A wireless network interruption detection method based on variational autoencoder

Abstract

The present invention discloses a wireless network interruption detection method based on a variational autoencoder, comprising the following steps: a first step: collecting key performance indicators (KPIs) of a wireless communication system and forming a data set; a second step: performing data preprocessing on the data set X to obtain a data set; a third step: using the feature mean distribution learned by VAE in the second step as a parameter for subsequent clustering to form a new data set V; a fourth step: specifying the number m of clusters and the maximum number N of iterations, and clustering the data set V using the K-means algorithm; and a fifth step: after clustering, defining labels according to the number of data contained in each cluster, and assigning the cluster with a small amount of data as the interruption type to obtain an output set. This method combines the performance advantages of VAE and the K-means algorithm (K-means), solves the problem of low accuracy of interruption detection methods based on unsupervised machine learning, and also performs well in the case of data imbalance.

Description

Technical Field

[0001] The present invention belongs to the technical field of network autonomous driving in wireless communications, and specifically relates to an interruption detection method suitable for a network fault intelligent detection and recovery system. Background Art

[0002] With the rapid development of mobile communication networks, network structures are becoming increasingly complex, and network parameters are increasing dramatically. Manually detecting and recovering from cell outages is extremely costly. Network autonomous driving, encompassing intelligent network deployment, intelligent network optimization, and intelligent network fault detection and recovery, is a key technology for addressing these issues. As a key component of network autonomous driving, cell outage detection plays a crucial role in intelligent network fault detection and recovery. Machine learning-based cell outage detection methods are widely used. Depending on the machine learning method employed, outage detection can be categorized into two types: supervised learning-based and unsupervised learning-based. Supervised learning-based outage detection methods generally consist of two phases: training and prediction. During the training phase, each training sample consists of input data and a label for the expected output. Through training, the supervised learning model learns the mapping between input data and output data, thereby achieving predictions. However, in wireless networks, only a small number of datasets have labels. For most areas without pre-labeled data, manual labeling is still required, resulting in the lack of universality of supervised learning-based outage detection. Furthermore, outages are rare events, and the amount of data related to outages is far less than normal data. In this case, the results obtained by directly using machine learning for classification will be biased towards the majority class, and the interruption detection performance will be significantly reduced due to the data imbalance problem.

[0003] To address these issues, existing research has introduced methods such as K-means, Diffusion Map, and Local Outlier Factor (LOF) to directly cluster or reduce the dimensionality of unlabeled interruption data, dividing the data into two clusters: normal data and interruption data. However, when the data imbalance is large, the interruption detection performance still needs to be improved.

[0004] To address the above issues, the present invention proposes an interruption detection method based on a variational auto-encoder (VAE). This method combines the performance advantages of VAE and K-means algorithm, solves the problem of low accuracy of interruption detection methods based on unsupervised machine learning, and also performs well in the case of data imbalance. Summary of the Invention

[0005] Technical Problem: This invention proposes a wireless network outage detection method based on a variational autoencoder. This method uses a VAE to extract features from unlabeled data and then uses the K-means algorithm to cluster the mean values ​​of the features extracted by the VAE for each data point. Compared to traditional unsupervised detection methods and data clustering methods, this method significantly improves outage detection performance.

[0006] Technical solution: The wireless network interruption detection method based on variational autoencoder of the present invention comprises the following steps:

[0007] Step 1: Collect the key performance indicators (KPIs) of wireless communication systems and form a data set Where M is the number of elements in the data set X, (x i ,y i ) is the i-th element in X, i = 1, 2, ..., M, x i ∈R K Indicates the K-dimensional key performance indicator (KPI) information reported by a user at a certain moment, y i is x i The label indicates the status of the base station serving the user and can be used to compare the results of subsequent unlabeled training.

[0008] Step 2: Preprocess the data set X to obtain the data set The dataset X′ is an unlabeled dataset used for training. The variational autoencoder VAE is used to learn the distribution characteristics of the dataset X′, so that the network can extract the characteristics of the dataset X′ and generate data with similar characteristics to the dataset X′. The distribution characteristics of the dataset X′ are recorded. The VAE consists of an encoder E and a decoder D.

[0009] Step 3: Use the feature mean distribution learned by VAE in the second step as the parameter for subsequent clustering to form a new dataset V;

[0010] Step 4: Specify the number of clusters m and the maximum number of iterations N, and use the K-means algorithm to cluster the data set V;

[0011] Step 5: After clustering is completed, define labels according to the number of data contained in each cluster, and the ones with less data are interrupt types to obtain the output set y i ′∈R 1 , x i is an element in the dataset X, and y′ is x i Corresponding to the predicted label, if y′∈{1, 2, 3}, then x i For interrupt data.

[0012] in,

[0013] The first step is to collect the key performance indicators (KPIs) of the wireless communication system and form a data set X, specifically:

[0014] The Minimization of Drive-Tests (MDT) method is used to obtain KPI information reported by users in the wireless communication system within t seconds. The KPIs include Block Error Rate (BLER), Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), Received Signal Strength Indicator (RSSI), Throughput, and Signal to Interference and Noise Ratio (SINR). i is x i The label indicates the status of the base station serving the user, which can be recorded and used for comparison of the results of unlabeled training. The base station status is divided into two categories: normal and interrupted. The interrupted base station is further divided into light interruption, moderate interruption and heavy interruption according to the degree of interruption. Therefore, y i Is a one-dimensional variable with a value range of y i ∈{0,1,2,3};y i =0 means the base station is in normal state and the received KPI value is within the normal range; i =1 means the base station is in a light interruption state. At this time, the received KPI value is slightly outside the normal range. i =2 means the base station is in a moderate interruption state, the base station performance is seriously degraded, and the received KPI value is far beyond the normal range; i =3 indicates that the base station is in a severe outage state and the received KPI value is abnormal.

[0015] The second step includes the following process:

[0016] Step 2.1, data preprocessing:

[0017] For the elements (x i ,y i ) is normalized to the maximum value. The range of KPI data after normalization is [-1,1]. Only the normalized elements (x i ′,y i ) i ' forms a new data set X', x i ′∈RK , the dataset X′ is the unlabeled dataset used for training, and the label y of the dataset X i It can be used to compare the results after training on the dataset X′.

[0018]

[0019] x ik ′ is the i-th element x in the normalized dataset X′ i The value of the k-th dimension KPI of ′;

[0020] Step 2.2, define the objective function of VAE and set the training parameters:

[0021] The objective function of VAE is as follows:

[0022]

[0023] Among them, KL(p(x,z)||q(x,z)) is the KL divergence (Kullback-Leibler Divergence) of two probability distributions p(x,z) and q(x,z), x is the real sample of the input, assuming that the real sample x is generated by the variable z, z is a hidden variable, an unobservable random variable, assuming that q(z) is the probability distribution of the hidden variable z, which obeys the standard normal distribution. p(z|x),q(x|z) are conditional normal distributions, that is, for a real sample x i ′, assuming there is a unique i ′’s posterior distribution p(z i |x i ′), and further assume that this distribution is an independent multivariate normal distribution.

[0024] Set the parameters required for subsequent model training: the maximum number of iterations n for training the encoder and decoder at a time ed , its value is determined by experimental experience; the batch size c for each training is determined by experimental experience; set the number of batches c for the current training now = 0, whose value increases with training and reaches a maximum value of c; the maximum number of model iterations E, whose value is determined by experimental experience; the number of model iterations e = 0. The learning rate α of the evolved symbolic momentum (Lion) optimizer, whose value is determined by experimental experience; the hidden layer activation function of both encoder E and decoder D uses the LeakyReLU function.

[0025] A neural network is a computational model composed of multiple interconnected neurons, each of which receives input signals and generates output signals. In a neural network, each neuron has a weight matrix W and a bias vector B, which determine how the neuron processes the input signal. The weight matrix W is used to measure the impact of the input signal on the neuron's output, and the bias vector B is used to adjust the neuron's activation threshold. The VAE encoder includes a five-layer neural network, and the decoder also includes a five-layer neural network. This method uses Gaussian distribution to initialize the weight matrix W of the encoder E and decoder D in the VAE model. E ,W D , and the bias vector B E 、B D The initial value of is set to 0;

[0026] Step 2.3, train the encoder E and decoder D:

[0027] Use the gradient descent method to minimize Equation (2) and update the weight matrix W of the encoder E and decoder D E ,W D and the bias vector B E 、B D Until the number of iterations is n ed , set the batch size c of the current training now =c now +1;

[0028] Step 2.4, update the latent space and its characteristics formed in the process of VAE extracting data features;

[0029] Any K-dimensional distribution can be generated by selecting a set of K variables that obey the normal distribution and mapping this distribution to a sufficiently complex function. Therefore, it can be assumed that the hidden variable z in the training process obeys the Gaussian distribution, and then a specific function is constructed to map our hidden space that obeys the Gaussian distribution to a more complex distribution. This is the training process of VAE. The encoder of VAE includes a five-layer neural network, and the decoder also includes a five-layer neural network. When the data set X′ is input into the encoder, the encoder finally outputs the mean μ and variance σ, which represent the data distribution characteristics of the original sample. For each element x in the data set X′ i ′, the encoder will generate a corresponding hidden variable z that obeys the Gaussian distribution i ′ to simulate element x i ′ corresponds to the distribution of hidden variables z i ′ has its own mean μ i and variance σ i , according to the encoder parameter update in step 2.3, reconstruct all x i ′ corresponds to the hidden variable zi ′.

[0030] Step 2.5, repeat steps 2.3 and 2.4 until c now =c.

[0031] Step 2.6, use the evolutionary symbolic momentum (Lion) optimizer to update the gradient of the gradient descent method used in step 2.3, and let c now =0, e=e+1.

[0032] Step 2.7, repeat steps 2.3, 2.4, 2.5, and 2.6 until the number of iterations of the model e=E.

[0033] Step 2.8, training generation, record the weight matrix W of encoder E and decoder D E ,W D and the bias vector B E 、B D , record all x i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i .

[0034] The third step is specifically as follows: according to each x in the recorded data set X' i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i Building a new dataset x i ∈R K ,μ i ∈R K ,σ i ∈R K .

[0035] The fourth step is specifically as follows:

[0036] In step 4.1, the number of clusters m=4 and the maximum number of iterations N are specified, and their values ​​are determined by experimental experience.

[0037] Step 4.2, select an initial cluster center for each cluster of the dataset V, the initial center of the first class (x j ′,μ j ,σ j ) is randomly selected, and then the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the initial cluster center is calculated i (v) And with probability

[0038]

[0039] Select the center point of the new class and get the second cluster point (x l ′,μ l ,σ l ).

[0040] Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first two cluster centers i (v) And with probability

[0041]

[0042] Select the center point of the new class and get the third cluster point (x b ′,μ b ,σ b ).

[0043] Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first three cluster centers i (v) And with probability

[0044]

[0045] Select the center point of the new class and get the fourth cluster point (x g ′,μ g ,σ g ).

[0046] Finally, the set W = {(x j ′,μ j ,σ j ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}.

[0047] Step 4.3: Assign the remaining elements of the dataset V to the nearest neighbor cluster according to the minimum distance principle. Specifically:

[0048] Calculate the minimum Euclidean distance from the latent variable mean of the remaining elements of the data set V to the latent variable mean of the four cluster points, And classify each element into the cluster point that can obtain the minimum Euclidean distance.

[0049] Step 4.4, all elements are classified into the set W = {(x j ′,μj ,σ j ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}

[0050] These four cluster points are removed. The cluster centers are updated using the sample mean of each cluster. The elements in the set W are replaced.

[0051] Specifically:

[0052] For the points classified into the first cluster (x j ′,μ j ,σ j ) consists of the set V of all elements j , take V j The mean of all elements in the latent variable U j It is V j The number of elements in , let The cluster center is updated. Repeat the same operation for other cluster points.

[0053] Step 4.5: Repeat steps 4.3 and 4.4 until the cluster center no longer changes or the maximum number of iterations N is reached.

[0054] Step 4.6, output the final cluster center set and 4 cluster divisions.

[0055] The fifth step is to define labels according to the number of data in each cluster after clustering is completed. The data with less amount is interrupt type, and the label is y′∈{1, 2, 3}, and the data with more amount is normal type, and y′∈{0}. The final output set is y i ′∈R 1 , x i is an element in the dataset X, and y′ is x i Corresponding to the predicted label, if y′∈{1, 2, 3}, then x i For interrupt data.

[0056] Beneficial Effects: The interruption detection method based on a variational autoencoder proposed in this invention offers the following advantages: Compared to previous unsupervised clustering methods, this method leverages the feature extraction capabilities of VAEs to record the latent vector distribution of each data point, and then uses the K-means algorithm for secondary clustering, thereby improving interruption detection accuracy. Furthermore, the method has the advantage of specifying the number of clusters, making it particularly effective when dealing with multiple interruption types and data imbalance. DETAILED DESCRIPTION

[0057] An example of the interruption detection method based on the variational autoencoder of the present invention is:

[0058] Step 1: Collect key performance indicators (KPIs) of wireless communication systems and form a data set x i ∈R 5 ; where 1000000 is the number of elements in the dataset X, (x i ,y i ) is the i-th element in X, i = 1, 2, ..., M, x i ∈R 5 Indicates the 5-dimensional key performance indicator KPI information reported by a user at a certain time, y i is x i The label indicates the status of the base station serving the user, which can be used to compare the results of subsequent unlabeled training. Specifically:

[0059] The Minimization of Drive-Tests (MDT) method is used to obtain KPI information reported by users in the wireless communication system within 10 seconds. During this period, interruption data is uploaded for 0.1 seconds to form a data set. x i ∈R 5 ; where (x i ,y i ) is the i-th element in X, i = 1, 2, ..., M; x i ∈R 5 Indicates the five-dimensional key performance indicator (KPI) information reported by a user at a certain moment. The KPIs include block error rate (BLER), reference signal received power (RSRP), reference signal received quality (RSRQ), received signal strength indicator (RSSI), throughput, and signal to interference and noise ratio (SINR). iis x i The label indicates the status of the base station serving the user. The base station status is divided into normal and interrupted. The interrupted base station is further divided into light interruption, moderate interruption and heavy interruption according to the degree of interruption. Therefore, y i Is a one-dimensional variable with a value range of y i ∈{0,1,2,3};y i =0 means the base station is in normal state and the received KPI value is within the normal range; i =1 means the base station is in a light interruption state. At this time, the received KPI value is slightly outside the normal range. i =2 means the base station is in a moderate interruption state, the base station performance is seriously degraded, and the received KPI value is far beyond the normal range; i =3 indicates that the base station is in a severe outage state and the received KPI value is abnormal.

[0060] The second step includes the following process:

[0061] Step 2.1, data preprocessing: the elements (x i ,y i ) is normalized to the maximum value. The range of KPI data after normalization is [-1,1]. Only the normalized elements (x i ′,y i ) i ' forms a new data set X', x i ∈R 5 , dataset X′ is the unlabeled dataset used for training, and dataset X can be used for result comparison.

[0062]

[0063] x ik ′ is the i-th element x in the normalized dataset X′ i The value of the k-th dimension KPI of ′;

[0064] Step 2.2, define the objective function of VAE and set the training parameters.

[0065] The objective function of VAE is as follows:

[0066]

[0067] Here, we assume that q(z), which is also the probability distribution of the latent variable z, is a standard normal distribution. p(z|x),q(x|z) is a conditional normal distribution, that is, for a real sample x i ′, assuming there is a unique i′’s posterior distribution p(z i |x i ′), and further assume that this distribution is an independent multivariate normal distribution.

[0068] Set the parameters required for subsequent model training: the maximum number of iterations n for training the encoder and decoder at a time ed = 10, whose value is determined by experimental experience; the batch size c for each training run is 640, whose value is determined by experimental experience; the maximum number of model iterations E is 1000; the number of model iterations e is 0. The learning rate α of the evolved symbolic momentum (Lion) optimizer is 0.0001, whose value is determined by experimental experience; the hidden layer activation function of both encoder E and decoder D uses the LeakyReLU function.

[0069] A neural network is a computational model composed of multiple interconnected neurons, each of which receives input signals and generates output signals. In a neural network, each neuron has a weight matrix W and a bias vector B, which determine how the neuron processes the input signal. The weight matrix W is used to measure the impact of the input signal on the neuron's output, and the bias vector B is used to adjust the neuron's activation threshold. The VAE encoder and decoder each include a five-layer neural network. This method uses a Gaussian distribution to initialize the weight matrix W of the encoder E and decoder D in the VAE model. E ,W D , and the bias vector B E 、B D The initial value is set to 0.

[0070] Step 2.3, train the encoder E and decoder D,

[0071] Use the gradient descent method to minimize Equation (2) and update the weight matrix W of the encoder E and decoder D E ,W D and the bias vector B E 、B D Until the number of iterations is n ed , set the current batch number c now =c now +1;

[0072] Step 2.4, update the latent space and its characteristics formed in the process of VAE extracting data features;

[0073] Any K-dimensional distribution can be generated by selecting a set of K variables that obey the normal distribution and mapping this distribution to a sufficiently complex function. Therefore, it can be assumed that the hidden variable z in the training process obeys the Gaussian distribution, and then a specific function is constructed to map our hidden space that obeys the Gaussian distribution to a more complex distribution. This is the training process of VAE. The decoder of the VAE encoder consists of five layers of neural networks. When the data set X' is input into the encoder, the encoder finally outputs the mean μ and variance σ, which represent the data distribution characteristics of the original sample. For each element x in the data set X' i ′, the encoder will generate a corresponding hidden variable z that obeys the Gaussian distribution i ′ to simulate element x i ′ corresponds to the distribution of hidden variables z i ′ has its own mean μ i and variance σ i , according to the encoder parameter update in step 2.3, reconstruct all x i ′ corresponds to the hidden variable z i ′.

[0074] Step 2.5, repeat steps 2.3 and 2.4 until c now =c

[0075] Step 2.6, use the evolutionary symbolic momentum (Lion) optimizer to update the gradient of the gradient descent method used in step 2.3, and let c now =0, e=e+1;

[0076] Step 2.7, repeat steps 2.3, 2.4, 2.5, and 2.6 until the number of iterations of the model e = E;

[0077] Step 2.8, training generation, record the weight matrix W of encoder E and decoder D E ,W D and the bias vector B E 、B D , record all x i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i .

[0078] The third step is specifically as follows: according to each x in the recorded data set X' i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i Building a new dataset x i ∈R 5 ,μi ∈R 5 ,σ i ∈R 5 .

[0079] The fourth step is specifically as follows:

[0080] In step 4.1, the number of clusters m=4 and the maximum number of iterations N are specified, and their values ​​are determined by experimental experience.

[0081] Step 4.2, select an initial cluster center for each cluster of the dataset V, the initial center of the first class (x j ′,μ j ,σ j ) is randomly selected, and then the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the initial cluster center is calculated i (v) And with probability

[0082]

[0083] Select the center point of the new class and get the second cluster point (x l ′,μ l ,σ l ).

[0084] Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first two cluster centers i (v) And with probability

[0085]

[0086] Select the center point of the new class and get the third cluster point (x b ′,μ b ,σ b ).

[0087] Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first three cluster centers i (v) And with probability

[0088]

[0089] Select the center point of the new class and get the fourth cluster point (x g ′,μ g ,σ g ).

[0090] Finally, the set W = {(x j ′,μ j ,σj ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}.

[0091] Step 4.3: Assign the remaining elements of the dataset V to the nearest neighbor cluster according to the minimum distance principle. Specifically:

[0092] Calculate the minimum Euclidean distance from the latent variable mean of the remaining elements of the data set V to the latent variable mean of the four cluster points, And classify each element into the cluster point that can obtain the minimum Euclidean distance.

[0093] Step 4.4, all elements are classified into the set W = {(x j ′,μ j ,σ j ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}. Update the cluster centers using the sample mean of each cluster. Replace the elements in set W.

[0094] Specifically:

[0095] For the points classified into the first cluster (x j ′,μ j ,σ j ) consists of the set V of all elements j , take V j The mean of all elements in the latent variable U j It is V j The number of elements in , let The cluster center is updated. Repeat the same operation for other cluster points.

[0096] Step 4.5: Repeat steps 4.3 and 4.4 until the cluster center no longer changes or the maximum number of iterations N is reached.

[0097] Step 4.6, output the final cluster center set and 4 cluster divisions.

[0098] The fifth step is to define labels according to the number of data in each cluster after clustering is completed. The data with less amount is interrupt type, and the label is y′∈{1, 2, 3}, and the data with more amount is normal type, and y′∈{0}. The final output set is y i ′∈R 1 , x i is an element in the dataset X, and y′ is x i Corresponding to the predicted label, if y′∈{1, 2, 3}, then x i For interrupt data.

[0099] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A wireless network interruption detection method based on variational autoencoder, characterized in that The steps include: Step 1: Collect key performance indicators (KPIs) of wireless communication systems and form a data set Where M is the number of elements in the data set X, (x i ,y i ) is the i-th element in X, i = 1, 2, ..., M, x i ∈R K Indicates the K-dimensional key performance indicator (KPI) information reported by a user at a certain moment, y i is x i The label indicates the status of the base station serving the user, which is used for comparison of the results of subsequent unlabeled training; Step 2: Preprocess the data set X to obtain the data set The dataset X′ is an unlabeled dataset used for training. The variational autoencoder (VAE) is used to learn the distribution characteristics of the dataset X′, so that the network can extract the characteristics of the dataset X′ and generate data with similar characteristics to the dataset X′. The distribution characteristics of the dataset X′ are recorded. The VAE consists of an encoder E and a decoder D. Step 3: Use the feature mean distribution learned by VAE in the second step as the parameter for subsequent clustering to form a new data set. i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i Building a new dataset Step 4: Specify the number of clusters m and the maximum number of iterations N, and use the K-means algorithm to cluster the data set V; Step 5: After clustering is completed, the labels are defined according to the number of data contained in each cluster. The ones with less data are interruption types, and the labels are y′∈{1, 2, 3}, and the ones with more data are normal types, and y′∈{0}. Finally, the output set is obtained. x i is an element in the dataset X, and y′ is x i Corresponding to the predicted label, R 1 Represents a one-dimensional real number.

2. The wireless network interruption detection method based on variational autoencoder according to claim 1 is characterized in that The first step is to collect the key performance indicators (KPIs) of the wireless communication system and form a data set X, specifically: The Minimization of Drive Tests (MDT) method is used to obtain KPI information reported by users in the wireless communication system within t seconds. The KPIs include block error rate (BLER), reference signal received power (RSRP), reference signal received quality (RSRQ), received signal strength indicator (RSSI), throughput, and signal-to-interference-plus-noise ratio (SINR).

3. The wireless network interruption detection method based on variational autoencoder according to claim 2 is characterized in that The status of the base station is divided into two categories: normal and interrupted. The interrupted base station is further divided into light interruption, medium interruption and heavy interruption according to the degree of interruption. i Is a one-dimensional variable with a value range of y i ∈{0,1,2,3};y i =0 means the base station is in normal state and the received KPI value is within the normal range; i =1 means the base station is in a light interruption state. At this time, the received KPI value is slightly outside the normal range. i =2 means the base station is in a moderate interruption state, the base station performance is seriously degraded, and the received KPI value is far beyond the normal range; i =3 indicates that the base station is in a severe outage state and the received KPI value is abnormal.

4. The wireless network interruption detection method based on variational autoencoder according to claim 1 is characterized in that The second step includes the following process: Step 2.1, data preprocessing: For the elements (x i ,y i ) to normalize the maximum value. The range of KPI data after normalization is [-1,1], only retain the normalized elements (x i ′,y i ) i 'Form a new data set X' ， Dataset X′ is the unlabeled dataset used for training, and the label y of dataset X is i Can be used as a comparison of the results after training on the dataset X′; x ik ′ is the i-th element x in the normalized dataset X′ i The value of the k-th dimension KPI of ′; Step 2.2, define the objective function of VAE and set the training parameters: The objective function of VAE is as follows: Among them, KL(p(x,z)||q(x,z)) is the KL divergence of two probability distributions p(x,z) and q(x,z), x is the real sample of the input, and the real sample x is generated by the variable z, z is a hidden variable, and is an unobservable random variable; let q(z) be the probability distribution of the hidden variable z, which obeys the standard normal distribution; p(z|x),q(x|z) are conditional normal distributions, that is, for a real sample x i ′, suppose there is a unique i ′’s posterior distribution p(z i |x i ′), and further assume that this distribution is an independent multivariate normal distribution; Set the parameters required for subsequent model training: the maximum number of iterations n for training the encoder and decoder at a time ed , its value is determined by experimental experience; the batch size c for each training is determined by experimental experience; set the number of batches c for the current training now =0, its value increases with training and its maximum value is c; the maximum number of model iterations E, its value is determined by experimental experience; the number of model iterations e=0, the learning rate α of the evolutionary symbolic momentum Lion optimizer, its value is determined by experimental experience; the hidden layer activation function of encoder E and decoder D both use the LeakyReLU function; A neural network is a computational model composed of multiple interconnected neurons. Each neuron receives input signals and generates output signals. In a neural network, each neuron has a weight matrix W and a bias vector B, which determine how the neuron processes the input signal. The weight matrix W is used to measure the impact of the input signal on the neuron's output, and the bias vector B is used to adjust the neuron's activation threshold. The VAE encoder includes a five-layer neural network, and the decoder also includes a five-layer neural network. Gaussian distribution is used to initialize the weight matrix W of the encoder E and decoder D in the VAE model. E ,W D , and the bias vector B E 、B D The initial value of is set to 0; Step 2.3, train the encoder E and decoder D: Use the gradient descent method to minimize Equation (2) and update the weight matrix W of the encoder E and decoder D E ,W D and the bias vector B E 、B D Until the number of iterations is n ed , set the batch size c of the current training now =c now +1; Step 2.4, update the latent space and its characteristics formed in the process of VAE extracting data features; For an arbitrary K-dimensional distribution, it can be generated by selecting a set of K variables that obey the normal distribution and mapping this distribution to a sufficiently complex function, so that the hidden variable z in the training process obeys the Gaussian distribution, and then constructing a specific function to map the latent space that obeys the Gaussian distribution to a more complex distribution. This is the training process of VAE; the encoder of VAE includes a five-layer neural network, and the decoder also includes a five-layer neural network. When the data set X′ is input into the encoder, the encoder finally outputs the mean μ and variance σ, which represent the data distribution characteristics of the original sample; for each element x in the data set X′ i ′, the encoder will generate a corresponding hidden variable z that obeys the Gaussian distribution i ′ to simulate element x i ′ corresponds to the distribution of hidden variables z i ′ has its own mean μ i and variance σ i , according to the encoder parameter update in step 2.3, reconstruct all x i ′ corresponds to the hidden variable z i '; Step 2.5, repeat steps 2.3 and 2.4 until c now =c; Step 2.6, use the evolutionary symbolic momentum Lion optimizer to update the gradient of the gradient descent method used in step 2.3, and let c now =0, e=e+1; Step 2.7, repeat steps 2.3, 2.4, 2.5, and 2.6 until the number of iterations of the model e = E; Step 2.8, training generation, record the weight matrix W of encoder E and decoder D E ,W D and the bias vector B E 、B D , record all x i ′ corresponds to the hidden variable z i The mean μ of ′ i and variance σ i .

5. The wireless network interruption detection method based on variational autoencoder according to claim 1 is characterized in that The fourth step is specifically as follows: Step 4.1: Set the number of clusters m = 4 and the maximum number of iterations N, whose values ​​are determined by experimental experience; Step 4.2, select an initial cluster center for each cluster of the dataset V, the initial center of the first class (x j ′,μ j ,σ j ) is randomly selected, and then the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the initial cluster center is calculated i (v) And with probability Select the center point of the new class and get the second cluster point (x l ′,μ l ,σ l ); Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first two cluster centers i (v) And with probability Select the center point of the new class and get the third cluster point (x b ′,μ b ,σ b ); Then calculate the farthest Euclidean distance d from the latent variable mean of each data point that has not been selected to the first three cluster centers i (v) And with probability Select the center point of the new class and get the fourth cluster point (x g ′,μ g ,σ g ); Finally, the set W = {(x j ′,μ j ,σ j ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}; Step 4.3: Assign the remaining elements of the dataset V to the nearest neighbor cluster according to the minimum distance principle, specifically: Calculate the minimum Euclidean distance from the latent variable mean of the remaining elements of the data set V to the latent variable mean of the four cluster points, And classify each element into the cluster point that can obtain the minimum Euclidean distance; Step 4.4, all elements are classified into the set W = {(x j ′,μ j ,σ j ),(x l ′,μ l ,σ l ),(x b ′,μ b ,σ b ),(x g ′,μ g ,σ g )}, use the sample mean of each cluster to update the cluster center, and replace the elements in the set W, specifically: For the points classified into the first cluster (x j ′,μ j ,σ j ) consists of the set V of all elements j , take V j The mean of all elements in the latent variable U j It is V j The number of elements in , let The cluster center is updated, and the same operation is performed on other cluster points; Step 4.5: Repeat steps 4.3 and 4.4 until the cluster center no longer changes or the maximum number of iterations N is reached; Step 4.6, output the final cluster center set and 4 cluster divisions.

Images