Task intention-oriented data link virtual network resource arrangement method

Abstract

The invention discloses a task intention-oriented data link virtual network resource arrangement method, and belongs to the technical field of data link networks. When facing scattered and heterogeneous data link virtual network resources, a traditional manual management mode is difficult to deal with the complexity, flexibility and real-time requirements of resource management, and consequently, the change of a task cannot be quickly responded in a task execution process. In order to solve the problem, the invention provides the task intention-oriented data link virtual network resource arrangement method, through a data driving mode, the virtual network resources and the task intention are subjected to quantitative modeling, and an artificial intelligence technology is used for learning and optimization, so that intelligent arrangement of the virtual network resources is realized. According to the scheme, multiple key problems in virtual network resource management are solved, including how to quickly generate an arrangement scheme meeting specific tactical task requirements on the basis of heterogeneous resources, and resource combinations can be flexibly adjusted according to different intentions of tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of data link network, and in particular relates to a data link virtual network resource arrangement method oriented to task intent. Background Art

[0002] With the development of modern information technology, data link network has become the core of information transmission and control platform. With the complexity and diversification of application environment, especially in the fast-changing real-time environment, the control core needs to respond quickly to various control requirements. These requirements are usually reflected in the form of different mission intents, and each mission intent has its specific resource requirements and performance indicators. In order to build a networking solution suitable for data link to meet the needs of mission intent, it is necessary to give full play to the ability of information sharing in the control process: change the information flow, reduce the information flow links, and flow directly to the required units; ensure that each unit obtains the required relevant resources in a very short time or at the same time, and use resources in time; through real-time sharing of information, use available resources nearby to achieve the purpose; change the coordination method between units, and carry out more random coordination and adaptive coordination; change the application of control methods and methods, and implement more leapfrog control, decentralized control and temporary control to improve control efficiency; realize accurate and efficient resource addressing to avoid redundancy and waste of resources. Under the above background and drive, the data link network can be implemented in a service-oriented and virtualized manner. The fundamental purpose and driving force of the development of the data link network is to provide the best service for the needs of battlefield applications and to provide flexible and changeable communication services covering the entire area for control needs. Therefore, it is an inevitable trend in the future development of data link networks to design a service-centric network architecture based on the foothold and starting point of providing customized services for mission intent and tasks in a flexible, fast and diversified manner.

[0003] The complexity and dynamic variability of tasks require that the data link network be able to schedule a large number of heterogeneous virtualized resources in a short period of time to meet the needs of different tasks. At the same time, the application of network function virtualization (NFV) technology in data link networks has enabled traditional hardware network functions to be gradually virtualized into software operations, improving the flexibility and scalability of resources. However, the resource heterogeneity and distribution complexity brought about by virtualization, as well as how to match these virtual network functions with specific task intentions, have become key challenges. In actual tasks, different tasks often have their own specific intentions and requirements, and traditional manual methods are difficult to match virtualized resources with these task intentions in a timely and accurate manner. Moreover, as the task environment continues to change, tasks need to adjust resource configuration in real time to maintain efficient task execution. The existing manual scheduling method is not only time-consuming, but also prone to problems such as low resource utilization and delayed response.

[0004] Therefore, how to characterize and associate various types of data link virtual network resources and mission intents, use intelligent methods to generate appropriate data link virtual network resource orchestration plans for different mission intents, meet the functional and performance indicator requirements of the mission intent, and flexibly combine existing tactical resources to complete the mission is a technical problem that technical personnel in this field currently need to solve. Summary of the invention

[0005] In order to address the shortcomings of virtual network resources in data link network systems oriented to task intentions, such as spatial dispersion, functional heterogeneity, difficult management, and cumbersome manual management, a data-driven approach is used to quantitatively model virtual network resources and task intentions. The virtual network resource orchestration schemes used in historical mission events are used as prior data. Through the learning of artificial intelligence methods, it is possible to recommend and generate appropriate network virtual network resource orchestration schemes for different tasks.

[0006] In order to achieve the above object, the present invention adopts the following technical solution, a data link virtual network resource arrangement method oriented to task intent, the method comprising the following steps:

[0007] Step 1: Data-driven data link virtual network resources and mission intent modeling;

[0008] Step 1.1: Based on the common descriptive features of the data link virtual network resources in the mission environment, data-driven modeling is performed on different virtual network resources; the content includes: resource category, resource location, resource size, and throughput performance. A virtual network resource metadata database is generated based on the modeling content of different virtual network resources. Any virtual network resource in the mission environment should be included in the metadata database;

[0009] Step 1.2: Carry out data-driven modeling of different task intentions, including: intention type, intention duration, functional performance, and generate a task intention metadata database based on the modeling content of different task intentions. The task intentions used in historical tasks and the task intentions that may be used should be included in the metadata database;

[0010] Step 1.3: Model the virtual network resource orchestration schemes used in historical tasks, that is, the resource orchestration schemes used in historical tasks to achieve a certain intention, each of which includes a series of virtual network resources;

[0011] Step 2: Establish a vector representation of virtual network resources based on the historical orchestration scheme sequence;

[0012] In natural language processing related tasks, using multi-dimensional vectors to represent vocabulary and measure similarity is a widely used method; of course, using one-hot encoding to represent each virtual network resource, the dimension, sparsity, and expressiveness of the one-hot vector all bring great difficulties to storage and calculation; therefore, technicians in this field usually use word embedding technology to convert vocabulary into continuous low-dimensional vector representation, solving the problems of dimensionality curse and lack of expressiveness. This method uses similar embedding technology to establish a low-dimensional vector representation of virtual network resources based on the virtual network resource scheduling schemes used in the history of different task intentions and the context in which different virtual network resources appear in the same task, and at the same time learns the probability of different virtual network resources being selected as scheduling schemes at the same time to complete the task;

[0013] Step 3: Learn the selection rates of different virtual network resources in different tasks based on metadata;

[0014] The vector representation of the virtual network resources learned in step 2 is obtained only through context-aware training, without considering the task intent and the content information of the virtual network resources; based on a large amount of historical data consisting of orchestration events, resource features, and intent features, the probability of a specific virtual network resource being selected in different task intents is trained. The present invention not only considers the use of the attributes of resource metadata features and intent metadata features themselves, but also considers the interactive attributes between features to train artificial intelligence models. Feature interaction combines two or more features and performs a series of operations to achieve nonlinear transformation of the sample space, which can improve the nonlinear ability of the model and achieve effective prediction of different feature combinations.

[0015] Step 4: Generate a preliminary arrangement plan based on the vector similarity of virtual network resources;

[0016] After obtaining the vector representation of the virtual network resources containing context information in step 2, the vectors of the virtual network resources selected by each task intention are averaged to obtain the vector representation of each task intention, that is, the task intention is described by historical experience, and the vector representation of each task intention contains the information of all selected virtual network resources; establish the correlation between the task intention and the virtual network resources, and select the virtual resources whose distance from the vector representation of the task intention in the vector space is less than the set threshold, and the distance is calculated by the cosine similarity index;

[0017] Step 5: Sort the selection rates of the virtual network resources in the preliminary orchestration plan to generate a final orchestration plan;

[0018] After learning the selection rate of the specific virtual network resource in step 3 in the realization of the specific task intent, the selection rate of the resource by the intent is calculated by inputting the content features of the virtual network resource and the content features of the task intent; according to the preliminary arrangement plan obtained in step 4, the virtual network resources contained in the plan are paired with the task intent one by one, and their selection rates are calculated. Finally, the top k items are selected as the final arrangement results of the task intent according to the ranking of the selection rates of all virtual network resources.

[0019] Furthermore, the modeling of virtual network resources, task intent, and historical orchestration scheme in step 1 is specifically expressed as follows:

[0020] Step 1A: Model virtual network resources. Virtual network resources are where r j represents the jth virtual network resource metadata sample, m1 is the total number of samples in the resource metadata database, and the feature vector of the virtual network resource metadata sample r is defined as: Where n1 represents the characteristic dimension of resource metadata;

[0021] Step 1B: Model the task intent. The task intent is where i j represents the jth task intent metadata sample, m2 is the total number of samples in the metadata database, and the feature vector of task intent metadata sample i is defined as Where n2 represents the feature dimension of intent metadata;

[0022] Step 1C: Model the virtual network resource orchestration scheme used in the historical tasks of task intention i. The historical orchestration scheme is in represents the j-th scheduling scheme of task intention i, m i The total number of scheduling plans for historical tasks of intention i.

[0023] Furthermore, the learning process of the vector representation of the virtual network resources in step 2 specifically includes the following steps:

[0024] Step 2A: Use one-hot encoding to create a vector of dimension m1 for each virtual network resource metadata sample r. Each vector contains only one 1 at a different position, and the rest are all 0s.

[0025] Step 2B: Construct a training dataset based on the virtual network resource orchestration scheme in the historical tasks, and use the continuous bag-of-words model to generate context and target resources, such as Figure 1 As shown in the figure, for each resource sample in the arrangement scheme, the two consecutive resources before and after it are selected as the context, and the resource itself is selected as the target to construct multiple samples; the input of the sample is the one-hot vector of the context word, and the output is the one-hot vector of the target word;

[0026] Step 2C: Use the samples generated in step 2B to train a single hidden layer neural network. After training, the parameter matrix between the input layer and the hidden layer and the parameter matrix between the hidden layer and the output layer are obtained, whose dimensions are m1×e and e×m1 respectively, where e is the embedding dimension; the jth dimension of the parameter matrix input to the hidden layer is used as the embedding vector E of the jth virtual network resource. j .

[0027] Furthermore, the method for predicting the selection rate in step 3 specifically includes the following steps:

[0028] Step 3A: Construct a dataset for training selection rate prediction. In the historical orchestration schemes of the task intent generated in step 1, each orchestration scheme includes multiple orchestration events consisting of intent and resources; each orchestration event consists of an n3-dimensional vector, n3=n1+n2, that is, it includes n1-dimensional resource metadata features and n2-dimensional intent metadata features, expressed as (X, y), where X=[X r ,X i ], y∈{0,1} indicates whether the resource is selected by the intent; the goal is to learn the nonlinear mapping function f: Predict the probability of selection for a given resource and intent; use the improved ONN model to calculate the selection rate. The improved ONN model scheme is as follows:

[0029] Step 3B: Embed the sparse feature x in the input feature X, expressed as:

[0030] [e 1 ,e 2 ,...,e o ]=[V 1 x,V 2 x,...,V o x] (1)

[0031] where e j is the jth embedding vector of feature x, V j represents the j-th embedding matrix of feature x, o represents the number of feature interactions; the position j of the embedding vector is calculated by the following mapping function:

[0032] Γ(k,s)=j (2)

[0033] Where k represents the index of the operation, s represents the index of the feature, and the function Γ maps the two to the operation index of the feature to achieve interactive operation-aware feature embedding;

[0034] Step 3C: After obtaining the sparse feature embedding, the feature interaction layer uses the inner product operation to convert the calculation of the interactive feature factor into the product of two embedding vectors. The input of this layer is the interactive operation-aware feature embedding vector obtained in step 3B; the output of the feature interaction layer is the embedded feature x e and the interaction feature x o , x e for:

[0035]

[0036] in, represents the copy operation of the jth feature, the interactive feature x o for:

[0037]

[0038] Among them, p j,k The inner product result of the j-th feature and the k-th feature is:

[0039]

[0040] in, Represents the inner product operation of the j-th feature and the k-th feature;

[0041] Step 3D: Embed feature x e , interaction feature x o And the preprocessed dense features x d Together they are fed into a fully connected layer, and the output of the jth fully connected layer is calculated as follows:

[0042]

[0043] in, represents the weight matrix of the jth fully connected layer, represents the bias vector of the jth fully connected layer, Relu(·) and Sigmoid(·) represent activation functions, and h is the number of hidden layers;

[0044] Step 3E: Use the training data constructed in step 3A to train and improve the ONN model according to steps 3B, 3C and gradient descent methods until the accuracy index is met.

[0045] Furthermore, the preliminary arrangement plan is generated in step 4, and the specific steps include:

[0046] Step 4A: Obtain the embedding vector E of all virtual network resources according to the method of step 2;

[0047] Step 4B: For task intention i, the set of all resources selected in its historical events is in represents the total number of resources selected by task intention i; the embedding vector of task intention i is calculated as:

[0048]

[0049] Step 4C: Assume that k×ρ preliminary arrangement results are to be obtained, i.e., ρ times the required recommendation results, E r is the embedding vector of the virtual network resource, and the embedding vector of each virtual network resource r in the calculation space is compared with E i The distance is calculated by the following formula:

[0050]

[0051] Step 4D: Select the first k×ρ virtual resources with the smallest CosSim value as the preliminary scheduling result of intention i.

[0052] Furthermore, the final arrangement result is generated in step 5, and the specific steps include:

[0053] Step 5A: Obtain k×ρ virtual network resources closest to task intention i according to the method described in step 4;

[0054] Step 5B: Using the improved ONN model trained in step 3, calculate the selection rate of each virtual network resource r in the target intention i, which is defined as:

[0055] s(r,s)=ONN(X r ,X i ) (10)

[0056] Step 5C: Select the largest first k virtual network resources from the calculation results of k×ρ virtual network resources as the scheduling solution recommendation result for task intention i.

[0057] The invention has a number of significant beneficial effects. First, the invention reduces the complexity of manual intervention by introducing intelligent and automated scheduling methods, and realizes fast and efficient virtual network resource scheduling. Secondly, based on the intelligent recommendation mechanism of historical task data, the system can automatically match the best resource combination according to task requirements, improve resource utilization and task completion. In addition, from the technical implementation, the invention adopts embedded vector representation in the field of natural language processing, which significantly improves the expression ability and computational efficiency of the model. By introducing the feature interaction method, the modeling ability of the nonlinear relationship between complex tasks and resource features is further enhanced. Finally, the method can adjust the resource scheduling scheme in real time according to the changes in the task, and has extremely high flexibility and adaptability, especially suitable for highly dynamic and rapidly changing environments. In the field of data link networks, the invention shows a wide range of applicability. In different types of tasks, the system can automatically generate the optimal virtual network resource scheduling scheme to ensure the efficient execution of tasks in different environments. The system can adjust communication links, sensors, interference resources, etc. in real time according to the task intent, and improve the flexibility and execution of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the method for constructing a training dataset for learning vector representations of virtual network resources.

[0059] Figure 2 Schematic diagram of the neural network structure used to learn the vector representation of virtual network resources.

[0060] Figure 3 The figure is a schematic diagram of the ONN neural network structure used in the present invention for predicting the selection rate of virtual resources.

[0061] Figure 4 A schematic diagram of the process of training two neural network models in the present invention.

[0062] Figure 5 Schematic diagram of the training status of the Word2vec network model under different embedding dimensions.

[0063] Figure 6 Schematic diagram of the visualization results of the embedding vectors of the same resource type.

[0064] Figure 7 Schematic diagram of the visualization results of the embedding vectors of resources selected for the same task intent.

[0065] Figure 8 A schematic diagram of a specific solution for the task intent-oriented virtual network resource orchestration of the present invention.

[0066] Fig. 9 A diagram showing the comparison of arrangement accuracy under different embedding dimensions.

[0067] Fig.10 Schematic diagram of the comparison of orchestration NDCG performance under different embedding dimensions.

[0068] Fig.11 Schematic diagram of the accuracy comparison between the Word2vec method and different Top@N methods.

[0069] Fig.12 Schematic diagram of NDCG comparison with Word2vec method under different Top@N. DETAILED DESCRIPTION

[0070] When faced with distributed and heterogeneous data link virtual network resources, traditional manual management methods are difficult to cope with the complexity, flexibility and real-time requirements of resource management, resulting in the inability to quickly respond to changes in tasks during task execution. To address this problem, the present invention proposes a data link virtual network resource orchestration method oriented to task intent, which quantitatively models virtual network resources and task intent in a data-driven manner, and uses artificial intelligence technology for learning and optimization, thereby realizing intelligent orchestration of virtual network resources. This solution solves several key problems in virtual network resource management, including how to quickly generate an orchestration plan that meets the requirements of specific tactical tasks based on heterogeneous resources, and can flexibly adjust resource combinations according to different task intents.

[0071] The specific implementation method involves multiple key steps, including data modeling, resource vector representation generation, selection probability training, preliminary orchestration scheme generation, and final orchestration scheme optimization. Through these steps, the system can dynamically and automatically generate adaptive virtual network resource orchestration schemes according to mission intent to meet the real-time changes of different tactical requirements.

[0072] Step 1: Generate a resource list based on the available virtual network resources in the data link network, such as detection system (IDS, Intrusion Detection System), firewall (Firewall), flow shaper (Flow Shaper), Web application firewall (WAF, Web Application Firewall), access control list (ACL, Access Control List), virtual private network (VPN, Virtual Private Network), load balancer (Load Balancer), intrusion prevention system (IPS, Intrusion Prevention System), etc. Perform data-driven feature collection on each network resource and create a virtual network resource metadata database, as shown in Table 1.

[0073] Table 1 Data link virtual network resource characteristics

[0074]

[0075] Step 2: Generate a list of task intentions based on the specific requirements of the task. Each intention has different scopes and constraints. Perform data-driven feature collection on each intention and establish an intention metadata database, as shown in Table 2.

[0076] Table 2 Task intention characteristics

[0077]

[0078] Step 3: Follow the steps below: Figure 4 The training process shown in the figure trains the ONN model for resource selection rate prediction and the Word2vec model for generating preliminary scheduling plans. Specifically, the resource features, such as resource category, location, processing delay, etc., are combined with the features of task intent, such as type, duration, task priority, etc. to form the input features of the training, and the historical events, that is, whether the intent has ever selected the resource, are used as the output target to train the ONN neural network. The Word2vec neural network is trained using the historical task list, such as IDS, ACL, VPN, as a sequence. The training performance under different embedding dimensions is shown in Figure 2. Figure 5 As shown in the figure, the two trained models are saved as offline models respectively.

[0079] Step 4: Test the performance of the offline model, such as Figure 6 As shown in , by visualizing the embedding vectors of virtual resources of the same category through dimensionality reduction, it can be observed that most virtual network resources belonging to the same category are embedded in similar spaces. Figure 7 As shown in the figure, the embedding vectors of virtual network resources that have been selected by different task intents are visualized by dimensionality reduction. It can be observed that the network virtual resources that have been selected by the same task intent are embedded in a relatively close space, indicating that the trained embedding model can well capture the similarities between virtual network resources.

[0080] Step 5: Follow Figure 8 The prediction process shown in the figure, for the target task intent, first search the database for its historical task events, find all its selected virtual network resources, use the trained offline model in step 3 to calculate the embedding vector of each virtual network resource, average all the embedding vectors, and get the embedding vector of the task intent. Search in the embedding space for the k×ρ virtual resources with the smallest cosine similarity of the specific vector.

[0081] Step 6: Follow Figure 5The prediction process shown in the figure searches for the features of the target task intent in the metadata database, and at the same time searches for the features of the virtual network resources searched in step 4, and combines them one by one into a feature vector. The feature vector is input into the pre-trained offline ONN model, which gives a predicted score of the selection rate for each virtual network resource. The k×ρ virtual network resources obtained in step 4, i.e., the preliminary arrangement plan, are sorted according to their predicted scores, and the top k virtual network resources are taken as the output result of this arrangement.

[0082] Step 7: Use the test set that is not part of the training data set to perform performance tests on the trained model. Fig. 9 The embedding performance under different embedding dimensions is shown. It can be observed that when the embedding dimension is below 150, the model accuracy increases with the increase of dimension, indicating that the increase of dimension can accommodate more model information. When the embedding dimension is above 150, the model accuracy converges and even decreases to a certain extent, indicating that excessive increase of dimension does not improve detection performance, but may lead to problems such as overfitting. Therefore, it is appropriate to select an embedding dimension of 150 for this implementation case. Fig.10 The Normalized Discounted Cumulative Gain (NDCG) shown in the article is a commonly used evaluation indicator in the field of recommendation systems. Figure 8 Similar.

[0083] Step 8: Also using the test set, compare the performance of the model proposed in the present invention with the Word2vec-based arrangement method, as shown in Figure 11-12 As shown in Figure 2, eW represents the performance of the Word2vec model when the embedding dimension is e, and eH represents the performance of the solution proposed in the present invention. It can be observed that in any given embedding dimension, the method proposed in the present invention can achieve higher accuracy and NDCG indicators, that is, it can generate more accurate arrangement solutions and better guarantee the execution of given task intentions.

Claims

1. A data link virtual network resource arrangement method oriented to task intent, the method comprising the following steps: Step 1: Data-driven data link virtual network resources and mission intent modeling; Step 1.1: Based on the common descriptive characteristics of the data link virtual network resources in the mission environment, data-driven modeling of different virtual network resources is carried out; the content includes: Resource category, resource location, resource size, and throughput performance. Generate a virtual network resource metadata base based on different virtual network resource modeling contents. Any virtual network resource in the task environment should be included in the metadata base. Step 1.2: Carry out data-driven modeling of different task intentions, including: intention type, intention duration, functional performance, and generate a task intention metadata database based on the modeling content of different task intentions. The task intentions used in historical tasks and the task intentions that may be used should be included in the metadata database; Step 1.3: Model the virtual network resource orchestration schemes used in historical tasks, that is, the resource orchestration schemes used in historical tasks to achieve a certain intention, each of which includes a series of virtual network resources; Step 2: Establish a vector representation of virtual network resources based on the historical orchestration scheme sequence; According to the virtual network resource orchestration schemes adopted in the history of different task intentions, a low-dimensional vector representation of virtual network resources is established according to the context in which different virtual network resources appear in the same task, and the probability of different virtual network resources being selected as orchestration schemes to complete the task is learned at the same time; Step 3: Learn the selection rates of different virtual network resources in different tasks based on metadata; The vector representation of the virtual network resources learned in step 2 is trained only through context-aware methods, without considering the task intent and the content information of the virtual network resources; based on a large amount of historical data consisting of orchestration events, resource features, and intent features, the probability of a specific virtual network resource being selected in different task intents is trained; Step 4: Generate a preliminary arrangement plan based on the vector similarity of virtual network resources; After obtaining the vector representation of the virtual network resources containing context information in step 2, the vectors of the virtual network resources selected by each task intention are averaged to obtain the vector representation of each task intention, that is, the task intention is described by historical experience, and the vector representation of each task intention contains the information of all selected virtual network resources; establish the correlation between the task intention and the virtual network resources, and select the virtual resources whose distance from the vector representation of the task intention in the vector space is less than the set threshold, and the distance is calculated by the cosine similarity index; Step 5: Sort the selection rates of the virtual network resources in the preliminary orchestration plan to generate a final orchestration plan; After learning the selection rate of the specific virtual network resource in step 3 in the realization of the specific task intent, the selection rate of the resource by the intent is calculated by inputting the content features of the virtual network resource and the content features of the task intent; according to the preliminary arrangement plan obtained in step 4, the virtual network resources contained in the plan are paired with the task intent one by one, and their selection rates are calculated. Finally, the top k items are selected as the final arrangement results of the task intent according to the ranking of the selection rates of all virtual network resources.

2. A data link virtual network resource arrangement method oriented to task intent as claimed in claim 1, characterized in that: The modeling of virtual network resources, task intent, and historical orchestration scheme in step 1 is specifically expressed as follows: Step 1A: Model virtual network resources. Virtual network resources are where r j represents the jth virtual network resource metadata sample, m1 is the total number of samples in the resource metadata database, and the feature vector of the virtual network resource metadata sample r is defined as: Where n1 represents the characteristic dimension of resource metadata; Step 1B: Model the task intent. The task intent is where i j represents the jth task intent metadata sample, m2 is the total number of samples in the metadata database, and the feature vector of task intent metadata sample i is defined as Where n2 represents the feature dimension of intent metadata; Step 1C: Model the virtual network resource orchestration scheme used in the historical tasks of task intention i. The historical orchestration scheme is in represents the j-th scheduling scheme of task intention i, m i The total number of scheduling plans for historical tasks of intention i.

3. A data link virtual network resource arrangement method oriented to task intent as claimed in claim 2, characterized in that: The learning process of the vector representation of the virtual network resources in step 2 specifically includes the following steps: Step 2A: Use one-hot encoding to create a vector of dimension m1 for each virtual network resource metadata sample r. Each vector contains only one 1 at a different position, and the rest are all 0s. Step 2B: Construct a training data set based on the virtual network resource orchestration scheme in the historical tasks, use the continuous bag-of-words model to generate context and target resources, and for each resource sample in the orchestration scheme, select the two consecutive resources before and after it as the context and the resource sample itself as the target to construct multiple samples; the input of the sample is the one-hot vector of the context word, and the output is the one-hot vector of the target word; Step 2C: Use the samples generated in step 2B to train a single hidden layer neural network. After training, the parameter matrix between the input layer and the hidden layer and the parameter matrix between the hidden layer and the output layer are obtained, whose dimensions are m1×e and e×m1 respectively, where e is the embedding dimension; the jth dimension of the parameter matrix input to the hidden layer is used as the embedding vector E of the jth virtual network resource. j .

4. A method for arranging data link virtual network resources based on task intent as claimed in claim 3, characterized in that: The method for predicting the selection rate in step 3 specifically comprises the following steps: Step 3A: Construct a dataset for training selection rate prediction. In the historical orchestration schemes of the task intent generated in step 1, each orchestration scheme includes multiple orchestration events consisting of intent and resources; each orchestration event consists of an n3-dimensional vector, n3=n1+n2, that is, it includes n1-dimensional resource metadata features and n2-dimensional intent metadata features, expressed as (X, y), where X=[X r ,X i ], y∈{0,1} indicates whether the resource is selected by the intent; the goal is to learn a nonlinear mapping function Predict the probability of selection for a given resource and intent; use the improved ONN model to calculate the selection rate. The improved ONN model scheme is as follows: Step 3B: Embed the sparse feature x in the input feature X, expressed as: [e 1 ,e 2 ,...,e o ]=[V 1 x,V 2 x,...,V o x] (1) where e j is the jth embedding vector of feature x, V j represents the j-th embedding matrix of feature x, o represents the number of feature interactions; the position j of the embedding vector is calculated by the following mapping function: Γ(k,s)=j (2) Where k represents the index of the operation, s represents the index of the feature, and the function Γ maps the two to the operation index of the feature to achieve interactive operation-aware feature embedding; Step 3C: After obtaining the sparse feature embedding, the feature interaction layer uses the inner product operation to convert the calculation of the interactive feature factor into the product of two embedding vectors. The input of this layer is the interactive operation-aware feature embedding vector obtained in step 3B; the output of the feature interaction layer is the embedded feature x e and the interaction feature x o , x e for: in, represents the copy operation of the jth feature, the interactive feature x o for: Among them, p j,k The inner product result of the j-th feature and the k-th feature is: in, Represents the inner product operation of the j-th feature and the k-th feature; Step 3D: Embed feature x e , interaction feature x o And the preprocessed dense features x d Together they are fed into a fully connected layer, and the output of the jth fully connected layer is calculated as follows: in, represents the weight matrix of the jth fully connected layer, represents the bias vector of the jth fully connected layer, Relu(·) and Sigmoid(·) represent activation functions, and h is the number of hidden layers; Step 3E: Use the training data constructed in step 3A to train and improve the ONN model according to steps 3B, 3C and gradient descent methods until the accuracy index is met.

5. A method for arranging data link virtual network resources based on task intent as claimed in claim 4, characterized in that: The preliminary arrangement plan is generated in step 4, and the specific steps include: Step 4A: Obtain the embedding vector E of all virtual network resources according to the method of step 2; Step 4B: For task intention i, the set of all resources selected in its historical events is in represents the total number of resources selected by task intention i; the embedding vector of task intention i is calculated as: Step 4C: Assume that k×ρ preliminary arrangement results are to be obtained, i.e., ρ times the required recommendation results, E r is the embedding vector of the virtual network resource, and the embedding vector of each virtual network resource r in the calculation space is compared with E i The distance is calculated by the following formula: Step 4D: Select the first k×ρ virtual resources with the smallest CosSim value as the preliminary scheduling result of intention i.

6. A method for arranging data link virtual network resources based on task intent as claimed in claim 5, characterized in that: The final arrangement result in step 5 is generated, and the specific steps include: Step 5A: Obtain k×ρ virtual network resources closest to task intention i according to the method described in step 4; Step 5B: Using the improved ONN model trained in step 3, calculate the selection rate of each virtual network resource r in the target intention i, which is defined as: s(r,s)=ONN(X r ,X i ) (10) Step 5C: Select the largest first k virtual network resources from the calculation results of k×ρ virtual network resources as the scheduling solution recommendation result for task intention i.

Images