Network load balancing evaluation method based on deep learning model

Abstract

The invention relates to the technical field of network communication, and discloses a network load balancing evaluation method based on a deep learning model, which comprises the following steps: S1, data preprocessing: collecting flow data, and carrying out normalization, noise reduction and enhancement processing; s2, manifold modeling: dimension reduction is performed on high-dimensional flow data, and a local structure is reserved; s3, feature extraction: extracting optimized low-dimensional features, and reducing redundancy; s4, performance evaluation: comprehensively evaluating throughput, delay and balance; s5, anomaly detection: detecting abnormal traffic and outputting a result; s6, feedback optimization: dynamically adjusting a load strategy, and improving the performance. Through deep learning, variational information bottleneck and manifold learning, accurate modeling and optimization of dynamic network traffic features are realized, the extracted features accurately reflect the traffic change trend, the adaptive capacity of the model to sudden and abnormal traffic is enhanced, performance reduction caused by feature distortion in a traditional method is effectively avoided, and the method has a good application prospect. And reliable support is provided for a load balancing strategy.

Description

Technical Field

[0001] The present invention relates to the field of network communication technologies, and specifically to a network load balancing evaluation method based on a deep learning model. Background Art

[0002] Load balancing is a key technology in the field of network communication, mainly used to reasonably distribute network traffic to different servers or nodes to improve resource utilization efficiency and network performance. Currently, most traditional load balancing evaluation methods are based on static rules or empirical formulas, and formulate load distribution strategies by analyzing simple features of traffic (such as traffic volume, number of session connections, etc.). However, with the expansion of network scale and the increase in traffic complexity, traditional methods are difficult to accurately reflect the non-linear distribution characteristics of traffic in a dynamic network environment. At the same time, although some load balancing methods based on machine learning have been applied in performance prediction and traffic classification, there are still problems of low evaluation efficiency and poor adaptability when dealing with high-dimensional, multi-objective and dynamically changing network traffic.

[0003] However, the application of existing technologies in complex network environments still has obvious deficiencies. First, for the modeling of high-dimensional network traffic characteristics, most methods fail to effectively capture the inherent non-linear distribution of data, resulting in inaccurate evaluation results. Second, in the evaluation of multi-objective performance indicators, traditional methods usually tend to optimize a single objective (such as throughput or latency), lacking systematic consideration of the trade-off between multiple objectives. In addition, in the face of burst traffic and abnormal traffic, existing evaluation methods are difficult to effectively adapt to traffic dynamic changes, resulting in insufficient robustness and distorted evaluation results. Therefore, there is an urgent need for a technical solution that can accurately evaluate and dynamically optimize load balancing strategies in complex dynamic network environments to comprehensively improve the resource utilization efficiency and service quality of the system. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a network load balancing evaluation method based on a deep learning model, which solves the problems of insufficient accuracy and adaptability in evaluating load balancing strategies in complex dynamic network environments.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A network load balancing evaluation method based on a deep learning model, including the following steps: S1. Data preprocessing: Collect network traffic feature data, and perform normalization, noise reduction, and feature enhancement processing on the data; S2. Flow feature manifold modeling: Use manifold learning methods to perform dimensionality reduction processing on high-dimensional network traffic feature data, map it to a low-dimensional embedding space, and retain the local neighborhood structure of the traffic data; S3. Variational Information Bottleneck Feature Extraction: Extract low-dimensional features through a deep learning model, optimize the correlation of features with the load balancing objective, and reduce redundant information; S4. Multi-objective Performance Evaluation: Based on the extracted low-dimensional features, optimize the evaluation of the throughput, average latency, and node utilization balance of the load balancing strategy; S5. Abnormal Traffic Detection: Based on a deep learning model, model the traffic distribution, detect traffic anomalies, and output the detection results; S6. Dynamic Feedback and Optimization: Dynamically adjust the load distribution strategy according to the performance evaluation results to optimize the system performance.

[0006] Preferably, the S2 step further includes the following steps: S21. Construct a similarity matrix to characterize the similarity between network traffic feature data; S22. Use a manifold learning algorithm to map high-dimensional features to a low-dimensional embedding space by optimizing the distance relationship between embedding points.

[0007] Preferably, the S3 step further includes the following steps: S31. Use a deep learning model to extract low-dimensional latent features; S32. Associate the extracted features with the load balancing objective and retain task-related information; S33. Use a regularization method to constrain the low-dimensional feature distribution and reduce redundant information.

[0008] Preferably, in the S31 step, the deep learning model includes: An encoder, which is used to convert the input network traffic feature data into low-dimensional latent features and achieve feature extraction by learning the latent distribution; A decoder, which is used to generate the prediction results of the load balancing objective performance metrics according to the low-dimensional latent features.

[0009] Preferably, the S4 step further includes the following steps: S41. Calculate the throughput of the load balancing strategy; S42. Calculate the average latency of the load balancing strategy; S43. Evaluate the dispersion degree of node utilization and determine the utilization balance.

[0010] Preferably, the S4 step uses a non-dominated sorting genetic algorithm for multi-objective performance evaluation, specifically including the following steps: S44. Initialize the population and randomly generate candidate solutions; S45. Calculate the fitness values of the candidate solutions in the population, including throughput, latency, and utilization balance; S46. Stratify and sort the candidate solutions according to the non-dominance relationship; S47. Generate the next generation of candidate solutions through crossover and mutation operations, and repeat the iteration until convergence.

[0011] Preferably, the S5 step further includes the following steps: S51. Establish a deep learning model, assume that the traffic characteristics follow a latent distribution, and use the model to learn the parameters of the traffic distribution; S52. According to the distribution probability value of the traffic characteristics, when the distribution probability value is lower than the set threshold, it is determined as abnormal traffic.

[0012] Preferably, in the S52 step, the abnormal traffic adopts a dynamic adjustment of the load balancing strategy detection result: S521. Adjust the distribution weight of the abnormal traffic request; S522. Update the regularization parameter of the low-dimensional feature extraction model.

[0013] Preferably, the S6 step further includes the following steps: S61. Periodically evaluate the performance of the current load distribution strategy; S62. Adjust the weight parameters of the low-dimensional feature extraction model according to the evaluation result; S63. Update the load distribution strategy to adapt to the dynamic changes of network traffic.

[0014] Preferably, in the S6 step, an optimal solution that meets the maximization of throughput and the minimization of latency is selected and optimized, and the balance of node utilization is adjusted.

[0015] The present invention provides a network load balancing evaluation method based on a deep learning model. It has the following beneficial effects: 1. By combining the deep learning model with the variational information bottleneck theory, the present invention accurately models and optimizes the highly dynamic and complex network traffic characteristics. At the same time, through manifold learning, it captures the intrinsic non-linear distribution of traffic characteristics and removes redundant information, enabling the extracted features to effectively reflect the change trend of network traffic, enhancing the model's adaptability to burst traffic and abnormal traffic, thereby showing higher robustness in dynamic network scenarios and avoiding the performance degradation problems caused by feature distortion or distribution changes in traditional evaluation methods, providing precise support for the load balancing strategy.

[0016] 2. The present invention uses a non-dominated sorting genetic algorithm to perform multi-objective performance optimization on throughput, latency, and node utilization balance. By generating a Pareto optimal solution set, it solves the conflicts among multi-objective performance indicators, ensures the reasonable allocation of network resources and the balance of system performance. At the same time, combined with the weighted processing of abnormal traffic detection results, it reduces the interference of abnormal data on performance results, improves the stability and accuracy of multi-objective optimization, meets diverse network requirements, and enhances the robustness of the algorithm.

[0017] 3. The present invention realizes the closed-loop optimization of the network load balancing strategy through a dynamic feedback mechanism. Combined with an online learning method, according to multi-objective performance evaluation and abnormal detection results, it adjusts the load allocation strategy and feature extraction weights in real time, ensures the rapid response of the network system to dynamic traffic changes, thereby improving the utilization efficiency of network resources. It also preferentially guarantees the traffic allocation of critical tasks in abnormal node or resource-constrained scenarios, thus enhancing the adaptive ability and quality of service in complex network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is the flowchart of the method steps of the present invention; Figure 2 is the expanded flowchart of step S2 of the present invention; Figure 3 is the expanded flowchart of step S3 of the present invention; Figure 4 is the expanded flowchart of step S4 of the present invention; Figure 5 is the flowchart of the multi-objective performance evaluation steps of the present invention; Figure 6 is the expanded flowchart of step S5 of the present invention; Figure 7 is the expanded flowchart of step S6 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Please refer to the attached Figure 1 , the embodiment of the present invention provides a network load balancing evaluation method based on a deep learning model, including the following steps: S1. Data preprocessing: Collect network traffic feature data, and perform normalization, noise reduction, and feature enhancement processing on the data; S2. Flow Feature Manifold Modeling: Use manifold learning methods to reduce the dimensionality of high-dimensional network traffic feature data, map it to a low-dimensional embedding space, and preserve the local neighborhood structure of the traffic data; S3. Variational Information Bottleneck Feature Extraction: Extract low-dimensional features through a deep learning model, optimize the relevance of the features to the load balancing objective, and reduce redundant information; S4. Multi-objective Performance Evaluation: Based on the extracted low-dimensional features, optimize the evaluation of the throughput, average latency, and node utilization balance of the load balancing strategy; S5. Abnormal Traffic Detection: Model the traffic distribution based on a deep learning model, detect traffic anomalies, and output the detection results; S6. Dynamic Feedback and Optimization: Dynamically adjust the load distribution strategy according to the performance evaluation results to optimize the system performance.

[0021] Specifically, the present invention analyzes the dynamic characteristics of network traffic through a deep learning model, and combines manifold learning and variational information bottleneck theory to achieve the dimensionality reduction and optimized expression of high-dimensional traffic features. Specifically, first, the collected network traffic features are preprocessed to remove redundant data and noise, and standardized to adapt to the input requirements of subsequent models. Then, a manifold learning method is used to perform dimensionality reduction modeling on the high-dimensional feature data, map the complex traffic data to a low-dimensional space, and at the same time preserve its local geometric structure, thereby effectively reducing redundant information and capturing the internal patterns of the data.

[0022] Based on the dimensionality-reduced features, the variational information bottleneck feature extraction module further optimizes the feature representation, and realizes the effective expression of the features for the load balancing objective by maximizing the information related to the target and minimizing redundancy. Subsequently, the multi-objective performance evaluation module quantitatively evaluates the key performance indicators of the load balancing strategy based on the extracted features, including throughput, latency, and node utilization balance, etc. The performance evaluation results can not only reveal the advantages and disadvantages of the current strategy, but also provide data support for dynamic optimization.

[0023] For the possible burst traffic and abnormal data in the network environment, the present invention designs an abnormal traffic detection module, models the traffic distribution through a variational autoencoder, detects traffic anomalies, and feeds the detection results back to the performance evaluation module to enhance the adaptability of the model to dynamic scenarios. On this basis, the dynamic feedback and optimization module adjusts the load distribution strategy according to the evaluation results to further improve the utilization efficiency of network resources and the load balancing performance.

[0024] Among them, step S1 is the basic link for implementing subsequent steps. Its main purpose is to perform operations such as cleaning, denoising, and normalizing the collected network traffic feature data to improve data quality and the effectiveness of model input. Through data preprocessing, redundant and noise interferences can be effectively removed, the feature expressiveness of the input data can be enhanced, and more adaptable input data can be provided for subsequent traffic feature modeling and optimization steps. It should be noted that this step is not just a simple data processing process, but also fully considers the complexity and dynamics of high-dimensional network traffic feature data.

[0025] In this embodiment, the data preprocessing includes the following technical contents: As an option, the data preprocessing of the present invention first collects network traffic feature data from a real network environment or a public network traffic dataset. Exemplarily, these data may include conventional traffic features such as traffic size, time interval of data packets, source address, and destination address.

[0026] It should be noted that the collected network traffic feature data often contains a large amount of noise or incomplete data. To ensure the accuracy and consistency of the data, the collected data is cleaned in this embodiment. Specifically, by setting a missing value threshold, data samples with incomplete traffic features are removed, and data containing obvious incorrect feature values (such as abnormally high traffic values) are deleted or corrected. For example, the Z-score method in statistics can be used to detect and remove outliers, and its calculation formula is: where, is the sample value, is the mean of the feature, is the standard deviation of the feature. For samples that satisfy , they are marked as outliers and removed from the dataset.

[0027] In some embodiments, to address the possible impact of inconsistent eigenvalue scales on model input, the data is normalized. Specifically, the data feature values are mapped to the interval [0, 1] or [-1, 1] to make the numerical ranges of different features consistent. Exemplarily, the min-max normalization method can be used, and its calculation formula is: where, and respectively represent the minimum and maximum values of a specific feature. As another option, the mean normalization method can also be used to make the mean of the feature values 0 and the variance 1.

[0028] As a possible implementation method, feature enhancement is also performed in this embodiment to improve the data's ability to express dynamic loads. Specifically, the moving mean and variance of traffic features are calculated using a sliding window method in the time series dimension to capture the dynamic trend of traffic changes. For example, the sliding window size can be set to 5 seconds, 10 seconds, or adjusted according to the actual network scenario. For example, the moving average calculation formula is: At the same time, calculate the corresponding moving variance: In some embodiments, in order to further enhance the spatial expression ability of features, the structural information of the network topology can also be combined. Specifically, a weighted graph structure of the network topology is constructed based on the source address and destination address of the data packet, and indicators such as the degree, weighted degree or centrality of the node are used as new features. For example, the degree of each network node can be defined as the number of neighboring nodes it is connected to, which is used to indicate the concentration of network traffic.

[0029] It should be noted that in the preprocessed data, redundant features have been removed and the time series and spatial expression of traffic features have been enhanced. These processed data will be directly input into the next step of traffic feature modeling module as the input of subsequent dimensionality reduction operations.

[0030] It is understandable that although the data preprocessing step seems to be a basic operation, its quality directly determines the accuracy of the subsequent deep learning model and the efficiency of feature extraction. Therefore, in some possible implementations, the preprocessing parameters, such as the cleaning threshold, the normalization range, and the feature enhancement method, can be adjusted according to the specific network environment and the complexity of the traffic characteristics.

[0031] Please refer to the attached Figure 2 , step S2 further comprises the following steps: S21, constructing a similarity matrix to represent the similarity between network traffic feature data; S22. Using the manifold learning algorithm, high-dimensional features are mapped to low-dimensional embedding space by optimizing the distance relationship between embedded points.

[0032] Specifically, in the network load balancing evaluation method of the present invention, traffic feature manifold modeling is an important step after data preprocessing and before entering deep learning feature extraction. The main task of this step is to map high-dimensional network traffic features to low-dimensional embedding space while retaining the intrinsic geometric structure and local neighborhood relationship of the data as much as possible. This dimensionality reduction process can effectively reduce redundant features and provide a computationally friendly input space for subsequent deep learning models.

[0033] It should be noted that this step is not a simple linear dimensionality reduction method, but a method that captures the non-linear distribution law of network traffic characteristics through manifold learning technology, making the features after dimensionality reduction more conform to the internal structure of the original data, especially suitable for high-dimensional data with complex distributions such as network traffic characteristics.

[0034] In this embodiment, the manifold modeling of traffic characteristics includes the following technical contents: As an option, the present invention uses Laplacian eigenmaps in the manifold learning method to reduce the dimensionality of high-dimensional features. Laplacian eigenmaps constructs a similarity matrix between data samples to reflect the relationship between samples in the high-dimensional space, and finally realizes the construction of a low-dimensional embedding space.

[0035] Specifically, first construct a similarity matrix for the sample set after data preprocessing. Exemplarily, the similarity between sample points is calculated using a Gaussian kernel function, and the formula is as follows: Where, and respectively represent two traffic characteristic data samples, is the Euclidean distance between them, is the bandwidth parameter of the kernel function. In some embodiments, The value of can be adaptively adjusted according to the characteristics of the data distribution.

[0036] As a possible implementation, in order to further reduce the computational complexity, a sparsification strategy can be introduced into the similarity matrix, that is, only the nearest neighbor relationship of each sample is retained, and the similarity values of other irrelevant samples are set to zero. This can effectively reduce the interference of non-local relationships on the embedding results.

[0037] After calculating the similarity matrix, construct a normalized Laplacian matrix . It should be noted that the construction formula of the normalized Laplacian matrix is: Where, is a diagonal matrix, and its diagonal element is the degree of sample In another form, it is expressed as: Based on the Laplacian matrix, in this embodiment, the squared distance loss between low-dimensional embedding points is minimized to optimize the representation of the embedding space. The specific optimization objective is: Where, represents sample Representation in the low-dimensional embedding space. The result of this optimization process is to construct a low-dimensional embedding space such that the local structure of the high-dimensional data is preserved as much as possible in the low-dimensional space.

[0038] Exemplarily, in some embodiments, the target optimization can be accomplished by matrix factorization. By calculating the eigenvalues and eigenvectors of the Laplacian matrix, the samples are mapped to the first d principal components of the eigenvector space to obtain a low-dimensional embedding representation.

[0039] It should be noted that the low-dimensional features obtained after dimensionality reduction not only reduce the dimension of the original data but also retain the local geometric relationship of the traffic features. This feature is crucial for subsequent variational information bottleneck feature extraction because it provides input features with high correlation and simplicity.

[0040] As another option, other methods of manifold learning, such as t-SNE or ISOMAP, can also be combined to complete the dimensionality reduction of traffic features. In some scenarios where the network traffic distribution is relatively complex, these methods may provide better dimensionality reduction results.

[0041] It can be understood that the traffic feature manifold modeling step significantly improves the data representation ability through the structured dimensionality reduction of high-dimensional data and guarantees the computational efficiency of the subsequent deep learning feature extraction step. In practical applications, a suitable manifold learning method can be selected according to the data scale, computing resources, and target requirements.

[0042] Please refer to Appendix Figure 3 , Step S3 further includes the following steps: S31. Use a deep learning model to extract low-dimensional latent features; S32. Associate the extracted features with the load balancing target and retain task-related information; S33. Use a regularization method to constrain the low-dimensional feature distribution and reduce redundant information; In step S31, the deep learning model includes: An encoder, which is used to convert the input network traffic feature data into low-dimensional latent features and achieve feature extraction by learning the latent distribution; A decoder, which is used to generate a prediction result of the load balancing target performance metric according to the low-dimensional latent features.

[0043] Specifically, in the network load balancing evaluation method of the present invention, variational information bottleneck feature extraction is a key step to further optimize and extract the core features directly related to the load balancing target from the traffic features after dimensionality reduction. This step combines the theoretical frameworks of deep learning and information theory. By optimizing the correlation and redundancy of the features, the extracted features can not only accurately reflect the dynamic characteristics of network traffic but also provide precise support for the load balancing target.

[0044] It should be noted that this step takes the result of traffic feature manifold modeling as input, constructs the latent distribution of features using a deep learning model, and optimizes the features by maximizing the mutual information and regularization constraints. This process not only improves the representation ability of the features but also provides efficient and accurate input data for subsequent multi-objective performance evaluation.

[0045] In this embodiment, variational information bottleneck feature extraction includes the following technical content: As an option, the present invention adopts the variational information bottleneck method to optimize the information correlation between the input features and the load balancing objective. Specifically, the goal is to maximize the correlation between the dimensionality-reduced features and the objective while minimizing the redundancy information between the features and the original input . The optimization objective can be expressed as: where the first term represents the predictive ability of the low-dimensional features for the objective , and the second term is a regularization term used to constrain the redundancy of the features, being a hyperparameter for adjusting the weights of the two.

[0046] In a possible implementation, to calculate the mutual information , the present invention adopts the variational inference method and represents it as the KL divergence between the posterior distribution and the prior distribution : where is modeled by the encoder of the deep learning model, is usually taken as the standard normal distribution .

[0047] Specifically, in this embodiment, the encoder maps the dimensionality-reduced features XXX to the latent space ZZZ, which is expressed as: where and represent the mean and standard deviation of the latent distribution respectively, both modeled by a deep neural network; is the standard normal distribution noise used to enhance the robustness of the model.

[0048] As a possible implementation, the decoder generates the prediction result of the load balancing objective based on the latent features . The goal of the decoder is to maximize , that is, maximizing the prediction accuracy given the potential features. It should be noted that this process is carried out through supervised learning, and the error between the output of the decoder and the actual target value is used to optimize the parameters of the entire model. The error between the output of the decoder and the actual target value is used to optimize the parameters of the entire model.

[0049] In some embodiments, in order to further improve the efficiency of feature extraction, the regularization term: can be used to introduce weight penalties to control the complexity of the latent distribution. Exemplarily, the relationship between feature correlation and redundancy control can be balanced by adjusting the value. It should be noted that the variational information bottleneck method can remove a large amount of redundant information unrelated to the target during the feature extraction process, making the extracted features more targeted and compact. In some possible implementation manners, the network structures of the encoder and the decoder can also be optimized. For example, by introducing a convolutional neural network to enhance the ability to capture the spatial distribution of traffic features, or by introducing a long short-term memory network to process time series features.

[0050] As another option, a target weighting strategy can also be added at the decoder output end to achieve a trade-off between multiple load balancing targets (such as throughput, latency, and node utilization balance). For example, when the load balancing strategy focuses more on throughput, the loss weight of the throughput-related target can be increased, so that the model is more inclined to extract features related to throughput.

[0051] It can be understood that the core of variational information bottleneck feature extraction lies in constructing the latent distribution of low-dimensional features through a deep learning model, while optimizing the effectiveness and redundancy of the features. The optimized results of the dimensionality-reduced features will be directly used in the subsequent multi-objective performance evaluation module to support the core capabilities of the entire load balancing evaluation method.

[0052] Please refer to Appendix Figure 4 - Appendix Figure 5 , step S4 further includes the following steps: S41. Calculate the throughput of the load balancing strategy; S42. Calculate the average latency of the load balancing strategy; S43. Evaluate the discreteness of node utilization and determine the utilization balance; Step S4 uses the non-dominated sorting genetic algorithm for multi-objective performance evaluation, specifically including the following steps: S44. Initialize the population and randomly generate candidate solutions; S45. Calculate the fitness values of the candidate solutions in the population, including throughput, latency, and utilization balance; S46. Stratify and sort the candidate solutions according to the non-dominance relationship; S47. Generate the next generation of candidate solutions through crossover and mutation operations, and repeat the iteration until convergence.

[0053] Specifically, in the network load balancing evaluation method of the present invention, multi-objective performance evaluation is a key step in quantifying the current load balancing strategy based on the low-dimensional features extracted by variational information bottleneck. This step comprehensively evaluates the actual performance of the network load balancing strategy by constructing multiple performance indicators, and provides a decision-making basis for subsequent dynamic optimization. It should be noted that the focus of multi-objective performance evaluation is to comprehensively measure the three core indicators of system throughput, average latency, and node utilization balance. Through the combination of deep learning and optimization algorithms, this step can effectively evaluate the load balancing strategy in a complex network environment.

[0054] In this embodiment, the multi-objective performance evaluation includes the following technical contents: As an option, this embodiment constructs multiple key performance indicators to quantify the effect of the load balancing strategy. Specifically, the throughput indicator measures the total amount of data processed by the network system per unit time and is an important measure of system performance; the average latency indicator is used to represent the average time from the issuance to the completion of a user request, reflecting the service response efficiency; the node utilization balance indicator evaluates the load distribution of each node to ensure the fairness of resource allocation.

[0055] The calculation of throughput is based on the extracted low-dimensional features and the load distribution strategy, and is defined as: Wherein, represents the time window, is the time of the low-dimensional features. It should be noted that the features used in the throughput calculation are obtained from the optimized results of the variational information bottleneck module and can effectively capture the dynamic characteristics of network traffic.

[0056] As another option, the calculation formula for the average latency is: Wherein, represents the average latency at time . Exemplarily, the latency can be calculated through the relationship model between the traffic characteristics and the load distribution strategy.

[0057] The node utilization balance is quantified by calculating the dispersion degree of the loads of all nodes. The specific formula is as follows: Wherein, is the total number of network nodes, is the utilization rate of node , represents the average utilization rate of all nodes. It should be noted that the higher the balance of node utilization rates, the better the fairness of resource allocation.

[0058] In a possible implementation manner, in order to comprehensively evaluate these three performance metrics, the present invention uses a non-dominated sorting genetic algorithm to solve the multi-objective optimization problem. The non-dominated sorting genetic algorithm maps the multi-objective performance metrics into the solution space, performs non-dominated sorting and selection on the candidate solutions, and finally generates a Pareto optimal solution set.

[0059] Specifically, the implementation of the non-dominated sorting genetic algorithm includes the following steps: First, initialize the population and randomly generate initial solutions. Each individual in the population corresponds to a load distribution strategy, and its fitness value is calculated through low-dimensional features , including throughput, latency, and balance.

[0060] Then, perform hierarchical sorting on the population according to the non-dominated relationship, and calculate the crowding distance for each solution to measure the diversity of the solutions in the solution space. The non-dominated relationship is defined as: solution non-dominated solution , if and only if is not inferior to in all objectives, and is superior to in at least one objective.

[0061] Next, generate the next generation population through selection, crossover, and mutation operations, and continuously iterate until the stop condition is met. The finally output Pareto optimal solution set is the set of the best load distribution strategies under different objective weights.

[0062] It can be understood that the non-dominated sorting genetic algorithm can effectively balance the weight conflicts between multiple performance metrics and provide a rich selection space for the dynamic optimization of the load balancing strategy. In some embodiments, the parameters of the non-dominated sorting genetic algorithm can also be adjusted, such as the population size, the maximum number of iterations, etc., to adapt to the requirements of different network environments.

[0063] It should be further noted that the multi-objective performance evaluation module in the present invention not only supports the evaluation of static network traffic, but also can achieve the adaptive evaluation of the dynamic network environment by continuously updating the extracted low-dimensional features and performance metrics. For example, during the traffic peak period, the module can adjust the weight of the throughput metric and preferentially select the strategy with high throughput to ensure the quality of service.

[0064] Please refer to the appendix Figure 6, step S5 further includes the following steps: S51. Establish a deep learning model, assume that the traffic characteristics follow a latent distribution, and use the model to learn the parameters of the traffic distribution; S52. According to the distribution probability value of the traffic characteristics, when the distribution probability value is lower than the set threshold, it is determined as abnormal traffic; In step S52, the abnormal traffic adopts the detection result of the dynamic adjustment load balancing strategy: S521. Adjust the distribution weight of the abnormal traffic request; S522. Update the regularization parameter of the low-dimensional feature extraction model.

[0065] Specifically, in the network load balancing evaluation method of the present invention, the abnormal traffic detection is an important step to ensure the robustness of the evaluation model in a complex network environment. This step models the latent distribution of the traffic characteristics through a deep learning model, real-time detects and marks the abnormal traffic, and feeds the detection result back to the multi-objective performance evaluation module to reduce the interference of abnormal data on the load balancing evaluation result. It should be noted that the abnormal traffic detection is not only used to mark the abnormal data, but also provides a basis for dynamic feedback optimization to ensure that the system can maintain stability and reliability in an abnormal network environment.

[0066] In this embodiment, the abnormal traffic detection includes the following technical contents: As an option, the present invention models the distribution of the traffic characteristics based on the Variational Autoencoder (VAE). The core of VAE is to parametrically model the latent distribution of the traffic characteristics through a deep neural network, so as to learn the mean and variance of the latent distribution. In a specific implementation, the traffic characteristic data is mapped to the latent space, and the distribution of the traffic characteristics is defined in the latent space.

[0067] As a possible implementation manner, the decoder reconstructs the input features according to the latent variables, and evaluates the learning effect of the model through the reconstruction error. It should be noted that the larger the reconstruction error, the lower the matching degree between the input data and the latent distribution, and it may be abnormal traffic.

[0068] In this embodiment, the judgment criterion for abnormal traffic is based on the calculation of the log-likelihood value of the latent distribution. Exemplarily, the formula for the log-likelihood value is: where is the dimension of the latent variable, and are the mean and variance of the i-th dimension of the latent variable respectively. If the log-likelihood value is lower than the preset threshold, the corresponding traffic characteristic is determined as abnormal traffic.

[0069] As an alternative, it is also possible to determine whether the traffic is abnormal by the mean square error (MSE) of the reconstruction error, and the calculation formula is: where n is the feature dimension. If the MSE is greater than the threshold, then this traffic sample is considered abnormal traffic.

[0070] It should be noted that in this embodiment, the abnormal traffic detection module not only outputs the label of the abnormal traffic, but also feeds back the abnormal detection result to the multi-objective performance evaluation module. In the evaluation module, the abnormal points are weighted to reduce the impact of the abnormal traffic on the overall evaluation result of the system. Specifically, the reliability of the evaluation result can be ensured by reducing the weight of the abnormal traffic or directly removing the abnormal samples.

[0071] In some embodiments, in order to improve the real-time performance of the detection, the VAE can adopt a lightweight structure to meet the requirements of the high-dynamic network environment. Exemplarily, the inference speed and computational efficiency of the VAE can be optimized by methods such as model pruning, quantization, or knowledge distillation.

[0072] It should be further noted that the abnormal traffic detection module in the present invention can also be linked with the dynamic feedback optimization module. When a large-scale abnormal traffic is detected, the system will adjust the load distribution strategy in real time and give priority to processing non-abnormal traffic, thereby ensuring the quality of service of the normal network traffic.

[0073] It can be understood that the core of the abnormal traffic detection step lies in accurately modeling the traffic feature distribution through a deep learning model, so as to achieve efficient abnormal detection and evaluation interference control. This design not only improves the adaptability of the present invention in a complex network environment, but also provides a stable basis for subsequent dynamic optimization. As an extension, this module can also combine the abnormal traffic characteristics of specific scenarios and continuously optimize the model parameters through online learning or transfer learning to adapt to the dynamically changing network environment.

[0074] Please refer to Appendix Figure 7 , and step S6 further includes the following steps: S61. Periodically evaluate the performance of the current load distribution strategy; S62. Adjust the weight parameters of the low-dimensional feature extraction model according to the evaluation result; S63. Update the load distribution strategy to adapt to the dynamic changes of the network traffic; In step S6, an optimal solution that meets the maximization of throughput and the minimization of latency is selected and optimized, and the balance of node utilization is adjusted.

[0075] Specifically, in the network load balancing evaluation method of the present invention, dynamic feedback and optimization is the last step of the entire method and the core link to achieve dynamic adjustment and continuous optimization of load balancing. Based on the results of multi-objective performance evaluation and abnormal traffic detection, this step adjusts the load distribution strategy in real time to ensure that the network system can adapt to the dynamic changes in traffic. It should be noted that this step is not limited to optimizing the current strategy, but also realizes the cyclic improvement of the entire system through a closed-loop mechanism, further improving the utilization efficiency of network resources and service quality.

[0076] In this embodiment, the dynamic feedback and optimization include the following technical contents: As an option, the present invention uses a non-dominated sorting genetic algorithm to generate an optimized solution set of the load distribution strategy. Specifically, according to the throughput, delay, and node utilization balance metrics calculated in the multi-objective performance evaluation, the traffic distribution weights of each node are dynamically adjusted. It should be noted that the multi-objective requirements of the system are always considered during the optimization process to balance the conflicts between different performance metrics.

[0077] Specifically, the implementation steps of the non-dominated sorting genetic algorithm include the following aspects. First, initialize the population, randomly generate the initial load distribution strategy, and each strategy corresponds to an individual. The fitness value of the individual is determined by the multi-objective performance metrics where: represents the system throughput; represents the system average delay; represents the node utilization balance.

[0078] It should be noted that the size of the initial population and the maximum number of iterations are set according to the actual network size and computing resources.

[0079] As a possible implementation, after the population initialization, the algorithm sorts the candidate solutions through the non-dominated relationship. The non-dominated relationship is defined as: solution is not inferior to solution in all objectives, and is superior to in at least one objective, then is non-dominated . Through non-dominated sorting, the candidate solutions are divided into multiple levels, and the solutions with higher non-dominated levels are preferentially selected.

[0080] In the crossover and mutation operations, this embodiment adopts a strategy adjustment method based on weighted distribution. Exemplarily, for the scenario with throughput priority, the weight of the high-throughput strategy will be increased during the crossover operation; while in the latency-sensitive scenario, the mutation operation will tend to reduce the load deviation between nodes, thereby improving the response time.

[0081] It should be noted that during the iterative process, the optimization process will be dynamically adjusted by continuously combining the abnormal traffic detection results. When a large-scale abnormal traffic is detected, the system will dynamically adjust the weights of the optimization objectives to give priority to ensuring the quality of service of non-abnormal traffic. For example, by reducing the priority of abnormal traffic, more resources can be allocated to normal traffic.

[0082] When outputting the optimization results, the system selects the load distribution strategy according to the final Pareto optimal solution set. As an option, the strategy selection can be achieved through a multi-objective weighted function, for example: where, , , represents the weight coefficient of the performance metric. It should be noted that the setting of the weights can be flexibly adjusted according to the requirements of specific scenarios. For example, during peak periods, the throughput weight can be preferentially increased, and during low-load scenarios, the balance weight can be increased.

[0083] In a possible implementation manner, the present invention realizes closed-loop optimization through a dynamic feedback mechanism. Specifically, after each optimization, the system will re-evaluate the performance metrics of the current load distribution strategy and adjust the weight parameters of the feature extraction model according to the evaluation results. For example, when the evaluation results show that the current strategy has obvious deficiencies in terms of latency, the system will enhance the weights of the latency-related features in the low-dimensional features, thereby providing more targeted input data for the next round of optimization.

[0084] It can be understood that the dynamic feedback and optimization are not only a simple strategy adjustment process, but also a closed-loop improvement mechanism throughout the entire system. In some embodiments, the dynamic feedback also combines online learning algorithms to enhance the system's adaptability to dynamic traffic changes by real-time updating the model weights and parameters. For example, by introducing a reinforcement learning model, the optimization process can be modeled as a state-action-reward problem, and the optimal load distribution strategy can be learned through continuous interactions.

[0085] It should be further noted that this step fully considers the limitations of network resources and the fairness of load distribution during the optimization process. For example, in the scenario where node resources are limited, the system will give priority to ensuring the traffic allocation of critical tasks while reducing the processing volume of low-priority traffic. This design makes the present invention applicable not only to general network environments, but also to complex and highly dynamic scenarios.

[0086] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A network load balancing evaluation method based on a deep learning model, characterized in that: The following steps are involved: S1. Data preprocessing: Collect network traffic feature data, normalize, reduce noise and enhance features of the data; S2. Traffic feature manifold modeling: Use manifold learning methods to reduce the dimensionality of high-dimensional network traffic feature data, map it to a low-dimensional embedding space, and retain the local neighborhood structure of the traffic data; S3, Variational Information Bottleneck Feature Extraction: Extract low-dimensional features through deep learning models, optimize the relevance of features to load balancing goals, and reduce redundant information; S4, multi-objective performance evaluation: based on the extracted low-dimensional features, the throughput, average delay and node utilization balance of the load balancing strategy are optimized and evaluated; S5. Abnormal traffic detection: Model the traffic distribution based on the deep learning model, detect traffic anomalies and output the detection results; S6. Dynamic feedback and optimization: Dynamically adjust the load distribution strategy based on the performance evaluation results to optimize system performance.

2. The network load balancing evaluation method based on deep learning model according to claim 1 is characterized in that: The S2 step further comprises the following steps: S21, constructing a similarity matrix to represent the similarity between network traffic feature data; S22. Using the manifold learning algorithm, high-dimensional features are mapped to low-dimensional embedding space by optimizing the distance relationship between embedded points.

3. The network load balancing evaluation method based on deep learning model according to claim 1 is characterized in that: The S3 step further includes the following steps: S31. Extract low-dimensional latent features using deep learning models; S32, associating the extracted features with the load balancing target and retaining task related information; S33. Use regularization methods to constrain the low-dimensional feature distribution and reduce redundant information.

4. The network load balancing evaluation method based on deep learning model according to claim 3 is characterized in that: In the step S31, the deep learning model includes: The encoder is used to convert the input network traffic feature data into low-dimensional latent features and realize feature extraction by learning the latent distribution; A decoder is used to generate prediction results of load balancing target performance indicators based on low-dimensional latent features.

5. The network load balancing evaluation method based on deep learning model according to claim 1 is characterized in that: The S4 step further comprises the following steps: S41, calculating the throughput of the load balancing strategy; S42, calculating the average delay of the load balancing strategy; S43. Evaluate the discrete degree of node utilization and determine the utilization balance.

6. The network load balancing evaluation method based on deep learning model according to claim 5 is characterized in that: The S4 step uses a non-dominated sorting genetic algorithm to perform multi-objective performance evaluation, which specifically includes the following steps: S44, initialize the population and randomly generate candidate solutions; S45, calculating the fitness values ​​of the candidate solutions in the population, including throughput, delay, and utilization balance; S46, hierarchically sorting the candidate solutions according to the non-dominated relationship; S47. Generate the next generation of candidate solutions through crossover and mutation operations, and repeat the iteration until convergence.

7. The network load balancing evaluation method based on deep learning model according to claim 1 is characterized in that: The step S5 further comprises the following steps: S51. Establish a deep learning model, assume that the traffic characteristics obey the potential distribution, and use the model to learn the parameters of the traffic distribution; S52. According to the distribution probability value of the traffic characteristics, when the distribution probability value is lower than a set threshold, it is determined to be abnormal traffic.

8. The network load balancing evaluation method based on deep learning model according to claim 7 is characterized in that: In the step S52, the abnormal traffic adopts the dynamic adjustment load balancing strategy detection result: S521, adjusting the distribution weight of abnormal traffic requests; S522: Update the regularization parameters of the low-dimensional feature extraction model.

9. The network load balancing evaluation method based on deep learning model according to claim 1 is characterized in that: The step S6 further comprises the following steps: S61, performing periodic performance evaluation on the current load distribution strategy; S62, adjusting the weight parameters of the low-dimensional feature extraction model according to the evaluation results; S63. Update the load distribution strategy to adapt to the dynamic changes of network traffic.

10. The network load balancing evaluation method based on deep learning model according to claim 1, characterized in that: In the step S6, an optimal solution that meets the requirements of maximizing throughput and minimizing delay is selected, and the balance of node utilization is adjusted.

Images