5g-a test method and device based on cgan-pca

Abstract

The embodiment of the application provides a 5G-A test method and device based on CGAN-PCA, specifically: a first correspondence relationship between industrial scene parameters and network performance indexes is established based on a first artificial intelligence model; wherein the network performance indexes include throughput, delay and packet loss rate; a second correspondence relationship between the network performance indexes and performance evaluation features is established based on a second artificial intelligence model; target network performance indexes are acquired and generated according to the first correspondence relationship; target performance evaluation features are generated according to the target network performance indexes and the second correspondence relationship; and data quality is generated by performing multi-dimensional evaluation on the target network performance indexes and the target performance evaluation features. The application constructs a test data generation system with CGAN as a core architecture to realize effective mapping of test data on a real industrial scene in terms of structure, distribution, feature mode and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of testing, in particular to a 5G-A testing method and device based on CGAN-PCA. BACKGROUND

[0002] In recent years, 5G-A (5G-Advanced, 5G enhanced version) system as the evolution and upgrade of 5G, has significant enhancement in bandwidth, latency, connection reliability and device connection scale, provides strong support for efficient communication and cooperation of industrial internet, is an ideal choice to realize intelligentization and automation of industrial production, can support remote control, real-time monitoring of devices, fault warning and high-speed transmission of large-scale machine data, and promotes production modernization. It adopts network slicing technology to virtualize the physical network into multiple logical networks, reduces the cost and complexity of networking, and better adapts to the demand of flexible production; multi-connection technology allows the terminal to connect to multiple base stations at the same time, combined with advanced coding and modulation technology to improve data transmission reliability and spectrum utilization; zero-trust security architecture breaks the boundary between internal and external networks in traditional networks, and verifies and authorizes all access, effectively dealing with security threats brought by the blurring of 5G network boundaries.

[0003] In addition, ultra-wideband technology realizes high-precision positioning, and deterministic network technology guarantees deterministic data transmission, makes up for the shortcomings of existing 5G applications, and constructs a deterministic capability system from six dimensions of latency, space-time, reliability, business, security and system to empower industrial internet. For example, the project reduces energy consumption and cost through edge computing and network slicing technologies, and improves production and management efficiency, but it faces challenges such as electromagnetic interference, high-speed movement of devices, and large-scale concurrent connection of devices in industrial internet applications, which affect system performance, so comprehensive and strict testing is crucial, and the design and generation of test cases are key.

[0004] GANs (Generative Adversarial Networks, generative adversarial networks) as a powerful generative modeling framework, has been widely used in recent years due to its two-person minimax game structure, which consists of a generator and a discriminator. The two optimize each other in the game, the generator generates realistic samples to "deceive" the discriminator, and the discriminator distinguishes between real and generated samples, making it perform well in various generation tasks. With the development of various adaptations and extensions, applications are also expanding. In addition to traditional fields such as image generation and speech synthesis, StackGANs (Stacked Generative Adversarial Networks, stacked generative adversarial networks) focus on multi-stage high-resolution image generation, and TGAN (Conditional Tabular Generative Adversarial Networ, conditional tabular generative adversarial network) is designed specifically for synthesizing tabular data, showing wide adaptability.

[0005] But based on the traditional 5G-A test case generation method depends on rules and artificial experience, it is difficult to deal with complex and changeable industrial internet scene, test case coverage breadth and depth is limited, can not cover the potential fault mode, affect the test integrity and effectiveness. The traditional data generation method based on GAN (Generative Adversarial Network, generative adversarial network) has limitations in 5G-A test, lack of effective embedding mechanism of industrial communication scene parameters, training is unstable and prone to mode collapse or gradient disappearance, the architecture is not adapted to 5G-A high-dimensional, strong time sequence communication data, and lack of perfect generated data evaluation system. The traditional data feature extraction method also has some shortcomings in 5G-A test, which mainly focuses on macro trend prediction and cannot effectively extract communication performance related features. When feature dimensionality reduction, key information may be lost, and based on static assumption, it is difficult to adapt to dynamic data changes, reducing the practicality and timeliness of test data. SUMMARY

[0006] In view of the problems, the present application is proposed to provide a CGAN-PCA based 5G-A test method and device to overcome the problems or at least partially solve the problems, which includes:

[0007] The CGAN-PCA based 5G-A test constructs a first artificial intelligence model based on industrial scene parameters, which includes random noise vector and industrial scene condition label.

[0008] The steps include:

[0009] Based on the first artificial intelligence model, a first correspondence between the industrial scene parameters and network performance indicators is established, wherein the network performance indicators include throughput, delay and packet loss rate.

[0010] Based on the second artificial intelligence model, a second correspondence between the network performance indicators and performance evaluation features is established.

[0011] Obtain the target industrial scene parameters and generate the target network performance indicators according to the first correspondence.

[0012] Generate target performance evaluation features according to the target network performance indicators and the second correspondence.

[0013] Multi-dimensional evaluation is performed on the target network performance indicators and the target performance evaluation features to generate data quality.

[0014] Preferably, the step of establishing the first correspondence between the industrial scene parameters and the network performance indicators based on the first artificial intelligence model comprises:

[0015] The industrial scene parameters are input as conditional labels into a generator, and the network performance indicators are input as result labels into a discriminator to construct an adversarial network model.

[0016] The adversarial network model is updated alternately through an alternating updating mechanism to dynamically adjust training parameters, and the first correspondence is established through adversarial training of the adversarial network model.

[0017] Preferably, the step of constructing the adversarial network model, inputting the industrial scene parameters as conditional labels into a generator, and inputting the network performance indicators as result labels into a discriminator comprises:

[0018] The adversarial network model is constructed, and the concatenation of a random noise vector and an industrial scene conditional label is input into a generator to output a network performance indicator sample.

[0019] The concatenation of the network performance indicator sample and the conditional label is input into a discriminator to output a probability of authenticity.

[0020] Preferably, the step of updating the adversarial network model through an alternating updating mechanism to dynamically adjust training parameters, and establishing the first correspondence through adversarial training of the adversarial network model comprises:

[0021] The network parameters of the adversarial network model are initialized, and the training parameters are set.

[0022] The adversarial network model is updated through a loss function.

[0023] When the loss value of the adversarial network model reaches a predetermined stable threshold, the adversarial network model is normalized, and the first correspondence is established.

[0024] Preferably, the step of establishing the second correspondence between the network performance indicators and the data quality based on a second artificial intelligence model comprises:

[0025] The network performance indicators are subjected to feature extraction through performance evaluation feature analysis, and performance evaluation features with a predetermined variance contribution rate are selected to establish the second correspondence.

[0026] Preferably, the step of extracting features from the network performance indicators through performance evaluation feature analysis, and selecting performance evaluation features with a predetermined variance contribution rate to establish the second correspondence comprises:

[0027] The network performance indicators are subjected to standardization preprocessing to obtain preprocessed data.

[0028] The number of performance evaluation features is determined through cross-validation based on the preprocessed data, and performance evaluation features reflecting throughput quantile features, sample proportion features, and volatility features are extracted to establish the second correspondence.

[0029] Preferably, the step of performing multi-dimensional evaluation on the target network performance indicator and the target performance evaluation feature to generate data quality comprises:

[0030] According to the target network performance indicator and the target performance evaluation feature, and the distribution distance of real data in a high-dimensional feature space;

[0031] According to the distribution distance measurement, the difference degree on the probability distribution is generated;

[0032] According to the difference degree, the consistency level on local and global features is evaluated;

[0033] According to the consistency level, the data quality is generated.

[0034] To achieve the present application also includes a 5G-A test device based on CGAN-PCA, a first artificial intelligence model is constructed by industrial scene parameters, the industrial scene parameters include random noise vector and industrial scene condition label; comprising:

[0035] The first artificial intelligence model module is used to establish the first correspondence relationship between the industrial scene parameters and the network performance indicators based on the first artificial intelligence model; wherein the network performance indicators include throughput, delay and packet loss rate;

[0036] The second artificial intelligence model module is used to establish the second correspondence relationship between the network performance indicators and the performance evaluation features based on the second artificial intelligence model;

[0037] The target network performance indicator module is used to obtain the target industrial scene parameters and generate the target network performance indicators according to the first correspondence relationship;

[0038] The target performance evaluation feature module is used to generate the target performance evaluation features according to the target network performance indicators and the second correspondence relationship;

[0039] The data quality module is used to perform multi-dimensional evaluation on the target network performance indicators and the target performance evaluation features to generate data quality.

[0040] To achieve the present application also includes a computer electronic device, comprising a processor, a memory and a computer program stored on the memory and capable of running on the processor, when the computer program is executed by the processor, the steps of the 5G-A test method based on CGAN-PCA are realized.

[0041] To achieve the present application also includes a computer readable storage medium, the computer readable storage medium on which stores a computer program, the computer program is executed by the processor to achieve the steps of the CGAN-PCA based 5G-A test method.

[0042] The present application has the following advantages:

[0043] In the embodiments of the present application, the lack of dynamic driving mechanism in the prior art leads to insufficient data scene adaptability; the feature consistency evaluation system is imperfect, making it difficult to quantify data quality, which seriously restricts the reliability verification and performance optimization of 5G-A systems in industrial internet, and under the background of increasing complexity of test data demand of 5G-A systems in industrial internet, a test data generation method capable of dynamically adapting to variable industrial scenes and having quantitative evaluation capability is provided to meet the demand of high-fidelity and diversified data support, specifically: a first correspondence between the industrial scene parameters and network performance indicators is established based on a first artificial intelligence model; wherein the network performance indicators include throughput, latency and packet loss rate; a second correspondence between the network performance indicators and performance evaluation features is established based on a second artificial intelligence model; target industrial scene parameters are obtained and target network performance indicators are generated according to the first correspondence; target performance evaluation features are generated according to the target network performance indicators and the second correspondence; and the target network performance indicators and the target performance evaluation features are subjected to multidimensional evaluation to generate data quality. The present application is based on CGAN (Conditional Generative Adversarial Network) to design a test data generation model with industrial scene adaptability: a test data generation system with CGAN as the core architecture is constructed, and control parameters such as load type, network state and device behavior in industrial task scenes are introduced as generation conditions to realize effective mapping of test data on real industrial situations in terms of structure, distribution and feature mode. At the same time, by designing a high-robustness generator and discriminator network structure, stable generation of high-quality samples is realized in adversarial training, enhancing the generalization and adaptation ability of the model to complex industrial states, thereby improving the authenticity, representativeness and comprehensive coverage of the test data, providing basic data support for the reliability verification of 5G-A systems in industrial internet. PCA (Principal Component Analysis) feature extraction and multidimensional evaluation verification: combining classic dimension reduction methods such as PCA, the key performance indicators of 5G-A network, such as throughput, latency fluctuation and load state, are effectively extracted, and a cross-dimensional evaluation index system is established, including FID (Fréchet Inception Distance) based on distribution difference, KL divergence (Kullback-Leibler Divergence), MMD (Maximum Mean Discrepancy) and other measurement methods, as well as quantitative analysis methods such as cumulative variance contribution rate of PCA feature results, MSE (Mean Squared Error), R-squared coefficient and Euclidean distance.The evaluation framework can be used to reflect the consistency of the synthetic data and the real data in overall characteristics, probability distribution and structural similarity, and can also be used to guide the optimization and iteration of the generation model, and finally realize the comprehensive verification of the synthetic data in authenticity, information integrity and scene mapping ability. The test data generated by embedding the industrial scene parameter dynamic driving mechanism in the CGAN model can accurately match the complex and variable requirements of the industrial environment, solve the problem that the traditional method is difficult to cover the dynamic scene, significantly improve the coverage and applicability of the test case, and improve the scene adaptability of the test data. The application combines PCA feature extraction with FID, KL divergence, MMD and other multi-dimensional evaluation indexes to build a scientific and perfect evaluation system, which can comprehensively measure the quality of the generated data from the overall distribution, probability distribution and local features, and ensure the high consistency of the generated data with the real data, and realize the quantitative verification of the quality of the generated data. The application successfully solves the problem of dynamic generation of test data in the industrial internet scene of 5G-A system, provides strong data support for network performance optimization, reliability verification and intelligent decision-making, and helps to promote the wide application and actual deployment of 5G-A technology in the industrial field, so as to promote the landing of industrial internet application. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the description of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 is a step flow chart of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application;

[0046] Figure 2 is a generator structure schematic diagram of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application;

[0047] Figure 3 is an actual generator structure schematic diagram of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application;

[0048] Figure 4 is a discriminator structure schematic diagram of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application;

[0049] Figure 5 is an actual discriminator structure schematic diagram of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application;

[0050] Figure 6 is the learning rate 10 of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application -3 、10 -4 、10 -5 The corresponding loss function change diagram is shown in the figure.

[0051] Figure 7 is the corresponding loss function change diagram of the momentum decay rate 0.85, 0.9 of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application.

[0052] Figure 8 is the flow structure diagram of the CGAN-PCA-based 5G-A test method provided by an embodiment of the present application.

[0053] Figure 9 is the structure block diagram of the CGAN-PCA-based 5G-A test device provided by an embodiment of the present application.

[0054] Figure 10 is the structure diagram of a computer device provided by an embodiment of the present application.

[0055] 1, computer device; 2, external device; 3, processing unit; 4, bus; 5, network adapter; 6, I / O interface; 7, display; 8, memory; 9, random access memory; 10, cache memory; 11, storage system; 12, program / utility; 13, program module. DETAILED DESCRIPTION

[0056] In order to make the purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0057] The inventors found through analysis of the prior art that the 5G-A system, as an evolution and upgrade of the 5G system, not only significantly enhances bandwidth, latency, connection reliability and device connection scale, but also provides strong support for efficient communication and collaboration of the industrial internet. The advantages of 5G-A make it an ideal choice for realizing the intelligentization and automation of industrial production, supporting remote control, real-time monitoring of devices, fault warning and high-speed transmission of large-scale machine data, and promoting the modernization of production.

[0058] 5G-A system adopts network slicing technology to virtualize the physical network into multiple logical networks, thus reducing the networking cost and complexity, and better adapting to the flexible production demand. Multi-connection technology enables terminals to connect to multiple base stations simultaneously, while improving data transmission reliability and spectrum utilization through advanced coding and modulation techniques. The zero-trust security architecture breaks the boundaries between the internal and external networks of traditional networks, and verifies and authorizes all access, thus effectively addressing the security threats brought by the blurring of boundaries in 5G networks.

[0059] In addition, ultra-wideband technology realizes high-precision positioning, and deterministic network technology guarantees deterministic data transmission, making up for the shortcomings of existing 5G applications. The 5G / 5G-A system constructs a deterministic capability system from six dimensions to empower the industrial internet: latency determinism reduces latency through edge computing and network slicing; space-time determinism realizes precise positioning using high-precision satellite timing and Massive MIMO (Massive Multiple-Input Multiple-Output); reliable determinism improves data transmission reliability through redundant links and error correction coding; service determinism realizes intelligent scheduling with AI algorithms; secure determinism guarantees network security through encryption algorithms and enhanced security measures; and system determinism improves communication system stability through redundant servers and storage.

[0060] For example, through technologies such as edge computing and network slicing, not only energy consumption and cost are significantly reduced, but also production and management efficiency are improved. In the future, the deterministic technology of 5G-A will develop towards more precise fault handling, low-power high-precision positioning, multi-modal perception communication, etc., further promoting the application innovation of industrial internet.

[0061] However, 5G-A system faces many challenges in industrial internet applications, such as electromagnetic interference, high-speed movement of devices, and large-scale concurrent connection of devices, which will affect the performance of the system and cause communication interruption or data transmission errors. Therefore, to ensure the stable and reliable operation of 5G-A system in industrial internet, comprehensive and strict testing is essential. The design and generation of test cases are the key to testing work, which directly affects the coverage and effectiveness of testing.

[0062] In the prior art, GANs, as a powerful generative modeling framework, have been widely applied in recent years due to their unique two-player minimax game structure. A generative adversarial network consists of a generator and a discriminator, which optimize each other in the game process. The generator "cheats" the discriminator by generating realistic samples, while the discriminator tries to distinguish between real and generated samples. This training mechanism of the generative adversarial network makes it exhibit excellent performance in various generation tasks. With the in-depth development of generative adversarial networks, many different forms of adaptation and expansion have also emerged. For example, CategoricalGANs use unlabeled or partially labeled data to promote unsupervised and semi-supervised learning, improving the robustness of the model; GMAN further improves the diversity and performance of the generation process by introducing multiple discriminators; Mogren's C-RNN-GAN applies GAN to continuous data, especially in classical music generation, and performs well. In addition, Mescheder bridges the gap between variational autoencoders and generative adversarial networks through an adversarial variational Bayes method.

[0063] The application of generative adversarial networks is also expanding. In addition to its widespread application in traditional fields such as image generation and speech synthesis, StackGANs focus on multi-stage high-resolution image generation, further improving the quality of generation. TGAN is designed specifically for synthesizing tabular data, demonstrating the wide adaptability of GANs in handling different data formats and generation tasks.

[0064] Data feature extraction in the prior art, as an important part of data analysis, has received extensive attention and in-depth research from the academic community. For the prediction problem of power transformer feature parameters, a data preprocessing method based on discrete grey model is proposed, which successfully improves the prediction accuracy and solves the problems of randomness and non-equidistant data. A data feature-driven nuclear energy consumption prediction modeling method is proposed, emphasizing the importance of data features in the modeling process.

[0065] In the embodiments of the present application, the lack of dynamic driving mechanism in the prior art leads to insufficient data scene adaptability; the feature consistency evaluation system is imperfect, it is difficult to quantify data quality, which seriously restricts the reliability verification and performance optimization of 5G-A system in industrial internet, and under the background of increasing complexity of test data demand of 5G-A system in industrial internet, a test data generation method capable of dynamically adapting to variable industrial scenes and having quantitative evaluation capability is provided to meet the demand of high-fidelity and diversified data support, specifically: a first correspondence between the industrial scene parameters and network performance indicators is established based on a first artificial intelligence model; wherein the network performance indicators include throughput, delay and packet loss rate; a second correspondence between the network performance indicators and performance evaluation features is established based on a second artificial intelligence model; target industrial scene parameters are obtained and target network performance indicators are generated according to the first correspondence; target performance evaluation features are generated according to the target network performance indicators and the second correspondence; the target network performance indicators and the target performance evaluation features are evaluated in multiple dimensions to generate data quality. The present application is based on CGAN to design a test data generation model with industrial scene adaptability: a test data generation system with CGAN as the core architecture is constructed, and control parameters such as load type, network state and device behavior in industrial task scenes are introduced as generation conditions to realize effective mapping of test data on real industrial situations in terms of structure, distribution, feature mode, etc. At the same time, by designing a high-robustness generator and discriminator network structure, stable generation of high-quality samples is realized in adversarial training, enhancing the generalization and adaptation ability of the model to complex industrial states, thereby improving the authenticity, representativeness and comprehensive coverage of the test data, providing basic data support for the reliability verification of 5G-A system in industrial internet. PCA feature extraction and multi-dimensional evaluation verification: combining classic dimension reduction methods such as PCA, the key performance indicators of 5G-A network, such as throughput, delay fluctuation and load state, are effectively extracted, and a cross-dimensional evaluation index system is established, including FID, KL divergence, MMD and other measurement methods based on distribution difference, as well as quantitative analysis methods such as cumulative variance contribution rate, MSE, R-squared coefficient and Euclidean distance for PCA feature results. This evaluation framework can not only reflect the consistency of synthetic data and real data in overall features, probability distribution and structural similarity, but also guide the optimization and iteration of the generation model, ultimately realizing comprehensive verification of synthetic data in authenticity, information integrity and scene mapping ability. The present application embeds an industrial scene parameter dynamic driving mechanism in the CGAN model, and the generated test data can accurately match the complex and variable requirements of industrial environment, solving the problem that traditional methods are difficult to cover dynamic scenes, significantly improving the coverage and applicability of test cases, thereby improving the scene adaptability of test data.The application combines PCA feature extraction with FID, KL divergence, MMD and other multi-dimensional evaluation indexes to construct a scientific and perfect evaluation system, which can comprehensively measure the quality of generated data from multiple dimensions such as overall distribution, probability distribution and local features, ensure high consistency with real data, and realize quantitative verification of the quality of generated data. The application successfully solves the problem of dynamic generation of test data in the industrial internet scene of the 5G-A system, provides strong data support for network performance optimization, reliability verification and intelligent decision-making, helps to promote the wide application and actual deployment of 5G-A technology in the industrial field, and thus promotes the landing of industrial internet applications.

[0066] Reference Figure 1 , a step flow chart of the CGAN-PCA-based 5G-A test provided by an embodiment of the application is shown, and specifically includes the following steps:

[0067] S110, a first correspondence relationship between the industrial scene parameters and network performance indexes is established based on a first artificial intelligence model; wherein the network performance indexes include throughput, delay and packet loss rate;

[0068] S120, a second correspondence relationship between the network performance indexes and performance evaluation features is established based on a second artificial intelligence model;

[0069] S130, target network performance indexes are generated according to the first correspondence relationship based on target industrial scene parameters;

[0070] S140, target performance evaluation features are generated according to the target network performance indexes and the second correspondence relationship;

[0071] S150, the target network performance indexes and the target performance evaluation features are subjected to multi-dimensional evaluation to generate data quality.

[0072] In the following, the CGAN-PCA-based 5G-A test in the present exemplary embodiment will be further described.

[0073] In an embodiment of the application, the specific process of the step S110 of "establishing the first correspondence relationship between the industrial scene parameters and network performance indexes based on the first artificial intelligence model; wherein the network performance indexes include throughput, delay and packet loss rate" can be further described in combination with the following description.

[0074] As an example, when constructing the adversarial network model to establish the first correspondence relationship between the industrial scene parameters and the network performance indicators, the generator and the discriminator structures need to be designed respectively. The generator receives the industrial scene parameters as the conditional label, and maps the network performance indicators in the corresponding scene, such as throughput, delay, and packet loss rate, through a multi-layer neural network to generate prediction values. The discriminator inputs the real network performance indicators, i.e., the result label and the industrial scene parameters, and judges whether the input data is a real sample or a fake sample generated by the generator through feature extraction.

[0075] As an example, the alternating update mechanism is adopted in the training stage. First, the parameters of the discriminator are fixed, the loss function is the probability that the generated samples are misjudged as real, the parameters of the generator are optimized to make the output closer to the real distribution; then the parameters of the generator are fixed, the loss function is the accuracy of distinguishing real and generated samples, and the parameters of the discriminator are updated to enhance its discrimination ability. Through multiple rounds of adversarial iteration, when the generator and the discriminator reach Nash equilibrium, the generator can stably learn the nonlinear mapping relationship between the industrial scene parameters and the network performance indicators, i.e., the first correspondence relationship is formed. This process dynamically adapts to the parameter coupling characteristics of complex industrial environments and improves the mapping precision.

[0076] Step S110 specifically includes the following steps:

[0077] The adversarial network model is constructed, the industrial scene parameters are input into the generator as the conditional label, and the network performance indicators are input into the discriminator as the result label.

[0078] The adversarial network model is alternately updated and trained through the alternating update mechanism to dynamically adjust the training parameters, and the first correspondence relationship is established. The network performance indicators are test data.

[0079] In the embodiments of the present application, the network parameters of the adversarial network model are initialized and the training parameters are set; the adversarial network model is updated through the loss function; when the loss value of the adversarial network model reaches a predetermined stable threshold, the adversarial network model is normalized and the first correspondence relationship is established.

[0080] As an example, the concatenation of the random noise vector and the industrial scene conditional label is input into the generator, and is sequentially processed through three layers of full connection networks with different numbers of neurons to output network performance indicator samples containing key performance indicators; each layer of the full connection network is followed by a LeakyReLU activation function.

[0081] As an example, the concatenation of the network performance indicator samples and the conditional label is input into the discriminator, and is processed through four layers of discriminant full connection networks with different numbers of neurons to output a probability of authenticity; wherein each layer of the discriminant full connection network is followed by a Sigmoid activation function.

[0082] As an example, binary cross-entropy is used as the loss function; an optimizer with a preset initial learning rate is used for parameter updating; a training termination mechanism is triggered when the discriminator loss value reaches a predetermined stability threshold; batch normalization processing is implemented at the generator output layer, and a discriminator output probability threshold is set for sample screening.

[0083] As described in the following steps, initialize the network parameters of the generator and discriminator, set the training rounds and the initial learning rate of the optimizer; fix the generator parameters, update the discriminator parameters using the binary cross-entropy loss function; fix the discriminator parameters, update the generator parameters using the binary cross-entropy loss function; monitor the discriminator loss value in real time, and terminate the training process when the loss value fluctuation amplitude of consecutive multiple rounds is less than the preset stability threshold; add a batch normalization layer to the generator output layer to process the output data; set a screening threshold for the discriminator output probability to filter the generated samples; repeat the steps until the preset training rounds are completed or the early stopping condition is met.

[0084] In a specific embodiment, a generator model is constructed: the generator is the core of the entire model, and its input includes two parts: a random noise vector with a dimension of 100 and a conditional label in an industrial scenario, such as device density, service type, channel environment, etc. The schematic diagram and specific implementation of the generator are shown in Figure 2 and Figure 3 These input data will be processed through a three-layer fully connected network, and the network structure is 256, 512, and 1024 neurons in turn, each followed by a LeakyReLU activation function (negative slope of 0.2) for nonlinear mapping. Finally, the output of the generator will be test data samples that meet the characteristics of the 5G-A system, including key performance indicators such as throughput, latency, and packet loss rate.

[0085] In a specific embodiment, a discriminator model is constructed, and the role of the discriminator is to distinguish whether the input sample is a real sample or a generated sample. Its input is also the concatenation of data samples and conditional labels. The schematic diagram and specific implementation of the discriminator are shown in Figure 4 and Figure 5 The discriminator contains four fully connected networks with neuron counts of 512, 1024, 1024, and 1024 in turn, each followed by a LeakyReLU activation function. To prevent overfitting, a Dropout layer (dropout rate of 0.5) is also added to the second-to-last layer of the discriminator. The output layer of the discriminator uses a Sigmoid activation function, and the output value represents the probability that the sample is real data.

[0086] As an example, the training strategy of the generative adversarial model is optimized with the stability parameter. In terms of the training strategy, both the generator and the discriminator use binary cross-entropy as the loss function. The goal of the generator is to minimize the probability that the discriminator judges the generated sample as fake data, thereby "cheating" the discriminator; while the discriminator strives to maximize the discrimination accuracy of the real sample and the generated sample. The definition of the binary cross-entropy loss function is as follows:

[0087]

[0088] where y i represents the label of sample i (1 for real sample and 0 for generated sample), p i represents the probability that the sample is judged as positive, and N represents the number of samples.

[0089] During the training process, the optimizer uses the Adam algorithm, the initial learning rate is set to 1e-4, and the momentum parameters and are 0.9 and 0.999, respectively. To verify the rationality of the hyperparameters, a grid search experiment is conducted to compare three learning rates (1e-3, 1e-4, 1e-5) and two momentum decay rates (0.85 and 0.9). The specific changes are shown in Figure 6 and Figure 7

[0090] The results show that when the learning rate is 1e-4, the loss function has the most stable downward trend, and the convergence speed and stability are the best; when the momentum parameter is set to 0.9, compared with 0.85, the training shock amplitude is reduced by 37%, significantly improving the continuity and reliability of the training.

[0091] The entire training process uses an alternating update mechanism, that is, in each training period, the discriminator is updated first, and then the generator is updated to maintain the dynamic balance of adversarial training. At the same time, an early stopping mechanism is introduced, which automatically terminates the training when the loss value of the discriminator fluctuates by less than 1% for 10 consecutive rounds, thereby avoiding model overfitting or falling into a local optimal solution.

[0092] As an example, the quality of the generated data is guaranteed. In order to improve the quality of the generated data, the batch normalization mechanism is introduced in the output layer of the generator to unify the distribution characteristics of the data and enhance its reality and consistency in industrial scenarios. In addition, a confidence threshold of 0.45 is set for the output probability of the discriminator to filter the generated samples and eliminate potential outliers. This filtering strategy effectively guarantees the availability and structural rationality of the generated data, providing reliable support for subsequent modeling and performance testing.

[0093] ​In an embodiment of the present application, the specific process of "establishing the second correspondence relationship between the network performance indicators and the performance evaluation features based on the second artificial intelligence model" in step S120 can be further illustrated in combination with the following description.

[0094] Step S120 includes the following steps:

[0095] The performance evaluation features are principal components.

[0096] In an embodiment of the present application, the network performance indicators are standardized for preprocessing to obtain preprocessed data; the number of performance evaluation features is determined through cross-validation based on the preprocessed data, and performance evaluation features reflecting throughput quantile features, sample ratio features and volatility features are extracted to establish the second correspondence relationship.

[0097] In an embodiment of the present application, the specific process of "determining the number of performance evaluation features through cross-validation based on the preprocessed data, and extracting performance evaluation features reflecting throughput quantile features, sample ratio features and volatility features to establish the second correspondence relationship" can be further illustrated in combination with the following description.

[0098] As an example, data preprocessing and PCA feature extraction, in the data preprocessing stage, first, the original 5G-A test data is standardized to ensure that all feature dimensions are in the same scale, thereby avoiding the influence of dimension difference on the analysis result.

[0099] Subsequently, PCA is used for dimensionality reduction processing. The preliminary step of PCA is to set a large number of principal components to ensure that the main information of the data is retained. Through the cross-validation method, the number of principal components is gradually adjusted, and the reconstruction error under different numbers of principal components is calculated. According to the reconstruction error and the cumulative variance contribution rate, the number of principal components with a cumulative variance contribution rate of more than 95% is selected to ensure that the data after dimensionality reduction can effectively represent the characteristics of the original data. Finally, the key principal components are extracted through PCA to simplify the data structure and retain the most representative performance features.

[0100] As an example, feature extraction results and analysis, after PCA on the 5G-A test data, rich and valuable results are obtained, as shown in Table 1 below. These results help to deeply understand the network performance features contained in the data. From the loading matrix, each original feature shows significantly different coefficient distribution on different principal components.

[0101] Table 1: Loading matrix of PCA analysis

[0102]

[0103] PC1 mainly reflects the characteristic performance of throughput at the 25th percentile, with thput_25_dl and thput_25_ul coefficients of 0.456 and 0.439, respectively, indicating that PC1 effectively captures the comprehensive characteristics of uplink and downlink throughput under normal traffic. In addition, the contribution of total_thput_sample further strengthens the representativeness of PC1 for overall throughput.

[0104] PC2 reflects the relationship between total_thput_sample, thput_sample_ratio, thput_std_dl, and thput_std_ul. total_thput_sample and thput_sample_ratio show opposite trends, suggesting that as overall throughput increases, the sample ratio may decrease, which provides a basis for the regularity of network throughput distribution.

[0105] PC3 is mainly dominated by thput_std_dl (0.464) and thput_std_ul (0.499), reflecting the volatility of network uplink and downlink throughput. The large standard deviation coefficient indicates that PC3 can effectively evaluate the stability of network throughput at different times or regions.

[0106] PC4 mainly focuses on the coordinated changes between total_thput_sample (0.728) and thput_sample_ratio (0.670). When PC4 value increases, these two features change in the same direction, which is of reference significance for understanding the dynamic change pattern of network throughput.

[0107] PC5 is dominated by thput_75_dl (0.766) and thput_25_ul (-0.590), which show opposite trends. PC5 can be used to analyze the performance balance of the network under high load (downlink 75th percentile) and low load (uplink 25th percentile).

[0108] The five principal components extracted by PCA effectively integrate the original features, highlighting the throughput performance, volatility, and their interrelationships of 5G-A networks under different load conditions. This result not only simplifies the data structure but also provides data support for network performance optimization.

[0109] In a specific embodiment, the process of feature extraction and establishing the second correspondence relationship for network performance indicators requires three steps: standardization preprocessing, PCA dimension reduction and feature selection, and principal component analysis.

[0110] As an example, the network performance indicators (including throughput, latency, packet loss rate, etc.) are standardized preprocessed. By eliminating dimensional differences (such as converting different units of throughput Mbps and latency ms into dimensionless data), it ensures that all indicators participate in analysis on the same scale, avoids interference with the feature extraction result due to the difference in value range, and lays a data foundation for subsequent PCA analysis.

[0111] As an example, the number of performance evaluation features is determined based on the preprocessed data through cross-validation. A large number of principal components is initially set to retain the core information of the original data, and then the number of principal components is adjusted repeatedly through cross-validation: the reconstruction error under different numbers is calculated, and the cumulative variance contribution rate is tracked at the same time, and finally the number of principal components with a cumulative variance contribution rate of more than 95% is selected, such as the 5 principal components in the example. While simplifying the data structure, it ensures that the data after dimension reduction can still effectively represent the original network performance features.

[0112] As an example, performance evaluation features reflecting specific attributes are extracted to establish the second correspondence relationship. Combined with PCA load matrix analysis, the five principal components correspond to key features: PC1 is dominated by thput_25_dl (0.456) and thput_25_ul (0.439), reflecting the comprehensive features of uplink and downlink throughput at the 25th percentile; PC2 reflects the sample proportion feature through the inverse change of total_thput_ratio and thput_sample_ratio; PC3 is dominated by thput_std_dl (0.464) and thput_std_ul (0.499), capturing the throughput volatility feature; PC4 focuses on the synergistic change of total_thput_ratio and thput_sample_ratio, and PC5 reflects the performance balance under high / low load through the inverse relationship between thput_75_dl and thput_25_ul. These principal components together constitute the performance evaluation feature set, and their mapping relationship with the original network performance indicators is the second correspondence relationship, which provides structured feature basis for accurately analyzing network performance rules.

[0113] In an embodiment of the present application, the specific process of "obtaining target industrial scene parameters and generating target network performance indicators according to the first corresponding relationship" in step S130 can be further illustrated in combination with the following description.

[0114] In a specific embodiment, when obtaining target industrial scene parameters, key parameter data under the scene needs to be collected, including device running state, data transmission scale, task type and priority, environmental interference factors, etc., to form a structured parameter set. The parameter set is input into the generator of the trained adversarial network model, and the generator outputs the corresponding target network performance indicator prediction value through internal neural network operation according to the first corresponding relationship established in advance, covering throughput, latency and packet loss rate.

[0115] During the generation process, the model uses the nonlinear mapping ability learned in the training stage to accurately adapt to the characteristics of the target scene parameters, ensuring that the output performance indicators can reflect the real running situation of the network under the scene, providing quantitative basis for subsequent network optimization, resource allocation, etc.

[0116] In an embodiment of the present application, the specific process of "generating target performance evaluation features according to the target network performance indicators and the second corresponding relationship" in step S140 can be further illustrated in combination with the following description.

[0117] As an example, when generating target performance evaluation features according to target network performance indicators and the second corresponding relationship, first, the target network performance indicators (throughput, latency, packet loss rate, etc.) are standardized and preprocessed to eliminate dimensional differences and converted into preprocessed data that meet the requirements of PCA analysis. Based on the established second corresponding relationship, the principal component extraction model determined through cross-validation is called to perform feature mapping on the preprocessed data. Specifically, the weight relationship between each principal component in the PCA load matrix and the original performance indicators is used to calculate the principal component scores under the target scene: PC1 generates an evaluation value reflecting the throughput characteristics under general load by weighted combination of the 25th percentile of the target downlink and uplink throughputs (thput_25_dl, thput_25_ul); PC2 outputs the sample distribution feature indicator by combining the inverse correlation between the total throughput ratio (total_thput_ratio) and the sample ratio (thput_sample_ratio); PC3 calculates the volatility of network throughput according to the standard deviations of the downlink and uplink throughputs (thput_std_dl, thput_std_ul); PC4 generates a dynamic mode feature by the synergistic change of the total throughput ratio and the sample ratio; and PC5 evaluates the high-low load balance state based on the inverse relationship between the 75th percentile of the downlink and the 25th percentile of the uplink.

[0118] As an example, the generated set of target performance evaluation features not only retains more than 95% of the information in the original indicators, but also simplifies the feature structure by principal component integration, providing accurate and efficient analysis basis for subsequent network performance evaluation and optimization.

[0119] In an embodiment of the present application, the specific process of "evaluating the data quality of the target network performance indicators and the target performance evaluation features in multiple dimensions" described in step S150 can be described in combination with the following description.

[0120] Step S150 includes the following steps: according to the target network performance indicators and the target performance evaluation features, and the distribution distance of the real data in the high-dimensional feature space; according to the distribution distance measurement to generate the difference degree on the probability distribution; according to the difference degree to evaluate the consistency level on the local and global features; and according to the consistency level to generate the data quality.

[0121] As an example, the distribution distance of the generated data and the real data in the high-dimensional feature space is calculated; based on the calculation results of the above steps, the difference degree of the two in the probability distribution is measured; according to the measurement results of the above steps, the consistency level of the two in the local and global features is evaluated; and the data quality score is generated by integrating the above evaluation results.

[0122] As an example, in order to ensure the quality of the generated data, the CGAN generated data evaluation adopts multiple evaluation methods. First, FID is used to measure the distribution similarity of the generated data and the real data in the high-dimensional feature space. The lower the FID value, the higher the quality of the generated data, and the closer the distribution to the real data. The calculation method of FID value is:

[0123]

[0124] where x represents the feature set of the real data, g represents the feature set of the generated data, represents the mean vector of the real data features, represents the mean vector of the generated data features, represents the covariance matrix of the real data features, represents the covariance matrix of the generated data features. Then, KL divergence is used to evaluate the difference between the generated data and the real data in the probability distribution. The smaller the KL divergence value, the closer the distribution of the generated data to the real data. The calculation method of KL value is:

[0125]

[0126] where, represents the probability distribution of the real data, Q represents the probability distribution of the generated data, p(x) represents the probability of sample x under the real data distribution, p(x) represents the probability of sample x under the real data distribution,

[0127] Finally, MMD is used to measure the difference between generated data and real data in local and global features, and the lower the MMD value, the more consistent the generated data is with the real data in global distribution and local structure. The calculation method of MMD value is:

[0128]

[0129] where, p(x) represents the probability of sample x under the real data distribution, p(x) represents the probability of sample x under the real data distribution,

[0130] As an example, the selected measure of feature extraction is CEVR, which is used to evaluate the ability of principal components to retain original information, representing the cumulative contribution of each principal component to the total variance. Generally, principal components with cumulative contribution rate reaching more than 95% are selected to ensure that the main features of the data are retained as much as possible, while noise and redundant information are discarded. The calculation method of CEVR is:

[0131]

[0132] where k represents the number of retained principal components, g represents the total number of original data features, λi represents the eigenvalue of the ith principal component, σi represents the variance of the ith original feature, σj represents the variance of the jth original feature, CEVR represents the cumulative variance contribution rate of the first k principal components.

[0133] MSE measures the reconstruction error between the reduced data and the original data, reflecting the degree of information loss. The smaller the value, the better the reduction effect, and the original data structure is better preserved. The specific formula is as follows:

[0134]

[0135] where n represents the number of samples, and p represents the number of features, xi,j represents the jth feature value of the ith sample of the original data, xi,j represents the jth feature value of the ith sample of the original data,

[0136] R-squared reflects the fitting degree between the reduced data and the original data, and the closer the value to 1, the more information is retained. It is an important reference for evaluating the effectiveness of PCA, but it should be comprehensively judged in combination with other indicators to avoid excessive dependence. The calculation method of R-squared is:

[0137]

[0138] wherein, denotes the data value after dimensionality reduction reconstruction, denotes the original data value, denotes the mean value of the original data, n denotes the sample number, and p denotes the feature number.

[0139] The Euclidean distance measures the spatial difference between the original data and the reduced data. The smaller the distance, the higher the similarity of the data after dimensionality reduction, indicating that PCA effectively preserves the original structure and characteristics.

[0140] Through these comprehensive evaluation results, the principal component number and feature extraction process are further optimized to ensure that the data after dimensionality reduction can effectively support subsequent data analysis and modeling tasks.

[0141] As an example, the GAN-generated data evaluation results are shown in Table 2, which shows that the FID values of each feature are generally low, indicating that the GAN-generated data is close to the real data distribution in the high-dimensional feature space. Among them, thput_sample_ratio (0.00217), thput_std_dl (0.00554), etc. are particularly outstanding, indicating that the model can effectively simulate the throughput-related features. However, the FID of total_thput_sample is relatively high, which needs to be further analyzed in combination with other indicators.

[0142] Table 2 FID values corresponding to each feature value

[0143]

[0144] From Table 3, the KL divergence of most features is small, reflecting that the probability distribution difference between the GAN-generated data and the real data is small, and the fitting effect is good. The KL value of thput_75_dl is relatively high, which may be due to the complex distribution and difficulty in network fitting.

[0145] Table 3 KL divergence corresponding to each feature value

[0146]

[0147] Table 4 shows that the MMD values of all features are low, indicating that the GAN has preserved the overall data distribution consistency while also retaining the details of local features. In particular, thput_sample_ratio (0.00352) performs best, further verifying the quality of the model's generation.

[0148] Table 4 MMD values corresponding to each feature value

[0149]

[0150] Through the evaluation of the three indicators of FID, KL divergence, and MMD, we conclude that the 5G-A test data generated by GAN has high similarity in overall feature distribution, probability distribution, and local features with real data. These evaluation results show that the GAN model performs well in generating close-to-real 5G-A network data, providing high-quality synthetic data support for subsequent research and application.

[0151] The present scheme proposes an innovative 5G-A system test data generation and evaluation method, which solves the problem of dynamic generation of test data in the industrial internet scene through the synergistic effect of conditional generative adversarial network and principal component analysis. The method first constructs an industrial scene perception type CGAN model, taking device density, business type and other key parameters as conditional input, and realizes on-demand generation of test data. Then, through PCA technology, the core features of the data are extracted, and a multi-dimensional evaluation system is established to ensure the authenticity of the generated data in key indicators such as throughput and latency. The entire technical process forms a closed-loop optimization mechanism of "generation-extraction-evaluation".

[0152] In a specific embodiment, the core innovation of the present application is reflected in three aspects: first, an industrial scene parameter dynamic driving mechanism is designed, which makes the generated data accurately match the complex and changing needs of the industrial environment through the embedding of conditional labels; second, a stable and efficient adversarial training strategy is developed, which combines alternating update with early stopping mechanism to significantly improve the stability of model training; finally, a multi-dimensional evaluation system based on PCA features is constructed, which realizes scientific verification of the quality of generated data through quantitative indicators such as FID and KL divergence. These innovations together solve the problems of insufficient scene adaptability and imperfect evaluation system of traditional methods. The implementation process is divided into four key stages: the model construction stage completes the network design of the generator and discriminator, among which the generator adopts a three-layer fully connected structure and the discriminator introduces Dropout to prevent overfitting; the training optimization stage dynamically adjusts the learning rate and momentum parameters to make the model loss function converge quickly; the feature extraction stage performs PCA dimensionality reduction on the generated data, retaining principal components with a variance contribution rate of more than 95%; the quality evaluation stage performs comprehensive verification from the dimensions of distribution distance and probability difference.

[0153] Among them, Dropout is a random inactivation, which is a commonly used regularization technique in deep learning, mainly used to prevent neural network overfitting. Its working principle is: in the model training process, a part of neurons (and their connections) are temporarily "inactivated" (i.e. set to 0), and the set of neurons inactivated each iteration changes randomly.

[0154] This mechanism forces the network to learn more robust features — features that are not dependent on the presence of particular neurons, thus reducing co-adaptation between neurons and enhancing the model's ability to generalize. After training, dropout is not performed in the prediction phase, and all neurons participate in the computation.

[0155] Dropout is widely used in fully connected networks, convolutional neural networks (CNNs), and other types of networks, and is a classic method for improving model stability.

[0156] In a specific embodiment, the present application is based on a CGAN-based industrial scene perception data generation method: for the first time, CGAN is introduced into industrial communication testing, by splicing scene parameters (such as device density, service type, channel environment) with generator input, a controllable and dynamically driven data generation mechanism is realized, effectively improving the scene adaptability of generated data in key indicators such as throughput, latency, and packet loss rate. Stable and efficient adversarial training strategy: adopt a training method combining alternating update and early stopping mechanism, automatically terminate training through loss fluctuation monitoring, prevent overfitting, and introduce batch normalization and confidence filtering to significantly improve the quality and reliability of generated data. Feature dimension reduction scheme for network performance optimization: combined with PCA principal component analysis technology, extract core performance features such as throughput and latency from original 5G-A test data, ensure the representativeness and high fidelity of compressed data, and provide support for subsequent model decision-making and system evaluation.

[0157] For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts are described in the method embodiment.

[0158] Referring to Figure 9 , a 5G-A test device based on CGAN-PCA is shown, which comprises the following modules:

[0159] The first artificial intelligence model module 910 is configured to establish a first correspondence between the industrial scene parameters and the network performance indicators based on a first artificial intelligence model; wherein the network performance indicators include throughput, latency, and packet loss rate.

[0160] The second artificial intelligence model module 920 is configured to establish a second correspondence between the network performance indicators and the performance evaluation features based on a second artificial intelligence model.

[0161] The target network performance indicator module 930 is configured to obtain target industrial scene parameters and generate target network performance indicators according to the first correspondence.

[0162] The target performance evaluation feature module 940 is configured to generate a target performance evaluation feature according to the target network performance indicator and the second correspondence.

[0163] The data quality module 950 is configured to perform multi-dimensional evaluation on the target network performance indicator and the target performance evaluation feature to generate data quality.

[0164] In an embodiment of the present application, the first artificial intelligence model module 910 includes:

[0165] The adversarial network model submodule is configured to construct an adversarial network model, input the industrial scene parameter as a conditional label into a generator, and input the network performance indicator as a result label into a discriminator.

[0166] The first correspondence submodule is configured to perform adversarial training on the adversarial network model by dynamically adjusting training parameters through an alternating update mechanism and establish the first correspondence.

[0167] In an embodiment of the present application, the adversarial network model submodule includes:

[0168] The network performance indicator sample output submodule is configured to construct an adversarial network model, input a splicing of a random noise vector and an industrial scene conditional label into a generator, and output a network performance indicator sample.

[0169] The authenticity probability output submodule is configured to input a splicing of the network performance indicator sample and a conditional label into a discriminator and output an authenticity probability.

[0170] In an embodiment of the present application, the first correspondence submodule includes:

[0171] The training parameter submodule is configured to initialize network parameters of the adversarial network model and set training parameters.

[0172] The update parameter submodule is configured to update parameters of the adversarial network model through a loss function.

[0173] The normalization processing submodule is configured to perform normalization processing on the adversarial network model and establish the first correspondence when a loss value of the adversarial network model reaches a predetermined stable threshold.

[0174] In an embodiment of the present application, the second artificial intelligence model module includes:

[0175] The second correspondence submodule is configured to perform feature extraction on the network performance indicator by performance evaluation feature analysis, select performance evaluation features with a predetermined variance contribution rate, and establish the second correspondence.

[0176] In an embodiment of the present application, the second correspondence submodule includes:

[0177] a preprocessing data submodule configured to perform standardization preprocessing on the network performance indicators to obtain preprocessing data;

[0178] a feature extraction submodule configured to determine a performance evaluation feature quantity by cross validation according to the preprocessing data, and extract performance evaluation features reflecting throughput quantile features, sample ratio features and volatility features to establish the second correspondence.

[0179] In an embodiment of the present application, the data quality module comprises:

[0180] a distribution distance submodule configured to determine a distribution distance of the target network performance indicators and the target performance evaluation features in a high-dimensional feature space according to the target network performance indicators and the target performance evaluation features;

[0181] a difference degree submodule configured to generate a difference degree on a probability distribution according to the distribution distance measurement;

[0182] a consistency level submodule configured to evaluate a consistency level on local and global features according to the difference degree;

[0183] a data quality submodule configured to generate data quality according to the consistency level.

[0184] The computer device 1 is in the form of a general-purpose computing device, and the components of the computer device 1 can include but are not limited to one or more processors or processing units 3, a memory 8, and a bus 4 connecting different system components including the memory 8 and the processing unit 3.

[0185] The bus 4 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor or local bus using any of a variety of bus architectures. For example, these architectures include but are not limited to an industry standard architecture (ISA) bus, a microchannel architecture (MAC) bus, an enhanced ISA bus, an audio video electronics standards association (VESA) local bus, and a peripheral component interconnect (PCI) bus.

[0186] The computer device 1 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 1 and includes both volatile and nonvolatile media, removable and non-removable media.

[0187] Memory 8 can include computer system readable media in the form of volatile memory, such as random access memory 9 and / or cache memory 10. Computer device 1 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 11 can be used for reading from and writing to non-removable, non-volatile magnetic media (typically called a "hard drive"). Although Figure 10 Although not shown in FIG. 10, a disk drive can be used for reading from or writing to a non-removable, non-volatile magnetic media (e.g., a "hard drive"). Although not shown in FIG. 10, an optical disk drive can be used for reading from or writing to a non-removable, non-volatile optical media such as an optical disk (e.g., a CD-ROM, DVD-ROM, etc.). In such instances, each drive can be connected to bus 4 by one or more data media interfaces. The drives and their associated computer-readable media provide non-volatile storage of computer-readable instructions, data structures, program modules, and other data for computer device 1.

[0188] Program / utility 12, having a set (at least one) of program modules 13, can be stored in, for example, memory by way of example, and not limitation, includes an operating system, one or more application programs, other program modules 13, and program data, each of which or a combination can include implementation of a network environment. Program modules 13 generally carry out the functions and / or methodologies of embodiments described herein.

[0189] Computer device 1 can also communicate with one or more external devices 2 such as a keyboard, a pointing device, a display 7, a camera, etc.; one or more devices that enable a user to interact with computer device 1; and / or one or more devices that enable computer device 1 to communicate with one or more other computer devices. Such communication can be via I / O interface 6. Additionally, computer device 1 can communicate with one or more networks such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet) via network adapter 5. As Figure 10 illustrated, network adapter 5 can communicate with the other components of computer device 1 via bus 4. It should be understood that although not shown, other hardware and / or software components that are commonly used in computing can also be included in computer device 1 such as, for example, a non-transitory computer-readable medium that stores the one or more program modules 13. Figure 10 It should be understood that, although not shown in FIG. 10, other hardware and / or software components could be used in conjunction with computer device 1. These components, which can be implemented as program modules 13, are meant to be illustrative and not restrictive, as such components can be added to or removed from computer device 1, and / or can be implemented as an integrated element of computer device 1. To that end, such other components are meant to be within the scope of computer device 1, as that term is used in this document.

[0190] Processing unit 3 can execute various program applications and data processing by running programs stored in memory 8, such as implementing the CGAN-PCA based 5G-A testing provided by embodiments of the present application.

[0191] That is, the above processing unit 3 implements the following when executing the above program: establishing a first correspondence relationship between the industrial scene parameters and network performance indicators based on a first artificial intelligence model; wherein the network performance indicators include throughput, latency and packet loss rate; establishing a second correspondence relationship between the network performance indicators and performance evaluation features based on a second artificial intelligence model; obtaining target industrial scene parameters and generating target network performance indicators according to the first correspondence relationship; generating target performance evaluation features according to the target network performance indicators and the second correspondence relationship; and performing multi-dimensional evaluation on the target network performance indicators and the target performance evaluation features to generate data quality.

[0192] In the embodiments of the present application, the present application also provides a computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the CGAN-PCA based 5G-A test as provided in all embodiments of the present application.

[0193] That is, the above processing unit 3 implements the following when executing the above program: establishing a first correspondence relationship between the industrial scene parameters and network performance indicators based on a first artificial intelligence model; wherein the network performance indicators include throughput, latency and packet loss rate; establishing a second correspondence relationship between the network performance indicators and performance evaluation features based on a second artificial intelligence model; obtaining target industrial scene parameters and generating target network performance indicators according to the first correspondence relationship; generating target performance evaluation features according to the target network performance indicators and the second correspondence relationship; and performing multi-dimensional evaluation on the target network performance indicators and the target performance evaluation features to generate data quality.

[0194] Any combination of one or more computer readable medium can be employed. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In this document, the computer readable storage medium can be any tangible medium that contains, or stores a program for use by or in connection with an instruction execution system, apparatus, or device.

[0195] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0196] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the operator's computer, partially on the operator's computer, as a standalone software package, partially on the operator's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the operator's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider). The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

[0197] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0198] Finally, it is to be understood that the phraseology or terminology such as "first" and "second" etc. used herein is merely intended to differentiate one entity or operation from another entity or operation, without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0199] The above describes the CGAN-PCA-based 5G-A test method and device provided by the present application in detail, and the principles and implementation manners of the present application are described by using specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed, and the above description of the present application should not be understood as a limitation.

Claims

1. A 5G-A test method based on CGAN-PCA, characterized in that, A first artificial intelligence model is constructed by industrial scene parameters, the industrial scene parameters including a random noise vector and an industrial scene condition label; The steps include; A first correspondence between the industrial scene parameters and network performance indicators is established based on the first artificial intelligence model, wherein the network performance indicators include throughput, latency, and packet loss rate; an adversarial network model is constructed, the industrial scene parameters are input into a generator as condition labels, and the network performance indicators are input into a discriminator as result labels; the adversarial network model is subjected to adversarial training by dynamically adjusting training parameters using an alternating update mechanism and the first correspondence is established; A second correspondence between the network performance indicators and performance evaluation features is established based on a second artificial intelligence model; feature extraction is performed on the network performance indicators using performance evaluation feature analysis, and performance evaluation features with a predetermined variance contribution rate are selected to establish the second correspondence; Target network performance indicators are generated from target industrial scene parameters based on the first correspondence; Target performance evaluation features are generated from the target network performance indicators based on the second correspondence; Data quality is generated by multi-dimensional evaluation of the target network performance indicators and the target performance evaluation features; a distribution distance of the target network performance indicators and the target performance evaluation features in a high-dimensional feature space is determined; a difference degree on a probability distribution is generated based on the distribution distance measurement; a consistency level on local and global features is evaluated based on the difference degree; and data quality is generated based on the consistency level. 2.The CGAN-PCA based 5G-A testing method of claim 1, wherein, The step of constructing an adversarial network model, inputting the industrial scene parameters into a generator as condition labels, and inputting the network performance indicators into a discriminator as result labels includes: An adversarial network model is constructed, a random noise vector and an industrial scene condition label are input into a generator, and network performance indicator samples are output; The network performance indicator samples and the condition labels are input into a discriminator, and authenticity probability is output. 3.The CGAN-PCA based 5G-A testing method of claim 1, wherein, The step of dynamically adjusting training parameters using an alternating update mechanism, performing adversarial training on the adversarial network model, and establishing the first correspondence includes: The network parameters of the adversarial network model are initialized and the training parameters are set; The adversarial network model is updated by a loss function; When the adversarial network model loss value reaches a predetermined stable threshold, the adversarial network model is normalized and the first correspondence is established. 4.The CGAN-PCA based 5G-A testing method of claim 1, wherein, The step of performing feature extraction on the network performance indicators using performance evaluation feature analysis, and selecting performance evaluation features with a predetermined variance contribution rate to establish the second correspondence includes: The network performance indicators are standardized to obtain preprocessed data; The number of performance evaluation features is determined by cross-validation based on the preprocessed data, and performance evaluation features reflecting throughput quantile features, sample proportion features, and volatility features are extracted to establish the second correspondence.

5. A 5G-A test device based on CGAN-PCA, characterized by, The first artificial intelligence model is constructed by industrial scene parameters, the industrial scene parameters include a random noise vector and an industrial scene condition label; comprising: A first artificial intelligence model module is configured to establish a first correspondence between the industrial scene parameters and network performance indicators based on the first artificial intelligence model; wherein the network performance indicators include throughput, latency, and packet loss rate; an adversarial network model is constructed, the industrial scene parameters are input into a generator as condition labels, and the network performance indicators are input into a discriminator as result labels; the adversarial network model is subjected to adversarial training by dynamically adjusting training parameters through an alternating update mechanism and the first correspondence is established; A second artificial intelligence model module is configured to establish a second correspondence between the network performance indicators and performance evaluation features based on a second artificial intelligence model; the network performance indicators are subjected to feature extraction by performance evaluation feature analysis, and performance evaluation features with a predetermined variance contribution rate are selected to establish the second correspondence; A target network performance indicator module is configured to obtain target industrial scene parameters and generate target network performance indicators based on the first correspondence; A target performance evaluation feature module is configured to generate target performance evaluation features based on the target network performance indicators and the second correspondence; A data quality module is configured to perform multi-dimensional evaluation on the target network performance indicators and the target performance evaluation features to generate data quality; the target network performance indicators and the target performance evaluation features, and the distribution distance of real data in a high-dimensional feature space; the difference degree on the probability distribution is generated based on the distribution distance measurement; the consistency level on local and global features is evaluated based on the difference degree; and the data quality is generated based on the consistency level.

6. A computer electronic device, comprising: A processor, a memory, and a computer program stored on the memory and capable of running on the processor are included, and the computer program is executed by the processor to implement the steps of the CGAN-PCA based 5G-A test method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the steps of the CGAN-PCA based 5G-A test method according to any one of claims 1 to 4.

Images