A wireless communication network fault diagnosis method based on a generative adversarial network

Abstract

The application relates to the technical field of wireless mobile communication networks, and discloses a wireless communication network fault diagnosis method based on a generative adversarial network, which comprises the following steps: inputting collected network fault real data; performing data processing and extracting key performance indicators (KPI) of equipment; designing an NS-DCGAN, inputting the KPI of the equipment into the NS-DCGAN, and generating virtual fault data under different network states; supplementing external environment characteristic data to the virtual fault data by using a clustering algorithm; combining the virtual data set and the real data set, and constructing a wireless communication network fault sample data set D; dividing the data set D into an internal equipment indicator part D1 and an external environment characteristic part D2; designing an EAFD-GhostNet, inputting D1 and D2 into the EAFD-GhostNet; performing wireless communication network fault classification diagnosis, and outputting a diagnosis result. The application can solve the problems of insufficient fault samples and uneven data distribution, can realize short-time and high-efficiency multi-dimensional diagnosis and analysis of faults, and can effectively improve the accuracy of fault classification diagnosis.

Description

Technical Field

[0001] This invention relates to the field of wireless mobile communication network technology, and in particular to self-organizing network technology. Specifically, it discloses a method for fault diagnosis of wireless communication networks based on generative adversarial networks, which is oriented towards cross-virtual and real integration based on a generative adversarial sample library. Background Technology

[0002] With the rapid development of 5G and 6G technologies, mobile communication network structures are becoming increasingly complex, and network operation and maintenance costs are rising daily. SON (Self-Organized Networks) technology can greatly improve the automation and intelligence level of wireless mobile communication networks, effectively reducing operating costs.

[0003] Fault diagnosis, as a crucial component of SON (Signal-Oriented Network) technology, primarily aims to analyze reported network parameters, flexibly provide fault warnings and diagnoses, and ensure the Quality of Service (QoS) of the communication network. In wireless communication networks, nodes are constantly moving, network conditions are complex and prone to change, faults are easily detected, and they need to be diagnosed and processed quickly. Simultaneously, network equipment has limited computing resources, requiring the system to rapidly perform fault diagnosis under these resource constraints. Therefore, the design of fault diagnosis algorithms must fully consider frequent network changes and resource limitations. Furthermore, existing communication network fault samples suffer from insufficient sample size and uneven distribution, which significantly hinders the training and design of fault diagnosis algorithms. Existing technologies have the following shortcomings:

[0004] (1) Existing fault diagnosis designs are based on fixed node communication networks with stable network structures and link resources, which cannot be directly adapted to wireless mobile communication networks with scarce computing resources.

[0005] (2) Existing fault diagnosis designs are difficult to implement and have poor results when there are insufficient communication fault samples and uneven data distribution.

[0006] (3) Existing fault diagnosis designs for wireless mobile communication networks mostly consider the system equipment itself, and pay less attention to objective external environmental factors. Summary of the Invention

[0007] To address the aforementioned technical problems in the existing technology, this invention proposes a fault diagnosis method for wireless communication networks based on generative adversarial networks, comprising the following steps:

[0008] Step 1: Collect real network fault data from heterogeneous wireless communication network environments, including key performance indicator (KPI) data of internal devices and external environmental characteristic data;

[0009] Step 2: Perform data processing, including data cleaning and filling, to form a real dataset, and separate and extract the key performance indicators (KPIs) data of the equipment;

[0010] Step 3: Design NS-DCGAN. Input the key performance indicators (KPIs) of the device into NS-DCGAN. Based on the adversarial approach of generator and discriminator, learn the existing data distribution and generate virtual fault data under different network conditions.

[0011] Step 4: Use a clustering algorithm to supplement the virtual fault data with external environmental feature data to obtain a virtual dataset;

[0012] Step 5: Merge the virtual dataset with the real dataset to construct the wireless communication network fault sample dataset D;

[0013] Step 6: Divide the dataset D into an internal device indicator part D1 and an external environment feature part D2;

[0014] Step 7: Design a dual-path environmental adaptive fault diagnosis neural network EAFD-GhostNet. After processing the data of the internal equipment indicator part D1, input it into Path1 of the environmental adaptive fault diagnosis neural network EAFD-GhostNet for feature extraction. After processing the data of the external environment feature part D2, input it into Path2 of the environmental adaptive fault diagnosis neural network EAFD-GhostNet for encoding. Connect the data encoded by Path2 with the data extracted by Path1 to perform wireless communication network fault classification and diagnosis, and output the diagnosis results.

[0015] Furthermore, the key performance indicator (KPI) data of the equipment extracted in step 2 includes:

[0016] Collect KPI data using time aggregation levels of different lengths;

[0017] Define S = [KPI1, KPI2, KPI3, ..., KPI m ] represents the input vector for different network states, and S is a vector containing m KPIs;

[0018] Define C = [FC1, FC2, FC3, ..., FC n ] represents the n distinct states of the network;

[0019] Within time T, the network experiences a fault state FC. i State samples Represented as:

[0020]

[0021] For the selected KPIi The normalization operation is as follows:

[0022]

[0023] Among them, KPI i Let i be the normalized i-th type of KPI.

[0024] Furthermore, the NS-DCGAN described in step 3 includes a generator G and a discriminator D, and the activation function Nselu(x) is:

[0025]

[0026] The optimized objective function is:

[0027] Discriminator:

[0028] Generator:

[0029] Among them, P data P represents the true data distribution. z This represents the distribution that the noise follows.

[0030] The NS-DCGAN is designed using the following steps:

[0031] Step 3.1: The generator randomly generates j noise samples Z, which are then processed by a deep convolutional network to generate generator samples.

[0032] p g (z):{z (1) ,…,z (j)} (6)

[0033] Step 3.2: Randomly sample j real communication network fault samples from the real communication network fault dataset to form a real fault sample:

[0034] p data (s):{s (1) ,…,s (j)} (7)

[0035] Step 3.3: Fix the generator state and set the generator sample p. g (z) and real fault sample p data (s) Input discriminator, trained through convolution operations and activation functions, based on real fault samples p data (s) and generator sample p g The difference between (z) is used to obtain the discriminant's loss. Based on the optimized objective function, the discriminant's loss is designed as follows:

[0036]

[0037] Step 3.4: Update the parameters using the stochastic gradient algorithm as follows:

[0038]

[0039] Where η is the learning rate, θ d These are the discriminator weight parameters;

[0040] Step 3.5: After the discriminator is updated, fix the discriminator, randomly input noise samples into the generator, and generate sample p. g (z):{z (1) ,…,z (j) Based on the discriminator's judgment result, the generator loss is obtained. Based on the optimized objective function, the generator loss is designed as follows:

[0041]

[0042] Step 3.6: Train and update the generator. The parameter update function is as follows:

[0043]

[0044] Where η is the learning rate, θ g These are the generator weight parameters;

[0045] Step 3.7: Following steps 3.1-3.6, the discriminator and generator are trained iteratively and alternately until the network converges or reaches a preset termination condition;

[0046] Step 3.8: Save the model after training is complete.

[0047] Furthermore, step 4 specifically includes: using real sample data as cluster centers, performing clustering based on Euclidean distance, and adding external environmental feature data to the virtual fault data, wherein the external environmental feature data includes weather conditions and ground conditions.

[0048] Furthermore, step 7 specifically includes:

[0049] Input the internal device indicator part D1 into path1, perform a convolution operation, and initially extract features;

[0050] Next, the data enters the G-bneck layer for feature extraction. The G-bneck layer replaces the original convolution, and linear operations replace some convolution operations. The specific formula is as follows:

[0051] y ij =φ i,j (y′ i (12)

[0052] Where, y′ i φ represents the features extracted by the initial convolution. i,j This represents the linear operation function used to generate the feature map;

[0053] Next, a 1×1 convolution kernel is used for data dimensionality reduction to achieve cross-channel feature aggregation;

[0054] Input the external environment feature part D2 into Path2, and perform one-hot encoding on the text features of the external environment feature part D2;

[0055] Concatenate the encoded data in Path2 with the data after feature extraction and processing in Path1;

[0056] The connected data is then fed into the fully connected layer (FC) for subsequent joint decision-making.

[0057] The fully connected data is processed by the softmax function to normalize the output, thus obtaining the fault diagnosis classification output.

[0058] The output results are compared with the label data, and the classification error is calculated using the multi-class cross-entropy function, as shown in the formula:

[0059]

[0060] Where M represents the number of categories, N is the number of samples, and y ic The sign function is p, which takes the value 1 if the sample equals the true class, and 0 otherwise. ic The predicted probability that the observed sample belongs to category c;

[0061] Based on the classification error, the backpropagation process uses the Adam optimizer to optimize parameters;

[0062] Repeat the above process until the preset termination condition is reached, then stop training and save the model;

[0063] By inputting real-time wireless mobile communication network data into the saved model, fault diagnosis results can be obtained.

[0064] This invention is based on the complex and rapidly changing wireless mobile communication network environment. Based on the generative adversarial approach, it innovatively designs the NS-DCGAN generative adversarial network to generate virtual communication network fault samples and construct a virtual-real combined sample library to solve the problems of insufficient fault samples and uneven data distribution.

[0065] This invention is designed for complex and ever-changing network environments. Based on artificial intelligence and deep learning technologies, it can achieve rapid and efficient multi-dimensional diagnosis and analysis of faults.

[0066] This invention fully incorporates system-related internal equipment indicators and objective external environmental characteristics, processing and analyzing both separately to retain highly correlated features to the greatest extent possible, effectively improving the accuracy of fault classification and diagnosis. Furthermore, considering the scarcity of communication network resources, this invention designs a lightweight, dual-path, environment-adaptive fault diagnosis neural network. Previously generated virtual fault samples are merged with real samples, and the EAFD-GhostNet algorithm is used to achieve fault classification and diagnosis of wireless communication networks. Attached Figure Description

[0067] Figure 1 This is a flowchart of the present invention;

[0068] Figure 2 Here is a structural diagram of NS-DCGAN;

[0069] Figure 3 This is a diagram of the discriminator structure of NS-DCGAN;

[0070] Figure 4 The generator structure diagram of NS-DCGAN is shown below;

[0071] Figure 5 This is a structural diagram of EAFD-GhostNet. Detailed Implementation

[0072] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] The process of this invention is as follows Figure 1 As shown, it includes the following steps:

[0074] (1) Collect a small amount of real network fault data from a heterogeneous wireless mobile communication network environment as input data, including key performance indicator (KPI) data of internal equipment and external environmental characteristic data.

[0075] (2) Data processing is carried out, mainly including data cleaning and filling, forming a real dataset, and separating and extracting the key performance indicators (KPIs) of the equipment.

[0076] In communication networks, key performance indicators (KPIs) of internal equipment are data collected during equipment operation, related to the system's program status, and are characterized by large volume and high dimensionality. External environmental data, on the other hand, represents the objective reality of the external environment over a relatively stable period, unrelated to the system itself. However, the external environment is highly correlated with network failures; for example, mountainous terrain with numerous obstructions can cause link instability, and thunderstorms can lead to short circuits. Therefore, it is desirable to retain as much external environmental data as possible. Given these different characteristics, internal equipment KPIs and external environmental data are considered separately.

[0077] (3) Design NS-DCGAN (NSelu Deep Convolutional Generative Adversarial Network) to input the key performance indicators (KPIs) of internal devices into NS-DCGAN. Based on the adversarial approach of generator and discriminator, learn the existing data distribution to generate a large number of virtual mobile communication network fault data under different network conditions, thus solving the problems of insufficient fault samples and unbalanced data distribution.

[0078] (4) The generated virtual fault data does not include external environment data. Based on Euclidean distance, a clustering algorithm is used to supplement the virtual data with external environment feature data to obtain a virtual dataset.

[0079] (5) Merge the virtual dataset with the real dataset to construct a complete wireless communication network fault sample dataset D that combines virtual and real data.

[0080] (6) Divide the dataset D into internal equipment index part D1 and external environment feature part D2.

[0081] (7) A lightweight network design is adopted. For complex and rapidly changing wireless mobile communication networks with scarce resources, an innovative dual-path environmental adaptive fault diagnosis neural network (EAFD-GhostNet) is designed. After processing the D1 and D2 data, they are input into Path1 and Path2 of EAFD-GhostNet respectively. After feature extraction, the internal equipment indicators and the features of the external environment are connected together for comprehensive analysis and joint fault classification diagnosis of wireless communication networks, and the diagnosis results are output.

[0082] The data used in this invention includes internal system equipment performance data and objective external environmental data to determine seven different fault types. Equipment performance indicators (KPIs) include collected key equipment performance data such as Reference Signal Receiving Power (RSRP), Reference Signal Receiving Quality (RSRQ), Uplink Packet Drop-UpLink (PD_UL), and Downlink Packet Drop-DownLink (PD_DL). External environmental data includes weather conditions and terrain conditions, as shown in Table 1.

[0083] Table 1. Fault Diagnosis Related Parameters

[0084]

[0085]

[0086] The intelligent communication network fault diagnosis method based on generative adversarial networks proposed in this invention can be roughly divided into two parts: virtual communication network fault sample generation based on NS-DCGAN generative adversarial network and communication network fault classification and diagnosis based on dual path EAFD-GhostNet. Specifically, it includes four stages: data sample processing, virtual fault sample generation, virtual data feature supplementation, and communication network fault classification and diagnosis.

[0087] 1. Data sample processing.

[0088] In a wireless network environment, internal device performance data is collected. These real, small-sample fault data form the input data vector, which consists of all relevant KPI device performance data of the object under study. Depending on the time granularity required for the communication network fault diagnosis process, KPI data can be collected using different time aggregation levels (such as seconds, minutes, hours, etc.).

[0089] Define S = [KPI1, KPI2, KPI3, ..., KPI m Let S be the input vector for different network states, and let S be a vector containing m KPIs. KPI values ​​not in the interval [0, 1] should be normalized to ensure that all variables are within the expected interval and that all indicators are on the same order of magnitude.

[0090] Define C = [FC1, FC2, FC3, ..., FC n] represents n different states of the network, such as the network working normally or experiencing a certain type of fault, such as link failure or interference.

[0091] During a certain period of time T, the network experienced a fault state FC. i State samples during this period It is expressed as follows:

[0092]

[0093] For the selected KPI i The normalization operation is as follows:

[0094]

[0095] Among them, KPI i Refers to the i-th normalized KPI.

[0096] 2. When there are insufficient fault samples in wireless mobile communication networks, virtual fault samples are generated based on the NS-DCGAN network.

[0097] The NS-DCGAN generative adversarial network mainly consists of two modules: a generator G and a discriminator D. The NS-DCGAN network structure is as follows: Figure 2 As shown.

[0098] Deep convolutional networks often use the ReLU function. When the input is less than 0 during forward propagation, neurons are inactive and "kill" gradients during backward propagation. Experiments have shown that as the number of iterations increases, the training of generative adversarial networks may oscillate, leading to non-convergence or even overfitting, resulting in poor sample generation. Therefore, this invention designs a novel activation function, Nselu. While addressing the above problems, it automatically normalizes samples to a mean of 0 and a unit variance, achieving self-normalization. Furthermore, it sets an upper limit threshold for the input x; when the input exceeds a certain value, feature learning stops, effectively reducing the possibility of network oscillations in later stages.

[0099]

[0100] To enhance the accuracy and robustness of the discriminator and enable the generative adversarial network to reach Nash equilibrium as quickly as possible, a regularization optimization is performed on the discriminator's objective function based on parameter gradients, and a special penalty term is added to the discrimination result. Therefore, the optimized objective function of the NS-DCGAN network is:

[0101] Discriminator:

[0102] Generator:

[0103] Among them, P dataP represents the true data distribution. z This represents the distribution that the noise follows.

[0104] 1) The generator randomly generates j noise samples Z, such as Gaussian noise, which are input into the generator and then processed by a deep convolutional network to generate fault samples:

[0105] p g (z):{z (1) ,…,z (j)} (6)

[0106] 2) Randomly sample j real communication network fault samples from the real communication network fault dataset.

[0107] p data (s):{s (1) ,…,s (j)} (7)

[0108] 3) Fix the generator state and set the generator sample p g (z) and real fault sample p data (s) Input the discriminator, train it through convolution operations and activation functions, and obtain the discriminator's loss based on the difference between real fault samples and generator samples. Based on the improved objective function, the discriminator loss is designed as follows:

[0109]

[0110] 4) Then, the parameters are updated using the stochastic gradient algorithm as follows, where η is the learning rate and θ is the learning rate. d Discriminator weight parameters:

[0111]

[0112] Discriminator structure as follows Figure 3 As shown.

[0113] 5) After the discriminator is updated, fix the discriminator, randomly input noise samples into the generator, and generate sample p. g (z):{z (1) ,…,z (j) Based on the discriminator's judgment result, the generator loss is obtained. Based on the improved objective function, the generator loss is designed as follows:

[0114]

[0115] 6) Train and update the generator using the following parameter update function, where η is the learning rate and θ is the parameter update function. g For generator weight parameters:

[0116]

[0117] Generator structure as follows Figure 4 As shown.

[0118] 7) Following the above process, the discriminator and generator are trained alternately and iteratively until the network converges or reaches the pre-set termination condition.

[0119] Based on the above design, and using the proposed activation function and objective function, this invention proposes the NS-DCGAN network, the specific architecture of which is shown in Table 2.

[0120] Table 2 NS-DCGAN Network Architecture

[0121]

[0122] 8) After training, save the model and run the generator to generate virtual communication network fault samples.

[0123] 3. Supplement and improve the characteristics of the external environment based on clustering.

[0124] Real-world communication network fault sample data includes objective external information such as weather and ground conditions, while the generated virtual data only includes equipment specifications and does not include the aforementioned external environmental features. Therefore, based on real samples, clustering is performed using Euclidean distance with real sample data as cluster centers, and external environmental features are added to the annotations of virtual fault samples. In this method, external environmental feature data includes weather and ground conditions.

[0125] 4. Fault classification and diagnosis of EAFD-GhostNet communication network based on dual paths.

[0126] Real-world and virtual fault samples from communication networks are merged into a new dataset D. This dataset includes internal device performance data and corresponding external environmental characteristics over a specific time period. Internal device performance indicators generally exhibit internal correlations, showing strong interrelationships and are mostly numerical. External environmental characteristics, on the other hand, are objective realities that tend to stabilize over a period, with lower correlations and non-numerical forms, differing from the characteristics of internal device performance indicators. Furthermore, based on historical experience, external environmental factors significantly impact system faults in wireless mobile communication networks, exhibiting high correlation; therefore, it is desirable to retain as many external environmental factors as possible. Thus, a dual-path processing approach is needed for both internal device performance indicators and external environmental characteristics, taking both into account.

[0127] This invention targets a communication network fault dataset D that combines internal equipment indicators with external environmental characteristics. It performs external environment adaptive optimization to increase the number of effective features, thereby further improving the accuracy of mobile communication network fault diagnosis. Based on a lightweight neural network, an improved and optimized dual-path EAFD-GhostNet (Environmental Adaptive Fault Diagnosis GhostNet) is proposed, with the structure as follows: Figure 5 As shown.

[0128] 1) Separate the internal equipment indicator data and external environmental characteristics into D1 and D2 in a one-to-one correspondence.

[0129] 2) Input dataset D1 into path1, perform convolution operation, and initially extract features;

[0130] 3) The data then enters the G-bneck layer for feature extraction. Using G-bneck instead of the original convolution, and employing simple linear operations to replace some convolution operations, effectively reduces the number of network parameters and improves network computational efficiency.

[0131] y ij =φ i,j (y′ i (12)

[0132] y′ i φ represents the features extracted by the initial convolution. i,j This represents a linear operation function used to generate more feature maps.

[0133] 4) Next, a 1×1 convolution kernel is used to reduce the dimensionality of the data, thereby achieving cross-channel feature aggregation;

[0134] 5) Input the external environment feature dataset into Path2;

[0135] 6) Perform one-hot encoding on the text features of the external environment feature dataset D2, and input it into the next step to preserve the external environment features to the greatest extent.

[0136] 7) The encoded data in Path2 is concatenated with the feature-extracted data in Path1.

[0137] 8) Input the concatenated data into the fully connected layer (FC) for subsequent joint decision-making;

[0138] 9) The fully connected data is processed by the softmax function to normalize the output, thereby obtaining the fault diagnosis classification output;

[0139] 10) Compare the output results with the label data, and calculate the classification error using the multi-class cross-entropy function:

[0140]

[0141] Where M represents the number of categories, N is the number of samples, and y ic The sign function (0 or 1) is used; it takes the value 1 if the sample equals the true class, and 0 otherwise. ic The predicted probability that the observed sample belongs to category c.

[0142] 11) Based on the classification error, the Adam optimizer is used for parameter optimization during backpropagation.

[0143] 12) Repeat the above training process until the preset termination condition is reached, then stop training and save the model.

[0144] 13) By inputting real-time wireless mobile communication network data into the saved model, the fault diagnosis results can be obtained.

[0145] For complex and rapidly changing wireless mobile communication network environments, this invention proposes a fault diagnosis method for wireless communication networks based on generative adversarial networks, with the following advantages:

[0146] This invention addresses the issue of insufficient fault samples in wireless mobile communication networks by generating fault samples based on generative adversarial thinking. It constructs a virtual-real combined generative adversarial sample library, which supplements the number of samples while solving the problem of uneven distribution of fault samples.

[0147] This invention addresses the network convergence problem during sample generation by proposing novel network objective and activation functions and adjusting the generative adversarial network architecture to facilitate rapid attainment of Nash equilibrium, thereby generating more accurate virtual communication network fault samples.

[0148] This invention fully considers the characteristics of wireless mobile communication network data, integrates internal system data and external environmental features, further enriches the data dimensions, and retains external features that are highly correlated with faults to the greatest extent, thereby further improving the accuracy of fault diagnosis.

[0149] This invention takes into account the insufficient resources of wireless mobile communication networks, adopts a lightweight network design with fewer network parameters and less computational resource consumption, and can complete the classification and diagnosis of mobile communication network faults more efficiently and quickly.

[0150] This invention proposes a novel nonlinear activation function, Nselu, which helps reduce oscillations during deep generative adversarial network (GAN) training and accelerates network convergence. It also proposes an innovative method for generating fault samples in wireless mobile communication networks based on NS-DCGAN, constructing a virtual-real combined sample library. By employing the proposed Nselu activation function, the objective and update functions of the GAN are improved, and an NS-DCGAN GAN is constructed, effectively improving the efficiency of fault sample generation in communication networks. This invention focuses on complex and rapidly changing wireless mobile communication networks, addressing the challenges of limited bandwidth and computing resources, and designs a lightweight fault classification and diagnosis network. Furthermore, considering the fault characteristics of wireless mobile communication networks, this invention comprehensively analyzes internal system equipment performance data and external environmental features, adopting a dual-path architecture to innovatively improve the lightweight network through dual-path design, proposing a lightweight dual-path environmentally adaptive fault diagnosis neural network, EAFD-GhostNet.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fault diagnosis of wireless communication networks based on generative adversarial networks, characterized in that, Includes the following steps: Step 1: Collect real-world network fault data from a heterogeneous wireless communication network environment, including key performance indicators of internal devices. KPI Data and external environmental characteristics data; Step 2: Perform data processing, including data cleaning and filling, to form a realistic dataset, and separate and extract the key performance indicators of the equipment. KPI data; Step 3, Design NS - DCGAN The key performance indicators of the equipment KPI Data input NS - DCGAN Based on an adversarial approach involving generators and discriminators, the system learns the existing data distribution and generates virtual fault data under different network conditions. NS-DCGAN Including generators G and discriminator D Activation function Nselu ( x )for: （3） The optimized objective function is: Discriminator: (4) Generator: (5) in, P data Represents the true data distribution. P z This represents the distribution that the noise follows. The design is carried out through the following steps. NS - DCGAN : Step 3.1: Generator generates random data. j noise sample Z Generator samples are generated using a deep convolutional network: (6) Step 3.2: Randomly sample from real communication network fault datasets. j A real communication network fault sample is formed, which constitutes a real fault sample: (7) Step 3.3: Fix the generator state and set the generator sample. and real fault samples The input discriminator is trained using convolution operations and activation functions, based on real fault samples. and generator samples The difference between them is obtained by the discriminator loss Based on the optimized objective function, the discriminator loss The design is as follows: (8) Step 3.4: Update the parameters using the stochastic gradient algorithm as follows: (9) in, For learning rate, These are the discriminator weight parameters; Step 3.5: After the discriminator is updated, fix the discriminator, randomly input noise samples into the generator, and generate samples. Based on the discriminator's judgment result, the generator is obtained. loss Based on the optimized objective function, the generator loss The design is as follows: (10) Step 3.6: Train and update the generator. The parameter update function is as follows: (11) in, For learning rate, These are the generator weight parameters; Step 3.7: Following steps 3.1-3.6, the discriminator and generator are trained iteratively and alternately until the network converges or reaches a preset termination condition; Step 3.8: Save the model after training is complete; Step 4: Use a clustering algorithm to supplement the virtual fault data with external environmental feature data to obtain a virtual dataset; Step 5: Merge the virtual dataset with the real dataset to construct a wireless communication network fault sample dataset. D ; Step 6: Transfer the dataset D Divided into internal equipment indicators D 1 and External Environmental Characteristics D 2; Step 7: Design a dual-path environmentally adaptive fault diagnosis neural network EAFD - GhostNet Internal equipment specifications D After data processing, the data is input into the environmental adaptive fault diagnosis neural network. EAFD - GhostNet of Path 1. Perform feature extraction, extracting the external environment features. D After data processing, the data is input into the environmental adaptive fault diagnosis neural network. EAFD - GhostNet of Path 2. Encode; will pass through Path 2. Encoded data and the processed data Path 1. Connect the data after feature extraction, perform wireless communication network fault classification and diagnosis, and output the diagnosis results.

2. The method for fault diagnosis of wireless communication networks based on generative adversarial networks according to claim 1, characterized in that, Key performance indicators of the extraction equipment described in step 2 KPI The data includes: Collect using time aggregation levels of different lengths KPI data; definition , representing the input vector for different network states. S contain m indivual KPI ; definition , representing the network n A different state; In time T Internally, the network is experiencing a failure. FC i State samples Represented as: （1） For the selected KPI i The normalization operation is as follows: （2） in, KPI i For the normalized first i kind KPI .

3. The method for fault diagnosis of wireless communication networks based on generative adversarial networks according to claim 2, characterized in that, Step 4 specifically includes: using real sample data as cluster centers, performing clustering based on Euclidean distance, and adding external environmental feature data to the virtual fault data, including weather conditions and ground conditions.

4. The method for fault diagnosis of wireless communication networks based on generative adversarial networks according to claim 3, characterized in that, Step 7 specifically includes: Internal equipment specifications D 1 input path 1. Perform convolution operations to initially extract features; Then, the data enters G - bneck Feature extraction is performed in layers, using G - bneck Instead of the original convolution, a linear operation is used to replace part of the convolution operation. The specific formula is as follows: (12) in, This represents the features extracted from the initial convolution. This represents the linear operation function used to generate the feature map; Next, a 1×1 convolution kernel is used for data dimensionality reduction to achieve cross-channel feature aggregation; External environmental characteristics D 2 inputs Path 2. External environmental characteristics D The text features of 2 are encoded using one-hot encoding; Will Path The encoded data in 2 and Path 1. Connect the data after feature extraction and processing; The concatenated data is then input into the fully connected layer. FC This is for use in subsequent joint decision-making; After full connection, the data passes through softmax The function normalizes the output results to obtain the fault diagnosis classification output; The output results are compared with the label data, and the classification error is calculated using the multi-class cross-entropy function, as shown in the formula: loss ： (13) in, M Represents the number of categories, N For the number of samples, This represents the sign function, which takes the value 1 if the sample equals the true class, and 0 otherwise. The observed sample belongs to the category c The predicted probability; Based on classification error, the backpropagation process uses Adam The optimizer performs parameter optimization; Repeat the above process until the preset termination condition is reached, then stop training and save the model; By inputting real-time wireless mobile communication network data into the saved model, fault diagnosis results can be obtained.

Images