A method, device and storage medium for predicting the performance of a Siamese network based on a graph neural network

Abstract

The present invention discloses a method, device and storage medium for predicting the performance of a twin network based on a graph neural network, belonging to the technical field of communication networks. The method includes: preprocessing the obtained historical information of the physical communication network to obtain a dependency graph data set; initializing the states of each node in the dependency graph data set to obtain the initial hidden states of each node; using the initial hidden states to train a pre-constructed network performance prediction model to obtain a trained network performance prediction model, wherein the network performance prediction model collaborates through an attention mechanism, a message passing neural network and a gated recurrent unit to obtain iteratively updated hidden state information and adjust the model parameters; inputting the obtained real-time information of the physical communication network into the trained network performance prediction model to obtain a performance prediction result, improving the stability, accuracy and efficiency of the performance prediction of the digital twin communication network.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication networks, and particularly relates to a twin network performance prediction method, device and storage medium based on a graph neural network. Background Art

[0002] In recent years, with continuous technological innovation in the communication field, new network paradigms such as software-defined network (SDN) and digital twin network (DTN) have emerged, promoting the transformation and upgrading of traditional networks. Among them, DTN is a network system that creates a virtual twin of a physical network entity in a digital way and can interact and map with the physical network entity in real time, used to expand the functions of the physical network. At present, the digital twinning of communication networks is in the exploration stage. Due to the characteristics of dynamic changes and large environmental impact at the physical level of communication networks, the internal structure and data are complex, increasing the difficulty of network performance prediction and ultimately making it difficult to achieve digital twinning.

[0003] Due to its powerful ability to process graph-structured data, the graph neural network (GNN) further enhances the accuracy of performance prediction and shows great application potential in the digital twin of communication networks. At present, many methods for network performance prediction using graph neural networks have emerged at home and abroad.

[0004] In the prior art, there is a method that uses graph spatial domain convolution to process the message passing process of graph network nodes, conducts relationship reasoning between network information, and realizes accurate prediction of network performance such as delay, jitter and packet loss rate. However, the features extracted by the model are relatively one-sided and the generalization performance is poor. There is also a method in the prior art that trains a new lightweight graph network (GLN) in the way of federated learning to predict the network performance indicators of the slice network, improving the efficiency of model training. However, there is still room for further improvement in the balance between model performance and training efficiency. To sum up, the existing network performance prediction technologies have limitations in dealing with dynamic and changeable network environments and are difficult to meet the requirements of high fidelity and low latency of digital twins. Summary of the Invention

[0005] The present invention aims to provide a twin network performance prediction method, device and storage medium based on a graph neural network. By performing feature selection and feature integration on the training set data to select the best features, and by combining the attention mechanism, GRU unit and message passing neural network (MPNN), the stability of the hidden state update process of graph-structured data is improved, thereby improving the stability, accuracy and efficiency of communication network performance prediction to adapt to the rapidly changing communication service requirements.

[0006] To achieve the above object, the present invention is implemented by the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting the performance of a Siamese network based on a graph neural network, including:

[0008] Obtain real-time relevant network information of a physical communication network and input it into a trained network performance prediction model to obtain a performance prediction result;

[0009] The process of determining the trained network performance prediction model includes:

[0010] Preprocess the historical relevant network information of the physical communication network obtained by the digital twin communication network to obtain a dependency graph dataset, where the dependency graph dataset includes link, path, and queue information;

[0011] Initialize the states of each link, path, and queue in the dependency graph dataset respectively to obtain the initial hidden states of each link, path, and queue;

[0012] Use the initial hidden states to train a pre-constructed network performance prediction model to obtain a trained network performance prediction model, where the network performance prediction model collaboratively works through an attention mechanism, a message passing neural network, and a gated recurrent unit to obtain iteratively updated hidden state information and adjusts the model parameters according to the hidden state information.

[0013] Optionally, the preprocessing of the historical relevant network information of the physical communication network obtained by the digital twin communication network to obtain a dependency graph dataset includes:

[0014] Normalize the read original network samples to obtain a standardized dataset;

[0015] Convert the original network samples in the standardized dataset into network dependency graphs to obtain a network dependency graph dataset, where the network dependency graph includes a set of links , a set of queues on the output ports of network devices and a set of source-to-destination paths ;

[0016] , where represents the th link in a set of links, represents the number of links in a set of links;

[0017] , where represents the th queue in a set of queues, represents the number of queues in a set of queues;

[0018] , where Indicates the th path in a set of paths, Indicates the number of paths in a set of paths, represented by a series of tuples , , represents the number of elements in represents the th queue along the path, represents the th link along the path.

[0019] Optionally, the state initialization is respectively performed on each link, path, and queue state in the dependency graph dataset to obtain the initial hidden states of each link, path, and queue, including:

[0020] Path state initialization: Create a first state vector by concatenating traffic characteristics, packet quantity characteristics, and zero padding , where the traffic characteristic represents the data traffic passing through the path, the packet quantity characteristic represents the number of packets generated on the path, and zero padding is used to form a feature vector of a predetermined dimension;

[0021] And / or, link state initialization: Create a second state vector by concatenating capacity characteristics, policy characteristics, and zero padding , where the capacity characteristic represents the bandwidth capacity of the link, the policy characteristic represents the queue scheduling policy on the link, and zero padding is used to meet the dimension requirements;

[0022] And / or, queue state initialization: Create a third state vector by concatenating size characteristics, priority characteristics, weight characteristics, and zero padding , where the size characteristic represents the current number of packets in the queue, the priority characteristic represents the packet priority, the weight characteristic represents the queue weight related to the scheduling policy, and zero padding is used to meet the dimension requirements.

[0023] Optionally, the pre-constructed network performance prediction model includes: a GRUCell gated recurrent unit layer, an Attention attention layer, a message passing recurrent module, and a Readout readout layer connected in sequence; the message passing recurrent module includes an MPNN message passing neural network layer and an RNN recurrent neural network layer connected to each other.

[0024] Optionally, using the initial hidden state to train the pre-constructed network performance prediction model to obtain a trained network performance prediction model, including:

[0025] Define a path recurrent unit, a link recurrent unit, and a queue recurrent unit respectively through the GRUCell gated recurrent unit layer;

[0026] Through the combination of the Attention attention layer, the MPNN message passing neural network layer, and the RNN recurrent neural network layer, iteratively update the hidden state information of the path recurrent unit, the link recurrent unit, and the queue recurrent unit from the initial hidden state through the message passing process to obtain the global state of the physical communication network;

[0027] Output a performance prediction value through the Readout readout layer according to the updated path hidden state;

[0028] Train the network performance prediction model according to the mean square error MSE between the performance prediction value and the true value, and set the model hyperparameters. Among them, the training objective of the network performance prediction model is expressed as:

[0029] ,

[0030] In the formula: is the actual performance value of the i-th dependency graph sample, is the performance prediction value of the i-th dependency graph sample, is the number of samples in the dependency graph dataset, and i represents the serial number of the sample in the dependency graph dataset.

[0031] Optionally, the method further includes: selecting a test set from the dependency graph dataset, inputting the test set into the network performance prediction models generated in every n rounds of training and the last round of training, and calculating the mean absolute error MAE, the mean absolute percentage error MAPE, and the correlation coefficient between the predicted value and the true value of the network performance prediction model to evaluate the training effect of the model and find better hyperparameters to obtain the best network performance prediction model, where:

[0032] Calculate the mean absolute error MAE, the mean absolute percentage error MAPE, and the correlation coefficient between the predicted value and the true value of the network performance prediction model. The calculation formula is:

[0033] ,

[0034] ,

[0035] ,

[0036] In the formula: is the average value of the actual performance values of n dependency graph samples.

[0037] Optionally, the iterative update of the hidden state information encoded for the path loop unit, link loop unit, and queue loop unit through the message passing process includes: in each iteration, the network performance prediction model performs message passing in the following three stages:

[0038] Queue and link to path: Collect the information of queues and links included in each path, and calculate the temporary state generated by the attention mechanism according to the link hidden state and queue hidden state at a fixed moment, and update to obtain a new path hidden state and a new path temporary hidden state using the GRU unit in the RNN recurrent neural network layer with and the path hidden state at this moment as inputs; ;

[0039] Path to queue: Aggregate the temporary hidden state information of queues and all paths containing queues , sum all the temporary hidden state information to obtain the sum of queue temporary hidden state information , and update to obtain a new queue hidden state and a new queue temporary hidden state using the GRU unit in the RNN recurrent neural network with the sum of queue temporary hidden state information and queue hidden state as inputs;

[0040] Queue to link: Update to obtain a new link hidden state using the GRU unit in the RNN recurrent neural network with the new queue temporary hidden state and link hidden state as inputs.

[0041] In a second aspect, the present invention provides a twin network performance prediction device based on a graph neural network, including:

[0042] A performance prediction result acquisition module: used to obtain real-time relevant network information of the physical communication network, input it into the trained network performance prediction model, and obtain a performance prediction result;

[0043] Among them, the determination process of the trained network performance prediction model includes:

[0044] Preprocess the historical relevant network information of the physical communication network obtained from the digital twin communication network to obtain a dependency graph dataset, where the dependency graph dataset includes link, path, and queue information;

[0045] Initialize the features of each link, path, and queue status in the dependency graph dataset respectively to obtain the initial hidden states of each link, path, and queue;

[0046] Use the initial hidden states to train a pre-constructed network performance prediction model to obtain a trained network performance prediction model. Among them, the network performance prediction model collaborates through an attention mechanism, a message passing neural network, and a gated recurrent unit to obtain iteratively updated hidden state information, and adjusts the model parameters according to the hidden state information.

[0047] In a third aspect, the present invention provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the graph neural network-based twin network performance prediction method described in any step of the first aspect.

[0048] In a fourth aspect, the present invention provides an electronic terminal, including a processor and a memory connected to the processor. A computer program is stored in the memory. When the computer program is executed by the processor, it executes the graph neural network-based twin network performance prediction method described in any step of the first aspect.

[0049] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: By preprocessing the data collected from the digital twin communication network system, a suitable dataset is obtained for the training and evaluation of the network performance prediction model, and the node features of the graph structure data are initialized. The main work is feature selection and feature integration to select the best features. Then, by constructing a graph neural network GNN model including an attention mechanism, GRU units, and a message passing neural network MPNN, the stability of the hidden state update process of the graph structure data is improved, and at the same time, the generalization performance of the model is improved. Under the condition that the physical communication network is dynamically variable, the stability, accuracy, and efficiency of the digital twin communication network performance prediction can be improved; During the model training process, an optimizer is used to efficiently optimize the model parameters and prevent overfitting. Then, hyperparameter tuning technology is used to evaluate the training effect of the model and optimize the hyperparameters in a timely manner, which can further improve the model prediction accuracy and reduce the running time. The network performance prediction model proposed by the present invention shows higher accuracy compared with traditional models and relatively advanced models when processing the same dataset, and also shows good generalization performance when processing various types of test sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The flowchart of the digital twin communication network performance prediction method in Embodiment 1 of the present invention is shown;

[0051] Figure 2 The flowchart of the digital twin communication network performance prediction method in Embodiment 2 of the present invention is shown;

[0052] Figure 3 The following shows the architecture diagram of the digital twin communication network system in an embodiment of the present invention;

[0053] Figure 4 The following shows the internal structure diagram of the network performance prediction model in an embodiment of the present invention;

[0054] Figure 5 The following shows the schematic diagram of the change in the loss value during the training and validation process in an embodiment of the present invention;

[0055] Figure 6 The following shows the schematic diagram of the comparison result of the cumulative distribution function of the relative error of different test sets in an embodiment of the present invention;

[0056] Figure 7 The following shows the schematic diagram of the comparison result of the prediction performance of different models in an embodiment of the present invention;

[0057] Figure 8 The following shows the internal structure diagram of a computer device in an embodiment of the present invention. Detailed implementation manners

[0058] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0059] Embodiment 1

[0060] As Figure 1 shown, this embodiment provides a method for predicting the performance of a twin network based on a graph neural network, including:

[0061] Obtain real-time relevant network information of the physical communication network and input it into the trained network performance prediction model to obtain a performance prediction result;

[0062] Among them, the determination process of the trained network performance prediction model includes:

[0063] Preprocess the historical relevant network information of the physical communication network obtained from the digital twin communication network to obtain a dependency graph data set, where the dependency graph data set includes link, path, and queue information;

[0064] Initialize the states of each link, path, and queue in the dependency graph data set respectively to obtain the initial hidden states of each link, path, and queue;

[0065] Use the initial hidden state to train a pre - constructed network performance prediction model to obtain a trained network performance prediction model. Among them, the network performance prediction model collaborates with an attention mechanism, a message passing neural network, and a gated recurrent unit to obtain iteratively updated hidden state information, and adjusts model parameters according to the hidden state information.

[0066] By further feature selection and integration of the graph - structured data in the historical dataset, removing the data that has a greater impact on the final result to obtain a more refined training set, and training a network performance prediction model that includes an attention mechanism, a message passing neural network, and a gated recurrent unit with the training set after feature initialization, the stability of the hidden state update process of the graph - structured data can be improved, making the output prediction results more accurate, thereby improving the stability and accuracy of digital twin communication network performance prediction.

[0067] Embodiment 2

[0068] Based on Embodiment 1, the following design is also made in this embodiment, as Figure 2 shown, the digital twin communication network performance prediction method is divided into the following steps.

[0069] Step 1: Construct the digital twin communication network system architecture, and use the historical data obtained by the data collection and transmission module as the dataset for training the network performance prediction model;

[0070] Step 2: Obtain the original network dataset, process the original network data into a dependency graph structure through dependency libraries such as TensorFlow and networkx, and perform normalization pre - processing on the data;

[0071] Step 3: Initialize node features, define node attributes, select the features that have a greater impact on the model performance in the dataset, and construct a feature matrix through feature integration as the input of the network performance prediction model to improve the operation efficiency;

[0072] Step 4: Construct a graph neural network GNN model as the network performance prediction model, and apply this model to digital twin communication network performance prediction;

[0073] Step 5: Train the graph neural network GNN model, define the loss function, optimizer, and related hyperparameters.

[0074] Step 6: Evaluate the trained graph neural network GNN model, select the mean square error MSE, mean absolute error MAE, mean absolute percentage error MAPE between the predicted value and the true value, and the correlation coefficient As a performance metric for evaluating the prediction performance of the algorithm. According to the evaluation results of the model, hyperparameter tuning techniques are used to further optimize the parameters of the model, and the optimal model is selected for the digital twin communication network performance prediction task;

[0075] Step 7: Sequentially transmit the data information collected by the digital twin system in real time to the optimal performance prediction module, so as to realize the real-time prediction of the intelligent communication network performance.

[0076] In step 1, the present invention first constructs the digital twin communication network system architecture.

[0077] The digital twin communication network system architecture is as Figure 3 shown. The whole system consists of three layers, namely the physical network layer, the twin network layer and the network application layer. The physical network layer is similar to the data plane of the software-defined network SDN, and contains the actual information of the network such as network topology, traffic, device status, etc. The twin network layer contains a data sharing warehouse for storing and managing the data collected from the physical network layer, a twin network model for digitally modeling the physical network, algorithm models such as prediction algorithms and optimization algorithms, and network control components. The network application layer is the top layer of the system, contains the specific applications of the model, and is responsible for the overall management and innovation of the network. The system uses sensors, communication protocols, etc. to collect the relevant status information of the physical network, properly processes the massive redundant data through the data transmission module and the data processing module, and saves the processed data to the shared database in the twin layer. The network performance prediction model obtains historical data from the shared database as a data set for model training, and lays a foundation for further network optimization through performance prediction.

[0078] In step 2, since the collected data set, although containing relevant network information, is still not strictly graph-structured data and cannot be directly used for model training, it needs to be further processed. Using dependent libraries such as TensorFlow and networkx in the network library, read the original network samples and convert the original network graph into a directed graph to represent the dependency relationship of the network, forming strictly graph-structured data. At the same time, since there are large differences in the values between different features in the data set, it is necessary to perform normalization processing on some values. Normalization is a mathematical method to simplify the calculation amount, that is, to transform the dimensional expression through processing into a dimensionless expression to become a scalar. The data after normalization processing can eliminate the influence caused by different indicator units and their numerical orders of magnitude, reduce the evaluation error, and can also speed up the model training speed. Therefore, before training, normalize each index of the selected training set data, and the processing method is as follows:

[0079] ,

[0080] In the formula: is the normalized i-th feature data, is the i-th original feature data, and are the maximum and minimum values ​​in the original data set respectively, and i is.

[0081] At the same time, the network dependency graph is defined as a set of links and queues on output ports of network devices and a set of source-to-destination paths Composition, among which, , , , Represents the first link in a group of links Links, represents the number of links in a set of links, Represents the first queues, Indicates the number of queues in a group of queues, Represents the first Path, Indicates the number of paths in a set of paths. In addition, a path can be viewed as a sequence of multiple tuples, each tuple consisting of a queue on a port and the subsequent links, which is used to represent all queues and links contained in the path. Therefore, the path It can be further defined as: ,in, Indicates the path The number of tuples in Represents the first queues, Represents the first A link.

[0082] Finally, use the tf.data.Dataset.from_generator dependency library of TensorFlow to create a TensorFlow dataset, shuffle the new dataset, and selectively filter topological structures of a specific size.

[0083] In step 3, the sample feature dimensions contained in the network dataset used are too many, including some redundant features that have little impact on the final result. Therefore, node feature initialization is performed first, which helps to reduce the complexity of the feature space, thereby reducing the running time and improving the model efficiency.

[0084] Node feature initialization is the process of setting an initial feature representation for each node in a graph before the training of a graph neural network (GNN) model begins. This feature representation is usually a high-dimensional vector that contains various attributes and status information related to the node. In step 2, the composition of the newly generated dependency graph dataset has been defined, including link, path, and queue information, and these three objects contain many features such as data packets, traffic, and priority. Based on this, node feature initialization is performed.

[0085] The specific method for node feature initialization is as follows:

[0086] (1) Initialize the path state path_state by concatenating the traffic feature, the packets feature, and zero-padding tf.zeros to create a 32-dimensional state vector , where the traffic feature represents the data traffic passing through this path, the packets feature represents the number of data packets generated on this path, and the zero-padding is used to form a feature vector of a predetermined dimension.

[0087] (2) Initialize the link state link_state by concatenating the capacity feature, the policy feature, and zero-padding tf.zeros to create a 32-dimensional feature vector , where the capacity feature represents the bandwidth capacity of the link, the policy feature uses one-hot encoding to represent the queue scheduling policy on the link, such as WFQ (Weighted Fair Queuing), SP (Strict Priority), DRR (Deficit Round Robin), FIFO (First In First Out), etc., and finally, a tf.zeros zero-padding is used to meet the dimension requirements.

[0088] (3) Initialize the queue state queue_state by concatenating the size feature, the priority feature, the weight feature, and zero-padding tf.zeros to create a 32-dimensional feature vector , where the size feature represents the current number of data packets in the queue, the priority feature uses one-hot encoding to represent the packet priority, the weight feature represents the queue weight related to the scheduling policy, and the zero-padding meets the dimension requirements.

[0089] (4) Concatenate the state vectors. Use the concatenation function tf.concat of TensorFlow to concatenate the above feature vectors to form a complete node state representation and complete the node state initialization.

[0090] In step 4, a complete graph neural network GNN model is constructed as a network performance prediction model. Refer to Figure 4, The complete GNN model consists of multiple layers of neural networks, including a gated recurrent unit layer GRUCell, an attention layer Attention, a message passing recurrent module, and a readout layer Readout connected in sequence. Among them, the message passing recurrent module includes a mutually connected message passing neural network MPNN layer and a recurrent neural network RNN layer, which are used to execute and iterate the message passing steps.

[0091] In this embodiment, the graph neural network GNN model further includes an Initialization layer, which actually includes the node state initialization process in step 3 and is a preprocessing step before the model starts training. That is, the node state initialization function is integrated into the graph neural network GNN model. The initialization layer uses the call method to execute the forward propagation of the model. It accepts the input data of the model and returns the output of the model.

[0092] The gated recurrent unit layer is used to define the gated recurrent unit GRUCell, which includes three GRUCell instances: self.path_update, self.link_update, and self.queue_update. GRUCell is the basic unit of the gated recurrent unit GRU, equivalent to half a time step of GRU, but GRUCell provides higher flexibility and allows customization of the loop logic. Therefore, GRUCell is used instead of GRU. GRUCell contains an update gate and a reset gate, and these two gating mechanisms allow the model to control the flow of information at each time step. The formula expressions of its components are as follows:

[0093] Update gate:

[0094] ,

[0095] In the formula: is the output of the update gate at time step t, is the update gate weight matrix, is the previous hidden state and the current time step concatenation, is the sigmoid activation function.

[0096] Reset gate:

[0097] ,

[0098] In the formula: is the output of the reset gate at time step t, is the reset gate weight matrix.

[0099] Candidate hidden state:

[0100] ,

[0101] where: is the candidate hidden state, is the weight matrix, is the output of the reset gate and the previous hidden state element-wise product.

[0102] Final hidden state:

[0103] ,

[0104] where: is the final hidden state at time step t.

[0105] Here, a GRUCell instance is first defined and then integrated into the message passing process later.

[0106] The attention layer is used for custom attention calculation. Through the attention function, a fully connected dense layer Dense is used to calculate the attention scores of the input, and then the softmax function is applied to obtain the normalized attention weights.

[0107] The calculation formula for the attention weights is as follows:

[0108] ,

[0109] ,

[0110] where: represents the input matrix, is the bias term, tanh is the hyperbolic tangent function used to introduce non-linearity, scores are the obtained attention scores, softmax is the normalization function, axis equal to 1 means normalizing the scores along the second dimension of the sequence, and attention_weights are the finally obtained normalized attention weights.

[0111] In the message passing loop module, the MPNN (Message Passing Neural Network) layer and the RNN (Recurrent Neural Network) layer are fused to process graph-structured data and time series data. The MPNN aggregates the neighbor information of the nodes in the graph through message passing steps, collects the features from neighbor nodes during the iterative process of message passing, and continuously updates the state information. At the same time, the dynamically changing graph data is also a type of time series data. The GNN model uses the GRU cells in the RNN to remember historical information, thereby processing the node features with time attributes. In the message passing process of the graph neural network, the following principles are mainly followed:

[0112] (1) The state of a path depends on the states of all queues and links traversed by the path.

[0113] (2) The state of a link depends on the states of all queues injecting traffic into the link.

[0114] (3) The state of a queue depends on the states of all paths injecting traffic into the queue.

[0115] The above principles are expressed by the following mathematical formulas:

[0116] ,

[0117] Where: , and are some unknown functions related to the queue state, link state, and path state respectively, represents the hidden state of queue , represents the hidden state of link , represents the hidden state of path , is the set of queues injecting traffic into link , represents the number of elements in path . The state update function and update process in the message passing process of this model will be described in more detail below.

[0118] First, in step 3, the initialization of the node hidden state is completed, and the output includes , and three hidden states. Then, through the message passing loop module, the message passing stage is started. Referring to Figure 4 , during the message passing process, there are T iteration processes, and in each iteration, the model performs message passing in three stages:

[0119] (1) Queue and link to path: Each path collects the information of the queues and links it passes through, and calculates the temporary state generated by the attention mechanism based on the link hidden state at the passing moment and the queue hidden state . Then, using and the path hidden state at this moment as inputs, the gated recurrent unit GRUCell is used to update to obtain the new path temporary hidden state and the new path hidden state .

[0120] (2) Path to Queue: Aggregate the queue and the temporary hidden state information of all paths passing through the queue For all Sum them up to get the sum of the queue's temporary hidden state information Then, according to and the queue hidden state Update through the gated recurrent unit GRUCell to obtain the new queue hidden state as well as the new queue's temporary hidden state .

[0121] (3) Queue to Link: According to the new queue's temporary hidden state and the link hidden state Update through the gated recurrent unit GRUCell to obtain the new link hidden state .

[0122] It can be seen that during the message passing process, the state of a path depends on the queues and links it passes through, while the states of queues and links in turn depend on the paths containing them. Therefore, there is a cyclic dependency relationship. Through multiple iterations, the state of each component gradually stabilizes and can ultimately reflect the global state of the entire network.

[0123] The readout layer of the GNN model is actually a multi-layer perceptron MLP, composed of multiple fully connected layers. The first layer is the input layer, defined using the dependent library tensor flow modular input layer tf.keras.layers.Input in TensorFlow, specifying the shape of the input data as 32. The second and third layers are two fully connected dense layers with the same number of units, using the SELU activation function and applying regularization to prevent the model from overfitting. The fourth layer is also a dense layer but without an activation function, used to directly output the predicted value. The entire Readout layer uses the tensor flow modular sequence model tf.keras.Sequential to construct a linearly stacked neural network, and outputs the predicted value of the performance according to the path state after passing through a series of dense layers and non-linear activation functions.

[0124] In step 5, train the designed graph neural network GNN model. Define the loss function of the model as the mean squared error MSE between the predicted value and the true value, set the optimizer and hyperparameters, and use minimizing the mean squared error as the training objective. The calculation formula of the mean squared error MSE is as follows:

[0125] ,

[0126] In the formula: is the actual value of the performance of the i-th dependency graph sample, is the performance prediction value of the i-th dependency graph sample, and n is the number of samples in the dependency graph dataset.

[0127] In step 6, the trained graph neural network GNN model is evaluated, and the hyperparameters of the model are further tuned according to the evaluation results to ensure the best performance of the trained model. The mean square error MSE, mean absolute error MAE, mean absolute percentage error MAPE, and correlation coefficient between the predicted value and the true value of the communication network performance parameters are selected as the performance metrics to evaluate the prediction performance of the model. Among them, MAE, MAPE, and The calculation formulas are as follows:

[0128] ,

[0129] ,

[0130] ,

[0131] In the formula: is the average of the actual performance values of n dependency graph samples. MAE and MAPE reflect the deviation between the predicted value and the true value. MAE is not sensitive to outliers and is suitable for robust error measurement, while MAPE is suitable for comparing datasets of different scales. is the goodness of fit of the model. The closer its value is to 1, the stronger the explanatory ability of the model. After training the model, it is necessary to evaluate the training effect of the model with the evaluation set or test set, judge whether the training is overfitting while looking for better hyperparameters, and then retrain the model to finally obtain the model with the best prediction performance.

[0132] In step 7, the data information collected by the digital twin system in real time is transmitted to the performance prediction module in sequence, so as to realize the efficient and accurate prediction of the intelligent communication network performance.

[0133] In this embodiment, the present invention uses TensorFlow as the deep learning framework to develop and implement the proposed twin network performance prediction model based on graph neural network. To verify the constructed model, an open-source network dataset generated by a network simulator is selected. In this example, the dataset generated by the OMNet++ network simulator is selected. It contains three datasets. The first is the Nsfnet dataset with approximately 20,000 samples composed of a 14-node topology communication network and a 24-node topology communication network. The second is the Geant dataset with approximately 1,000 samples composed of a 22-node topology communication network. The third is the Abilene hybrid dataset with approximately 1,000 samples composed of various real-world network topologies and real traffic matrices. All datasets contain relevant network information such as network topology, routing configuration, partial network configuration (traffic matrix, routing, queue scheduling policy), and some measurable network performance parameters (such as end-to-end delay).

[0134] During the experiment, the end-to-end delay of data packets from the source to the destination is mainly used as the network performance prediction target. The Nsfnet dataset is selected for training, and the 20,000 samples are divided into a train training set, an eval validation set, and a test test set according to the ratio of 8:1:1. The optimizer for training adopts the Adam algorithm for adaptive moment estimation, sets the initial learning rate to 0.001, sets the number of iteration steps to 400,000 steps, the decay rate of the learning rate to 0.7 and decays once every 40,000 steps, sets the batch size of the training data to 32, and the regularization coefficient of the readout layer is set to 0.001. A checkpoint is saved every 20 minutes during training and validation is performed on the validation set once. Figure 5 Describes the change in the loss function, i.e., the mean square error between the true value and the predicted value, on the training and validation sets during the model training process, which is Figure 5 It can be seen that as the number of training steps increases, the mean prediction error of the model on the validation set becomes smaller and smaller, and gradually converges. Finally, the mean square error stabilizes at about 0.008. At the same time, the loss functions on the training set and the validation set decrease synchronously, and there is no overfitting phenomenon.

[0135] Figure 6 Describes the test results of the trained GNN model on different test sets, mainly the cumulative distribution function (CDF) graph of the relative error (MRE) between the predicted value and the true value. Among them, the calculation formula of MRE is:

[0136] .

[0137] CDF represents the probability that a random variable is less than or equal to a specific value, and the formula is expressed as:

[0138] ,

[0139] In the formula: is the value of the cumulative distribution function, is a random variable, is a certain value that the random variable may take. In this experiment, the relative error between the predicted value and the true value is used as the random variable. The cumulative distribution function CDF is the sum of discrete variables and presents a trapezoidal structure. The more concentrated the trapezoidal line is at the center point position, that is, the relative error is concentrated near 0, the better the prediction effect. As can be seen from Figure 6 the model has good test results for each dataset. Among them, the Nsfnet dataset is the topology type seen during training and has the best effect. The Geant dataset has a larger network scale than the training set, and the test effect is still good. The Abilene dataset contains various real network topologies and real traffic matrices, and the test results on this dataset are slightly less accurate than the previous two. Generally speaking, the model shows good generalization performance. The specific experimental results of other evaluation indicators on the three datasets are shown in Table 1.

[0140] Table 1 Performance of the model on different test sets

[0141] Dataset MAE MSE MAPE r² / % Nsfnet 0.0482 0.0099 0.0251 99.47 Geant 0.0702 0.0126 0.0302 98.32 Abilene 0.0712 0.0093 0.0351 97.12

[0142] For further verification, the same Nsfnet dataset is selected as the training set, and the Abilene dataset is selected as the test set. The model proposed in the present invention is called the gated recurrent-attention GRU-AT model, and it is compared with the multi-layer perceptron MLP model and the routing network RouteNet model. Figure 7 describes the comparison of the network performance prediction results of different models. Among them, the MLP model consists of a four-layer network of non-linear activation nodes (an input layer, an output layer, and two hidden layers), takes the traffic matrix and queue and link configuration information as inputs, and models the end-to-end delay through the multi-layer perceptron MLP. The RouteNet model is similar to the model of the present invention and also uses a graph neural network model to perform network modeling by executing two-stage message passing between links and paths in the network.

[0143] The specific experimental results of the three network performance prediction models are shown in Table 2. Since the network-related information of the test set, including topology, traffic matrix, etc., is completely different from that of the training set, the prediction errors of the MLP model and the RouteNet model are relatively large, indicating poor generalization performance, while the GRU-AT model can still maintain high accuracy. In terms of the time consumed for each iteration of training, the internal structure of the MLP model is simple, and the time required for each step of training is the shortest. The GRU-AT model requires less time than the RouteNet model because the extracted features are more representative and the GRU and attention mechanisms are used in the message passing process.

[0144] Table 2 Performance Comparison of Different Models

[0145] Model MAE MSE MAPE r² / % Time / s MLP 0.295 16.31 0.712 3.112 0.112 RouteNet 0.351 14.45 0.351 7.268 0.281 GRU-AT 0.071 0.009 0.035 97.12 0.175

[0146] Starting from the network performance prediction of the digital twin communication network, this invention analyzes the advantages and disadvantages of several network performance prediction methods commonly used today. On this basis, a twin network performance prediction model based on graph neural network is proposed, which solves some defects in today's network performance prediction methods to a certain extent and improves the operation efficiency and prediction accuracy. Therefore, the digital twin communication network performance prediction model proposed in this invention is feasible.

[0147] Embodiment 3

[0148] This embodiment provides a twin network performance prediction device based on graph neural network, including:

[0149] Performance prediction result acquisition module: used to acquire the real-time relevant network information of the physical communication network, input it into the trained network performance prediction model, and obtain the performance prediction result;

[0150] Among them, the determination process of the trained network performance prediction model includes:

[0151] Preprocess the historical relevant network information of the physical communication network obtained from the digital twin communication network to obtain a dependency graph dataset, where the dependency graph dataset includes link, path, and queue information;

[0152] Initialize the features of each link, path, and queue status in the dependency graph dataset to obtain the initial hidden states of each link, path, and queue;

[0153] Use the initial hidden states to train a pre-constructed network performance prediction model to obtain a trained network performance prediction model, where the network performance prediction model collaborates with an attention mechanism, a message passing neural network, and a gated recurrent unit to obtain iteratively updated hidden state information and adjusts the model parameters according to the hidden state information.

[0154] Example 4

[0155] This embodiment provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the twin network performance prediction method based on a graph neural network as described in any step of Embodiment 2.

[0156] Example 5

[0157] An embodiment of the present invention further provides a computer device, which may be a server, and its internal structure diagram may be as Figure 8 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the data obtained and generated in the method for a robot to autonomously enter a packaging container. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements the method of the aforementioned Embodiment 2.

[0158] Those skilled in the art can understand that Figure 8 the structure shown in

[0159] is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component layout.

[0160] Those skilled in the art should understand that the embodiments of the present application may be provided as a method, a system, or a computer program product. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc. that contain computer-usable program code.

[0161] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus systems, and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device generate means for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0162] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0163] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0164] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the claims of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A twin network performance prediction method based on graph neural network, characterized in that: include: Obtain real-time relevant network information of the physical communication network, input it into the trained network performance prediction model, and obtain performance prediction results; The process of determining the trained network performance prediction model includes: Preprocessing the historical network information related to the physical communication network obtained by the digital twin communication network to obtain a dependency graph data set, wherein the dependency graph data set includes link, path and queue information; Initializing the features of each link, path and queue in the dependency graph data set to obtain the initial hidden state of each link, path and queue; Using the initial hidden state to train a pre-built network performance prediction model to obtain a trained network performance prediction model, wherein the network performance prediction model obtains iteratively updated hidden state information through the collaborative work of an attention mechanism, a message passing neural network, and a gated recurrent unit, and adjusts model parameters according to the hidden state information; The physical communication network history related network information obtained by the digital twin communication network is preprocessed to obtain a dependency graph data set, including: Normalize the read original network samples to obtain a standardized data set; The original network samples in the standardized data set are converted into a network dependency graph to obtain a dependency graph data set, wherein the network dependency graph includes a set of links , a set of queues on the output ports of a network device and a set of source-to-destination paths .

2. The twin network performance prediction method based on graph neural network according to claim 1 is characterized in that: The link ,queue and the source to destination path The expressions are as follows: , where Represents the first link in a group of links Links, represents the number of links in a set of links; , where Represents the first queues, Indicates the number of queues in a group of queues; , where Represents the first Path, Represents the number of paths in a set of paths, represented by a series of tuples , , represent The number of elements in Represents the first queues, Represents the first Links; The state of each link, path and queue information in the dependency graph data set is initialized to obtain the initial hidden state of each link, path and queue, including: Path state initialization: Create a first state vector by connecting traffic characteristics, packet number characteristics, and zero padding , where the traffic feature represents the data traffic passing through the path, the packet number feature represents the number of packets generated on the path, and zero padding is used to form a feature vector of a predetermined dimension; And / or, Link State Initialization: Create a second state vector by connecting capacity characteristics, policy characteristics, and zero padding , where the capacity feature represents the bandwidth capacity of the link, the policy feature represents the queue scheduling policy on the link, and zero padding is used to meet the dimensionality requirements; And / or, queue state initialization: create a third state vector by concatenating the size feature, priority feature, weight feature, and zero padding , where the size feature represents the current number of packets in the queue, the priority feature represents the packet priority, the weight feature represents the queue weight associated with the scheduling policy, and zero padding is used to meet the dimensionality requirement.

3. The twin network performance prediction method based on graph neural network according to claim 2 is characterized in that: The pre-built network performance prediction model includes: a GRUCell gated recurrent unit layer, an Attention layer, a message passing loop module and a Readout layer connected in sequence; the message passing loop module includes an MPNN message passing neural network layer and an RNN recurrent neural network layer connected to each other.

4. The twin network performance prediction method based on graph neural network according to claim 3 is characterized in that: The using the initial hidden state to train the pre-built network performance prediction model to obtain a trained network performance prediction model includes: A path cycle unit, a link cycle unit and a queue cycle unit are defined respectively through the GRUCell gated cycle unit layer; In combination with the Attention layer, the MPNN message passing neural network layer and the RNN recurrent neural network layer, the hidden state information of the path cycle unit, the link cycle unit and the queue cycle unit is iteratively updated starting from the initial hidden state through a message passing process to obtain the global state of the physical communication network; Outputting the performance prediction value according to the updated path hidden state through the Readout readout layer; The network performance prediction model is trained according to the mean square error MSE between the performance prediction value and the true value, and the model hyperparameters are set, wherein the training objective of the network performance prediction model is It is expressed as: ， Where: is the actual performance value of the i-th dependency graph sample, is the performance prediction value of the i-th dependency graph sample, is the number of samples in the dependency graph dataset, and i represents the sequence number of the sample in the dependency graph dataset.

5. The twin network performance prediction method based on graph neural network according to claim 4 is characterized in that: The method further includes: selecting a test set from the dependency graph data set, inputting the test set into the network performance prediction model generated by each n rounds of training and the last round of training, and calculating the mean absolute error (MAE), mean absolute percentage error (MAPE) and correlation coefficient between the predicted value and the true value of the network performance prediction model. Evaluate the model training effect, including: Calculate the mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient between the predicted value and the true value of the network performance prediction model , the calculation formula is: ， ， ， Where: is the average of the actual performance values ​​of n dependency graph samples.

6. The twin network performance prediction method based on graph neural network according to claim 5 is characterized in that: The iterative updating of the hidden state information of the path cycle unit, the link cycle unit and the queue cycle unit through the message passing process includes: in each iteration, the network performance prediction model performs the following three stages of message passing: Queue and link to path: collects information about the queues and links contained in each path, and hides the status of the links at a fixed time. and queue hidden state Compute the temporary state generated by the attention mechanism ,according to and the path hidden state at that moment As input, the GRU unit in the RNN recurrent neural network layer is used to update the hidden state of the new path and the new path temporarily hides the state ; Path to Queue: aggregates temporary hidden state information of the queue and all paths containing the queue , temporarily hide all status information After summing, we get the sum of temporary hidden state information of the queue. , based on the sum of the temporarily hidden state information of the queue and queue hidden status The new queue hidden state is obtained by updating the GRU unit in the RNN recurrent neural network and the new queue temporarily hides the state ; Queue to link: Temporarily hide state based on new queue and link hidden status , the new link hidden state is obtained by updating the GRU unit in the RNN recurrent neural network .

7. A twin network performance prediction device based on graph neural network, characterized in that: include: Performance prediction result acquisition module: used to obtain real-time relevant network information of the physical communication network, input it into the trained network performance prediction model, and obtain the performance prediction result; The process of determining the trained network performance prediction model includes: Preprocessing the historical network information related to the physical communication network obtained by the digital twin communication network to obtain a dependency graph data set, wherein the dependency graph data set includes link, path and queue information; Initializing the features of each link, path and queue in the dependency graph data set to obtain the initial hidden state of each link, path and queue; Using the initial hidden state to train a pre-built network performance prediction model to obtain a trained network performance prediction model, wherein the network performance prediction model obtains iteratively updated hidden state information through the collaborative work of an attention mechanism, a message passing neural network, and a gated recurrent unit, and adjusts model parameters according to the hidden state information; The physical communication network history related network information obtained by the digital twin communication network is preprocessed to obtain a dependency graph data set, including: Normalize the read original network samples to obtain a standardized data set; The original network samples in the standardized data set are converted into a network dependency graph to obtain a dependency graph data set, wherein the network dependency graph includes a set of links , a set of queues on the output ports of a network device and a set of source-to-destination paths .

8. A computer storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the twin network performance prediction method based on graph neural network as described in any one of claims 1 to 6 is implemented.

9. An electronic terminal, characterized in that: It includes a processor and a memory connected to the processor, wherein a computer program is stored in the memory. When the computer program is executed by the processor, the twin network performance prediction method based on graph neural network as described in any one of claims 1 to 6 is executed.

Images