Correlation weighted BPNN classification method based on TSN configuration detection
By using the correlation-weighted BPNN classification method in TSN network configuration detection, strong correlation characteristics are screened and BPNN network is optimized, which solves the problems of traditional algorithms with long detection time, low accuracy and poor generalization effect, and achieves fast and highly accurate TSN network configuration detection.
Patent Information
- Application Number
- CN202411797572.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-03-25
AI Technical Summary
Traditional algorithms have problems such as long detection time, low detection accuracy and poor generalization effect in TSN network configuration detection.
A correlation-weighted BPNN classification method based on TSN configuration detection is proposed. By constructing a network topology model, using Spearman algorithm to screen strong correlation features, and combining GA algorithm to optimize the initial weight and bias of the BPNN network, Spearman-WMVote-BPNN and Spearman-GA-BPNN models are constructed.
It achieves rapid training and high classification accuracy, improves detection speed and generalization capabilities, can efficiently detect the feasibility of TSN network configuration, and supports the development of a distributed, user-centric Web 3.0 architecture.
Smart Images
Figure CN119691558A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of TSN configuration, and in particular to a correlation weighted BPNN classification method based on TSN configuration detection. Background Art
[0002] Time-Sensitive Networking (TSN) technology helps ensure data consistency on distributed networks, avoiding data inconsistencies and conflicts. It also greatly enhances data security and reliability in the network, ensuring that users can trust the network to control their digital assets.
[0003] However, TSN configuration detection is a challenging task; traditional schedulability analysis methods are computationally intensive and time-consuming, and are not suitable for fast configuration of TSN networks.
[0004] Deep learning-based support for manufacturers and industries to predict their maintenance needs and reduce downtime; fast configuration detection methods based on machine learning (ML) and deep learning (DL) to improve the efficiency of configuration detection.
[0005] Mai T et al.'s On the use of supervised machine learning for assessing schedulability uses an ML algorithm to detect TSN configurations, which is 190 times faster than schedulability analysis, but the classification accuracy needs to be improved;
[0006] Mai,et al.’s Deep Learning to Predict the Feasibility of Priority-Based Ethernet Network Configurations uses an advanced graph neural network (GNN) model, but the model is somewhat robust to changes in the topology and traffic patterns of the training set, but the generalization classification accuracy is very low. Summary of the invention
[0007] In view of the shortcomings of the existing methods, the present invention solves the problems of long detection time, low detection accuracy and poor generalization effect existing in the traditional algorithms in TSN network configuration detection.
[0008] The technical solution adopted by the present invention is: a correlation weighted BPNN classification method based on TSN configuration detection comprises the following steps:
[0009] Step 1: construct a network topology model and collect parameter values corresponding to characteristic parameters in the network topology model;
[0010] As a preferred implementation manner of the present invention, the network topology model includes: Simple network topology and OEM network topology.
[0011] Step 2: Calculate the correlation between the characteristic parameters and the TSN configuration feasibility using a correlation algorithm to obtain the correlation characteristic parameters and perform normalization processing;
[0012] As a preferred implementation manner of the present invention, the correlation characteristic parameters include: key stream, audio stream, maximum load of all links, and link load Gini index.
[0013] As a preferred implementation of the present invention, the correlation algorithm is the Spearman algorithm.
[0014] Step 3: Take the correlation characteristic parameters in the schedulability scenario as input and the feasibility of TSN configuration as output, and use the WMVote-BPNN network for training;
[0015] As a preferred embodiment of the present invention, the WMVote-BPNN network includes:
[0016] Step 31, constructing a BP neural network;
[0017] Step 32: Initialize weights and biases using uniform distribution, normal distribution, Xavier, or He.
[0018] As a preferred embodiment of the present invention, He is sampled using normal distribution.
[0019] As a preferred embodiment of the present invention, Xavier adopts uniform distribution sampling.
[0020] Step 33, forward propagation;
[0021] Step 34, calculate the error and back propagate;
[0022] Step 35: updating weight and bias parameters;
[0023] Step 36: Repeat steps 32 to 36 until the stop condition is reached;
[0024] Step 37: Use weighted majority voting to vote on the prediction result labels of the four models. The specific voting classification formula is:
[0025]
[0026] Among them, s Uni 、s Nd 、s Xavier 、s HeThey are the classification results of the four model tests. W1, W2, W3, and W4 represent the accuracy ratios of the four models during training, and Label represents the feasibility of TSN configuration.
[0027] As a preferred implementation of the present invention, a GA algorithm is used to optimize the BPNN network hyperparameters.
[0028] As a preferred embodiment of the present invention, a relevance weighted BPNN classification system based on TSN configuration detection includes: a memory for storing instructions executable by a processor; and a processor for executing instructions to implement a relevance weighted BPNN classification method based on TSN configuration detection.
[0029] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and when the computer program code is executed by a processor, a correlation weighted BPNN classification method based on TSN configuration detection is implemented.
[0030] Beneficial effects of the present invention:
[0031] 1. The present invention proposes to use BPNN to detect TSN network configuration. First, a batch of strongly correlated features are screened out using the Spearman algorithm. Then, five of the features are selected as a strongly correlated feature subset in combination with the generalization requirements. Then, four neural network initialization methods are integrated through hard voting to construct the Spearman-WMVote-BPNN.
[0032] 2. The initial weights and biases of the BPNN network are optimized by the GA algorithm to complete the construction of the Spearman-GA-BPNN model. Through simulation comparison experiments, it is found that the Spearman-WMVote-BPNN model has the fastest training speed and higher classification accuracy.
[0033] 3. The Spearman-GA-BPNN model has the highest classification accuracy, the fastest detection speed and good generalization ability, which proves that the optimized BPNN model can meet the requirements of efficient detection of TSN network configuration feasibility, help ensure the high performance, data security and real-time communication of the network, and thus support the development of distributed, user-centric Web 3.0 architecture. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 is the Spearman-WMVote-BPNN flow chart of the present invention;
[0035] Figure 2 It is a Simple network topology;
[0036] Figure 3It is the OEM network topology;
[0037] Figure 4 This is a comparison chart of the Spearman-WMVote-BPNN test accuracy of the present invention;
[0038] Figure 5 is the change of loss during the training process of different initialization models of the present invention;
[0039] Figure 6 It is a comparison chart of the Spearman-GA-BPNN test accuracy of the present invention. DETAILED DESCRIPTION
[0040] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.
[0041] Scheduling analysis is a mathematical model used to derive the upper limit of performance indicators (such as communication delay, etc.), which can verify whether the computer system meets the real-time requirements; determine the applicable scheduling solution based on the TSN protocol that the topology can support; the specific schedulability methods are as follows:
[0042] First-in-first-out scheduling (FIFO): All flows belong to the same traffic class and have the same priority;
[0043] Manual classification priority scheduling (Manual): Classify traffic into audio stream, video stream, and command and control stream, and set the priority according to the criticality of the stream: command and control stream > audio stream > video stream;
[0044] "Concise Priority" (CP8) with eight priority levels: "Concise Priority" is the name of the priority assignment algorithm in RTaW-Pegase, which implements the optimal priority assignment algorithm and performs best for periodic transmission in switched Ethernet.
[0045] Manual classification with "Preshaping": The present invention uses manual classification with three priorities and instead applies a traffic shaping strategy called "in-transit preshaping" to all video streams; this strategy combines standard static priority scheduling with traffic shaping.
[0046] The feasibility label of each configuration in the dataset is determined by calculating the worst-case traversal time (WCTT) for each of the four scheduling methods. If and only if each flow meets the WCTT requirement of the system, the network configuration is feasible and the label is set to 1; if only one flow does not meet the WCTT requirement, the configuration is infeasible and the label is set to 0.
[0047] like Figure 1 As shown, the correlation weighted BPNN classification method based on TSN configuration detection includes the following steps:
[0048] Step 1: construct a network topology model and collect parameter values corresponding to characteristic parameters in the network topology model;
[0049] The network topology model of the present invention is constructed using an automotive electronic system, in which brakes and sensors are connected to an electronic control unit (ECU). During the interaction of each ECU, a large number of candidate configurations will be generated, and each configuration has a set of various types of flows. These flows constitute the initial raw data set of the experiment, which contains a total of 8966 samples and 212 feature parameters.
[0050] Network topology model such as Figure 2 and 3 As shown, Figure 2 The topology is Simple; the network consists of two switches and eight nodes; except for the 1Gbps on the link between switches, the data transmission rate on all links is 100Mbps to avoid the serious bottleneck that may occur in such a dumbbell topology. The packet switching delay is assumed to be at most 1.3us in the experiment. Figure 3 The topology shown is OEM, which consists of 5 switches and 16 nodes. Except for the 1Gbps link from ECU12 to Switch3, the data transmission rate of all links is 100Mbps.
[0051] by Figure 2 Taking as an example, the characteristic parameters include: key flow, audio flow, the maximum load of all links, the link load Gini index, the time to run a WCTT, queue size boundary ECU1FIFO, queue size boundary ECU2FIFO, queue size boundary ECU3FIFO, queue size boundary ECU4FIFO, queue size boundary ECU5FIFO, queue size boundary ECU6FIFO, queue size boundary ECU7FIFO, queue size boundary ECU8FIFO, queue size boundary ECU2FIFO, queue size boundary SW1FIFO, queue size boundary SW2FIFO, etc.
[0052] The parameter values corresponding to the characteristic parameters take the key stream, audio stream, and video stream as examples, and the unit of the parameter value is number of times.
[0053] During the experiment, the number of flows will change, but the ratio of flows will not change. The present invention uses a deadline constraint: the WCTT of each flow calculated by schedulability analysis must be less than the deadline of the flow.
[0054] Step 2: Calculate the correlation between the characteristic parameters and the TSN configuration feasibility using a correlation algorithm to obtain the correlation characteristic parameters and perform normalization processing;
[0055] Feature screening of the data set can screen out features that are strongly correlated with the target label (TSN configuration feasibility, with a value of 0 indicating infeasible and 1 indicating feasible), so that the model can obtain stronger classification capabilities; features closely related to the network configuration (TSN) characteristics, namely correlation feature parameters, are selected through Spearman correlation analysis; the purpose is to maximize the classification accuracy and meet the requirements of generalization to other topologies; finally, a group of five most relevant features are screened out, and the data samples are divided proportionally based on these five features; the division ratio of the training set and the test set is 4:1.
[0056] The Spearman correlation formula is:
[0057]
[0058] Among them, d i =x i -y i ,(1≤i≤N).
[0059] The correlation characteristic parameters include: key stream, audio stream, maximum load of all links, and link load Gini index.
[0060] The correlation coefficients of the corresponding correlation characteristic parameters under different schedulability scenarios are shown in Table 1;
[0061] Table 1 Spearman correlation coefficient between characteristic parameters of TSN network and target labels
[0062]
[0063] As can be seen from Table 1, there are certain differences in the degree of correlation between network features and target labels in different scheduling methods; the Spearman correlation coefficients of each correlation feature in Manual, CP8, and Preshaping are all greater than 0.5, which indicates that the selected features have a strong correlation with the target label; similarly, the Spearman correlation coefficients of each feature in FIFO are all greater than 0.3, indicating that the selected features also have a certain correlation with the target label; the link load Gini index is negatively correlated with the feasibility of network configuration, and the larger the absolute value of the Gini index (that is, the more unbalanced the load), the less feasible the network; therefore, the five selected correlation feature parameters are used as the input vector of the model to improve the detection effect of the model.
[0064] Step 3: Take the correlation characteristic parameters in the schedulability scenario as input and the feasibility of TSN configuration as output, and use the WMVote-BPNN network for training;
[0065] The construction of the WMVote-BPNN network includes:
[0066] Step 31, construct a BP neural network, including: input layer, hidden layer, output layer, the corresponding number of neurons is 5, 11 and 1 respectively;
[0067] Step 32, initializing weights and biases;
[0068] Four different initialization methods are selected to initialize the weights and biases of the connected neurons in each layer of the BP neural network; the integrated algorithm is to build multiple learners and then combine them through a certain strategy to complete the learning task, which can often obtain a learner that is significantly superior to a single learner. Here, the results of the four models are voted and integrated; the four initializations used for the weight W and bias B-ultrasound parameters are shown below.
[0069] Uniform distribution initialization (Uni): The initial values of weights and biases are randomly sampled from a uniform distribution in the interval [-1,1]. The uniform distribution initialization formula is:
[0070] W~U[-1,1],B~U[-1,1] (2)
[0071] Normal distribution initialization (Nd): In normal distribution initialization, the initial values of weights and biases are randomly sampled from a normal distribution (also called a Gaussian distribution). Normal distribution initialization formula:
[0072] W~N(0,1),B~N(0,1) (3)
[0073] Xavier initialization: is a method of adaptively initializing weights based on the number of input and output nodes. The goal is to keep the variance of activation values as equal as possible between different layers of the neural network. This helps avoid the problem of gradient vanishing or gradient exploding in deep neural networks, thereby improving the training effect of the network.
[0074] The formula for Xavier initialization based on uniform distribution sampling is as follows, where the biases are initialized to 0:
[0075]
[0076] The formula for Xavier initialization based on normal distribution is as follows:
[0077]
[0078] Where x is the number of input nodes, y is the number of output nodes, E(W) = 0; Xavier initialization based on uniform distribution sampling is used.
[0079] He initialization initializes weights according to the number of neurons in the previous layer of the network, which helps solve the gradient vanishing problem when training deep neural networks.
[0080] The He initialization formula based on uniform distribution sampling is as follows, where the biases are initialized to 0:
[0081]
[0082] The He initialization formula based on normal distribution sampling is as follows:
[0083]
[0084] Wherein, x is the number of input nodes, E(W)=0; in this embodiment, He initialization based on normal distribution sampling is used.
[0085] Step 33, forward propagation;
[0086] After the features of the training samples are input into the input layer of the network, the output value of each node is calculated layer by layer through the network; for the hidden layer and the output layer, an activation function (tansig) is used to convert the input value of the node into the output value; the formula of the tansig function is as follows:
[0087]
[0088] Formula (8) maps the input x to an output range between -1 and 1.
[0089] Step 34, calculate the error and back propagate;
[0090] Starting from the output layer, calculate the contribution of each node to the error, calculate the gradient of weights and biases based on the error contribution, and propagate the error backward to the previous layer;
[0091] This example uses the mean square error (MSE) loss function to calculate the network error, which measures the average square difference between the output of the neural network and the actual value. The MSE loss is calculated as follows:
[0092]
[0093] Where N is the number of samples, is the classification label of the neural network for the i-th sample, y i is the actual label (0 or 1) of the i-th sample.
[0094] Step 35: parameter update;
[0095] The weights and biases of the network are updated by formula (10) to gradually optimize the performance of the model. The formula is:
[0096]
[0097] Step 36: Repeat steps 32 to 36 until a stopping condition is met, such as reaching a maximum number of iterations (1000 times) or the error is reduced to a set threshold.
[0098] Step 37: Use the four trained BP neural network models to perform integrated voting classification on the test set. Use weighted majority voting to vote on the prediction result labels of the four models. The specific voting classification formula is:
[0099]
[0100] Among them, s Uni 、s Nd 、s Xavier 、s He They are the classification results of the four model tests. W1, W2, W3, and W4 represent the accuracy ratios of the four models during training. Label = 0 means that the TSN configuration is not feasible, and Label = 1 means that the TSN configuration is feasible.
[0101] Step 4: Use GA algorithm to optimize the hyper parameters of BP network;
[0102] Step 41, initialize GA population parameters: take all hyperparameters of BPNN: weights, bias values as population individuals; train BP neural network and obtain the error as fitness;
[0103] Step 42, selection operation: according to the fitness evaluation result, select a part of individuals as parents to generate the next generation of individuals;
[0104] Step 43, crossover operation: select a pair of individuals from the parent individuals and generate new individuals through crossover operation; the crossover operation simulates the gene recombination process and exchanges the chromosome parts of the two individuals.
[0105] Step 44, mutation operation: perform mutation operation on the newly generated individuals to introduce new gene combinations.
[0106] Step 45, repeat steps 41 to 44 until the stop condition is reached; that is, the maximum number of population updates (50 times) is reached.
[0107] The parameters obtained through the GA algorithm are substituted into the neural network and converted into the initial weights and biases of each layer of nodes; then the signal forward propagation and error back propagation are performed to continuously update the weights and biases until the number of iterations reaches the maximum number or the error meets the requirements.
[0108] In view of the problems of long detection time, low detection accuracy and poor generalization effect existing in traditional algorithms in TSN network configuration detection, the present invention proposes to use BPNN to detect TSN network configuration. Firstly, a batch of strongly correlated features are screened out by using the Spearman algorithm, and then five of the features are selected as the strongly correlated feature subset in combination with the generalization requirement. Secondly, two methods are proposed to construct two TSN network configuration detection models: Spearman-WMVote-BPNN and Spearman-GA-BPNN. The Spearman-WMVote-BPNN model was constructed by integrating four neural network initialization methods through hard voting; in addition, the initial weights and biases of the BP neural network were optimized by the GA algorithm to complete the construction of the Spearman-GA-BPNN model; through simulation comparison experiments, it was found that the Spearman-WMVote-BPNN model has the fastest training speed and high classification accuracy; the Spearman-GA-BPNN model also has the highest classification accuracy and the fastest detection speed as well as good generalization ability, which proves that the optimized BPNN model can meet the requirements of efficient detection of the feasibility of TSN network configuration, which helps to ensure the high performance, data security and real-time communication of the network, thereby supporting the development of distributed, user-centric Web 3.0 architecture.
[0109] Experimental process:
[0110] MATLAB software was used to complete data preprocessing, construction of Spearman-WMVote-BPNN model and Spearman-GA-BPNN model; the division of training set and test set adopted the same strategy as in the literature, that is, the ratio of training set to test set was 4:1; the iterative loss experimental results were illustrated using the FIFO mechanism.
[0111] The present invention uses accuracy ACC, true positive rate TPR, and true negative rate TNR to evaluate the model effect. The BPNN model parameters are shown in Table 2.
[0112] Table 2 BPNN model parameters
[0113]
[0114] Spearman-WMVote-BPNN simulation experiment:
[0115] The learning rate and number of training runs were set through repeated trials. The learning rate lr was 0.001 and the maximum number of training runs was 1000.
[0116] Use the Simple dataset for training and testing, and use the OEM dataset for generalization experiments; Figure 4 It can be seen that the model has the highest classification accuracy for the FIFO mechanism, up to 97.7%; the classification for the Preshaping mechanism is the worst, but it also reaches 94.4%; in addition, the trained model has strong generalization ability, and it still maintains a good classification effect when generalized to OEM, with the accuracy change within 1%.
[0117] Figure 5 The loss changes of models with different initializations during training are recorded. Figure 5 It can be seen that the four models basically converged in less than 6 times, among which the Spearman-Xavier-BPNN model converged the fastest and had the smallest initial loss.
[0118] Spearman-GA-BPNN simulation experiment:
[0119] Referring to the parameters in Table 2, the maximum number of iterations of BP neural network training is set to 1000;
[0120] The present invention uses the Simple dataset for training and testing, and uses the OEM dataset for generalization experiments; Figure 6 It can be seen that the model has a high classification accuracy of 97.9% for FIFO mechanism, and the classification of Preshaping mechanism is the worst, reaching 93.7%. In addition, it can be seen that the trained model has strong generalization ability and still maintains a high accuracy rate when generalized to OEM.
[0121] In order to verify the model effect, the proposed model is compared in terms of training time, detection time, ACC of different detection mechanisms, TPR and TNR; the results are shown in Table 3; it can be seen from the table that Spearman-WMVote-BPNN has the shortest training time, Spearman-GA-BPNN has the highest detection accuracy and the shortest detection time; the accuracy of different models in different mechanisms is generally: FIFO>Manual>CP8>Preshaping.
[0122] Table 3 Experimental results of two models
[0123]
[0124]
[0125] The present invention compares the detection accuracy of the two models of Spearman-WMVote-BPNN, Spearman-GA-BPNN and two models in other studies (K-means, KNN) by testing 1000 configurations, and the comparison is shown in Table 4; Table 4 shows the detection accuracy values of the four verification techniques;
[0126] The accuracy of the Spearman-GA-BPNN model and the Spearman-WMVote-BPNN model both exceeds that of K-means and KNN.
[0127] Table 4 Comparison of experimental results of multiple models
[0128]
[0129] The Spearman-WMVote-BPNN model of the present invention is an integrated framework that uses a weighted majority voting strategy combined with four different neural network initialization techniques; in contrast, the Spearman-GA-BPNN model uses the GA algorithm to refine the weight and bias initialization in the BPNN to improve performance; computer simulations provide convincing evidence that these optimized BPNN models perform very well in classification accuracy, proving the effectiveness of feature selection through Spearman correlation analysis combined with generalization requirements; the Spearman-WMVote-BPNN model combines various initialization techniques while ensuring robustness while also having a high accuracy rate; on the other hand, the Spearman-GA-BPNN model establishes a new benchmark in classification accuracy, achieving an impressive 97.9% accuracy rate, and a detection delay of only 10.91 milliseconds; these findings confirm the potential of the improved BPNN model in meeting the needs of efficient TSN network configuration detection.
[0130] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.
Claims
1. A correlation-weighted BPNN classification method based on TSN configuration detection, characterized in that: The following steps are involved: Step 1: construct a network topology model and collect parameter values corresponding to characteristic parameters in the network topology model; Step 2: Calculate the correlation between the characteristic parameters and the TSN configuration feasibility using a correlation algorithm to obtain the correlation characteristic parameters and perform normalization processing; Step 3: Take the correlation characteristic parameters in the schedulability scenario as input and whether the TSN configuration is feasible as output, and use the WMVote-BPNN network for training.
2. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 1 is characterized in that: The WMVote-BPNN network includes: Step 31, constructing a BP neural network; Step 32: Initialize weights and biases using uniform distribution, normal distribution, Xavier, or He. Step 33, forward propagation; Step 34, calculate the error and back propagate; Step 35: weight and bias parameter update; Step 36, repeating steps 32 to 36 until a stop condition is reached; Step 37: Use weighted majority voting to vote on the prediction result labels of the four models. The specific voting classification formula is: Among them, s Uni 、s Nd 、s Xavier 、s He They are the classification results of the four model tests. W1, W2, W3, and W4 represent the accuracy ratios of the four models during training, and Label represents the feasibility of TSN configuration.
3. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 2 is characterized in that: He uses normal distribution sampling.
4. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 2 is characterized in that: Xavier uses uniform distribution sampling.
5. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 2 is characterized in that: The GA algorithm is used to optimize the BPNN network hyperparameters.
6. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 1, characterized in that: Network topology models include: Simple network topology and OEM network topology.
7. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 1 is characterized in that: The correlation characteristic parameters include: key stream, audio stream, maximum load of all links, and link load Gini index.
8. The correlation weighted BPNN classification method based on TSN configuration detection according to claim 1, characterized in that: The correlation algorithm is the Spearman algorithm.
9. A correlation-weighted BPNN classification system based on TSN configuration detection, characterized in that: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the relevance-weighted BPNN classification method based on TSN configuration detection as described in any one of claims 1-8.
10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements the correlation-weighted BPNN classification method based on TSN configuration detection as described in any one of claims 1 to 8.