TSN network configuration detection method and device based on data enhancement and GCN network

Abstract

The invention discloses a TSN network configuration detection method and device based on data enhancement and a GCN network, and the method comprises the steps: screening features through employing an MI index, and screening out the features with higher correlation degree; a CTGAN algorithm is introduced to carry out data enhancement on an original data set, and the proportion of feasible and infeasible configurations in FIFO, Manial and CP8 scheduling methods in the data set is calculated; expanding minority class samples in the data set through a generative adversarial network (GAN) after modal normalization and condition vector optimization; a GCN neural network is selected to carry out network configuration detection, features obtained through screening are used as node features, link loads are introduced to serve as edge features, adjacent nodes are gathered through two layers of graph convolution layers and normalization layers in a rolling mode, overfitting is reduced, and finally network configuration information is fully extracted through global average pooling GAP and a full connection layer MLP. According to the method, through CTGAN data enhancement and the GCN neural network, the classification capability of the model on the unbalanced data set is enhanced, and the classification accuracy is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of graph convolutional neural networks and time-sensitive networks, and in particular to a TSN network configuration detection method and device based on data enhancement and GCN networks. Background Art

[0002] The rapid development of next-generation information and communication technologies, such as the Industrial Internet of Things (IIoT), has brought tremendous opportunities and challenges to the manufacturing industry, driving the transformation of manufacturing enterprises in emerging economies. In this context, real-time communication capabilities have become crucial. While traditional Ethernet offers excellent bandwidth and versatility, its limitations in real-time and deterministic performance make it difficult to meet the stringent communication requirements of certain industrial scenarios with high quality of service (QoS) requirements, such as industrial automation and automotive electronics. To address this, Transmission Switching (TSN) technology has emerged. TSN is a family of protocols that define mechanisms and protocols that implement various QoS strategies, such as prioritization, traffic shaping, time-triggered traffic, and frame preemption. Through these protocols, TSN aims to introduce real-time and deterministic performance to traditional Ethernet, ensuring the timely transmission of high-priority traffic within the network.

[0003] To enhance the flexibility and convenience of TSN configuration, the IEEE 802.1 working group has proposed a series of protocols to optimize network configuration and management models. As networks expand in size and complexity, and as a result, deploying a large number of sensors on industrial networks makes TSN network configuration a challenging task. The accuracy of TSN network configuration directly impacts the performance and stability of the entire system. A configuration that does not meet system requirements can increase data transmission latency or even cause data loss, impacting the execution of critical tasks. Researching how to effectively test TSN network configurations to ensure they meet expected performance standards is fundamental to reliable system operation.

[0004] A major model-based approach to assessing whether a real-time system meets its timing constraints is schedulability analysis. Schedulability analysis, also known as feasibility analysis, worst-case analysis, response time analysis, or worst-case traversal time analysis (WCTT in networking), involves deriving upper bounds on performance metrics (such as communication delay and buffer utilization) from a mathematical model of the system.

[0005] The design of complex embedded systems, such as the electrical / electronic (E / E) architecture of automotive platforms, requires the use of design space exploration (DSE) algorithms. DSE algorithms typically involve three distinct steps: creating candidate solutions, configuring these solutions, and performing performance evaluation using throughput constraints or worst-case schedulability analysis. However, performance evaluation is a computationally intensive process, which severely limits the size of the search space that can be explored.

[0006] In TSN, the choice of a scheduling strategy is crucial for ensuring real-time data transmission, meeting timing constraints, and optimizing network performance. Its purpose is to limit traffic as it leaves the queue for link transmission, thereby improving the determinism of data transmission. Different strategies have different trade-offs in meeting transmission determinism requirements (such as time constraints). The three commonly used scheduling strategies are as follows: 1. First-in-first-out (FIFO) scheduling: There is only one traffic class, and all flows have the same priority.

[0007] 2. Manual Priority Scheduling: This method is considered a baseline solution, attempting to prioritize traffic based on its relative criticality. Traffic is categorized into three types: audio, video, and command and control. These three types of traffic are assigned different priorities, with the following relationship: command and control > audio > video.

[0008] 3. "Concise Priority" scheduling with eight priority levels (CP8): "Concise Priority" is the name of the priority assignment algorithm in RTaW-Pegase, which can assign the same type of traffic to different traffic classes. This algorithm implements the optimal priority assignment algorithm.

[0009] The above scheduling solutions belong to QoS strategies using static priority scheduling. They can be implemented with basic TSN network equipment that only provides eight priority levels. These three scheduling schemes are compatible with standard static priority scheduling analysis, that is, calculating the worst-case traversal time (WCTT) of each flow as the basis for judging the feasibility of each configuration.

[0010] To address the issue of efficiently detecting the feasibility of TSN network configurations, deep learning algorithms using graph neural network (GNN) models have made significant progress in classification detection. This approach uses a global vector containing graph information and the correlations between its components as the neural network input. This method demonstrates robustness in adapting to changes in the training set topology and traffic patterns. However, even after optimizing the training set construction and improving the basic GNN model, the average classification accuracy remains insufficient for certain scheduling methods, such as Concise Priority (CP). Summary of the Invention

[0011] The present invention proposes a TSN network configuration detection method, device, and storage medium based on data enhancement and GCN network, which can solve at least one of the technical problems in the background technology.

[0012] To achieve the above object, the present invention adopts the following technical solutions: A TSN network configuration detection method based on data enhancement and GCN network includes the following steps: (1) Preprocess the original network configuration data set, compare the WCTT of each flow with the system's time constraints on the flow, label the configuration feasibility, and screen out the seven most relevant features based on the MI value.

[0013] (2) The dataset is divided into a training set and a test set. The minority class samples in the training set are used as real samples for the CTGAN data augmentation algorithm. In the generative adversarial network after modal normalization and conditional vector optimization, the generator and the discriminator conduct game adversarial learning.

[0014] (3) The discriminator improves its ability to distinguish fake samples generated by the generator in each iterative update, and the generator improves the similarity between the generated fake samples and the real samples. After the iterative training is completed, the generated samples are added to the minority class samples.

[0015] (4) The expanded training set is sent to the constructed GCN network model with a topological structure consistent with the data set. The Gini coefficient is calculated from the load of each link and a total of seven feature data are obtained as node features through feature screening. The link load data in the network topology is used as edge features.

[0016] (5) Through the convolution aggregation of the graph convolution layer, the feasibility information of the network configuration is gradually extracted. The activation function and batch normalization layer are added after the convolution layer. The feature averaging of the global average pooling layer and the fully connected layer are then used to output the classification results.

[0017] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0018] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0019] From the above technical solution, it can be seen that the TSN network configuration detection method based on data enhancement and GCN network of the present invention proposes a graph convolutional neural network (GCN) model based on feature screening and conditional table generative adversarial network (CTGAN) data enhancement in the TSN environment to perform TSN network configuration detection, expand the data set to address the imbalance problem of the original data set, and use GCN to aggregate adjacent nodes to improve the accuracy of network configuration detection.

[0020] This paper aims to optimize TSN network configuration detection. It introduces the CTGAN algorithm to enhance the original configuration dataset, leveraging the game learning of a generative adversarial network (GAN) to oversample minority class samples in the dataset. This method expands the training set and enhances the model's sensitivity to minority class samples. A GCN neural network is used for TSN network configuration detection. Mutual information (MI) is used to filter out highly relevant features, and link load is introduced as an edge feature to help the model fully extract network configuration information during the convolutional aggregation of adjacent nodes.

[0021] This paper applies a graph convolutional neural network (GCN) model based on feature screening and data augmentation using a conditional table generative adversarial network (CTGAN) to TSN network configuration feasibility testing. This method uses the CTGAN algorithm to enhance the original dataset, calculating the ratio of feasible to infeasible configurations for the FIFO, Manual, and CP8 scheduling methods. It then augments the minority class samples in the dataset using a generative adversarial network (GAN) with modal normalization and conditional vector optimization. A GCN neural network is also used for network configuration testing. Features obtained through feature screening are used as node features, and link load is introduced as edge features. Two graph convolutional layers and a normalization layer are used to aggregate adjacent nodes and reduce overfitting. Finally, global average pooling (GAP) and fully connected layers (MLP) are used to fully extract network configuration information. This method, through CTGAN data augmentation and the GCN neural network, strengthens the model's classification capabilities for imbalanced datasets, further improving classification accuracy.

[0022] Specifically, this paper introduces the CTGAN algorithm to enhance the TSN network configuration dataset, leveraging the game learning of a generative adversarial network (GAN) to oversample minority class samples in the dataset. Furthermore, based on feature selection using MI, a GCN neural network is selected for network configuration detection. This method, through CTGAN data augmentation and the GCN neural network, strengthens the model's classification capabilities for imbalanced datasets, thereby improving classification accuracy.

[0023] The present invention calculates the ratio of feasible to infeasible configurations in the FIFO, Manual, and CP8 scheduling methods in the original dataset. According to the degree of imbalance in the scheduling method subsets, a generative adversarial network (GAN) after modal normalization and conditional vector optimization is introduced. The generator and discriminator are iteratively trained to expand the minority class samples in the dataset.

[0024] The present invention selects the GCN neural network to perform TSN network configuration detection on the expanded data set, uses the highly correlated features screened out by MI as node features, and introduces link load as edge features. Two layers of graph convolution layers convolute and aggregate adjacent node features, and use a normalization layer (BN) to reduce overfitting. Global average pooling (GAP) is used to reduce the node features to one dimension, and finally a fully connected layer (MLP) is used to fully extract the network configuration information to obtain the classification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is the topology used in the experiment of the present invention; Figure 2 is a flow chart of the MI-CTGAN-GCN model in an embodiment of the present invention; Figure 3 This is the process of data enhancement by CTGAN in the embodiment of the present invention; Figure 4 is the GCN network model constructed in the embodiment of the present invention; Figure 5 This is a comparison chart of the TSN network configuration accuracy of four models in the embodiment of the present invention; Figures 6(a) and 6(b) are TPR and TNR comparison diagrams of the three data augmentation models in the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0027] To address the class imbalance in the original dataset and improve the accuracy of deep learning algorithms in detecting TSN network configuration feasibility, this paper proposes a new model construction method. This method uses the MI metric to screen features and select those with high correlation. The Conditional Table Generative Adversarial Network (CTGAN) algorithm is then used to generate minority class samples, addressing the class imbalance in the dataset. Finally, a graph convolutional neural network (GCN) model is used to detect TSN network configurations, effectively enhancing the model's classification capabilities for imbalanced datasets and improving classification accuracy.

[0028] The technical solution adopted by the present invention is: a TSN network configuration detection method based on data enhancement and GCN network based on feature screening and data enhancement, the flow chart is as follows Figure 2 As shown, the following steps are included: (1) First, the network configuration raw data set is preprocessed. The WCTT of each flow is compared with the time constraint of the system on the flow, and the configuration feasibility is labeled. At the same time, the seven most relevant features are selected based on the MI value. The seven features represent the total number of flows in the network topology, the number of key flows, the number of audio flows, the number of video flows, the maximum load of the link, the time required to execute WCTT, and the link Gini coefficient calculated based on the load of each link.

[0029] (2) The original configuration dataset is divided into a test set and a training set in a ratio of 0.85:0.15. The percentage of feasible configurations in each scheduling method is calculated by the ratio of feasible and infeasible labels in the original dataset, that is, f The minority class configuration samples in the training data set are used as the objects for CTGAN data enhancement. The specific process is as follows Figure 3 As shown in Figure 2, CTGAN is a data augmentation method based on generative adversarial networks (GANs). GANs are deep learning models that enable unsupervised learning on complex distributions. GANs have two modules: the generator (G) and the discriminator (D). These two modules interact to learn and generate new data. Essentially, the goal is to ensure that the generator's mapping of random noise closely matches real data.

[0030] (3) Perform data normalization on the minority class samples that need data augmentation obtained in (2). Use one-hot vectors to normalize the discrete columns in the training set, and perform mode-specific normalization on the continuous columns with complex distributions. The steps are as follows: (3.1) For each continuous column in the dataset , use the variational Gaussian mixture model (VGM) to fit the Gaussian mixture and estimate the number of modes The learned Gaussian mixture can be expressed as:

[0031] in is the mathematical expectation of the pattern, is the weight of the pattern, is the standard deviation of the pattern.

[0032] (3.2) Yes Each value in , calculate the probability of it coming from each mode. The probability density calculation formula is:

[0033] (3.3) Sample a pattern from a given probability density and normalize the values ​​using the sampled pattern. Use a one-hot vector to represent The sampled mode and using scalar Represents a value within a pattern. The calculation formula is as follows:

[0034] (4) Introducing conditional vectors cond To indicate the condition To guide the generator G to generate samples. Convert to one-hot vector ,in ,for . Set i One-hot vector No. i The mask vector is ,for , the condition can be expressed by the mask vector as:

[0035] Then the condition vector cond The available mask vector can be expressed as: . cond Used to assist G in generating samples. G consists of two fully connected hidden layers and an output layer and uses the Relu activation function. During the training process, the standard MVN variable is input into G and mapped to generate a set of unique hot vectors. During training, by adding and The cross entropy between them is used to penalize the loss, so that G will give the given Copy to the generated After training, G can learn to generate samples that meet specific conditions, thereby generating false samples.

[0036] (5) The fake samples obtained in step (4) serve as one input to the discriminator D. The minority class samples in the original dataset generate real samples, which participate in the game learning between G and D and serve as the other input of D. D consists of two fully connected hidden layers and an output layer. Leaky relu is used as the activation function, and dropout is added after the hidden layer to reduce overfitting of the model. The fake samples and real samples are sent to D for discrimination, and the discrimination result of the samples is output.

[0037] (6) Calculate two loss functions based on the discriminant results and , and back propagation alternately iterates to update the network model parameters of G and D. The two are constructed as follows:

[0038]

[0039] in m is the number of samples in the batch, x represents the real sample, represents a false sample, D represents the output of the discriminator, H is the cross entropy score, which is always positive. D's network model improves its ability to distinguish fake samples from real samples with each iteration. G improves the similarity between generated fake samples and real samples with each iteration to confuse the discriminator. After iterative training, G generates data augmentation samples that are highly similar to real samples.

[0040] (7) Use the data augmentation samples obtained in step (6) together with the training set in the original dataset as the expanded training set to train the constructed GCN network model. The network topology corresponding to GCN is as follows Figure 1 As shown in the figure, the network consists of 5 switches and 14 nodes, namely electronic control units (ECUs). Except for the link from ECU12 to Switch3, which is 1Gbps, the data transmission rate of all links is 100Mbps. The core idea of ​​GCN is to aggregate the information of neighboring nodes to update the features of each node. The Gini coefficient is calculated from the load of each link and a total of seven feature data are obtained as node features through feature screening, and the link load data in the network topology is used as edge features. GCN consists of two layers of graph convolution layers, and a batch normalization layer (BN) is added to reduce overfitting, and ReLU is used as the activation function to provide more efficient gradient descent and back propagation. The convolution layer gradually extracts information about the feasibility of the configuration by convolution of each network configuration data. The GCN model is as follows Figure 4 As shown. The expression of the ReLU activation function is as follows:

[0041] (8) After the network configuration feature information in the graph is aggregated by convolution in step (7), the graph is subjected to global average pooling (GAP) to average all node features. In this process, the seven-dimensional features of each node in the topology are converted into one-dimensional features. , the formula is as follows:

[0042] in for The global eigenvector of for The number of nodes in for Middle n The feature vector of each node. The pooled features are mapped to the output dimension through MLP to obtain the classification result. The cross entropy loss function updates the parameters of the GCN classification model through back propagation. The cross entropy loss function formula is as follows:

[0043] in N is the number of categories. Since this task is a binary classification task, , is the true label, is the predicted probability of the model.

[0044] (9) The GCN network model obtained by training the training set after data enhancement is used to classify the network configuration feasibility of the test sample set in the original data set, and the final classification results are output.

[0045] The following is a simulation experiment: Environment Configuration: The original dataset was collected from an automotive electronics system. The flow characteristics, their proportions, and the network topology were provided by an automotive OEM. During data transmission between ECUs, configurations to be tested are generated, consisting of different types of traffic flows. This experiment partitioned the original dataset into training and test sets with a ratio of 0.85:0.15.

[0046] The generator and discriminator in CTGAN are composed of two fully connected hidden layers and one output layer. The dimension of the fully connected hidden layer is 256, the learning rate of the Adam optimizer is 0.0002, and 600 iterations are performed. The first graph convolution layer of the GCN model has an input feature dimension of 7 and an output feature dimension of 32. The second graph convolution layer has an input feature dimension of 32 and an output feature dimension of 64. The batch size is 32, the learning rate is set to 0.001, and 350 iterations are performed. The present invention generates 3642, 248, and 2466 samples for the FIFO, CP8, and Manual scheduling methods respectively by using the CTGAN algorithm. f The value is balanced to 0.5, that is, the number of feasible and infeasible configurations is the same, and they are added to the training set to train the model.

[0047] Simulation results: Figure 5The accuracy of the GCN model for three scheduling methods was compared without data augmentation, and with three different data augmentation methods: Borderline-SMOTE (BS), standard GAN, and CTGAN. Previous research indicates that using a GNN model on the same topology achieves a maximum detection accuracy of 90.5% for the "Concise Priority" (CP) scheduling method. This demonstrates that the GCN model constructed in this paper significantly improves the accuracy of previous algorithms, with even greater improvements after data augmentation. The GCN model without data augmentation achieves an accuracy of 91.41% for the CP8 scheduling method, and this accuracy is further improved to 92.56% after using CTGAN data augmentation. The classification accuracy of all three scheduling algorithms demonstrates a significant improvement after data augmentation, with the CTGAN algorithm showing a clear advantage over the Borderline-SMOTE and standard GAN algorithms. The CTGAN-GCN model achieves an accuracy of 96.28% for the FIFO scheduling method. This result shows that the training set obtained by the CTGAN data augmentation algorithm can enhance the classification ability of the GCN model for imbalanced data sets.

[0048] Figures 6(a) and 6(b) compare the true positive rate (TPR) and true negative rate (TNR) for the three data augmentation methods. The positive examples in the FIFO and Manual scheduling methods, as well as the negative examples in the CP8 scheduling method, are all minority class examples. The simulation results comparing the true negative rates of the FIFO and Manual scheduling methods and the true negative rate of CP8 show that the CTGAN data augmentation algorithm can most significantly enhance the model's classification ability for minority class examples, and is most effective in improving the model's classification detection TPR and TNR.

[0049] Table 1

[0050] Table 1 shows the statistics of precision, recall, F1 value, and Kappa coefficient (K value) for the three scheduling methods and the corresponding four models. It can be seen that after data augmentation, the model performance improved, with varying degrees of increase in all four statistical metrics. The CTGAN-GCN algorithm achieved the most significant model improvement, with the most significant increases in precision, F1 value, and K value. The occasional decrease in recall is due to the model sacrificing some ability to classify majority class samples in order to improve its ability to classify minority class samples. This result shows that the CTGAN-GCN algorithm helps alleviate the class imbalance problem in the dataset, thereby helping the model capture more features of minority class samples, reducing overfitting of majority class samples, and reducing the randomness of model predictions.

[0051] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0052] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0053] In another embodiment provided by the present application, a computer program product comprising instructions is further provided. When the computer is run on the computer, the computer executes any of the TSN network configuration detection methods based on data enhancement and GCN network in the above embodiments.

[0054] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above methods.

[0055] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0056] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0057] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A TSN network configuration detection method based on data enhancement and GCN network, characterized in that: The following steps are included: S1. Obtain the original network configuration data set and perform data preprocessing. Use the MI index to filter features and select features that meet the required correlation. S2. Divide the original data set into training set samples that need data enhancement in proportion; S3, through the generative adversarial network GAN (CTGAN) after modal normalization and conditional vector optimization, the training set samples divided from the original dataset are expanded; S4. Select the GCN neural network for network configuration detection and use the expanded training set for training. The features obtained by filtering the MI indicator are used as node features, and the link load is introduced as the edge feature. Two layers of graph convolution and normalization layers are used to aggregate adjacent nodes and reduce overfitting. Finally, global average pooling (GAP) and fully connected layer (MLP) are used to fully extract the network configuration information.

2. The TSN network configuration detection method based on data enhancement and GCN network according to claim 1 is characterized in that: Step S1 specifically includes: First, the original network configuration data set is preprocessed, the WCTT of each flow is compared with the system's time constraint on the flow, the configuration feasibility is labeled, and the most relevant features are screened out based on the MI value.

3. The TSN network configuration detection method based on data enhancement and GCN network according to claim 2 is characterized in that: Step S2 specifically includes: The original configuration data set is divided into a test set and a training set in a ratio of 0.85:0.15; The percentage of feasible configurations in each scheduling method is calculated by the ratio of feasible and infeasible configuration labels in the original dataset, that is, f The minority class configuration samples in the training dataset are used as the objects for CTGAN data augmentation.

4. The TSN network configuration detection method based on data enhancement and GCN network according to claim 3 is characterized in that: Step S3 includes, Normalize the discrete columns in the training set using one-hot vectors and perform modal normalization on continuous columns with complex distributions as follows: (3.1) For each continuous column in the dataset , use the variational Gaussian mixture model VGM to fit the Gaussian mixture and estimate the number of modes ; The learned Gaussian mixture is represented as: in for Column No. values, For the The mathematical expectation of the pattern, For the The weight of the pattern, For the The standard deviation of each mode; (3.2) Yes Each value in , calculate the probability of it coming from each mode, and the probability density calculation formula is: (3.3) Sample a pattern from a given probability density and normalize the values ​​using the sampled pattern; use a one-hot vector to represent The sampled mode and using scalar Represents a value within a pattern; The calculation formula is as follows: 。 5. The TSN network configuration detection method based on data enhancement and GCN network according to claim 4 is characterized in that: Step S3 also includes, Introducing conditional vectors cond To indicate the condition To guide the generator G to generate samples, the discrete columns in the dataset Convert to one-hot vector ,in ,for ; Set up i One-hot vector No. i The mask vector is ,for , the condition is represented by the mask vector as: Then the condition vector cond The mask vector is expressed as: ; cond Used to assist the generator G in generating samples. The generator G consists of two fully connected hidden layers and one output layer and uses the Relu activation function; During training, the standard MVN variables are fed into the generator G and mapped to produce a set of unique hot vectors ; During training, by adding and The cross entropy between them is used to penalize the loss, so that the generator G will give the given Copy to the generated In the above, the trained generator G can learn to generate samples that meet specific conditions, thereby generating false samples; The obtained fake samples are used as an input of the discriminator D, and the minority class samples in the original data set generate real samples, participate in the game learning between the generator G and the discriminator D, and serve as another input of the discriminator D; The discriminator D consists of two fully connected hidden layers and an output layer. It uses leaky relu as the activation function and adds dropout after the hidden layer to reduce overfitting of the model. The fake samples and real samples are sent to the discriminator D for discrimination, and the discrimination result of the samples is output. Calculate two loss functions through the discrimination results and , and back propagate alternately and iteratively update the network model parameters of the generator G and the discriminator D. The two are constructed as follows: in m is the number of samples in the batch, x represents the real sample, represents a false sample, D represents the output of the discriminator, H is the cross entropy score, which is always positive. The network model of the discriminator D improves its ability to distinguish between false samples and real samples in each iterative update. The generator G improves the similarity between the generated false samples and real samples in each iteration to confuse the discriminator. The generator G after iterative training is used to generate data augmentation samples that are highly similar to real samples.

6. The TSN network configuration detection method based on data enhancement and GCN network according to claim 5 is characterized in that: S4 includes, Use the obtained data augmentation samples together with the training set in the original dataset as the expanded training set to train the constructed GCN network model; The GCN network model aggregates information from neighboring nodes to update the characteristics of each node. The Gini coefficient is calculated from the load of each link, and a total of seven feature data are obtained through feature screening as node features. The link load data in the network topology is used as edge features. GCN consists of two layers of graph convolutional layers, and adds a batch normalization layer BN to reduce overfitting. It uses ReLU as the activation function to provide more efficient gradient descent and backpropagation. The convolution layer gradually extracts information about the feasibility of each configuration by convolution of each network configuration data. The expression of the ReLU activation function is as follows: After the network configuration feature information is aggregated through convolution, the graph is subjected to global average pooling (GAP) to average the features of all nodes. In this process, the seven-dimensional features of each node in the topology are converted into one-dimensional features. For the graph , the formula is as follows: in for The global eigenvector of for The number of nodes in for Middle n The feature vector of each node is mapped to the output dimension through MLP to obtain the classification result. The cross entropy loss function updates the parameters of the GCN classification model through back propagation. The cross entropy loss function formula is as follows: in N is the number of categories. Since this task is a binary classification task, , is the true label, is the predicted probability of the model.

7. The TSN network configuration detection method based on data enhancement and GCN network according to claim 6 is characterized in that: The seven features represent the total number of traffic in the network topology, the number of key flows, the number of audio flows, the number of video flows, the maximum link load, the time required to perform WCTT, and the link Gini coefficient calculated based on the load of each link.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images