the final cache classification model. The method improves the particle swarm optimization algorithm and the SVM classification algorithm, improves the accuracy of the algorithm, reduces the amount of calculation, reduces the complexity of the model, and improves the speed of model classification.

SVM的基本概念是使用高维空间来寻找一个线性边界或超平面，用两类正负样本进行二分类，此方法可以很好的解决二分类问题，SVM被建模为一种二次规划问题，在训练中存在全局最优解，同时，SVM训练中加入惩罚参数可以最大化分类边际，因此有着十分优秀的泛化能力。另外，因为松弛变量拓宽了错误的边界，SVM对于异常值是鲁棒的。The basic concept of SVM is to use high-dimensional space to find a linear boundary or hyperplane, and use two types of positive and negative samples for binary classification. This method can solve the binary classification problem very well. SVM is modeled as a quadratic programming problem , there is a global optimal solution in the training, and at the same time, adding a penalty parameter to the SVM training can maximize the classification margin, so it has a very good generalization ability. In addition, SVM is robust to outliers because slack variables widen the margin of error.

经过归一化处理后的数据集D＝{(x₁,y₁),(x₂,y₂),…,(x_n,y_n)}。此时影响SVM分类器的分类精度的主要参数变量为的惩罚参数C和径向基核函数的参数σ的取值，惩罚因子C的取值决定了分类器的泛化能力，径向基核函数的参数σ则决定了线性不可分样本数据映射到高维空间后的径向作用范围。本发明使用特征结构排布优化后的粒子群优化算法对进行针对性参数C和σ二元寻优，无需遍历范围内所有参数点，使用较少的时间即可获得近似的全局最优解。The normalized dataset D={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _n ,y _n )}. At this time, the main parameter variables that affect the classification accuracy of the SVM classifier are the value of the penalty parameter C and the parameter σ of the radial basis kernel function. The value of the penalty factor C determines the generalization ability of the classifier, and the radial basis kernel The parameter σ of the function determines the radial action range after the linear inseparable sample data is mapped to the high-dimensional space. The invention uses the particle swarm optimization algorithm after the characteristic structure arrangement and optimization to carry out the binary optimization of the targeted parameters C and σ, without traversing all parameter points in the range, and the approximate global optimal solution can be obtained in less time.

粒子群优化算法的基本思路是初始化一组随机解，通过迭代寻找最优解。每个粒子都有一个表示在当前解空间位置的属性X_i＝{x_i1，x_i2,…,x_in}，并由评价函数计算其适应度，同时，每个粒子有一个速度V_i＝{v_i1,v_i2,…,v_in}决定其运动的方向和距离。粒子之间通过共享最优粒子的信息，在当前解空间中搜索最优解The basic idea of particle swarm optimization algorithm is to initialize a set of random solutions and find the optimal solution through iteration. Each particle has an attribute X _{i =} {x _i1 , x _i2 , _. {v _i1 ,v _i2 ,…,v _in } determine the direction and distance of its motion. By sharing the information of the optimal particle, the particles search for the optimal solution in the current solution space

根据粒子群优化算法优化支持向量机模型参数的具体实现方法是：将粒子群中的粒子设计为一个二进制位串。如图2所示，其数据主要包含三部分，分别为参数C，参数σ和特征掩码，粒子的前n_C位代表参数C，中间n_σ位代表参数σ，最后n_F位代表特征掩码，其中n_C和n_σ根据精度要求来决定，n_F根据数据集的特征数决定。基于特征排布优化粒子群算法的SVM参数寻优流程为：The specific implementation method of optimizing the parameters of the support vector machine model according to the particle swarm optimization algorithm is: design the particles in the particle swarm as a binary bit string. As shown in Figure 2, the data mainly includes three parts, namely parameter C, parameter σ and feature mask. The first n _C bits of the particle represent parameter C, the middle n _σ bits represent parameter σ, and the last n _F bits represent feature mask. code, where n _C and n _σ are determined according to the accuracy requirements, and n _F is determined according to the number of features of the data set. The SVM parameter optimization process based on the feature arrangement optimization particle swarm algorithm is as follows:

Step1、初始化粒子群；Step1, initialize the particle swarm;

Step2、根据粒子编码方式,将每个粒子转化为二进制表达，转化依据为SVM的参数C和σ及所选择的特征子集,然后调用SVM算法进行学习训练,测试并记录分类精度,根据式

计算粒子适应度,其中ω_α表示分类精度的权重，ω_f表示特征数目倒数的权重；Step2. According to the particle encoding method, convert each particle into a binary expression, and the conversion is based on the parameters C and σ of the SVM and the selected feature subset, and then call the SVM algorithm for learning and training, test and record the classification accuracy, according to the formula

Calculate the particle fitness, where ω _α represents the weight of the classification accuracy, and ω _f represents the weight of the inverse of the number of features;

Step3、根据粒子适应度，更新粒子的局部最优值和全局最优值；Step3. According to the particle fitness, update the local optimal value and the global optimal value of the particle;

Step4、根据式v_id＝v_id+c₁×raid×(p_id-x_id)+c₂×raid×(p_gd-x_id)，Step4. According to the formula v _id = v _id +c ₁ ×raid×(p _id -x _id )+c ₂ ×raid×(p _gd -x _id ),

更新粒子的速度；update the speed of the particle;

Step5、根据粒子的初始位置及粒子速度，更新粒子当前位置，判断是否达到最大迭代次数N，若达到，则输出当前最优的特征子集,参数C、σ及分类精度，否则返回Step2，继续迭代。Step5. According to the initial position and velocity of the particle, update the current position of the particle, and judge whether the maximum number of iterations N is reached. If so, output the current optimal feature subset, parameters C, σ and classification accuracy, otherwise return to Step2 and continue iterate.

上述过程建立了星上缓存决策模型，有效提高了缓存决策的准确率，提升了星上缓存空间的利用效率，仿真试验结果如图3所示。The above process establishes an on-board cache decision model, which effectively improves the accuracy of cache decision-making and improves the utilization efficiency of on-board cache space. The simulation test results are shown in Figure 3.

三、使用缓存决策模型和最大存活时间进行星上缓存替换，适应卫星动态性。3. Use the cache decision model and the maximum survival time to replace the on-board cache to adapt to the dynamics of the satellite.

得到离线训练的SVM缓存分类模型后，进行星上缓存替换，当在线组件收到用户请求的数据包对象g时，使用SVM智能替换组件判断是否将所请求的对象的副本复制到卫星网络缓存中。SVM分类器根据对象是否会被再次访问，将g的标签y_i标注为0,1，2。与此同时，使用SDN计算拓扑切换时间，将SDN发送的拓扑切换时间t作为缓存权重。每当SVM将对象标注为y_i≥0，则使其存活时间增加一个t值，内容在缓存堆栈内按照存活时间排序，存活时间为0时，当新的值得缓存文件到达时可以直接丢弃，分类准确度如图4所示。After obtaining the offline-trained SVM cache classification model, on-board cache replacement is performed. When the online component receives the data packet object g requested by the user, the SVM intelligent replacement component is used to determine whether to copy the copy of the requested object to the satellite network cache. . The SVM classifier labels the label _yi of g as 0, 1, 2 according to whether the object will be visited again. At the same time, the topology switching time is calculated using SDN, and the topology switching time t sent by the SDN is used as the cache weight. Whenever the SVM marks the object as y _i ≥ 0, its survival time is increased by a value of t, and the contents are sorted according to the survival time in the cache stack. When the survival time is 0, when a new worthwhile cache file arrives, it can be discarded directly. The classification accuracy is shown in Figure 4.

本发明提出一种基于SVM的多维缓存决策模型，可以兼顾内容的流行度，节点位置，缓存时间，文件大小等特征，同时使用粒子特征排布优化算法进行模型参数寻优，尽最大可能降低计算复杂度和路由器间的交互频率，节约了节点的计算资源和存储资源，因此该算法非常适合用在计算和存储资源有限的天地一体化智能网络中。同时由于SVM机器学习算法模型构建时间长，复杂度高，所以同时设计了离线模型训练和在线缓存决策相结合的实现方法。最后，针对卫星网络动态性高，拓扑时变的特点导致的缓存替换滞后性，设计了一种卫星网络自适应缓存替换策略。设计最大存活时间，保证卫星在切换时也具备较高的缓存命中率。The invention proposes a multi-dimensional cache decision-making model based on SVM, which can take into account the popularity of content, node location, cache time, file size and other characteristics, and at the same time uses the particle feature arrangement optimization algorithm to optimize model parameters, reducing the calculation as much as possible. The complexity and the interaction frequency between routers save the computing resources and storage resources of nodes, so the algorithm is very suitable for use in the intelligent network of space-earth integration with limited computing and storage resources. At the same time, due to the long construction time and high complexity of the SVM machine learning algorithm model, an implementation method combining offline model training and online cache decision-making is designed. Finally, aiming at the hysteresis of cache replacement caused by the characteristics of high dynamics and time-varying topology of satellite network, an adaptive cache replacement strategy of satellite network is designed. The maximum survival time is designed to ensure that the satellite also has a high cache hit rate during handover.

下面结合实验仿真对本发明做进一步说明：The present invention is further described below in conjunction with experimental simulation:

1、实验环境搭建1. Setting up the experimental environment

本实验使用matlab对星地一体化网络进行建模，仿真模型包括三颗高轨卫星，24颗walker星座低轨卫星和16个地面站。其中walker星座卫星的轨道高度为1400km，轨道倾角52°，分为3个轨道平面，每个轨道平面分布8颗卫星。Walker星座卫星真实运行周期约为120分钟，本实验对其进行等比缩放，以120S为一个卫星运行周期，卫星的拓扑结构不变，拓扑切换周期为10秒/次。SDN控制器放置于高轨卫星，主要作用是进行日志收集和全局路由控制，仿真中兴趣包发送由16个地面站同时进行，数据包则在低轨卫星节点中传输，地面站负责最后一跳接收。This experiment uses matlab to model the satellite-ground integrated network. The simulation model includes three high-orbit satellites, 24 low-orbit satellites in the Walker constellation and 16 ground stations. Among them, the orbital height of the Walker constellation satellites is 1400km, and the orbital inclination angle is 52°. The real operation period of the Walker constellation satellite is about 120 minutes. In this experiment, it is scaled proportionally. Taking 120S as a satellite operation period, the topology of the satellite remains unchanged, and the topology switching period is 10 seconds/time. The SDN controller is placed on the high-orbit satellite, and its main function is to collect logs and control global routing. In the simulation, interest packets are sent simultaneously by 16 ground stations, while data packets are transmitted in low-orbit satellite nodes, and the ground station is responsible for the last hop. take over.

卫星网络中的内容总请求根据Zipf分布函数

进行建模。其中P(r)为资源r的被请求频率，α为Zipf分布参数，C是内容被请求的次数。设计使用的100个资源文件分别放置于各低轨卫星网络节点，单个资源文件的大小为1-10MB范围内的随机固定值，全部资源文件总大小为800MB。为探究卫星节点缓存能力对缓存策略性能的影响，节点缓存能力的取值为50-300MB。同时，我们通过实验观察了不同Zipf分布指数对缓存命中率的影响，Zipf分布指数的取值为0.8-1.3，默认值为1。同时，本发明使用兴趣包发送频率模拟网络负载对缓存命中率的影响，网络链路带宽设置为20Mbps，兴趣包请求频率在10-100个/秒范围内变化，默认值为50.具体实验参数的设置如表2所示：Total requests for content in the satellite network according to the Zipf distribution function

model. where P(r) is the requested frequency of resource r, α is the Zipf distribution parameter, and C is the number of times the content is requested. The 100 resource files used in the design are placed in each low-orbit satellite network node. The size of a single resource file is a random fixed value in the range of 1-10MB, and the total size of all resource files is 800MB. In order to explore the impact of satellite node caching capability on the performance of the caching strategy, the node caching capability is 50-300MB. At the same time, we observed the impact of different Zipf distribution indices on the cache hit rate through experiments. The value of the Zipf distribution index is 0.8-1.3, and the default value is 1. At the same time, the present invention uses the sending frequency of Interest packets to simulate the impact of network load on the cache hit rate, the network link bandwidth is set to 20Mbps, the interest packet request frequency varies within the range of 10-100/second, and the default value is 50. Specific experimental parameters The settings are shown in Table 2:

表2仿真参数设置Table 2 Simulation parameter settings

实验平台搭建完成后，在实验环境中生成符合Zipf分布，且参数α＝1的500个请求。将单一节点缓存大小设置为200MB，使用得到的日志文件对节点内容缓存进行数据特征选取和数据集构建。After the experimental platform is built, 500 requests that conform to the Zipf distribution and parameter α=1 are generated in the experimental environment. Set the single node cache size to 200MB, and use the obtained log files to perform data feature selection and data set construction for the node content cache.

2、实验评估2. Experimental evaluation

1)比较AOPSO算法的优化性能。1) Compare the optimization performance of AOPSO algorithm.

使用三种寻优算法与本发明使用的aopso优化算法比较，分别为网格法，遗传算法，粒子群算法，通过表3可以看出，与GA，PSO相比，使用本发明AOPSO算法进行本模型的参数寻优时，适应度的遍历范围更广，在有限的进化代数内遍历了更多解空间，耗时少，且更加接近最优解。由于参数C代表的惩罚参数决定了分类器模型的泛化能力，过小的C将带来的结果是支持向量的宽度不足，导致分类器泛化能力不足，过大则会造成过拟合，进行数据判别时不能有效识别靠近支持向量的数据。两种情况都会导致分类精确度降低。综上，本实验提出的粒子群优化算法更适合的本离散模型的参数求解，具体对比见表3：Compared with the aopso optimization algorithm used in the present invention, three kinds of optimization algorithms are used, namely grid method, genetic algorithm, particle swarm optimization algorithm. It can be seen from Table 3 that compared with GA and PSO, the AOPSO algorithm of the present invention is used for this optimization. When the parameters of the model are optimized, the traversal range of fitness is wider, and more solution space is traversed within a limited evolutionary algebra, which takes less time and is closer to the optimal solution. Since the penalty parameter represented by parameter C determines the generalization ability of the classifier model, if C is too small, the result is that the width of the support vector is insufficient, resulting in insufficient generalization ability of the classifier, while too large C will cause overfitting. Data close to the support vector cannot be effectively identified when performing data discrimination. Both cases can lead to lower classification accuracy. In summary, the particle swarm optimization algorithm proposed in this experiment is more suitable for solving the parameters of this discrete model. The specific comparison is shown in Table 3:

表3四种算法生成模型参数对比Table 3 Comparison of parameters of four algorithms to generate models

2)比较卫星网络环境下的数据传输时延2) Compare the data transmission delay in the satellite network environment

为了与本发明中的SDN-SVM缓存策略进行对比，选择四种缓存策略进行对比，选择向下缓存策略LCD作为独立缓存策略，选择Prob作为协作缓存中概率模型的经典方案，CRCache作为协作缓存中考虑内容流行度的典型算法，LCE则作为基本通用方案。四种缓存策略的替换方案都选择LRU最近最少使用算法。In order to compare with the SDN-SVM caching strategy in the present invention, four kinds of caching strategies are selected for comparison, the downward caching strategy LCD is selected as the independent caching strategy, Prob is selected as the classical scheme of the probability model in the cooperative cache, and CRCache is selected as the Considering the typical algorithm of content popularity, LCE is used as a basic general scheme. The alternatives of the four caching strategies all choose the LRU least recently used algorithm.

为了探究卫星节点动态性对算法性能的影响，记录五种算法在一个完整仿真周期内，兴趣包发送频率为20个/秒，其它参数均为默认时，共计2400个兴趣包获取数据包的传输时延。In order to explore the impact of satellite node dynamics on the performance of the algorithm, five algorithms were recorded in a complete simulation cycle, and the sending frequency of Interest packets was 20 per second. When other parameters were default, a total of 2400 Interest packets were obtained to obtain the transmission of data packets. time delay.

实验结果如图5所示，对其进行数值分析可以看出，各卫星节点开始进行网络内缓存后，随着仿真时间的增加，数据传输的平均时延稳步降低，在仿真的中后期，平均时延保持基本稳定，这是由于卫星节点缓存已满，网络内缓存对数据传输时延的优化能力已达阈值。与此同时，由于卫星周期性进行拓扑切换，导致上一拓扑的热点资源大面积失效，缓存新的热点资源存在滞后性，使CRCache算法在拓扑切换的节点处表现不佳。本发明的SDN-SVM算法由于引入了最大存活时间概念，在拓扑切换后可以更快的替换掉非热点资源，使卫星拓扑切换造成的震荡较小，平均时延可以稳定维持在较低位置。平均时延计算结果见表4，由表可知，SDN-SVM算法的平均时延最低。从上述分析中得到的比较结果是，SDN-SVM算法在卫星网络中运行时，可将平均数据传输时延保持在较低位置，且相比其它算法稳定性更高。The experimental results are shown in Figure 5. The numerical analysis shows that after each satellite node starts to cache in the network, with the increase of simulation time, the average delay of data transmission decreases steadily. The delay remains basically stable. This is because the satellite node cache is full, and the network cache's ability to optimize the data transmission delay has reached the threshold. At the same time, due to the periodic topology switching of satellites, the hotspot resources of the previous topology are invalid in a large area, and there is a lag in caching new hotspot resources, which makes the CRCache algorithm perform poorly at the nodes of topology switching. Since the SDN-SVM algorithm of the present invention introduces the concept of maximum survival time, non-hotspot resources can be replaced more quickly after topology switching, so that the vibration caused by satellite topology switching is small, and the average delay can be stably maintained at a low position. The average delay calculation results are shown in Table 4. It can be seen from the table that the average delay of the SDN-SVM algorithm is the lowest. The comparison result obtained from the above analysis is that when the SDN-SVM algorithm runs in the satellite network, the average data transmission delay can be kept at a low position, and it is more stable than other algorithms.

表4五种算法的在单个仿真周期内的平均时延Table 4. Average delay of five algorithms in a single simulation cycle

综上，本发明提出的天地一体化卫星网络节点的智能缓存策略，通过集中式SDN控制器，对卫星网络中资源文件的多维特征进行记录，使用支持向量机建立分类预测模型，从全局角度做出缓存决策，利用SDN的优势，为ICN提供全局缓存决策方案。同时为了适应卫星网络的动态变化，设计一种新的缓存替换策略，显著提升了缓存空间的利用效率。实验结果表明该技术能够对卫星节点的缓存情况进行准确预测，且该分类模型有着很好的稳定性和准确率，显著提高了卫星节点的缓存空间利用率，降低了数据传输延迟。To sum up, the intelligent caching strategy of the satellite network nodes of the space-ground integration proposed by the present invention records the multi-dimensional features of the resource files in the satellite network through the centralized SDN controller, and uses the support vector machine to establish a classification prediction model, which is done from a global perspective. Make caching decisions and take advantage of SDN to provide a global caching decision solution for ICN. At the same time, in order to adapt to the dynamic changes of the satellite network, a new cache replacement strategy is designed, which significantly improves the utilization efficiency of the cache space. The experimental results show that the technology can accurately predict the cache situation of satellite nodes, and the classification model has good stability and accuracy, which significantly improves the cache space utilization of satellite nodes and reduces data transmission delay.

以上所述，仅为本发明创造较佳的具体实施方式，但本发明创造的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明创造披露的技术范围内，根据本发明创造的技术方案及其发明构思加以等同替换或改变，都应涵盖在本发明创造的保护范围之内。The above is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. The equivalent replacement or modification of the created technical solution and its inventive concept shall be included within the protection scope of the present invention.