Method for adaptive dynamic deployment of virtual network elements based on isomorphic network element service demand characterization

Abstract

The application discloses a heterogeneous network element service demand representation and adaptive dynamic deployment method of virtual network elements, which comprises the following steps: obtaining historical data set information, and generating classification sample vector, service representation vector and resource use vector; performing fine classification based on a KNN model of double linear interpolation distance weighting, and outputting classification sample category data; predicting initial network resource demand based on a neural network model of an evolution algorithm; training an attribute prediction model for all virtual network elements of a given service based on a weight improved random forest regression model, and constructing a network resource demand distribution atlas; constructing nodes and links of the network resource demand distribution atlas based on a virtual network mapping algorithm of a particle swarm optimization, outputting an optimal virtual network mapping scheme from the mapping model of the two-stage underlying physical network, and optimizing dynamic migration of the virtual network elements based on a load balancing migration strategy. Through the application, the virtual network elements can reduce invalid migration.

Description

Technical Field

[0001] This invention relates to the field of information network technology, specifically to a method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements. Background Technology

[0002] Network function virtualization (NFV) is an important research direction in the future of networking. NFV technology integrates network function software onto industry-standard server, switch, and storage hardware, replacing traditional dedicated physical devices, thereby reducing costs, energy consumption, and complexity. NFV also possesses capabilities such as software and hardware decoupling, flexible network function deployment, and dynamic service provisioning.

[0003] The key method for implementing NFV is virtual machine placement (VMP). VMP refers to the process of allocating virtual machines (VMs) to physical servers (PSs). It can be understood as the process of using virtualization technology in a cloud data center to abstract the resources of the PS into corresponding virtual resources, which are then allocated to independent VMs.

[0004] Dynamic virtual machine migration enables the online migration of virtual machines between different hosts, offering numerous benefits, especially when hosts are under low or overloaded conditions. Therefore, migration operations make resource management within the data center more flexible. However, online virtual machine migration can negatively impact the performance of applications running on these virtual machines. Since providing adequate quality of service between cloud service providers and their users is crucial, dynamic virtual machine migration should prioritize optimizing the number of migrations.

[0005] Different deployment strategies during virtual machine (VM) deployment can lead to VM requests being mapped to different Service Level Servers (PSs), directly impacting VM performance, system resource utilization, load balancing, and other metrics. Typically, VM deployment results are measured by whether the corresponding Service Level Agreement (SLA) is met. Furthermore, excessive resource consumption can lead to SLA violations and PS load imbalance, while low resource utilization results in significant resource waste and energy consumption. Therefore, appropriate virtual network element (VM) deployment methods are necessary. Currently, most VM deployment algorithms can only optimize for a specific target. Secondly, due to differences in resource attributes between PSs, not all PSs can satisfy a particular VM resource request. A good deployment algorithm can more comprehensively characterize the effectiveness of virtual network mapping, thereby improving operator revenue while reducing mapping costs. Therefore, current VM deployment algorithms require further optimization.

[0006] In existing technologies, Hu Ronghui and Wang Ruitong, Research on Virtual Machine Cluster Migration Strategy in Cloud Environment [J]. Computer Technology and Development, 2017, 27(11):4, although it improves system resource utilization, reduces migration overhead, achieves the goal of reducing power consumption, and has some improvement in reducing SLA defaults, it also avoids the clustering effect caused by the migration of most virtual machines to the same server. However, the following technical problems still exist:

[0007] (1) Only the migration scheduling order was considered without quantifying the specific resource requirements of the virtual network element, and the final migration scheduling scheme still has room for improvement.

[0008] (2) Only the initial resource requirements of virtual network elements are considered, and the dynamic scaling of virtual network element resource requirements is rarely considered. In actual deployment, the resource usage during business operation is not stable. Summary of the Invention

[0009] The technical problem to be solved by this invention is to address the fact that existing technologies do not take into account the differences in resource attributes between different physical machines and the changes in network in a virtual machine deployment environment.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] A method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements includes the following steps:

[0012] S100: Obtain historical dataset information and generate classification sample vectors, business representation vectors, and resource usage vectors;

[0013] S200, the classification sample vector is simplified, and then the simplified classification sample vector is refined by a KNN model based on bilinear interpolation distance weighting, and the classification sample category data is output;

[0014] S300, based on the business representation vector, combined with the classification sample category data, and based on the neural network model of the evolutionary algorithm, predict the initial network resource requirements;

[0015] S400, based on the resource usage vector and the classification sample category data, and combined with the weight-based improved random forest regression model, for a given service, an attribute prediction model is trained for all its virtual network elements, and the attribute prediction results of each attribute prediction model are aggregated to form a dynamic attribute view of the virtual network elements of the entire network service.

[0016] S500 uses deep packet inspection to extract feature information based on network resource requirements, and constructs a network resource requirement distribution map based on transfer learning under reinforcement learning.

[0017] S600, based on the particle swarm optimization virtual network mapping algorithm, constructs the nodes and links of the network resource demand distribution map, and the underlying physical network mapping model of the two stages, and outputs the optimal virtual network mapping scheme.

[0018] S700 optimizes the dynamic migration of virtual network elements based on a load-balancing migration strategy.

[0019] Advantages: By employing a KNN model based on bilinear interpolation distance weighting for refined classification, it addresses the differences in resource attributes between different physical machines, thus satisfying the resource requests of specific virtual network elements. Through dynamic resource usage vectors generated from historical datasets, combined with a weighted improved random forest regression model, an attribute prediction model is trained for all virtual network elements for a given service, providing a basis for dynamic adjustment of resource allocation methods.

[0020] In one embodiment of the present invention, the step of outputting the classified sample category data includes the following steps:

[0021] S210, In the reduced classification sample vector, obtain the difference in the number of samples based on the number of samples in the class with the most samples and the number of samples in each minority class, and determine the number of interpolations required for each minority class.

[0022] S220, for each minority sample class, obtain the sum of the Euclidean distances from each sample to each non-sample in that class, and the mean of the Euclidean distances;

[0023] S230, for each minority sample class in step S220, obtain the probability that each sample will be sampled;

[0024] S240, each time a portion of reference samples are selected from the sample set of the minority sample class, the interpolation position for interpolating the minority sample class is obtained according to the bilinear interpolation method, and the reference samples are interpolated at the interpolation position in combination with the interpolation number.

[0025] S250, After interpolating the minority sample classes, the distance between the test sample used for classification and the training samples in the training set for training the KNN model is obtained, and the samples are sorted in ascending order of distance. The first k samples are selected and their distance weights are obtained.

[0026] S260, using a classification voting function with the distance weights, the distance weights are obtained and categorized by category, the category with the largest superimposed distance weight is selected, and the classification category is output;

[0027] The process of obtaining the probability of each sample being sampled includes the following steps:

[0028] S231, obtain the first number of samples in a minority sample class whose Euclidean distance to other samples in the same class is less than the mean distance;

[0029] S232, obtain the number of second samples in the minority sample class whose Euclidean distance to non-samples is less than the mean distance;

[0030] S233, based on the first number of samples, the second number of samples, the weight of samples in this class and the weight of samples outside this class, obtain the sampling weight of samples in the minority sample class;

[0031] S234, normalize the sampling weights to obtain the probability of each sample in the minority sample class being sampled.

[0032] In one embodiment of the present invention, the interpolation position is obtained by the following formula:

[0033]

[0034] In the formula, RSS ii d represents the interpolation position of the ii-th degree. n P represents the physical distance from the sampled sample to the nth reference sample. iR Let d1, d2, d3, and d4 represent the feature values ​​of the nth reference sample corresponding to the sampled sample in the ii-th interpolation, and let d1, d2, d3, and d4 represent the Euclidean distances from the sampled sample in the ii-th interpolation to the 1st, 2nd, 3rd, and 4th reference samples, respectively.

[0035] The distance weight is obtained using the following formula:

[0036]

[0037] In the formula, W iii Represented as distance weight, Td iii Td represents the distance between the sample to be tested and its i-th nearest neighbor, Td1 represents the distance among the k nearest neighbors, and Td represents the distance between the k nearest neighbors. k It is represented as the farthest distance among the k nearest neighbors;

[0038] The classification categories are obtained using the following formula:

[0039]

[0040] In the formula, c x Let V represent the category corresponding to the sample x to be tested, V represent the current set of samples to be tested, L represent the set of all sample categories, and N represent the set of all sample categories. k (X) represents the set of the k nearest neighbors of sample X in the training set, class(c xiii ) represents the training sample x iii The category, I(V = class(c xiii )) represents an exponential function that returns 1 when it is true and 0 otherwise.

[0041] In one embodiment of the present invention, the prediction of initial network resource requirements includes:

[0042] S310, Based on the service representation vector, construct a service dataset sample based on "output service quality performance - actual resource usage"; and divide the service dataset sample into a service data training set and a service data test set according to a certain ratio, and obtain the optimal weights and thresholds of the neural network through training iteration;

[0043] S320, an improved evolutionary algorithm is used to optimize the neural network, a neural network model based on the evolutionary algorithm is built, and the neural network model based on the evolutionary algorithm is trained and tested using the business data training set and the business data test set; and the neural network model based on the evolutionary algorithm after training and testing predicts the initial network resource requirements based on the business representation vector and the classification sample category data.

[0044] In the improved evolutionary algorithm, the reciprocal of the squared error is used as the fitness function to measure the fitness of individuals in the neural network population. The formula is as follows:

[0045]

[0046]

[0047] In the formula, E(jj) represents the error of the jj-th iteration. Let y represent the overall output of the (mm-1)th neuron in the jj-th iteration, w represent the weight vector, u represent the input vector, F(jj) represent the fitness function, and y represent the weight vector. jj Represented as the theoretical output value;

[0048] In the improved evolutionary algorithm, the individuals in the population are reordered to obtain the probability of an individual being selected after the reordering. This probability is obtained using the following formula:

[0049]

[0050] p = s(1 - p0) b-1 ;

[0051] In the formula, p0 represents the probability of the optimal individual being selected, s represents the value of the optimal individual after standardization, p represents the probability of an individual being selected in the reordered population, b represents the position of the N'th individual after the population is reordered, and a represents the population size in the evolutionary algorithm.

[0052] In the improved evolutionary algorithm, the change in root fitness during iteration adjusts the change in crossover probability, which is then obtained using the following formula:

[0053]

[0054] In the formula, P j Let P be the crossover probability. jmax Let P be the maximum crossover probability. jmin Let F represent the minimum crossover probability, and let F represent the maximum fitness of the two individuals undergoing crossover in the population. mean n represents the average fitness of the entire population. max n' represents the maximum number of iterations of the evolution operator during its operation; n' represents the number of iterations of the evolution algorithm during its current operation.

[0055] In one embodiment of the present invention, in step S400, during the training phase of the decision tree of the improved random forest regression model, the prediction performance of each decision tree is evaluated, and higher weights are assigned to decision trees that can accurately predict a small number of samples. The final dynamic attribute prediction results of the virtual network elements are obtained through weighted voting. Based on the dynamic attribute prediction results of the virtual network elements, the regression analysis method is used to aggregate different attributes to form a dynamic attribute view of the virtual network elements for the entire network service.

[0056] The dynamic attribute prediction result of the virtual network element is obtained through the following formula:

[0057]

[0058] In the formula, H(T) represents the dynamic attribute prediction result of the virtual network element, T is the number of decision trees, I(·) is the indicator function, and f tt (T) represents the prediction result of the t-th decision tree. The operator represents a comparison of the values ​​on both sides, where 'y' represents the actual value and 'ω' represents the true value. tt Let t be the voting weight of the t-th decision tree. The value of the indicator function is 1 when the prediction result of the decision tree is true, and 0 otherwise.

[0059] In one embodiment of the present invention, the output of the optimal virtual network mapping scheme includes:

[0060] S610 establishes multi-dimensional constraints on business type, time validity, and business priority;

[0061] S620 sets the particle swarm size N' and the maximum number of iterations MG for the algorithm; the initial positions X of the particles are randomly generated. i With speed V i ;

[0062] S630, calculate the fitness f(X) of all particles. i ), thus obtaining the globally optimal initial position X gb With the optimal initial position X of each individual particle pb ;

[0063] S640, For particles that satisfy the multidimensional constraints, perform velocity and position updates, and randomly select candidate nodes of the bottom network during position updates; for particles that do not satisfy the multidimensional constraints, regenerate their position and velocity parameters.

[0064] S650, for each particle in the particle swarm, if f(X) i ) <f(X pb ), then X pb =X i If f(X) pb ) <f(X gb ), then X gb =X pb ;

[0065] S660, check the current iteration number. If it is less than the maximum iteration number MG, proceed to step S640; otherwise, proceed to step S670.

[0066] S670 outputs the optimal virtual network mapping scheme and its fitness value;

[0067] Where, f(X) pb f(X) represents the optimal initial position for each individual particle.gb ) represents the fitness of the globally optimal initial position.

[0068] In one embodiment of the present invention, the position and velocity of particles in a particle swarm are defined as follows:

[0069] Definition 1, Particle position: The position of a particle Defined as the i-th mapping scheme, D represents the number of virtual network nodes contained in the virtual network request. Take a positive integer, whose value represents the number of the underlying network node selected by the j-th virtual node from the list of candidate nodes in the underlying network, j = 1, 2, 3, ..., D;

[0070] Definition 2, Particle velocity: The velocity of a particle The adjustment decision is defined as the i-th mapping scheme, where, This represents the node mapping selected by the j-th virtual node from the list of candidate nodes in the underlying network;

[0071] Definition 3, Subtraction Θ, is used to obtain the difference between two mapping schemes;

[0072] Definition 4, Addition Used to obtain adjustment decisions for the mapping scheme;

[0073] Definition 5, Multiplication Used to obtain a new mapping scheme;

[0074] The position and velocity update formulas for the particle swarm optimization algorithm after definition are as follows:

[0075]

[0076]

[0077] In the formula, V i+1 Defined as the adjustment decision for the (i+1)th mapping scheme; X i+1 Defined as the (i+1)th mapping scheme; X gb Defined as the best historical location in the neighborhood; X pb Defined as its own historical best position, P1, P2 and P3 are constants, and P1+P2+P3=1.

[0078] In one embodiment of the present invention, step S700 includes:

[0079] The S710 monitors the global load and the proportion of high and low load nodes, dynamically adjusts status thresholds, assesses the status of node resources, and determines the timing of migration.

[0080] S720 selects virtual machines to be migrated based on the dependency of virtual machines and node resources and the contribution of virtual machines.

[0081] In one embodiment of the present invention, the dynamic adjustment of the state threshold includes:

[0082] Set the high load ratio range (minLr) h ,maxLr h ) and low load ratio range (minLr) l ,maxLr l This is used to determine when to adjust the state threshold and when to turn nodes on or off;

[0083] When global load When, if minLr l <Lr l <maxLr l or minLr h <Lr h <maxLr h At this point, the state threshold needs to be adjusted, and the adjustment formula is as follows:

[0084]

[0085] In the formula, Thr h Represented as the high-order threshold; Thr l Represented as the low-order threshold; Thr′ h Thr′ l These represent the high-order and low-order thresholds before adjustment, respectively; τ h and τ l These represent the adjustment factors for the high and low thresholds, respectively; Ld dc Represented as global load; Lr h This indicates a high load ratio; Lr l This indicates a low load ratio; (ω) l ,ω h This represents the load balancing range generated based on the historical load of the data center during this period.

[0086] When Ld dc <ω l And Lr l >maxLr l When the data center load is low, it is necessary to select some lightly loaded nodes that need to be migrated. After the lightly loaded nodes complete the migration operation, the nodes will be shut down.

[0087] When Ld dc >ω h And Lr h >maxLr h When the data center load is high, start a new node.

[0088] When assessing the status of node resources, the product exponential weighting method is used to evaluate the load over m periods. The formula for the product exponential weighting method is as follows:

[0089]

[0090] In the formula, x t For the load in period t, w t y represents the weighting coefficient for option t. If the load is overloaded, then T1 is the number of periods when the load exceeds the high threshold, and T2 is the number of periods when the load is below the high threshold. If the load is lightly loaded, then T1 is the number of periods when the load value is below the low threshold, and T2 is the number of periods when the load value is above the low threshold. y represents the state of resource saving.

[0091] In one embodiment of the present invention, determining the migration timing includes:

[0092] S711, collect load information of each node, obtain the global load degree and the proportion of high and low load nodes within the monitoring period, and determine whether the status threshold needs to be updated. If yes, proceed to step S712; otherwise, proceed to step S713.

[0093] S712 analyzes the global load and the proportion of high and low load nodes, and adjusts and updates the status thresholds.

[0094] S713, Based on the status threshold, determine the load status of each node; if the node load index is higher than the high threshold, enable the overload assessment of the node load index, and proceed to step S714; if the node load is lower than the low threshold, enable the light load assessment of the node load, and proceed to step S715; if neither the low threshold nor the high threshold is met, the node is considered a suitable load node, and proceed to step S717.

[0095] S714, query the load index of the node that exceeds the high-level threshold for m periods and calculate its evaluation value. If the evaluation value reaches the overload requirement, determine that the node is overloaded for this index type, and update the high-load node matrix accordingly; go to step S713 and continue the loop.

[0096] S715, query the load of nodes with load below the low threshold for m periods and calculate their evaluation values. If the evaluation values ​​within m periods all meet the light load requirement, then determine that the node is a light load node, add it to the list of low load nodes, go to step S713, and continue the loop.

[0097] S716 sorts the list of low-load nodes in ascending order of load level, combines it with the global load level, and determines whether it is within the load balancing range. If Ld dc <ω l Then calculate the set of lightly loaded nodes to be migrated, if Ld dc >ω h And Lr h >maxLrh If so, then a node will be enabled;

[0098] S717 outputs the high-load node matrix and the low-load node matrix.

[0099] Compared with existing technologies, the beneficial effects of this invention are as follows: Considering the differences in resource attributes among different physical machines, not all physical machines can meet the resource requests of specific virtual network elements. Therefore, a refined service classification is performed using a KNN model based on bilinear interpolation distance weighting. Then, an initial network resource demand model is generated using a neural network model based on an evolutionary algorithm. Considering the changes in the network under real-world virtual machine deployment environments, and addressing the dynamic scaling of resource usage during actual operation, a dynamic attribute prediction model for virtual network elements is generated using a weighted improved random forest regression model, providing a basis for the dynamic adjustment of resource allocation methods. Attached Figure Description

[0100] Figure 1 This is a flowchart illustrating a method for characterizing heterogeneous network element service requirements and adaptively deploying virtual network elements according to an embodiment of the present invention.

[0101] Figure 2 This is a flowchart illustrating the output classification sample category data in an embodiment of the present invention.

[0102] Figure 3 This is a flowchart illustrating the process of obtaining the probability of each sample being sampled, according to an embodiment of the present invention.

[0103] Figure 4 This is a flowchart illustrating the prediction of initial network resource requirements according to an embodiment of the present invention.

[0104] Figure 5 This is a flowchart illustrating the optimal virtual network mapping scheme in an embodiment of the present invention.

[0105] Figure 6 This is a flowchart illustrating the optimized dynamic migration of virtual network elements according to an embodiment of the present invention. Detailed Implementation

[0106] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0107] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0108] Please see Figure 1 As shown, this invention provides a method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements, characterized by the following steps:

[0109] S100: Obtain historical dataset information and generate classification sample vectors, business representation vectors, and resource usage vectors.

[0110] Specifically, by analyzing statistical characteristics of historical service traffic, such as the number of DNS (Domain Name System, DNS) and NTP (Network Time Protocol, NTP) requests, packet length, number of packets, and time intervals, a vector space model of historical data is generated as the classification sample vector. From the service quality requirements and resource requirements of the service on the access network, transmission network, and core network sides, a multi-dimensional vector representing the service is obtained, namely the service representation vector. The resource usage vector is a dynamic vector of resource usage generated from historical dataset information due to network fluctuations and other factors; specifically, it includes service requirement characteristics, virtual network element attributes, and historical resource usage of virtual network elements—dynamic attributes of virtual network elements.

[0111] S200, the classification sample vector is simplified, and then the simplified classification sample vector is refined by a KNN model based on bilinear interpolation distance weighting, and the classification sample category data is output.

[0112] Among these methods, a rough set reduction approach based on heuristic information is used to reduce the classification sample vector. Multi-label learning represents the ability of an attribute set to clearly classify samples into each label category through a lower approximation of neighborhood rough sets. Considering the curse of dimensionality caused by complex samples and high information redundancy, reduction of the classification sample vector is necessary. An extended neighborhood rough set feature selection method is applied to multi-label data. Specifically, consider a reduction of an attribute of A. If adding feature A' to attribute set B does not change the dependency of the decision attribute on conditional attribute B (i.e., the importance of feature A' is 0), then A' can be removed from B. Based on this, a forward greedy reduction algorithm iteratively selects the feature with the highest importance, terminating when the addition of the attribute no longer affects classification accuracy.

[0113] Please see Figures 1 to 3 As shown, in one embodiment of the present invention, this step uses a KNN model based on bilinear interpolation distance weighting to perform refined classification of services, that is, the KNN algorithm based on bilinear interpolation distance weighting. The KNN algorithm is the K-Nearest Neighbor classification algorithm.

[0114] Traditional KNN models may result in a situation where features from other categories appear frequently in the nearest neighbor range, while features from the same category appear less frequently, leading to classification errors. Therefore, a KNN model based on bilinear interpolation distance weighting can compensate for sample identification errors caused by error disturbances during data acquisition. Specifically, the process of outputting the classified sample category data includes the following steps:

[0115] S210, in the reduced classification sample vector, obtain the sample number difference based on the sample number of the class with the most samples and the sample number of each minority sample class, and determine the number of interpolations required for each minority sample class.

[0116] S220, for each minority sample class, obtain the sum of the Euclidean distances from each sample to each non-sample in that class, and the mean of the Euclidean distances.

[0117] In step S220, the sum of the Euclidean distances is obtained using the following formula:

[0118]

[0119] In step S220, the mean value of the Euclidean distance is obtained using the following formula:

[0120]

[0121] In the formula, x i' Let represent the i'th sample in the minority class, n" represent the number of samples in the minority class, and m' represent the number of samples in the non-minority class. i' E is expressed as the sum of Euclidean distances. mean Represented as the distance mean, y j' Let j' be the j'th sample of the non-minority sample class;

[0122] S230, for each minority sample class in step S220, obtain the probability that each sample will be sampled.

[0123] The process of obtaining the probability of each sample being sampled includes the following steps:

[0124] S231, obtain the first number of samples in the minority sample class whose Euclidean distance to other samples in the same class is less than the mean distance.

[0125] S232, obtain the second number of samples in the minority sample class whose Euclidean distance to non-samples is less than the mean distance.

[0126] S233, based on the first number of samples, the second number of samples, the weight of samples in this class and the weight of samples outside this class, obtain the sampling weight of samples in the minority sample class.

[0127] The sampling weight is obtained using the following formula:

[0128] P' i =k i1 ×w i1 +k i2 ×w i2 ;

[0129] In the formula, P' i Represented as sampling weight, k i1 Let k represent the first sample size. i2 Let w represent the second sample size. i1 w represents the weight of this class of samples. i2 This represents the weight of samples not belonging to this class. In this embodiment, w i1 Set to 0.3, w i2 Set it to 0.7.

[0130] S234, normalize the sampling weights to obtain the probability of each sample in the minority sample class being sampled.

[0131] S240, each time a portion of reference samples are selected from the sample set of the minority sample class, the interpolation position for interpolating the minority sample class is obtained according to the bilinear interpolation method, and the reference samples are interpolated at the interpolation position in combination with the interpolation number.

[0132] The interpolation position is obtained using the following formula:

[0133]

[0134] In the formula, RSS ii d represents the interpolation position of the ii-th degree. n P represents the physical distance from the sampled sample to the nth reference sample. iR Let d1, d2, d3, and d4 represent the feature values ​​of the nth reference sample corresponding to the sampled sample in the ii-th interpolation, and let d1, d2, d3, and d4 represent the Euclidean distances from the sampled sample in the ii-th interpolation to the 1st, 2nd, 3rd, and 4th reference samples, respectively. In this embodiment, the four similar samples with the closest Euclidean distance are used as reference samples.

[0135] S250, after interpolating the minority sample classes, obtain the distance between the test sample used for classification and the training samples in the training set for training the KNN model, sort them in ascending order of distance, select the first k samples, and obtain their distance weights.

[0136] The distance weight is obtained using the following formula:

[0137]

[0138] In the formula, W iii Represented as distance weight, Td iii Td represents the distance between the sample to be tested and its i-th nearest neighbor, Td1 represents the distance among the k nearest neighbors, and Td represents the distance between the k nearest neighbors. k It is represented as the farthest distance among the k nearest neighbors.

[0139] S260, using a classification voting function with the distance weights, the distance weights are obtained and categorized by category, the category with the largest superimposed distance weight is selected, and the classification category is output;

[0140] The classification categories are obtained using the following formula:

[0141]

[0142] In the formula, c x Let V represent the category corresponding to the sample x to be tested, V represent the current set of samples to be tested, L represent the set of all sample categories, and N represent the set of all sample categories. k (X) represents the set of the k nearest neighbors of sample X in the training set. Let x be the training sample iii Category This is represented as an exponential function, returning 1 when its value is true and 0 otherwise. That is, it calculates the class to which the sample belongs if the majority of the selected K nearest neighbors belong. The output classification is the class output by the KNN model based on bilinear interpolation distance weighting. `arg max` is represented as a function.

[0143] S300, based on the business representation vector, combined with the classification sample category data, and a neural network model based on the evolutionary algorithm, predict the initial network resource requirements.

[0144] Please see Figure 1 and Figure 4 As shown, in one embodiment of the present invention, by combining machine learning methods to perform feature mining on historical resource consumption datasets, a mathematical expression of business resource requirements is characterized, and a representation system for differentiated business requirements is established. In the initial stage of deploying space-ground integrated network services, resources are reserved according to the predefined resource requirements of each type of service. Network resources are reserved based on the resource requirement characteristics of the service quality requirements of the target business, avoiding resource waste or service performance degradation caused by excessive or insufficient resource reservations. The mathematical expression of the business's network resource requirements is transformed into a multivariate nonlinear regression analysis problem. Based on an evolutionary algorithm-based neural network model, the complex nonlinear relationship between service quality requirements and network resource requirements is simulated. Specifically, the prediction of initial network resource requirements includes:

[0145] S310, based on the service representation vector, construct a service dataset sample based on "output service quality performance - actual resource usage"; and divide the service dataset sample into a service data training set and a service data test set according to a certain ratio, and obtain the optimal weights and thresholds of the neural network through training iterations. Obtaining the optimal weights and thresholds of the neural network through training iterations improves the prediction accuracy of the neural network.

[0146] S320, an improved evolutionary algorithm is used to optimize the neural network, a neural network model based on the evolutionary algorithm is built, and the neural network model based on the evolutionary algorithm is trained and tested using the business data training set and the business data test set; and the neural network model based on the evolved algorithm after training and testing predicts the initial network resource requirements based on the business representation vector and the classification sample category data.

[0147] ① Initialize the population: Randomly generate a certain number of individuals as the initial population, usually by randomly generating some parameters or instances.

[0148] ②Evaluate fitness: For each individual, evaluate it according to the specific objective function or performance index of the problem it represents to obtain its fitness value.

[0149] ③ Selecting individuals: Individuals in the population are selected based on their fitness values. Various selection algorithms are usually used, such as roulette wheel selection, competitive selection, and sorting selection.

[0150] ④ Mutation operation: By performing gene mutation operations on certain individuals, a certain degree of randomness and diversity is introduced to avoid getting trapped in local optima.

[0151] ⑤ Crossover operation: For the selected individuals, crossover operation is used to generate new offspring individuals, thereby realizing the exchange and combination of genetic information.

[0152] ⑥ Population renewal: Through operations such as selection, mutation, and crossover, a certain number of offspring individuals are generated and added to the population to renew it.

[0153] ⑦ Termination condition detection: Repeat steps ②-⑥ to check whether the preset termination condition has been met, such as reaching the maximum number of iterations or the objective function value reaching a certain threshold.

[0154] ⑧ Output results: Output the optimal solution or set of optimal solutions, or provide an explanation and visualization of the optimal solution.

[0155] Individual particles are the search particles in the evolutionary algorithm, and the position of a particle represents a solution set, i.e., the weight matrix and bias matrix of the neural network. Through continuous iteration of the evolutionary algorithm, an optimal particle is eventually determined, corresponding to an optimal solution set. The iterative process is crossover and mutation, which is the improved crossover and mutation method described below.

[0156] This step employs an improved evolutionary algorithm to optimize the neural network, constructs a neural network model based on the evolutionary algorithm, and trains and tests the neural network model using the business data training set and the business data test set. The basic deep neural network, on the other hand, is trained using an error backpropagation algorithm. It has a simple structure, high plasticity, and strong data fitting ability. A deep neural network mainly consists of an input layer, hidden layers, and an output layer. During training, the neural network continuously adjusts the weights and thresholds between the input layer and hidden layers, as well as between the hidden layers and the output layer. Training stops when the neural network output value matches the target value or when the number of iterations is reached. This type of neural network has strong generalization ability.

[0157] This step uses the genetic algorithm from evolutionary algorithms to replace the backpropagation (BP) algorithm, optimizing the deep neural network through evolutionary optimization. Evolutionary algorithms are designed based on the evolutionary patterns of organisms in nature. Their working principle involves first encoding the input data, then performing selection, crossover, and mutation operations with certain probabilities until the individual with the highest fitness is selected as the target value output, after which the computation stops.

[0158] In the improved evolutionary algorithm, the reciprocal of the squared error is used as the fitness function to measure the fitness of individuals in the neural network population. The formula is as follows:

[0159]

[0160]

[0161] In the formula, E(jj) represents the error of the jj-th iteration. Let y represent the overall output of the (mm-1)th neuron in the jj-th iteration, w represent the weight vector, u represent the input vector, F(jj) represent the fitness function, and y represent the weight vector. jj This represents the theoretical output value.

[0162] Traditional evolutionary algorithms select individuals from the population randomly during operation, which can easily miss the optimal individual and introduce significant errors in actual computation. Therefore, in this embodiment, the selection operator is improved by reordering the population individuals to obtain the probability of selecting an individual after reordering. The probability of selecting an individual after reordering is obtained using the following formula:

[0163]

[0164] p = s(1 - p0) b-1 ;

[0165] In the formula, p0 represents the probability of the optimal individual being selected, s represents the value of the optimal individual after standardization, p represents the probability of an individual being selected in the population after reordering, b represents the position of the N”th individual after rearranging the population, and a represents the population size in the evolutionary algorithm.

[0166] Traditional evolutionary algorithms typically set the crossover probability to a constant between 0.3 and 0.8 during operation. A high crossover probability enhances the algorithm's global search capability but reduces chromosome fitness, while a low crossover probability decreases both global search capability and convergence speed. In the improved evolutionary algorithm, the crossover operator is modified, and the change in root fitness during iteration adjusts the crossover probability. The crossover probability is then obtained using the following formula:

[0167]

[0168] In the formula, P j Let P be the crossover probability. jmax Let P be the maximum crossover probability. jmin Let F represent the minimum crossover probability, and let F represent the maximum fitness of the two individuals undergoing crossover in the population. mean n represents the average fitness of the entire population. max denoted by , n' represents the maximum number of iterations of the evolution operator during its operation, and n' represents the number of iterations of the evolution algorithm in its current operation. In this embodiment, during the initialization of the evolution algorithm, the maximum crossover probability can be set to 0.8, and the minimum crossover probability can be set to 0.3.

[0169] In this step, the parameters of the deep neural network used are: 5 network layers, 100 hidden layer nodes, learning rate of 0.001, batch size of 64, and number of iterations of 200.

[0170] S400, based on the resource usage vector and the classification sample category data, and combined with the weight-based improved random forest regression model, for a given service, an attribute prediction model is trained for all its virtual network elements, and the attribute prediction results of each attribute prediction model are aggregated to form a dynamic attribute view of the virtual network elements of the entire network service.

[0171] This paper proposes an improved random forest regression model that incorporates business demand characteristics into the selection of optimal feature values ​​for weighted decision trees, thereby enhancing prediction accuracy. Traditional random forests treat all decision trees indiscriminately, resulting in decision trees with varying prediction performance receiving the same voting weight. To improve the prediction ability of random forests for minority class samples, a weighted improved random forest regression model is proposed. During the training phase, the model evaluates the prediction performance of each decision tree and assigns higher weights to decision trees that accurately predict minority samples. The final dynamic attribute prediction results for virtual network elements are obtained through weighted voting. Based on these prediction results, regression analysis is used to aggregate different attributes, forming a dynamic attribute view of virtual network elements across the entire network, providing a basis for dynamic adjustments to resource allocation methods.

[0172] The dynamic attribute prediction result of the virtual network element is obtained through the following formula:

[0173]

[0174] In the formula, H(T) represents the dynamic attribute prediction result of the virtual network element, T is the number of decision trees, I(·) is the indicator function, and f tt (T) represents the prediction result of the t-th decision tree. The operator represents a comparison of the values ​​on both sides, where 'y' represents the actual value and 'ω' represents the true value. tt Let be the voting weight of the t-th decision tree. The indicator function has a value of 1 when the decision tree's prediction is true, and 0 otherwise. In this embodiment, if... If the difference between the two sides, y, is less than 0.5%, then f is determined. tt The predicted value of (T) matches the actual value.

[0175] Let TP represent classifying a stable sample as stable, FN represent classifying a stable sample as unstable, FP represent classifying an unstable sample as stable, and TN represent classifying an unstable sample as unstable. In this embodiment, the accuracy P of each decision tree in predicting unstable samples is used. re and recall rate R ec The harmonic mean F1 is used as the weight of the tree, and the voting weight ω of each tree is used. t Defined as:

[0176]

[0177]

[0178]

[0179] The higher the F1 score, the better the prediction performance of the decision tree for minority class samples.

[0180] S500 uses deep packet inspection to extract feature information based on network resource requirements, and constructs a network resource requirement distribution map based on transfer learning under reinforcement learning.

[0181] In the context of non-stationary transmission environments within integrated space-ground networks, the complex resource demands across the entire network lead to difficulties in network element deployment. Therefore, it is necessary to construct a network resource demand distribution map for real-time monitoring of the global network view. Based on feature data related to network resource demands such as computing resources, routing resources, and port resources, deep packet inspection and other methods are employed to extract feature information. Using transfer learning under reinforcement learning, the network is first trained in a small sample source domain. Then, the weights of each node in the network are fine-tuned layer by layer. Next, some layer parameters are frozen, and finally, local layers and fully connected layers are updated to achieve target domain learning. This improves the algorithm's feature extraction capability under limited sample conditions, enabling dynamic and accurate perception of network resources.

[0182] S600, based on a particle swarm optimization-based virtual network mapping algorithm, constructs the nodes and links of the network resource demand distribution map, and the underlying physical network mapping model of these two stages, outputting the optimal virtual network mapping scheme.

[0183] Please see Figure 1 and Figure 5 As shown, in one embodiment of the present invention, the output of the optimal virtual network mapping scheme includes:

[0184] S610 establishes multi-dimensional constraints on business type, time validity, and business priority.

[0185] In this embodiment, particle swarm optimization (PSO) is used to optimize the virtual network mapping problem, improving the virtual network mapping request acceptance rate and the long-term average operating benefit of the underlying network, while reducing the solution time for virtual network mapping. Specifically, this includes:

[0186] S620 sets the particle swarm size N', the maximum number of iterations MG, and the initial position X of the randomly generated particles. i With speed V i .

[0187] S630, calculate the fitness f(X) of all particles. i ), thus obtaining the globally optimal initial position X gb With the optimal initial position X of each individual particle pb .

[0188] S640: For particles that satisfy the multidimensional constraints, perform velocity and position updates, randomly selecting candidate nodes from the underlying network during position updates. For particles that do not satisfy the multidimensional constraints, regenerate their position and velocity parameters.

[0189] S650, for each particle in the particle swarm, if f(X) i ) <f(X pb ), then X pb =X i If f(X) pb ) <f(X gb ), then X gb =X pb .

[0190] S660, check the current iteration number. If it is less than the maximum iteration number MG, proceed to step S640; otherwise, proceed to step S670.

[0191] S670 outputs the optimal virtual network mapping scheme and its fitness value.

[0192] Where, f(X) pb f(X) represents the optimal initial position for each individual particle. gb ) represents the fitness of the globally optimal initial position.

[0193] In step S600, the particle swarm optimization algorithm denotes the fitness function as f(X), where the position vector X represents a possible mapping scheme. During the calculation of the fitness function, multidimensional constraints are checked first. If the mapping scheme is feasible, the value of f(X) represents the cost of the virtual network mapping. If the mapping scheme is not feasible, the value of f(X) is set to +∞.

[0194] The algorithm utilizes an aggregation strategy to calculate the fitness function of particles and employs linear programming to derive the final mapping scheme. This solves the multi-objective optimization problem under the differentiated service requirements of space-ground integrated networks, optimizing the deployment of virtual network elements under the constraints of service requirements and resource limitations.

[0195] Specifically, the positions and velocities of particles in a particle swarm are defined as follows:

[0196] Definition 1, Particle position: The position of a particle Defined as the i-th mapping scheme, D represents the number of virtual network nodes contained in the virtual network request. Take a positive integer, whose value represents the number of the underlying network node selected by the j-th virtual node from the list of candidate nodes in the underlying network, j = 1, 2, 3, ..., D.

[0197] Definition 2, Particle velocity: The velocity of a particle The adjustment decision is defined as the i-th mapping scheme, where, Let represent the node mapping selected by the j-th virtual node from the candidate node list of the underlying network, where j = 1, 2, 3, ..., D. It is a binary variable, if This indicates that the j-th virtual node needs to reselect a node mapping from its underlying network candidate node list.

[0198] Definition 3, Subtraction Θ, is used to obtain the difference between two mapping schemes. For example, CΘD, if mapping scheme C and mapping scheme D have the same value in the same dimension, then the difference between them is 1, otherwise it is 0.

[0199] Definition 4, Addition Used to obtain adjustment decisions for the mapping scheme. For example, Here, EF and GH represent maintaining the values ​​of each dimension of F with probability E and maintaining the values ​​of each dimension of H with probability G, respectively, and E+G=1, (0≤E≤1), (0≤G≤1).

[0200] Definition 5, Multiplication Used to obtain a new mapping scheme. For example, Then, mapping scheme C adjusts its virtual node mapping scheme according to adjustment decision F.

[0201] The position and velocity update formulas for the particle swarm optimization algorithm after definition are as follows:

[0202]

[0203]

[0204] In the formula, V i+1 Defined as the adjustment decision for the (i+1)th mapping scheme; X i+1 Defined as the (i+1)th mapping scheme; X gb Defined as the best historical location in the neighborhood; X pb Defined as its own historical best position, P1, P2 and P3 are constants, and P1+P2+P3=1.

[0205] X i Defined as the i-th mapping scheme, V i Defined as the adjustment decision for the i-th mapping scheme.

[0206] S700 optimizes the dynamic migration of virtual network elements based on a load-balancing migration strategy.

[0207] Please see Figure 1 and Figure 6As shown, in one embodiment of the present invention, a load-balancing-based virtual network element dynamic migration strategy is used to reduce invalid migrations and improve the load balancing capabilities of the data center while ensuring service quality. The dynamic migration method used in this step is the mainstream memory pre-copy method. Specifically, it includes:

[0208] The S710 monitors the global load and the proportion of high and low load nodes, dynamically adjusts status thresholds, assesses the status of node resources, and determines the timing for migration.

[0209] S720 selects virtual machines to be migrated based on the dependency of virtual machines and node resources and the contribution of virtual machines.

[0210] The dynamically adjusted state threshold includes:

[0211] Set the high load ratio range (minLr) h ,maxLr h ) and low load ratio range (minLr) l ,maxLr l This is used to determine when to adjust the state threshold and when to turn nodes on or off.

[0212] When global load When, if minLr l <Lr l <maxLr l or minLr h <Lr h <maxLr h At this point, the state threshold needs to be adjusted, and the adjustment formula is as follows:

[0213]

[0214] In the formula, Thr h Represented as the high-order threshold, Thr l Represented as the low-order threshold, Thr′ h Thr′ l These represent the high-order and low-order thresholds before adjustment, respectively, τ h and τ l Ld represents the adjustment factors for the high and low thresholds, respectively. dc Represented as global load, Lr h Indicated as a high load ratio, Lr l This indicates a low load ratio. (ω) l ,ω h This represents the load balancing range generated based on the historical load of the data center during this period, within which the data center load is considered appropriate.

[0215] When Ld dc <ωl And Lr l >maxLr l When the data center load is low, it is necessary to select some lightly loaded nodes that need to be migrated. After the lightly loaded nodes complete the migration operation, the nodes will be shut down.

[0216] When Ld dc >ω h And Lr h >maxLr h When the data center is under high load, a new node is started.

[0217] Considering that load fluctuations occur briefly during node load changes, resulting in instantaneous load values ​​that affect the determination of migration trigger timing, a product-exponential weighted method is used to evaluate the load over m periods. Therefore:

[0218] When assessing the status of node resources, the product exponential weighting method is used to evaluate the load over m periods. The formula for the product exponential weighting method is as follows:

[0219]

[0220] In the formula, x t For the load in period t, w t Let T1 be the load factor for option t. If the load is overloaded, then T1 is the number of periods when the load exceeds the high threshold, and T2 is the number of periods when the load is below the high threshold. If the load is lightly loaded, then T1 is the number of periods when the load value is below the low threshold, and T2 is the number of periods when the load value is above the low threshold. y represents the state of resource saving.

[0221] Please see Figure 1 and Figure 6 As shown, in one embodiment of the present invention, determining the migration timing includes:

[0222] S711, collect load information of each node, obtain the global load degree and the proportion of high and low load nodes within the monitoring period, and determine whether the status threshold needs to be updated. If yes, proceed to step S712; otherwise, proceed to step S713.

[0223] S712 analyzes the global load and the proportion of high and low load nodes, and adjusts and updates the status thresholds.

[0224] S713, Based on the status threshold, determine the load status of each node; if the node load index is higher than the high threshold, enable the overload assessment of the node load index and proceed to step S714; if the node load is lower than the low threshold, enable the light load assessment of the node load and proceed to step S715; if neither the low threshold nor the high threshold is met, the node is considered a suitable load node, then proceed to step S717.

[0225] S714, query the load index of the node that exceeds the high-level threshold for m periods and calculate its evaluation value. If the evaluation value reaches the overload requirement, it is determined that the node is overloaded for this index type. Based on the updated high-load node matrix, go to step S713 and continue the loop.

[0226] S715, query the load of nodes with load below the low threshold for m periods and calculate their evaluation values. If the evaluation values ​​within m periods all meet the light load requirement, then determine that the node is a light load node, add it to the list of low load nodes, go to step S713, and continue the loop.

[0227] S716 sorts the list of low-load nodes in ascending order of load level, combines it with the global load level, and determines whether it is within the load balancing range. If Ld dc <ω l Then calculate the set of lightly loaded nodes to be migrated, if Ld dc >ω h And Lr h >maxLr h If so, then a node will be enabled;

[0228] S717 outputs the high-load node matrix and the low-load node matrix.

[0229] In step S720, the selection of virtual machines to be migrated can be described as ensuring that the overloaded resources of a node return to a suitable range after migrating a certain virtual machine, i.e., not exceeding the high threshold of this resource type. The selection of virtual machines to be migrated includes: obtaining information about high-load nodes from the list of high-load nodes obtained from node evaluation; obtaining a set of virtual machines to be screened that can relieve the overload of a node with a single migration, based on the overload type of this node; obtaining the resource dependency of each virtual machine in the set of virtual machines to be screened and sorting them in descending order; obtaining the memory bandwidth ratio and load contribution of each virtual machine; selecting a virtual machine as the virtual machine to be migrated and calculating its resource ratio, adding it to the list of virtual machines to be migrated; and selecting the destination node based on the resource matching degree between the virtual machine and the destination node and the migration cost, thereby achieving dynamic adjustment of virtual machines on high-load and low-load nodes. The selection of the destination node includes: obtaining a list of candidate nodes from the list of suitable nodes obtained from node evaluation, sorting them in ascending order of load; obtaining the candidate node list and calculating the matching degree between the virtual machine and the node, sorting them in descending order; calculating the migration cost according to the matching degree order; and selecting the node with the highest matching degree and the lowest migration cost as the destination node.

[0230] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0231] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements, characterized in that, Includes the following steps: S100: Obtain historical dataset information and generate classification sample vectors, business representation vectors, and resource usage vectors; among them, by analyzing the statistical characteristic attributes of historical business traffic: number of DNS and NTP requests, packet length, number of packets, and time interval, a vector space model of historical data is generated as the classification sample vector. S200, the classification sample vector is simplified, and then the simplified classification sample vector is refined by a KNN model based on bilinear interpolation distance weighting, and the classification sample category data is output; S300, based on the business representation vector, combined with the classification sample category data, and based on the neural network model of the evolutionary algorithm, predict the initial network resource requirements; S400, based on the resource usage vector and the classification sample category data, and combined with a weighted improved random forest regression model, for a given service, an attribute prediction model is trained for all its virtual network elements. The attribute prediction results of each attribute prediction model are aggregated, and regression analysis is used to aggregate different attributes to form a dynamic attribute view of virtual network elements for the entire network service. During the training phase, the decision tree of the improved random forest regression model evaluates the prediction performance of each decision tree and assigns higher weights to decision trees that can accurately predict a small number of samples. The final dynamic attribute prediction results of the virtual network elements are obtained through weighted voting. S500 uses deep packet inspection to extract feature information based on network resource requirements, and constructs a network resource requirement distribution map based on transfer learning under reinforcement learning. S600, based on the particle swarm optimization virtual network mapping algorithm, constructs the nodes and links of the network resource demand distribution map, and the underlying physical network mapping model of the two stages, and outputs the optimal virtual network mapping scheme. S700 optimizes the dynamic migration of virtual network elements based on a load-balancing migration strategy.

2. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 1, characterized in that, The output classification sample category data includes the following steps: S210, In the reduced classification sample vector, obtain the difference in the number of samples based on the number of samples in the class with the most samples and the number of samples in each minority class, and determine the number of interpolations required for each minority class. S220, for each minority sample class, obtain the sum of the Euclidean distances from each sample to each non-sample in that class, and the mean of the Euclidean distances; S230, for each minority sample class in step S220, obtain the probability that each sample will be sampled; S240, each time a portion of reference samples are selected from the sample set of the minority sample class, the interpolation position for interpolating the minority sample class is obtained according to the bilinear interpolation method, and the reference samples are interpolated at the interpolation position in combination with the interpolation number. S250, after interpolating the minority sample classes, obtain the distance between the test sample used for classification and the training samples in the training set for training the KNN model, and sort them in ascending order of distance, selecting the top... k For each sample, obtain its distance weight; S260, using a classification voting function with the distance weights, the distance weights are obtained and categorized by category, the category with the largest superimposed distance weight is selected, and the classification category is output; The process of obtaining the probability of each sample being sampled includes the following steps: S231, obtain the first number of samples in a minority sample class whose Euclidean distance to other samples in the same class is less than the mean distance; S232, obtain the second number of samples in the minority sample class whose Euclidean distance to non-samples is less than the mean distance; S233, based on the first number of samples, the second number of samples, the weight of samples in this class and the weight of samples outside this class, obtain the sampling weight of samples in the minority sample class; S234, normalize the sampling weights to obtain the probability of each sample in the minority sample class being sampled.

3. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 2, characterized in that, The interpolation position is obtained using the following formula: ； In the formula, Represented as the first Interpolation position of the degree Represented as the sampled sample up to the th The physical distance to each reference sample Represented as the first The sampled sample in the interpolation corresponds to the first... Feature values ​​of a reference sample, 、 、 、 Represented as the first The Euclidean distances from the sampled sample in the interpolation to the 1st, 2nd, 3rd, and 4th reference samples; The distance weight is obtained using the following formula: ； In the formula, Represented as distance weights, Represented as the sample to be tested and the first The distance of a neighbor, Represented as The closest distance among neighbors Represented as The furthest distance among the nearest neighbors; The classification categories are obtained using the following formula: ； In the formula, c x Represented as the sample to be tested x The corresponding categories, where V represents the current set of samples to be tested, and L represents the set of all sample categories. N k (X) represents the sample X in the training set. k The set of nearest neighbors, Represented as training samples x iii Category It is represented as an exponential function that returns 1 when its value is true and 0 otherwise.

4. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 1, characterized in that, The prediction of initial network resource requirements includes: S310, Based on the service representation vector, construct a service dataset sample based on "output service quality performance - actual resource usage"; and divide the service dataset sample into a service data training set and a service data test set according to a certain ratio, and obtain the optimal weights and thresholds of the neural network through training iteration; S320, an improved evolutionary algorithm is used to optimize the neural network, a neural network model based on the evolutionary algorithm is built, and the neural network model based on the evolutionary algorithm is trained and tested using the business data training set and the business data test set; and the neural network model based on the evolutionary algorithm after training and testing predicts the initial network resource requirements based on the business representation vector and the classification sample category data. In the improved evolutionary algorithm, the reciprocal of the squared error is used as the fitness function to measure the fitness of individuals in the neural network population. The formula is as follows: ； ； In the formula, Represented as the first Error of the next iteration Represented as the first In the nth iteration mm -The overall output of one neuron, w Represented as a weight vector, u Represented as an input vector, Represented as the fitness function, Represented as the theoretical output value; In the improved evolutionary algorithm, the individuals in the population are reordered to obtain the probability of an individual being selected after the reordering. This probability is obtained using the following formula: ； ； In the formula, This represents the probability that the optimal individual is selected. s This is represented as the standardized value of the optimal individual. This represents the probability of an individual being selected from the reordered population. b This represents the position of the N'th individual after rearranging the population. a This represents the population size in the evolutionary algorithm; In the improved evolutionary algorithm, the change in root fitness during iteration adjusts the change in crossover probability, which is then obtained using the following formula: ； In the formula, P j Represented as crossover probability, P jmax This is represented as the maximum crossover probability. P jmin This is represented as the minimum crossover probability. F This represents the maximum fitness of two individuals undergoing crossover in the population. F mean The fitness of the entire population is expressed as the average fitness. n max This represents the maximum number of iterations the evolution operator performs during its operation. This represents the number of iterations of the evolutionary algorithm in the current working process.

5. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 1, characterized in that, The dynamic attribute prediction result of the virtual network element is obtained by the following formula: ； In the formula, H(T) represents the dynamic attribute prediction result of the virtual network element, and T is the number of decision trees. For indicator functions, For the first tt The prediction results of the decision tree The operator represents a comparison of the values ​​on both sides. Represented as the true value, For the first tt The voting weights of the decision trees are such that the indicator function has a value of 1 when the prediction result of the decision tree is true, and 0 otherwise.

6. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 1, characterized in that, The optimal virtual network mapping scheme for output includes: S610 establishes multi-dimensional constraints on business type, time validity, and business priority; S620, sets the size of the particle swarm. N’ The maximum number of iterations the algorithm can execute MG Particles are randomly generated at their initial positions. X i With speed V i ; S630, calculates the fitness of all particles. To obtain the globally optimal initial position With the optimal initial position of each individual particle ; S640, For particles that satisfy the multidimensional constraints, perform velocity and position updates, and randomly select candidate nodes of the bottom network during position updates; for particles that do not satisfy the multidimensional constraints, regenerate their position and velocity parameters. S650, for each particle in the particle swarm, if... ,but ;if ,but ; S660, Check the current iteration count; if it is less than the maximum iteration count... MG If the condition is met, proceed to step S640; otherwise, proceed to step S670. S670 outputs the optimal virtual network mapping scheme and its fitness value; in, This represents the optimal initial position for each individual particle. The fitness is represented by the globally optimal initial position.

7. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 6, characterized in that, The position and velocity of particles in a particle swarm are defined as follows: Definition 1, Particle position: The position of a particle Defined as the first i One mapping scheme, D This is indicated as a virtual network request containing D A virtual network node, Take a positive integer, whose value is represented as the first... j Each virtual node selects an underlying network node number from the list of candidate nodes in the underlying network. j =1, 2, 3, ... D ; Definition 2, Particle velocity: The velocity of a particle Defined as the first i The adjustment decision for each mapping scheme, among which... Represented as the first j A virtual node is a node mapping selected from the list of candidate nodes in the underlying network; Definition 3, Subtraction This is used to obtain the differences between the two mapping schemes; Definition 4, Addition This is used to obtain adjustment decisions for the mapping scheme; Definition 5, Multiplication , used to obtain a new mapping scheme; The position and velocity update formulas for the particle swarm optimization algorithm after definition are as follows: ； ； In the formula, Defined as the adjustment decision for the (i+1)th mapping scheme; Defined as the first i +1 mapping scheme; Xgb Defined as the best historical location in the neighborhood; Xpb Defined as its best historical position, , and It is a constant, and + + =1.

8. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 1, characterized in that, Step S700 includes: The S710 monitors the global load and the proportion of high and low load nodes, dynamically adjusts status thresholds, assesses the status of node resources, and determines the timing of migration. S720 selects virtual machines to be migrated based on the dependency of virtual machines and node resources and the contribution of virtual machines.

9. The method for characterizing heterogeneous network element service requirements and adaptive dynamic deployment of virtual network elements according to claim 8, characterized in that, The determination of migration timing includes: S711, collect load information of each node, obtain the global load degree and the proportion of high and low load nodes within the monitoring time, and determine whether the status threshold needs to be updated. If yes, proceed to step S712; otherwise, proceed to step S713. S712 analyzes the global load and the proportion of high and low load nodes, and adjusts and updates the status thresholds. S713, Based on the status threshold, determine the load status of each node; if the node load index is higher than the high threshold, enable the overload assessment of the node load index, and proceed to step S714; if the node load is lower than the low threshold, enable the light load assessment of the node load, and proceed to step S715; if neither the low threshold nor the high threshold is met, the node is considered a suitable load node, and proceed to step S717. S714, performs load metrics on nodes exceeding high-level thresholds. m The query period is used to calculate the evaluation value. If the evaluation value reaches the overload requirement, the node is determined to be overloaded for this indicator type. The high-load node matrix is ​​updated accordingly. Proceed to step S713 and continue the loop. S715, performs processing on nodes with load levels below the low threshold. m Query the load factor for the period and calculate its evaluation value. m If all evaluation values ​​during the period meet the light load requirements, then this node is determined to be a light load node, added to the low load node list, and the process proceeds to step S713 to continue the loop. S716 sorts the list of low-load nodes in ascending order of load level, and combines this with the global load level to determine whether they are within the load balancing range. Then calculate the set of lightly loaded nodes to be migrated. and If so, then a node will be enabled; For global load, , This is the load balancing range generated based on the historical load level of the data center during this period. For high load ratio; S717 outputs the high-load node matrix and the low-load node matrix.

Images