CN114501525A

CN114501525A - Wireless network interruption detection method based on condition generation countermeasure network

Info

Publication number: CN114501525A
Application number: CN202210108134.XA
Authority: CN
Inventors: 潘志文; 葛旭; 刘楠; 尤肖虎
Original assignee: Southeast University; Network Communication and Security Zijinshan Laboratory
Current assignee: Southeast University; Network Communication and Security Zijinshan Laboratory
Priority date: 2022-01-28
Filing date: 2022-01-28
Publication date: 2022-05-13
Anticipated expiration: 2042-01-28
Also published as: CN114501525B

Abstract

The invention discloses a wireless network interruption detection method based on a condition generation countermeasure network, which can accurately detect interruption under the conditions of unbalanced data sets and overlapping data classes and correctly classify different types of interruption, and specifically comprises the following steps: the first step is as follows: collecting network key performance indexes and forming a data set S; the second step is that: training the improved condition by using a data set S to generate an antagonistic network, wherein the antagonistic network consists of a generator G and a discriminator D, and both G and D are of a fully-connected neural network structure; the third step: collecting wireless networks T₂KPI information reported by users in time, and the fourth step: synthesizing interrupt data by using the model obtained in the second step, and balancing a data set H; the fifth step: calculating the inter-class overlap index of each sample in the training set V after calibration; and a sixth step: training a human using training set V and inter-class overlap exponent set OAn Artificial Neural Network (ANN) for obtaining an interruption detection model; the seventh step: according to KPI information x reported by users in network in real time (x belongs to R)ⁿ) Interrupt detection is performed.

Description

Wireless network interruption detection method based on condition generation countermeasure network

Technical Field

The invention belongs to the technical field of network interruption detection in a wireless network, and particularly relates to a wireless network interruption detection method for generating a countermeasure network based on conditions.

Background

As one of key technologies of the wireless network self-healing technology, the interruption detection has important significance for improving the operation and maintenance efficiency of the wireless network and reducing the operation and maintenance cost. However, a wireless network outage is a small probability event, and the amount of outage characteristic data that can be collected is much smaller than normal data, resulting in a severe imbalance of data sets. Furthermore, when there is more than one type of interruption in the network, there is often more severe inter-class coincidence between the different types of interruption data. These factors all contribute to a reduction in the performance of wireless network outage detection. Therefore, how to improve the wireless network interruption detection performance under the conditions of data imbalance and data type overlap is an important problem.

Disclosure of Invention

The technical problem is as follows: in order to solve the problems, the invention discloses a wireless network interrupt detection method for generating a countermeasure network based on conditions, which can accurately detect interrupts under the conditions of unbalanced data sets and overlapping of data classes and correctly classify different types of interrupts.

The technical scheme is as follows: the invention relates to a wireless network interruption detection method based on condition generation countermeasure network, which comprises the following steps:

the first step is as follows: collecting key performance indicators KPI of a network, and forming a data set S;

the second step is that: training an improved condition by using a data set S to generate an antagonistic network CGAN-W, wherein the CGAN-W consists of a generator G and a discriminator D, and both G and D are of a fully-connected neural network structure;

the third step: collecting wireless networks T₂KPI information reported by users in time and stored as data set

Wherein N is_HRepresents the total number of samples in H, element (H)_w,y_w),w＝1,2,…,N_H，h_w∈RⁿAnd representing n-dimensional KPI information reported by the user, wherein the specific value of n can be determined by the operator according to the number of users and the network operation condition, and y_wIs h_wThe value of the label (1) is 0,1,2, and 3. After acquiring the data set H, entering a fourth step;

the fourth step: synthesizing interrupt data by using the CGAN-W model obtained in the second step, and balancing a data set H;

the fifth step: calculating the inter-class overlap index of each sample in the training set V after calibration;

and a sixth step: training an Artificial Neural Network (ANN) by using a training set V and an inter-class overlap exponent set O to obtain an interruption detection model;

the seventh step: according to KPI information x reported by users in network in real time (x belongs to R)ⁿ) Interrupt detection is performed.

Wherein the content of the first and second substances,

the first step is as follows: collecting key performance indicators KPIs of the network, and forming a data set S, which specifically comprises the following steps:

step 1.1, obtaining time T in wireless network₁KPI information reported by internal users;

step 1.2, storing KPI information reported by users as a data set

In the form of (a); wherein N is_SIs the number of the element in S, the ith element (x) in S_i,y_i)，i＝1,2…,N_S，x_i∈RⁿThe method comprises the steps that n-dimensional KPI information reported by a user at a certain time is represented, and the n-dimensional KPI information specifically comprises reference signal receiving power and signal-to-interference-and-noise ratio of a service cell and a neighbor cell; the value of n can be determined by an operator according to the number of users and the network operation condition; y is_iIs x_iThe label of (a), which represents the state of the base station serving the user, takes values of 0,1,2, 3; wherein, y_i0 represents that the base station is in a normal state and has the strongest communication capability; y is_i1 indicates that the base station is in a light interruption state, the communication capability is slightly reduced, and y _i2 means that the base station is in a medium interruption state, and the communication capability is seriously reduced, which may cause a communication failure phenomenon; y is_iWhen the base station completely loses communication capability, a large number of link connection failure events and user handover events are triggered to acquire S.

The second step is specifically as follows:

step 2.1, normalizing the data in the data set S according to the formula (1) to ensure that the final data is distributed between-1 and 1;

wherein the content of the first and second substances,

represents the ith data sample x_iThe value of d-dimension characteristic is 1,2, …, n, n represents sample x_iA feature dimension of (a);

representing the normalized sample, and turning to the step 2.2 after acquiring a normalized data set S;

step 2.2, dividing S into four subsets according to the sample labels in the data set S: s₀,S₁,S₂,S₃(ii) a Wherein the subset S₀The middle element is labeled 0 or y_iSample of 0, representing normal data; subset S₁,S₂,S₃The middle elements are samples with labels of 1,2 and 3 respectively and represent interrupt type data respectively; statistical subset S_kTotal number of middle samples, denoted N_kK 1,2,3, a subset S containing interrupt data is obtained₁,S₂,S₃Then, the step 2.3 is carried out;

step 2.3, D and G loss functions in the CGAN-W are defined as shown in the formula (2) and the formula (3) respectively:

wherein L is_DRepresents a discriminator loss, L_GRepresents a generator loss; m represents the total number of samples used for training CGAN-W; z is a radical of_jTable sampled l-dimensional random noise samples, z, following a standard normal distribution_j∈R^l，j＝1,2,…,m；x_jN-dimensional KPI information x reported on behalf of user_j∈Rⁿ；y_jRepresents a sample label; using interrupt subsets S₁,S₂,S₃Training the CGAN-W model to enable the CGAN-W model to learn the interrupt-like data characteristics, so that the sample (x)_j,y_j)∈S_k，G(z_j,y_j) Is input (z)_j,y_j) The generator outputs layer neuron output, namely a synthesized sample; d (x)_j,y_j) Is input as (x)_j,y_j) Then, the discriminator outputs the neuron output; d (G (z)_j,y_j),y_j) Is an input (G (z)_j,y_j),y_j) Then, the discriminator outputs the layer neuron output; after defining the loss function, turning to step 2.4;

step 2.4, setting parameters required by subsequent model training: the method specifically comprises the following steps: a learning rate α; the truncation coefficient c is used for limiting the weight range after the updating of the discriminator; the batch size m is used for setting the number of samples sampled in each round of training; number of discriminant training times n_disSetting the training times of D when G is trained once; the maximum iteration number iteration of the model; number of iterations t, t of the discriminator<n_dis(ii) a Number of model iterations iter, iter<iteration; wherein, alpha, c, m, n_disThe numerical value of iteration is determined by the operator; after the setting is finished, the step 2.5 is carried out;

step 2.5, randomly initializing a generator G and a discriminator D in the CGAN-W model to obtain a weight vector W_G,W_DAnd an offset vector b_G,b_D(ii) a The number of iterations iter of the initialized model is 0, and the judgment is madeThe iteration time t is 0; after the model initialization is finished, entering step 2.6;

step 2.6, from the interruption subset S_kRandomly sampling m samples to obtain a real sample set { (x)_1,y₁),(x_2,y₂),…,(x_m,y_m) }; wherein, the set { x₁,x₂,…,x_mKPI information of sample is represented as real ═ x₁,x₂,…,x_m}; set { y₁,y₂,…,y_mRepresents the corresponding sample label, denoted label ═ y₁,y₂,…,y_m}; from the l-dimensional random noise z (z ∈ R) that follows a standard normal distribution^l) M noise samples are sampled to form a set noise ═ z₁,z₂,…,z_m}; the set noise and the label set label are set to { y ═ y₁,y₂,…,y_mCombine, i.e. for each noise sample z_jAdding a label y_jObtaining the element (z)_j,y_j)，y_jE.g. label; go through the m noise samples sampled as this operation z₁,z₂,…,z_mGet the set { (z)₁,y₁),(z₂,y₂),…,(z_m,y_m) }; after sampling is finished, turning to the step 2.7;

step 2.7, set { (z)₁,y₁),(z₂,y₂),…,(z_m,y_m) Input generator, output composite sample set

Wherein the content of the first and second substances,

combining the sample set

And step 2.6, the label set label obtained in step (y) ═ y₁,y₂,…,y_mCombining, i.e. for each synthesized sample

Adding label y_j,y_jBelongs to label to obtain elements

Go through the m samples generated according to the operation

Obtaining a set of synthetic samples

After a synthetic sample set is obtained, the step 2.8 is carried out;

step 2.8, set the real samples { (x)₁,y₁),(x₂,y₂),…,(x_m,y_m) And the resultant sample set

As the input of the discriminator, the discriminator parameter W is updated by using the small batch stochastic gradient descent algorithm maximization formula (2)_D,b_D(ii) a Making the iteration time t of the discriminator t equal to t +1, and turning to step 2.9;

step 2.9, the updated weight coefficient W of the discriminator_DTruncating to a value between-c and c, i.e. | W_DC is less than or equal to l, wherein c is a truncation coefficient, and the specific numerical value can be determined by an operator;

step 2.10, repeat steps 2.6, 2.7, 2.8, 2.9 until the number of times of arbiter iteration t>n_dis(ii) a The procedure is shifted to step 2.11;

step 2.11, again from the i-dimensional random noise z (z ∈ R) that follows the standard normal distribution^l) M noise samples are sampled to form a set noise '═ z'₁,z′₂,…,z′_mAnd (3) comparing the set noise' with the label set label obtained in the step 2.6, wherein the label set label is { y ═ y }₁,y₂,…,y_mCombine, i.e. z 'for each noise sample'_jAdding a label y_j,y_jE.g. label to obtain element (z'_j,y_j) (ii) a Operating as such, the sampled m noise samples { z'₁,z′₂,…,z′_mObtaining a set { (z'₁,y₁),(z′₂,y₂),…,(z′_m,y_m) And inputting the data into a generator; updating generator parameters W using a small batch stochastic gradient descent algorithm minimization equation (3)_G,b_GMaking the iteration number iter of the model equal to iter +1, and turning to step 2.12;

step 2.12, if iter>iteration, ending training, recording generator G, and discriminator D weight vector W_{G_opt},W_{D_opt}Offset vector b_{G_opt},b_{D_opt}Entering the step 2.13, otherwise turning to the step 2.5, and starting a new round of training;

step 2.13, repeating the steps 2.3-2.12 until three interrupt subsets S₁,S₂,S₃The training is completed, and the CGAN-W model with the learned interrupt characteristics is obtained.

The fourth step is specifically as follows:

step 4.1, sampling n from the random noise z of dimension I obeying the standard normal distribution_genNoise samples forming a set

Wherein n is_genRepresents the number of minority samples expected to be synthesized, n_gen>0; because the number of the added synthetic samples can influence the imbalance proportion of the finally obtained training set samples and further influence the classification effect, the optimal n is searched by a grid search method_genA value; sampling label information to obtain a label set

Only a few classes of data are synthesized for the purpose of balancing the data set, so the labels only take values in the set {1,2,3}, i.e., y_r∈{1,2,3},r＝1,2,…,n_gen(ii) a Combining noise sets

And a set of labels

I.e. for the noise sample z_r(r＝1,2,…,n_gen) Adding a label y_r(r＝1,2,…,n_gen),y_r∈label_genObtaining the element (z)_r,y_r) (ii) a Go through all samples according to the operation to obtain a set

Turning to step 4.2;

step 4.2, assemble

An input generator according to W obtained in the second step_{G_opt},b_{G_opt}Calculation Generator output, note

Wherein the content of the first and second substances,

will be assembled

And tag collections

Combining, i.e. for producing samples

Adding a label y_r,y_r∈label_genTo obtain the element

Traversing all the generated samples according to the operation to obtain a generated data set

After the generated data set U is obtained, turning to the step 4.3;

and 4.3, merging the generated data set U and the original data set H to obtain a calibrated training set V ═ U ═ H.

The fifth step is specifically as follows:

step 5.1, for each piece of KPI information V in V_e，e＝1,2,…,N_V，N_VRepresents the total number of samples in V, V_e∈RⁿSelecting q pieces of KPI information closest to the KPI information, and forming a sample set Neigh ═ v₁,v₂,…,v_q}; turning to step 5.2;

step 5.2, calculating a neutralization sample v in the set Neigh_eThe number of samples with the same label is recorded as NN, wherein NN is more than or equal to 0 and less than or equal to q, and the step 5.3 is carried out;

step 5.3, if NN>0, then sample v_eInter-class overlap index of (a)_eNN/q; if NN is equal to 0, then sample v_eInter-class overlap index of (a)_eβ; wherein the content of the first and second substances,

for adjusting coefficients, for adjusting the samples v_eThe specific value of beta can be determined by an operator, and the inter-class overlap index o obtained by calculation_eIncorporating the set O, i.e. O ═ O & { O_e}；

And 5.4, repeating the step 5.1, the step 5.2 and the step 5.3 until all samples in the V are traversed.

The sixth step is specifically as follows:

step 6.1, determining an ANN loss function according to V and O, wherein the ANN loss function is shown as a formula (4)

Wherein N is_VRepresents the total number of samples, o, in the training set V_e(e＝1,2,…,N_V) Representative sample v_eOf (2) inter-class overlap index, y_egAs a function of sign, if sample v_eIs equal to g, then 1 is taken, otherwise 0, θ is taken_jRepresents the weight matrix and bias vector corresponding to the jth neuron of the output layer_gRepresenting a weight matrix and a bias vector corresponding to the g-th neuron of the output layer, using superscript T to represent transposition, defining a loss function, and then turning to the step 6.2;

step 6.2, solving the minimum value of the formula (4) by using a gradient descent algorithm to obtain a weight vector W of the ANN network_{ANN_opt}And bias b_{ANN_opt}。

The seventh step is specifically as follows:

step 7.1, inputting x into the ANN model obtained in the sixth step according to W_{ANN_opt}And b_{ANN_opt}Calculating output layer output pred, pred is equal to R⁴；

Step 7.2, calculate the final prediction tag

Where argmax denotes the vector (pred)¹,pred²,pred³,pred⁴) The subscript index corresponding to the medium maximum value is numbered from 0 if

And judging x as an interrupt sample and determining the specific interrupt type, otherwise, judging the x as a normal sample.

Has the advantages that: according to the interrupt detection method of the countermeasure network based on condition generation, under the condition that the problems of unbalance and inter-class overlap exist in the data set, a small number of classes of data are sampled through the CGAN-W model, meanwhile, the problems of data unbalance and inter-class overlap are solved through calculating inter-class overlap indexes and weighting training ANN, and the overall improvement of the interrupt detection performance is brought. Compared with the traditional classification method and the data up-sampling method, the method has the advantage that the interruption detection performance is obviously improved.

Drawings

FIG. 1 is a schematic diagram of a CGAN-W model.

FIG. 2 is a schematic diagram of the structure of ANN.

Detailed Description

The invention relates to a wireless network interruption detection method for generating a countermeasure network based on conditions, which is used for collecting two types of KPI information: the Reference Signal Received Power (RSRP) and signal to interference and noise ratio (SINR) are exemplified for explanation. An embodiment of the method is given below, the steps of which are all performed in a monitoring center for monitoring the operation of the network.

The first step is as follows: network KPIs are gathered and a data set S is formed. The method comprises the following steps:

(1) and (4) acquiring KPI information reported by users in 80s in the wireless network, and switching to the step (2).

(2) Storing KPI information reported by users as data sets

In the form of (1). Wherein N is_SIs the number of elements in S. I (i ═ 1,2 …, N) in S_S) An element (x)_i,y_i)，x_i∈RⁿAnd the n-dimensional KPI information reported by the user at a certain time is represented. In this embodiment, the value of n is 8, and the KPI information includes RSRP and SINR of the serving cell and the three nearest neighbor cells. KPI information is specifically represented as x_i＝{RSRP_sev,SINR_sev,RSRP_nei1,SINR_nei1,RSRP_nei2,SINR_nei2,RSRP_nei3,SINR_nei3}. Where subscript sev denotes the serving cell and subscripts nei1, nei2, nei3 denote the three nearest neighbor cells. y is_iIs x_iThe label of (b), which represents the state of the base station serving the user, takes values of 0,1,2, 3. Wherein, y_i0 indicates that the base station is in a normal (normal) state. y is_i1 indicates that the base station is in a light outage (degraded) state. y is_i2 indicates that the base station is in a medium interruption (shredded) state. y is_iAnd 3 indicates that the base station is in a severe outage (catatonic) state. After S is obtained, the second step is carried out.

The second step is that: CGAN-W is trained using data set S. The CGAN-W is composed of a generator G and a discriminator D. G and D are all full-connection neural network structures, and the specific structure is shown in figure 1. The method comprises the following steps:

(1) the data in S are normalized according to equation (1) so that the final data are distributed between-1 and 1. And (5) after the normalized data set S is obtained, switching to the process (2).

(2) Dividing S into four subsets according to the sample labels in S: s₀,S₁,S₂,S₃. Wherein the subset S₀The middle element is labeled 0 (i.e., y)_i0) sample, representing normal data. Subset S₁,S₂,S₃The middle element is a sample with

labels

1,2 and 3 respectively, and represents an interrupt type data respectively. Statistical subset S_k(k is 1,2,3) total number of samples, denoted as N_k(k is 1,2, 3). Obtaining a subset S containing interrupt data₁,S₂,S₃Thereafter, the process proceeds to the step (3).

(3) D and G loss functions in CGAN-W are defined as shown in the formulas (2) and (3) respectively. After the loss function is defined, the process proceeds to the flow (4).

(4) And setting parameters required by subsequent model training. The method specifically comprises the following steps: learning rate α is 0.0005; the truncation coefficient c is 0.01 and is used for limiting the weight range after the updating of the discriminator; the batch size m is 64 and is used for setting the number of samples sampled in each round of training; number of discriminant training times n_dis5, setting the number of times that D needs to be trained when G is trained once; the maximum iteration number iteration of the model is 20000; number of times t (t) of arbiter iterations<n_dis) (ii) a Number of model iterations iter (iter)<iteration). In addition, in this embodiment, G and D are all configured as a fully-connected neural network structure including three hidden layers. G, D hidden layer activation functions all use the leak relu function, i.e., leak relu (x) ═ max (0, x) + γ × min (0, x), and G output layer activation functions use the tanh function, i.e., tan h

The D output layer does not use the activation function. And entering the flow (5) after the setting is finished.

(5) Randomly initializing weight vector W of generator G and discriminator D in CGAN-W model_G,W_DAnd biasVector b_G,b_D. The number of iterations iter of the initialization model is 0, and the number of iterations t of the discriminator is 0. After the model initialization is completed, the flow (6) is entered.

(6) From the interrupt subset S_kRandomly sampling m samples in (k is 1,2,3) to obtain a real sample set { (x)_1,y₁),(x_2,y₂),…,(x_m,y_m)}. Wherein, the set { x₁,x₂,…,x_mKPI information of sample is represented as real ═ x₁,x₂,…,x_m}. Set { y₁,y₂,…,y_mRepresents the corresponding sample label, denoted label ═ y₁,y₂,…,y_m}. From 100-dimensional random noise z (z ∈ R) that follows a standard normal distribution¹⁰⁰) M noise samples are sampled to form a set noise ═ z₁,z₂,…,z_m}. The set noise and the label set label are set to { y ═ y₁,y₂,…,y_mCombine, i.e. for each noise sample z_j(j-1, 2, …, m), appending a label y_j(j＝1,2,…,m),y_jE.g. label, to obtain the element (z)_j,y_j). Go through the m noise samples sampled as this operation z₁,z₂,…,z_mGet the set { (z)₁,y₁),(z₂,y₂),…,(z_m,y_m)}. And after sampling is finished, the process is switched to the flow (7).

(7) Set { (z)₁,y₁),(z₂,y₂),…,(z_m,y_m) Input generator, output composite sample set

Wherein the content of the first and second substances,

combining the sample set

And the label set label obtained in the procedure (6) ═ y₁,y₂,…,y_mCombine, i.e. for each synthesized sample

Adding a label y_j(j＝1,2,…,m),y_jBelongs to label to obtain elements

Go through the m samples generated according to the operation

Obtaining a set of synthetic samples

After the synthesis sample set is obtained, the procedure proceeds to (8).

(8) Set the real samples { (x)_1,y₁),(x_2,y₂),…,(x_m,y_m) And the resultant sample set

As the input of the discriminator, the discriminator parameter W is updated by using the small batch stochastic gradient descent algorithm maximization formula (2)_D,b_D. Let the discriminator iteration number t be t +1, and proceed to the process (9).

(9) For updated weight coefficient W of discriminator_DCut off to make | W_D|≤0.01。

(10) And (5) repeating the steps (6), (7), (8) and (9) until the number t of times of iteration of the discriminator is greater than 5. Proceed to the flow (11).

(11) Again from 100-dimensional random noise z (z ∈ R) following a standard normal distribution¹⁰⁰) M noise samples are sampled to form a set noise '═ z'₁,z′₂,…,z′_m}. The set noise' and the label set label obtained in the flow (6) are set to { y ═ y₁,y₂,…,y_mCombine, i.e. for each noise sample z_j' (j-1, 2, …, m) with the addition of the label y_j(j＝1,2,…,m),y_jE.g. label, to obtain the element (z)_j′,y_j). Go through the m noise sampled according to the operationSample { z'₁,z′₂,…,z′_mObtaining a set { (z'₁,y₁),(z′₂,y₂),…,(z′_m,y_m) And input to the generator. Updating generator parameters W using a small batch stochastic gradient descent algorithm minimization equation (3)_G,b_G. Let the number of model iterations iter be iter + 1. Proceed to the procedure (12).

(12) If ter>20000, finish training. Record generator G, arbiter D weight vector W_{G_opt},W_{D_opt}Offset vector b_{G_opt},b_{D_opt}Then, the flow proceeds to (13). Otherwise, turning to the flow (5) and starting a new training round.

(13) Repeating the processes (3) to (12) until three interrupt subsets S₁,S₂,S₃The training is completed, and the CGAN-W model with the learned interrupt characteristics is obtained. And entering the third step.

The third step: KPI information reported by users in wireless network 8s is collected and stored as data set

Wherein N is_HRepresents the total number of samples in H. Element (h)_w,y_w),w＝1,2,…,N_H，h_w∈RⁿAnd the n-dimensional KPI information reported by the user is represented. In this embodiment, the value of n is 8, and the KPI information includes RSRP and SINR of the serving cell and the three nearest neighbor cells. KPI information is specifically expressed as h_w＝{RSRP_sev,SINR_sev,RSRP_nei1,SINR_nei1,RSRP_nei2,SINR_nei2,RSRP_nei3,SINR_nei3}。y_wIs h_wThe value of the label (2) is 0,1,2, 3. After the data set H is acquired, the fourth step is entered.

The fourth step: and synthesizing the interrupt data by using the CGAN-W model obtained in the second step, and balancing the data set H. The method comprises the following steps:

(1) from 100 dimensions following a standard normal distributionSampling n in random noise z_genNoise samples forming a set

Wherein n is_gen(n_gen>0) Representing the number of samples of the few classes expected to be synthesized. Because the number of the added synthetic samples can influence the imbalance proportion of the finally obtained training set samples and further influence the classification effect, the optimal n can be searched by a grid search method_genThe value is obtained. Sampling label information to obtain a label set

It should be noted that the method synthesizes only a few types of data to achieve the goal of balancing the data set. Thus, the label only takes values in the set {1,2,3}, i.e., y_r∈{1,2,3},r＝1,2,…,n_gen. Combining noise sets

And a set of labels

I.e. for the noise sample z_r(r＝1,2,…,n_gen) Adding a label y_r(r＝1,2,…,n_gen),y_r∈label_genObtaining the element (z)_r,y_r). Go through all samples according to the operation to obtain a set

And (4) transferring to the process (2).

(2) Will be assembled

Wherein the content of the first and second substances,

will be assembled

And tag collections

Combining, i.e. for producing samples

Adding label y_r(r＝1,2,…,n_gen),y_r∈label_genTo obtain the element

After the generated data set U is obtained, the process proceeds to the flow (3).

(3) And merging the generated data set U and the original data set H to obtain a calibrated training set V ═ U @ H. And entering the fifth step.

The fifth step: the inter-class overlap index for each sample in the training set V after calibration is calculated. The method comprises the following steps:

(1) for each piece of KPI information V in V_e(e＝1,2,…,N_V，N_VRepresenting the total number of samples in V), V_e∈R⁸Selecting 5 KPI information nearest to the Euclidean distance to form a sample set Neigh ═ v₁,v₂,…,v₅}. And (3) transferring to the process (2).

(2) Calculating a neutralization sample v in the set Neigh_eThe number of samples with the same label is recorded as NN (NN is more than or equal to 0 and less than or equal to 5). The process proceeds to the step (3).

(3) If NN>0，Then the sample v_eInter-class overlap index of (a)_eNN/5. If NN is equal to 0, then sample v_eInter-class overlap index of (a)_e0.05. Will calculate the resulting inter-class overlap index o_eIncorporating the set O, i.e. O ═ O & { O_e}。

(4) And repeating the processes (1), (2) and (3) until all the sample traversals in the V are completed. And entering the sixth step.

And a sixth step: and (4) training the ANN by using the training set V and the inter-class overlap index set O to obtain an interruption detection model. The method comprises the following steps:

(1) and determining an ANN loss function according to V and O. Specifically, the formula is shown in (4). In addition, in this embodiment, the ANN is set to have a three-layer fully-connected neural network structure, and the number of neurons in the input layer is a KPI feature dimension, that is, 8. The number of neurons in the output layer is the total number of sample categories, namely 4. The specific structure is shown in fig. 2. The hidden layer activation function uses the LeakyReLU function, and the output layer activation function uses the softmax function. After the setting is completed, the process proceeds to the flow (2).

(2) Solving the minimum value of the formula (4) by using a gradient descent algorithm to obtain a weight vector W of the ANN network_{ANN_opt}And bias b_{ANN_opt}. And entering the seventh step.

The seventh step: according to the real-time report of KPI information x (x belongs to R) by users in network⁸) Interrupt detection is performed. The method comprises the following steps:

(1) inputting x into ANN model obtained in the sixth step according to W_{ANN_opt}And b_{ANN_opt}Calculating output layer output pred, pred is equal to R⁴。

(2) Computing a final predicted label

Where argmax denotes the vector (pred)¹,pred²,pred³,pred⁴) The subscript index corresponding to the medium maximum value starts from 0. If it is

Judging x as an interrupt sample and determining the specific interrupt type, otherwise, judging x as a normal sample

The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims

1. A wireless network interruption detection method for generating a countermeasure network based on conditions is characterized by comprising the following steps:

the first step is as follows: collecting key performance indicators KPIs of a network, and forming a data set S;

Wherein N is_HRepresents the total number of samples in H, element (H)_w，y_w)，w＝1，2，...，N_H，h_w∈RⁿAnd representing n-dimensional KPI information reported by the user, wherein the specific value of n can be determined by the operator according to the number of users and the network operation condition, and y_wIs h_wThe value of the label is 0,1,2 and 3, and the fourth step is carried out after the data set H is obtained;

2. The method for detecting interruption of a wireless network for generating a countermeasure network based on a condition according to claim 1, wherein the first step is: the method comprises the following steps of collecting key performance indicators KPI of a network, and forming a data set S, specifically:

step 1.2, storing KPI information reported by users as a data set

In the form of (a); wherein N is_SIs the number of the element in S, the ith element (x) in S_i，y_i)，i＝1，2...，N_S，x_i∈RⁿThe method comprises the steps that n-dimensional KPI information reported by a user at a certain time is represented, and the n-dimensional KPI information specifically comprises reference signal receiving power and signal-to-interference-and-noise ratio of a service cell and a neighbor cell; the value of n can be determined by an operator according to the number of users and the network operation condition; y is_iIs x_iThe label of (b) indicates the state of the base station serving the user, and values are 0,1,2, and 3; wherein, y_i0 represents that the base station is in a normal state and has the strongest communication capability; y is_i1 indicates that the base station is in a light interruption state, the communication capability is slightly reduced, and y_i2 means that the base station is in a medium interruption state, and the communication capability is seriously reduced, which may cause a communication failure phenomenon; y is_iWhen the base station completely loses communication capability, a large number of link connection failure events and user handover events are triggered to acquire S.

3. The method for detecting interruption of a wireless network for generating a countermeasure network based on conditions as set forth in claim 1, wherein the second step is specifically:

step 2.1, normalizing the data in the data set S according to the formula (1) to enable the final data to be distributed between-1 and 1;

wherein the content of the first and second substances,

represents the ith data sample x_iThe value of d-dimension feature, d 1,2, n, n represents sample x_iA feature dimension of (a);

step 2.2, dividing S into four subsets according to the sample labels in the data set S: s₀，S₁，S₂，S₃(ii) a Wherein the subset S₀The middle element is a sample with a label of 0, namely yi is 0, and represents normal data; subset S₁，S₂，S₃The middle elements are samples with labels of 1,2 and 3 respectively and represent interrupt type data respectively; statistical subset S_kTotal number of middle samples, denoted N_kK 1,2,3, a subset S containing interrupt data is obtained₁，S₂，S₃Then, the step 2.3 is carried out;

wherein L is_DRepresents a discriminator loss, L_GRepresents a generator loss; m stands for trainingTotal number of samples of CGAN-W; z is a radical of_jTable sampled l-dimensional random noise samples, z, following a standard normal distribution_j∈R^l，j＝1，2，...，m；x_jN-dimensional KPI information x reported on behalf of user_j∈Rⁿ；y_jRepresents a sample label; using interrupt subsets S₁，S₂，S₃Training the CGAN-W model to enable the CGAN-W model to learn the interrupt-like data characteristics, so that the sample (x)_j，y_j)∈S_k，G(z_j，y_j) Is input (z)_j，y_j) The generator outputs layer neuron output, namely a synthesized sample; d (x)_j，y_j) Is input as (x)_j，y_j) Then, the discriminator outputs the neuron output; d (G (z)_j，y_j)，y_j) Is an input (G (z)_j，y_j)，y_j) Then, the discriminator outputs the neuron output; after defining the loss function, turning to step 2.4;

step 2.4, setting parameters required by subsequent model training: the method specifically comprises the following steps: a learning rate α; a truncation coefficient c for limiting the weight range of the updated discriminator; the batch size m is used for setting the number of samples sampled in each round of training; number of discriminant training times n_disSetting the number of times D needs to be trained when G is trained once; the maximum iteration number iteration of the model; the iterative times t of the discriminator, t < n_dis(ii) a The iteration number iter of the model is less than iteration; wherein, alpha, c, m, n_disThe numerical value of iteration is determined by the operator; after the setting is finished, the step 2.5 is carried out;

step 2.5, randomly initializing a generator G and a discriminator D in the CGAN-W model to obtain a weight vector W_G，W_DAnd an offset vector b_G，b_D(ii) a Initializing the iteration times iter of the model to be 0, and the iteration times t of the discriminator to be 0; after the model initialization is finished, entering step 2.6;

step 2.6, from the interruption subset S_kRandomly sampling m samples to obtain a real sample set { (x)₁，y₁)，(x₂，y₂)，...，(x_m，y_m) }; wherein, the set { x₁，x₂，...，x_mKPI information of sample is represented as real ═ x₁，x₂，...，x_m}; set { y₁，y₂，...，y_mRepresents the corresponding sample label, denoted label ═ y₁，y₂，...，y_m}; from the l-dimensional random noise z (z ∈ R) that follows a standard normal distribution^l) M noise samples are sampled to form a set noise ═ z₁，z₂，...，z_m}; the set noise and the label set label are set to { y ═ y₁，y₂，...，y_mCombine, i.e. for each noise sample z_jAdding a label y_jObtaining the element (z)_j，y_j)，y_jE.g. label; go through the m noise samples sampled as this operation z₁，z₂，...，z_mGet the set { (z)₁，y₁)，(z₂，y₂)，...，(z_m，y_m) }; after sampling is finished, turning to the step 2.7;

step 2.7, set { (z)₁，y₁)，(z₂，y₂)，...，(z_m，y_m) Input generator, output composite sample set

Wherein the content of the first and second substances,

combining the sample set

And step 2.6, the label set label obtained in step (y) ═ y₁，y₂，...，y_mCombining, i.e. for each synthesized sample

Adding a label y_j，y_jBelongs to label to obtain elements

Go through the m samples generated according to the operation

Obtaining a set of synthetic samples

After a synthetic sample set is obtained, the step 2.8 is carried out;

step 2.8, set the real samples { (x)₁，y₁)，(x₂，y₂)，...，(x_m，y_m) And the resultant sample set

As the input of the discriminator, the discriminator parameter W is updated by using the small batch stochastic gradient descent algorithm maximization formula (2)_D，b_D(ii) a Making the iteration time t of the discriminator t equal to t +1, and turning to step 2.9;

step 2.10, repeating steps 2.6, 2.7, 2.8 and 2.9 until the iteration times t of the discriminator is more than n_dis(ii) a The procedure is shifted to step 2.11;

step 2.11, again from the i-dimensional random noise z (z ∈ R) that follows the standard normal distribution^l) M noise samples are sampled to form a set noise '═ z'₁，z′₂，...，z′_mAnd (3) comparing the set noise' with the label set label obtained in the step 2.6, wherein the label set label is { y ═ y }₁，y₂，...，y_mCombine, i.e. z 'for each noise sample'_jAdding a label y_j，y_jE.g. label to obtain element (z'_j，y_j)；Operating as such, the sampled m noise samples { z'₁，z′₂，...，z′_mObtaining a set { (z'₁，y₁)，(z′₂，y₂)，...，(z′_m，y_m) And inputting the data into a generator; updating generator parameters W using a small batch stochastic gradient descent algorithm minimization equation (3)_G，b_GMaking the iteration number iter of the model equal to iter +1, and turning to step 2.12;

step 2.12, if iter > iteration, ending training, recording generator G, and discriminator D weight vector W_{G_opt}，W_{D_opt}Offset vector b_{G_opt}，b_{D_opt}Entering the step 2.13, otherwise turning to the step 2.5, and starting a new round of training;

step 2.13, repeating the steps 2.3-2.12 until three interrupt subsets S₁，S₂，S₃The training is completed, and the CGAN-W model with the learned interrupt characteristics is obtained.

4. The method for detecting interruption of a wireless network for generating a countermeasure network based on conditions as set forth in claim 1, wherein the fourth step is specifically:

Wherein n is_genRepresents the number of minority samples expected to be synthesized, n_genIs greater than 0; because the number of the added synthetic samples can influence the imbalance proportion of the finally obtained training set samples and further influence the classification effect, the optimal n is searched by a grid search method_genA value; sampling label information to obtain a label set

Only a few classes of data are synthesized for the purpose of balancing the data set, so the label only takes values in the set {1,2,3}, i.e., y_r∈{1，2，3}，r＝1，2，...，n_gen(ii) a Combining noise sets

And a set of labels

I.e. for the noise sample z_r(r＝1，2，...，n_gen) Adding the label y_r(r＝1，2，...，n_gen)，y_r∈label_genObtaining the element (z)_r，y_r) (ii) a Go through all samples according to the operation to obtain a set

Turning to the step 4.2;

step 4.2, assemble

An input generator according to W obtained in the second step_{G_opt}，b_{G_opt}Calculation Generator output, note

Wherein the content of the first and second substances,

r＝1，2，...，n_gen，

will be assembled

And tag collections

Combining, i.e. for producing samples

Adding a label y_r，y_r∈label_genTo obtain the element

After the generated data set U is obtained, turning to the step 4.3;

5. The method for detecting interruption of a wireless network for generating a countermeasure network based on conditions as set forth in claim 1, wherein the fifth step is specifically:

step 5.1, for each piece of KPI information V in V_e，e＝1，2，...，N_V，N_VRepresents the total number of samples in V, V_e∈RⁿSelecting q pieces of KPI information closest to the KPI information, and forming a sample set Neigh ═ v₁，v₂，...，v_q}; turning to step 5.2;

step 5.2, calculating a neutralization sample v in the set Neigh_eThe number of samples with the same label is recorded as NN, the NN is more than or equal to 0 and less than or equal to q, and the step 5.3 is carried out;

step 5.3, if NN > 0, then sample v_eInter-class overlap index of (a)_eNN/q; if NN is equal to 0, then sample v_eInter-class overlap index of (a)_eβ; wherein the content of the first and second substances,

for adjusting the coefficients, for adjusting the samples v_eClass (D)The specific value of beta can be determined by the operator, and the inter-class overlap index o obtained by calculation_eIncorporating the set O, i.e. O ═ O & { O_e}；

6. The method for detecting interruption of a wireless network for generating a countermeasure network based on conditions according to claim 1, wherein the sixth step is specifically:

Wherein N is_VRepresents the total number of samples, o, in the training set V_e(e＝1，2，...，N_V) Representative sample v_eOf (2) inter-class overlap index, y_egAs a function of sign, if sample v_eIs equal to g, then 1 is taken, otherwise 0, θ is taken_jRepresents the weight matrix and bias vector corresponding to the jth neuron of the output layer, theta_gRepresenting a weight matrix and a bias vector corresponding to the g-th neuron of the output layer, using superscript T to represent transposition, defining a loss function, and then turning to the step 6.2;

7. The method for detecting interruption of a wireless network generating a countermeasure network based on conditions according to claim 1, wherein the seventh step is specifically:

Step 7.2, calculate the final prediction tag

Where argmax denotes the vector (pred)¹，pred²，pred³，pred⁴) The subscript index corresponding to the medium maximum value is numbered from 0 if