A fault confirmation method for user-sparse communication systems

Abstract

The present invention provides a fault confirmation method for a sparsely populated communication system. The method uses the similarity between the KPIs of a faulty network and a normally operating network to partition the network. Only the key performance indicators (KPIs) most relevant to network fault confirmation within each partition are extracted as training sequences for subsequent machine learning. This reduces the method's requirements for the number of users in each partition and mitigates the adverse effects of uneven user distribution on wireless communication system fault confirmation. Due to the random sampling involved in feature extraction based on data similarity, stability is low. However, due to the low requirement for the number of users in a region, performance is good when faced with an uneven distribution of network users. This method reduces the requirements for the number of users in each partition and mitigates the adverse effects of uneven user distribution on wireless communication system fault confirmation.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to a communication system fault confirmation method based on cost-sensitive support vector machine integrated learning of different feature extraction methods. Background Art

[0002] During the operation and maintenance of wireless communication networks, deterioration of Key Performance Indicators (KPIs) is inevitable. As networks become increasingly heterogeneous and complex, technicians must identify the causes of performance degradation within a large number of observed variables, making manual network fault determination difficult. Currently, existing approaches to network fault determination use machine learning to address the difficulties of manual fault determination. Machine learning approaches to fault determination often consider ideal user distribution. However, in real-world wireless communication networks, due to factors such as topography and urban development, user terminals are often sparsely distributed, manifesting in low user density and uneven spatial distribution. This makes fault determination, which is already difficult, even more challenging. Existing approaches address this user sparsity challenge by focusing on addressing low user density while neglecting to consider scenarios with uneven user distribution, which is clearly unreasonable.

[0003] Therefore, how to efficiently find the cause of communication system performance degradation under the challenge of uneven user distribution in actual wireless communication networks has become an urgent technical problem to be solved and also a task for wireless communication system fault confirmation. Summary of the Invention

[0004] Technical problem: The present invention provides a fault confirmation method for wireless communication systems with uneven user distribution. This method uses the similarity between the KPIs of the faulty network and the normally operating network to partition the network, and only extracts the network key performance indicators that are most relevant to network fault confirmation in each partition as the training sequence for subsequent machine learning, thereby reducing the method's requirements on the number of partitioned users and alleviating the adverse effects of uneven user distribution on wireless communication system fault confirmation.

[0005] Technical solution: The present invention is a fault confirmation method for a user-sparse communication system. The method first divides the base station coverage area into six partitions evenly, and uses the Mahalanobis distance (MD) for each partition to measure the data similarity of the network key performance indicators between the faulty network and the normal operating network; the larger the Mahalanobis distance, the smaller the data similarity, and the network key performance indicators with the smallest similarity in each partition are extracted as training sequences for machine learning, and feature extraction is performed. Then, principal component analysis is used to extract features from the original faulty network key performance indicators; then, the sequences after feature extraction using two different methods and their corresponding labels are combined into two training sets, and machine learning training is performed separately; in terms of machine learning methods, a cost-sensitive support vector machine is selected, which assigns different error root cause classification costs to fault samples of different severities. The higher the fault severity, the relatively greater the cost of confirming the classification error, so that more reasonable attention is paid to the classification of fault root causes of different severities; finally, fault confirmation is completed through ensemble learning of two support vector machines, and a fault root cause classifier with stable output and accurate judgment is obtained.

[0006] The network key performance indicators used in the present invention include Signal-to-interference-plus-noise Ratios (SINR) and Reference Signal Received Power (RSRP).

[0007] This method considers the following network fault types that may occur in the network: excessive uptilt (EU), excessive downtilt (ED), and excessive reduced power (ERP).

[0008] The present invention specifically comprises the following steps:

[0009] Step 1: Generate network KPIs,

[0010] The reference signal received power RSRP and signal interference plus noise ratio SINR of network I user are generated by simulation when M groups of network failures occur, and the set of each group of data is defined as the network KPIs vector, where the m∈{1, 2, ..., M}th network KPIs vector is x m =[RSRP1,SINR1,...,RSRP i , SINR i ,...,RSP I , SINR I ], i∈{1, 2, ..., I}, where RSRP iand SINR i is the RSRP and SINR of the i-th user; another simulation generates the RSRP and SINR of the i-th user when the M_Normal group network is operating normally, and each group of data is formed into a set, where the m_Normal∈{1,...,M_Normal}-th network KPIs vector is in and are the RSRP and SINR of the i-th user when the network is operating normally;

[0011] Step 2: Calculate the similarity of partitioned data and perform feature extraction to form a training set, including:

[0012] For any vector p=(p1,p2,...,p l ) T and q=(q1,q2,...,q l ) T , and its Mahalanobis distance is defined as follows:

[0013]

[0014] cov(p,q)=E(p T q)-E(p)E(q)

[0015] where d M (·) is the Mahalanobis distance, E(·) and (·) T They represent expectation and transposition respectively, cov(p,q) is the covariance of p and q; suppose the base station coverage area contains N section sectors, compare the user data between sectors, and extract the sector data most relevant to fault confirmation as the machine learning training sequence; at the same time, to cope with the challenge of user sparsity, each sector is evenly divided into partitions, and the features of users in the partitions are extracted, so that the number of users required for feature extraction is smaller; the base station coverage area needs to be divided into N setnurn Partition and extract features, each sector needs to be divided into at least two, that is, N setnum ≥2·N section , considering that dividing too many regions will lead to too many features being extracted and the extraction efficiency is too low, we take N setnum =2·N section ;

[0016] When constructing a training set for machine learning, x m_1 As the feature vector, and the fault type corresponding to the m-th KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is

[0017] X MD ={(x 1_1, actual1),(x 2_1 , actual1),...,(x M_1 ,actua l 1)}

[0018] Among them actual m ∈{1, 2, 3} is the fault type corresponding to the mth KPIs vector. m = 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault;

[0019] Network fault types include: antenna tilt angle is too small EU, antenna tilt angle is too large ED, and excessive power reduction ERP;

[0020] Step 3: Principal component analysis feature extraction to form a training set, including the following process:

[0021] Principal component analysis (PCA) is the most commonly used data dimensionality reduction method in conjunction with support vector machines.

[0022] Set all x m Composition matrix X:

[0023]

[0024] Decentralize the matrix X, that is, subtract the mean of each column from the mean of the column to perform zero mean processing;

[0025] Calculate the covariance matrix R is a set of real numbers, and C is decomposed to obtain the eigenvalue λ of the covariance matrix c , and its corresponding eigenvector v c ; The obtained λ c Arrange them in descending order and extract n c eigenvalues; for n c The value of follows the following principles: first determine a threshold thre λ ∈(0, 1), which reflects the proportion of the extracted features in the total features. Considering that we want to reduce the data dimension and retain the principal components, we choose thre λ =0.95, n λ is the number of eigenvalues ​​of C, n c The value is the smallest integer that satisfies the following formula:

[0026]

[0027] Combine the eigenvectors corresponding to the extracted eigenvalues ​​into a matrix The data after feature extraction is Thus, principal component analysis feature extraction is achieved;

[0028] When constructing a training set for machine learning, divide X' into M x rows m_2 As the feature vector, and the fault type corresponding to the m-th KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is

[0029]

[0030] Among them actual m ∈{1, 2, 3} is the fault type corresponding to the mth KPIs vector. m = 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault;

[0031] Step 4: Training of cost-sensitive support vector machine, including the following process:

[0032] The cost-sensitive support vector machine used assigns different classification error costs to faults of different severity, considering three fault severity levels: severe faults, moderate faults, and minor faults. Humans can manually judge the severity of faults based on the definition of raw network key indicators, i.e., key network indicators without feature extraction.

[0033] In the second and third steps, two different training sets are obtained, namely X MD and X PcA , respectively train two cost-sensitive support vector machines. Due to their similarity, X PCA Just repeat X MD step;

[0034] In the cost-sensitive support vector machine, the cost matrix Cost is introduced to assign different costs to each sample misjudgment; different error classification costs are set for three fault samples with different severities, and c1∈R is defined as the error classification cost of severe faults; c2∈R is the error classification cost of moderate faults; and c3∈R is the error classification cost of minor faults. Since the greater the severity of the fault, the greater the negative impact on the network, a larger error classification cost is required to more reasonably deploy cost-sensitive machine learning, that is, c1≥c2≥c3

[0035]

[0036] Where Cost(k, j) represents the cost of classifying the kth class into the jth class, k, j∈{1, 2, 3, 4, 5, 6, 7, 8, 9}; a variety of existing hyperparameter optimization methods can be used to optimize the parameters of c1, c2, and c3. Bayesian optimization predicts the most likely optimal c1, c2, and c3 as the values ​​of the next calculation, and iterates N times. BY Finally, the optimal Usually the default iteration is 30 times, which does not take too much time and can improve the optimization effect, that is, N BY =30; Substitute the above formula to obtain the optimal cost matrix Cost ＊ , so that the classifier performance is optimal;

[0037] Find w k * 、b k * and Cost * Then, the cost-sensitive support vector machine is used to train the sample set X MD For each sequence x m_1 For classification, x m_1 Determined as class predicted ∈{1, 2, 3, 4, 5, 6, 7, 8, 9}, if and only if:

[0038]

[0039] Where P(predictedclass=j|x m_1 ) means when x m_1 When making classification judgments, the probability of judging as class j needs to be the output value w of the support vector machine j * ·x m_1 +b j * Mapping to probability value, using existing methods to achieve this goal; mapping the nine-category problem to a three-category problem, if class predicted ∈{1, 2, 3}, then it is finally determined to be an EU fault, class predicted =1; if class predicted ∈{4, 5, 6}, then it is finally determined to be an ED fault, class predicted =2; if class predicted ∈{7, 8, 9}, then it is finally determined to be an ERP failure, classp redicted =3,

[0040] Use F actual -score as a performance indicator, F actualThe larger the -score, the better the system performance. It is defined as follows:

[0041]

[0042]

[0043]

[0044] where n actual，predicted represents the number of samples that the cost-sensitive support vector machine classifies from actual∈{1, 2, 3} to predicted∈{1, 2, 3}, n actual，actual Indicates the number of samples that the cost-sensitive support vector machine judges the actual class to be the actual class;

[0045] Step 5: Classification decision based on ensemble learning

[0046] Training set X MD There is random sampling in the generation process, so the stability is low, but because the number of users in the area is low, the performance is good when facing uneven distribution of network users; training set X PCA There is no random sampling in the process, so the stability is high, but a large number of users are required. When faced with an uneven distribution of network users, the performance is poor. Therefore, ensemble learning is used to obtain a stable and efficient classifier. The specific steps are as follows:

[0047] (1) Collect key network indicators (KPIsx) of the network to be confirmed as faulty new =[RSRP 1_new , SINR 1_new ,...,RSRP i_new , SINR i_new ,...,RSRP I_new , SINR I_new ], i∈{1, 2, ..., I}; where RSRP i_new and SINR i_new are the RSRP and SINR of the i-th user in the network to be confirmed for fault;

[0048] (2) x new Substitute into the second and third steps respectively, perform feature extraction and obtain x new_1 、x new_2 ;

[0049] (3) Substitute the cost-sensitive support vector machine solved in step 4 and output P(predictedclass=j|x new_1 ) and P(predictedclass=j|x new_2 );

[0050] (4) Using ensemble learning, new Perform classification judgment, x m_1 Determined as class predicted_new ∈{1, 2, 3, 4, 5, 6, 7, 8, 9}, if and only if:

[0051]

[0052] If class predicted_new ∈{1, 2, 3}, it is finally determined to be an EU fault; if class predicted_new ∈{4, 5, 6}, it is finally determined to be an ED fault; if class predicted_new ∈{7, 8, 9}, it is finally determined to be an ERP failure.

[0053] The sectors are specifically divided as follows: In each sector of the base station, the bisector of the central angle of the sector is used as the dividing line to evenly divide it into two partitions, and the total coverage area is divided into N setnum Partitions, in each partition, use Mahalanobis distance to calculate the similarity between the KPIs of the faulty network and the normal network, and divide I users into regions, where there are I setnum The users are distributed in the setnum∈{1,...,N setnum} partition, whose set is According to UE setnum , for each eigenvector x m Divided into N setnum Partition feature vector The same division is also performed on the network KPIs when the network is operating normally, and the eigenvectors for each feature vector are obtained. Divided into N setnum Partition feature vector Calculate the Mahalanobis distance between the partition feature vectors of all M_Normal groups during normal operation and the sample partition feature vectors, and sum them as x m Index of the partition in setnum

[0054]

[0055] The larger the index of the partition, the lower the similarity between the partition and the normal operation of the network, the more it can represent the characteristics of the fault, and the greater the impact on the fault confirmation result. In order to deal with the problem of uneven distribution of network users, only some users' RSRP and S1NR are extracted from some areas that can represent the fault characteristics as the training sequence for subsequent machine learning. Consider extracting N setnum The data of three partitions in the partitions can reflect the fault characteristics without introducing too much redundant data; take N setnum x in each partitionm The maximum value of the Mahalanobis distance index is the lowest similarity, and the corresponding partition is setnum_1; the second largest Mahalanobis distance index corresponds to the partition setnum_2, and the third largest Mahalanobis distance index corresponds to the partition setnum_3; in the user set UE corresponding to setnum_1 setnum_1 Randomly select N1 users to form a set Because users in the same area are close to each other, their RSRP and SINR have statistical similarities, and no more than 8 and no less than 2 features are extracted, that is, 2≤N1<8; in the user set UE corresponding to setnum_2 setnum_2 Randomly select N2 users to form a set Because the similarity of the user set partition is higher than setnum_1, the higher the similarity, the smaller the impact of the partition on the confirmation result, and the number of extracted users should be less than or equal to N1, that is, 2≤N2≤N1; in the user set UE corresponding to setnum_3 setnum_3 Randomly select N3 users to form a set The same principle as N2, N3 should satisfy 2≤N3≤N2. The feature sequence after feature extraction is as follows:

[0056]

[0057] The fault severity is defined as follows:

[0058] A severe fault is defined as a fault that has a serious impact on network performance, specifically defined as a fault in which the SINR of more than 60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a moderate fault is defined as a fault that has a moderate negative impact on network performance, specifically defined as a fault in which, among the original network key indicators, the SINR of 40%-60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a minor fault is defined as a fault that has a slight negative impact on network performance, specifically defined as a fault in which the SINR of 20%-40% of users in the network coverage area is lower than 0dBm or the RSRP is lower than -110dBm.

[0059] According to the fault severity and fault type of the training sequence, the training sequence is divided into nine categories and marked with k∈{1, 2, 3, 4, 5, 6, 7, 8, 9}. When k=1, it means actual m =1 and serious fault; when k=2, it means actual m =1 and it is a moderate fault; when k=3, it means actual m =1 and it is a minor fault; when k=4, it means actual m =2 and serious fault; when k=5, it means actual m=2 and it is a medium fault; when k=6, it means actual m =2 and it is a minor fault; when k=7, it means actual m =3 and serious fault; when k=8, it means actual m =2 and it is a medium fault; when k=9, it means actual m =2 and it is a minor fault.

[0060] In the fourth step, two cost-sensitive support vector machines X are respectively MD and X PCA Conduct training,

[0061] Convert the multi-classification problem into multiple binary classification problems. For the k-th classifier, let y m Represents x m_1 The binary classification label, if x m_1 Belongs to the kth category, then y m =1, otherwise y m = -1, N is the total sample space X MD The number of training samples belonging to the kth class in ;

[0062] Based on the training set X MD The cost-sensitive support vector machine training is defined as the hyperplane normal vector and intercept of the k-th linear classifier as the classification boundary as w k and b k , then the two-classification problem is expressed as follows:

[0063]

[0064] Among them, ||·|| 2 Indicates the modulus of the vector, ξ m =1-(w k ·x m_1 +b k ) represents the sample x m_1 The distance to the kth class boundary, ξ=[ξ1,ξ2,…,ξ N ],V k is the k-th soft margin penalty parameter, and the optimal value is determined by the ten-fold cross validation method;

[0065] Through the Karush-Kuhn-Tucker condition and the Lagrange multiplier method, the above problem is transformed into the following problem:

[0066]

[0067] where α m , β m is the Lagrange multiplier, α=[α1,α2,…,α N], β=[β1, β2,...,β N ];

[0068] Use partial derivatives to solve the above optimization problem and calculate L(w k , b k ,ξ,α,β) respectively for w k , b k , and set the partial derivative of ξ to 0, we can get:

[0069]

[0070] Substituting Equation (8) into Equation (7), the multivariable optimization problem can be simplified to m The single variable optimization problem is solved by using the Sequential minimal optimization algorithm, and the optimal solution is α m * , substituting into formula (8) we can get w k and b k The optimal solution, that is, w k * and b k * :

[0071]

[0072] Among them, x t_1 For the training set X MD There exists such a t * >0, y t is x t_1 The corresponding label;

[0073] The X MD and X PCA The corresponding cost-sensitive support vector machine acquisition steps are as follows:

[0074]

[0075] (1) Use cost-sensitive support vector machine to train the sample set X MD mid-sequence x m_1 Make classification judgments;

[0076] (2) If class predicted =actual m , otherwise, m=m+1;

[0077] (3) If m≤M, repeat (2) and (3) to complete the classification decision for all training sequences. Otherwise, calculate the performance index F of the classifier under the current c1, c2, and c3. actual -score, and use Bayesian optimization to predict the optimal value c1 of c1, c2, and c3 * , c2 * , c3 * , update, n BY =n BY +1, c1=c1 * , c2=c2 * , c3=c3 * ;

[0078] (4) Reinitialization

[0079] (5) If n BY ≤N BY , repeat (2), (3), (4), (5), otherwise, find F in all combinations of c1, c2, c3 actual -score, and applies its corresponding c1, c2, c3 values ​​to the cost matrix Cost.

[0080] After the above steps, we can get the training set X MD The corresponding cost-sensitive support vector machine, X MD Change to X PCA , repeat the above steps to obtain the training set X PCA The corresponding cost-sensitive support vector machine.

[0081] Beneficial effects: The wireless communication network fault confirmation method of the present invention has the following technical effects:

[0082] 1) By utilizing the similarity between the KPIs of the faulty network and the normally operating network, only the key performance indicators of the network that are most relevant to network fault confirmation in each partition are extracted as the training sequence for subsequent machine learning, thereby reducing the requirement of this method on the number of users in the partition and alleviating the adverse impact of uneven user distribution on wireless communication system fault confirmation.

[0083] 2) Through ensemble learning, when faced with the challenge of uneven distribution of network users, it combines the improved fault confirmation accuracy brought by the Mahalanobis distance similarity feature extraction method with the stability of the PCA feature extraction method, ultimately obtaining a stable and efficient fault root cause classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a schematic diagram of a wireless communication system of the present invention;

[0085] Figure 2It is a flow chart of the wireless communication system fault confirmation method of the present invention. DETAILED DESCRIPTION

[0086] See also Figure 1-Figure 2 The specific implementation steps of the wireless communication system fault confirmation method based on cost-sensitive support vector machine ensemble learning of different feature extraction methods of the present invention are as follows:

[0087] exist Figure 1 In the wireless communication network shown, the downlink in the millimeter wave network is considered, and the network operating frequency is set to 80GHz. The base station coverage radius is 25m, and each base station is configured with three uniform rectangular arrays, namely N section =3, the size of each array antenna is N t =N az ×N el =8×16, where N az =8 is the number of antennas in the horizontal bit plane, N el = 16 is the number of antennas in the elevation plane. Each array spans 120° in the horizontal plane, evenly dividing the entire horizontal plane. The antennas in the array are configured as follows: the horizontal half-power beamwidth is 65°, the vertical half-power beamwidth is 35°, the horizontal sidelobe level is 25dB, and the vertical sidelobe level is 20dB. To simplify modeling and analysis, it is assumed that the user end uses a single antenna for communication. Assume that the users are evenly distributed in h U =8m horizontal plane, the base station is located at h BS =1.5m height, a single base station provides a hexagonal coverage area, and it is stipulated that each user is provided with communication services by the nearest base station, located at (x i ,y i ) is served by the closest base station and receives interference from the second- and third-closest base stations. In actual mobile communication systems, signal propagation paths are obstructed by hills, trees, buildings, and other factors, and shadow fading with a standard deviation of 4 dB is added.

[0088] As mentioned above, the present invention focuses on troubleshooting sparse data scenarios caused by uneven user distribution. Therefore, after establishing the network model, it is necessary to simulate scenarios with uneven network user distribution. Considering a base station coverage area as a hexagon, a sector-shaped area is delineated within the coverage area. The center of the sector is the base station location, and the central angle is θ = 80°. This sector is assumed to have no users, and the sector is rotated counterclockwise with Δθ = 10°, so that the area without users is traversed.

[0089] During normal network operation, the base station transmit power is set to 1W, and the antenna tilt is set to 40°. The following parameters are configured for network faults of varying severity and root causes: For a minor EU fault, the base station transmit power is set to 1W, and the antenna tilt is set to 35°; for a moderate EU fault, the base station transmit power is set to 1W, and the antenna tilt is set to 25°; for a severe EU fault, the base station transmit power is set to 1W, and the antenna tilt is set to 5°. For a minor ED fault, the base station transmit power is set to 1W, and the antenna tilt is set to 45°; for a moderate ED fault, the base station transmit power is set to 1W, and the antenna tilt is set to 50°; for a severe ED fault, the base station transmit power is set to 1W, and the antenna tilt is set to 55°. For a minor ERP fault, the base station transmit power is set to 0.5W, and the antenna tilt is set to 40°; for a moderate ERP fault, the base station transmit power is set to 0.2W, and the antenna tilt is set to 40°; and for a severe ERP fault, the base station transmit power is set to 0.07W, and the antenna tilt is set to 40°.

[0090] Step 1: Generate network KPIs.

[0091] Through simulation, we generate M=900 sets of RSRP and SINR of user I in the network when various network failures occur, and define the set of each set of data as a network KPIs vector, where the m∈{1, 2, ..., 900}th network KPIs vector is x m =[RSRP1,SINR1,...,RSRP i , SINR i ,...,RSRP I , SINR I ], i∈{1, 2, ..., I}, where RSRP i and SINR i are the RSRP and SINR of the i-th user.

[0092] In addition, the simulation generates the RSRP and SINR of I users when the network is operating normally and groups each set of data into a set, where the m_Normal∈{1,...,100}th network KPIs vector is x m_Normal =[RSRP 1_Normal , SINR 1_Normal ,...,RSPR i_Normal , SINR i_Normal ,...,RSPR I_Normal , SINR I_Normal ], where RSPR i_Normal and SINR i_Normal are the RSRP and SINR of the i-th user when the network is operating normally.

[0093] Step 2: Calculate the similarity of the partitioned data and perform feature extraction to form a training set. This step includes the following process:

[0094] For any vector p=(p1,p2,...,p l ) T and q=(q1,q2,...,q l ) T , and its Mahalanobis distance is defined as follows:

[0095]

[0096] cov(p,q)=E(p T q)-E(p)E(q) (6)

[0097] where d M (·) is the Mahalanobis distance, E(·) and (·) T They represent expectation and transposition respectively, and cov(p,q) is the covariance of p and q.

[0098] Assume that the base station coverage area contains N section = 3 sectors, this method compares the user data between each sector and extracts the sector data most relevant to fault confirmation as the machine learning training sequence. At the same time, to address the challenge of user sparsity, each sector is evenly divided into partitions, and the features of users in the partition are extracted, so that the number of users required for feature extraction is smaller. The base station coverage area needs to be further divided into N setnum To extract features from partitions, each sector must be divided into at least two, namely N setnum ≥2·N section Considering that dividing too many regions will result in too many features to be extracted and the extraction efficiency is too low. setnum =2·N section =6. The specific division is as follows: In each sector of the base station, the bisector of the sector's central angle is used as the dividing line to evenly divide it into two partitions, so the total coverage area is divided into 6 partitions. In each partition, the Mahalanobis distance is used to calculate the similarity between the KPIs of the faulty network and the normal network. Divide I users into regions, there are I setnum users are distributed in the setnum∈{1,...,6} partition, whose set is According to UE setnum , for each eigenvector x m Divided into six partition feature vectors Calculate the Mahalanobis distance of all users in the six partitions and take the maximum value as x m Index of the partition in setnum

[0099]

[0100] The larger the partition index, the lower the similarity between the partition and the normal operation of the network, the more it can represent the characteristics of the fault, and the greater the impact on the fault confirmation result. In order to deal with the problem of uneven distribution of network users, only some users' RSRP and SINR are extracted from the partitions that can represent the fault characteristics. As the training sequence for subsequent machine learning, this method considers extracting data from 3 of the 6 partitions, which can reflect the fault characteristics without introducing too much redundant data. Take x in the 6 partitions m The maximum value of the Mahalanobis distance index (i.e. the lowest similarity) corresponds to the partition setnum_1; the second largest Mahalanobis distance index corresponds to the partition setnum_2, and the third largest Mahalanobis distance index corresponds to the partition setnum_3. In the user set UE corresponding to setnum_1 setnum_1 Randomly select N1=5 users to form a set The user set UE corresponding to setnum_2 setnum_2 Randomly select N2=2 users to form a set The user set UE corresponding to setnum_3 setnum_3 Randomly select N3=2 users to form a set The feature sequence after feature extraction is as follows:

[0101]

[0102] When constructing a training set for machine learning, x m_1 As the feature vector, and the fault type corresponding to the mth KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is

[0103] X MD ={(x 1_1 , actual1),(x 2_1 , actual1),...,(x M_1 , actual1)}

[0104] Among them actual m ∈{1,2,3} is the fault type corresponding to the mth KPIs vector. m = 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault.

[0105] Step 3: Principal component analysis feature extraction to form a training set. This step includes the following process:

[0106] Principal Component Analysis (PCA) is the data dimensionality reduction method most commonly used in conjunction with support vector machines.

[0107] Set all x m Composition matrix X:

[0108]

[0109] Decentralize the matrix X, that is, subtract the mean of each column from the mean of the column to perform zero mean processing.

[0110] Calculate the covariance matrix And perform eigendecomposition on C to obtain the eigenvalue λ of the covariance matrix c , and its corresponding eigenvector v c . The obtained λ c Arrange them in descending order and extract n c eigenvalues. c The value of follows the following principles: first determine a threshold thre λ ∈(0, 1), which reflects the proportion of the extracted features in the total features. Considering that we want to reduce the data dimension and retain the principal components, we choose thre λ =0.95. n λ is the number of eigenvalues ​​of C, n c The value is the smallest integer that satisfies (20):

[0111]

[0112] Combine its corresponding eigenvectors into a matrix The data after feature extraction is This enables principal component analysis feature extraction.

[0113] When constructing a training set for machine learning, divide X' into M x rows m_2 As the feature vector, and the fault type corresponding to the mth KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is

[0114] X PCA ={(x 1_2 , actual1),(x 2_2 , actual1),...,(x M_2 , actual1)}

[0115] Among them actual m ∈{1,2,3} is the fault type corresponding to the mth KPIs vector. m= 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault.

[0116] Step 4: Training of cost-sensitive support vector machine. This step includes the following process:

[0117] The cost-sensitive support vector machine used in this invention assigns different classification error costs to faults of different severity. Three fault severity levels are considered: severe, moderate, and minor. Manual judgment of fault severity can be performed based on the definition of raw network key indicators (i.e., key network indicators without feature extraction). The specific definitions are as follows:

[0118] A severe fault is defined as a fault that has a serious impact on network performance, specifically defined as a fault in which the SINR of more than 60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a moderate fault is defined as a fault that has a moderate negative impact on network performance, specifically defined as a fault in which, among the original network key indicators, the SINR of 40%-60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a minor fault is defined as a fault that has a slight negative impact on network performance, specifically defined as a fault in which the SINR of 20%-40% of users in the network coverage area is lower than 0dBm or the RSRP is lower than -110dBm.

[0119] According to the fault severity and fault type of the training sequence, the training sequence can be divided into nine categories, marked with k∈{1, 2, 3, 4, 5, 6, 7, 8, 9}. When k=1, it means actual m =1 and serious fault; when k=2, it means actual m =1 and it is a moderate fault; when k=3, it means actual m =1 and it is a minor fault; when k=4, it means actual m =2 and serious fault; when k=5, it means actual m =2 and it is a medium fault; when k=6, it means actual m =2 and it is a minor fault; when k=7, it means actual m =3 and serious fault; when k=8, it means actual m =2 and it is a medium fault; when k=9, it means actual m =2 and it is a minor fault. According to the fault severity and fault type, in the training process of the cost-sensitive support vector machine, the present invention can be resolved into a nine-classification problem, and the output result is class predicted∈{1, 2, ..., 9}, and this method does not care about the severity of the fault, so it is also necessary to map the nine-classification problem into a three-classification problem, that is, the final judgment is EU fault, ED fault, or ERP fault.

[0120] As mentioned above, this method requires the application of cost-sensitive support vector machines to multi-classification problems, while the basic support vector machine is a binary classification problem. Therefore, the multi-classification problem needs to be converted into multiple binary classification problems for solution. Various existing methods can achieve this goal, such as the one-to-rest strategy, which marks one category in the multi-classification problem as a positive example and the remaining labels are set as negative examples.

[0121] In the second and third steps, two different training sets are obtained, namely X MD and X PCA , respectively train the two cost-sensitive support vector machines. Due to their similarity, only X MD As an example, the detailed process is shown. PCA Just repeat the steps.

[0122] Convert the multi-classification problem into multiple binary classification problems. For the k-th classifier, let y m Represents x m_1 The binary classification label, if x m_1 Belongs to the kth category, then y m =1, otherwise y m = -1, N is the total sample space X MD The number of training samples belonging to the kth class in .

[0123] Based on the training set X MD The cost-sensitive support vector machine training is defined as follows: The normal vector and intercept of the hyperplane as the classification boundary in the k-th linear classifier are w k and b k , then the two-classification problem is expressed as follows:

[0124]

[0125] Among them, ||·|| 2 Indicates the modulus of the vector, ξ m =1-(w k ·x m_1 +b k ) represents the sample x m_1 The distance to the kth class boundary, ξ=[ξ1,ξ2,…,ξ N ],V k is the k-th soft margin penalty parameter, and the optimal value can be determined by the ten-fold cross validation method.

[0126] Through the Karush-Kuhn-Tucker condition and the Lagrange multiplier method, the above problem can be transformed into the following problem:

[0127]

[0128] where α m , β m is the Lagrange multiplier, α=[α1,α2,…,α N ], β=[β1, β2,...,β N ].

[0129] Use partial derivatives to solve the above optimization problem and calculate L(w k , b k ,ξ,α,β) respectively for w k , b k , and set the partial derivative of ξ to 0, we can get:

[0130]

[0131] Substituting (23) into (22), the multivariable optimization problem can be simplified to m The single variable optimization problem can be solved by various existing algorithms, such as Sequential minimal optimization method, and the optimal solution is α m * Substituting back into (23) we can get w k and b k The optimal solution, that is, w k * and b k * :

[0132]

[0133] Among them, x t_1 For the training set X MD There exists such a t * >0, y t is x t_1 The corresponding label.

[0134] In the cost-sensitive support vector machine, the cost matrix Cost is introduced to give different costs to each sample misjudgment. This method needs to set different error classification costs for three fault samples with different severity levels, and define The cost of misjudging a serious failure; The misjudgment cost for a medium fault; is the misclassification cost for minor faults. Since greater fault severity has a greater negative impact on the network, a higher misclassification cost is required to more reasonably deploy cost-sensitive machine learning, i.e., c1 ≥ c2 ≥ c3.

[0135]

[0136] Where Cost(k, j) represents the cost of classifying the kth class into the jth class, k, j∈{1, 2, 3, 4, 5, 6, 7, 8, 9}. There are many existing hyperparameter optimization methods that can be used to optimize the parameters of c1, c2, and c3. For example, Bayesian optimization predicts the most likely optimal c1, c2, and c3 as the values ​​for the next calculation, and iterates N times. BY = 30 times, and finally find the optimal c1 * , c2 * , c3 * Substitute into (25) to obtain the optimal cost matrix Cost * , making the classifier performance optimal.

[0137] Find w k * 、b k * and Cost * Then, the cost-sensitive support vector machine is used to train the sample set X MD For each sequence x m_1 For classification, x m_1 Determined as class predicted ∈{1, 2, 3, 4, 5, 6, 7, 8, 9}, if and only if:

[0138]

[0139] Where P(predictedclass=j|x m_1 ) means when x m_1 When making classification judgments, the probability of judging as class j needs to be the output value w of the support vector machine j * ·x m_1 +b j * Mapping to probability values ​​can be achieved using existing methods, such as Platt Scaling. As mentioned earlier, this method does not care about the severity of the fault, so it is also necessary to map the nine-classification problem to a three-classification problem. If class predicted ∈{1, 2, 3}, then it is finally determined to be an EU fault, class predicted =1; if class predicted ∈{4, 5, 6}, then it is finally determined to be an ED fault, classpredicted =2; if class predicted ∈{7, 8, 9}, then it is finally determined to be an ERP failure, class predicted =3.

[0140] This method uses F actual -score as a performance indicator, F actual The larger the -score, the better the system performance. It is defined as follows:

[0141]

[0142]

[0143]

[0144] where n actual，predicted represents the number of samples that the cost-sensitive support vector machine classifies from actual∈{1, 2, 3} to predicted∈{1, 2, 3}, n actual，actual Indicates the number of samples that the cost-sensitive support vector machine judges the actual class to be the actual class.

[0145] The specific steps are as follows:

[0146] (1) Initialize m=1,

[0147] (2) Use cost-sensitive support vector machine to train the sequence x in the sample set X1 m_1 Make classification decisions.

[0148] (3) If class predicted =actual m , otherwise, m=m+1.

[0149] (4) If m≤M, repeat (2) and (3) to complete the classification decision for all training sequences. Otherwise, calculate the performance index F of the classifier under the current c1, c2, and c3. actual -score, and use Bayesian optimization to predict the optimal value c1 of c1, c2, and c3 * , c2 * , c3 * , update, n BY =n BY +1, c1=c1 * , c2=c2 * , c3=c3 * .

[0150] (5) Reinitialization

[0151] (6) If n BY ≤30, repeat (2), (3), (4), (5), otherwise, find F in all combinations of c1, c2, c3 actual -score, and applies its corresponding c1, c2, c3 values ​​to the cost matrix Cost.

[0152] Training set X MD The corresponding cost-sensitive support vector machine obtains the optimal solution of c1, c2, c3 as follows: c1 * =1.7178, c2 * =4.3065, c3 * =10.95.

[0153] After the above steps, we can get the training set X MD The corresponding cost-sensitive support vector machine, X MD Change to X PCA , repeat the above steps to obtain the training set X PCA The corresponding cost-sensitive support vector machine.

[0154] Training set X PCA The corresponding cost-sensitive support vector machine obtains the optimal solution of c1, c2, c3 as follows: c1 * =3.34, c2 * =6.14,,c3 * =10.753.

[0155] Step 5: Classification decision based on ensemble learning

[0156] Training set X MD There is random sampling in the generation process, so the stability is low, but because the number of users in the area is low, the performance is good when facing uneven distribution of network users. PCA There is no random sampling in the classification process, so the stability is high, but a large number of users are required. When faced with an uneven distribution of network users, the performance is poor. Therefore, ensemble learning is used to obtain a stable and efficient classifier. The specific steps are as follows:

[0157] (1) Collect key network indicators (KPIsx) of the network to be confirmed as faulty new =[RSRP 1_new , SINR 1_new ,...,RSRP i_net , SINR i_new ,...,RSRP I_new , SINR I_new], i∈{1, 2, ..., I}. Among them, RSRP i_new and SINR i_new are the RSRP and SINR of the i-th user in the network to be confirmed for fault.

[0158] (2) x new Substitute into the second and third steps respectively, perform feature extraction and obtain x new_1 、x new_2 .

[0159] (3) Substitute the cost-sensitive support vector machine solved in step 4 and output P(predictedclass=j|x new_1 ) and P(predictedclass=j|x new_2 ).

[0160] (4) Using ensemble learning, new Perform classification judgment, x m_1 Determined as class predicted_new ∈{1, 2, 3, 4, 5, 6, 7, 8, 9}, if and only if:

[0161]

[0162] If class predicted_new ∈{1, 2, 3}, it is finally determined to be an EU fault; if class predicted_new ∈{4, 5, 6}, it is finally determined to be an ED fault; if class predicted_new ∈{7, 8, 9}, it is finally determined to be an ERP failure.

[0163] The present invention provides a fault confirmation method for wireless communication systems with uneven user distribution. This method uses the similarity between the KPIs of a faulty network and a properly functioning network to partition the network. Only the key performance indicators (KPIs) most relevant to network fault confirmation within each partition are extracted as training sequences for subsequent machine learning. This reduces the method's requirement for the number of users in each partition and mitigates the adverse effects of uneven user distribution on wireless communication system fault confirmation. Because the feature extraction process based on data similarity involves random sampling, stability is low. However, due to the low requirement for the number of users within a region, performance is good when faced with uneven network user distribution. The commonly used PCA feature extraction process does not involve random sampling, resulting in high stability but requiring a large number of users, resulting in poor performance when faced with uneven network user distribution. Training sequences are obtained for two feature extraction methods, which are then trained separately and integrated to produce a fault root cause classifier with stable output and accurate judgment. In terms of machine learning method selection, a cost-sensitive support vector machine (CSVM) is chosen, enabling the algorithm to more effectively focus on classifying fault root causes of varying severity, further improving the accuracy of network fault confirmation.

Claims

1. A fault confirmation method for a user-sparse communication system, characterized in that: The method first divides the base station coverage area into six evenly divided areas, and uses Mahalanobis distance (MD) to measure the data similarity of key performance indicators between the faulty network and the normal network in each area. The larger the Mahalanobis distance, the smaller the data similarity. The network key performance indicators with the smallest similarity in each partition are extracted as training sequences for machine learning to perform feature extraction. Then, principal component analysis is used to extract features from the original faulty network key performance indicators. Then, the sequences after feature extraction by the two different methods and their corresponding labels are combined into two training sets, and machine learning training is performed separately. In terms of machine learning methods, a cost-sensitive support vector machine is selected, which assigns different error root cause classification costs to fault samples with different severities. The higher the fault severity, the greater the cost of confirming the classification error, which makes it more reasonable to focus on the classification of fault root causes of different severities. Finally, fault confirmation is completed through ensemble learning of two support vector machines, and a fault root cause classifier with stable output and accurate judgment is obtained.

2. A fault confirmation method for a user-sparse communication system according to claim 1, characterized in that: The method specifically comprises the following steps: Step 1: Generate network KPIs, The reference signal received power RSRP and signal interference plus noise ratio SINR of network I user are generated by simulation when M groups of network failures occur, and the set of each group of data is defined as the network KPIs vector, where the m∈{1,2,...,M}th network KPIs vector is x m =[RSRP1,SINR1,...,RSRP i ,SINR i ,...,RSRP I ,SINR I ], i∈{1, 2, ..., I}, where RSRP i and SINR i is the RSRP and SINR of the i-th user; another simulation generates the RSRP and SINR of the I-th user when the M_Normal group network is operating normally, and each group of data is formed into a set, where the m_Normal∈{1,...,M_Normal}-th network KPIs vector is RSRP i Normal and are the RSRP and SINR of the i-th user when the network is operating normally; Step 2: Calculate the similarity of partitioned data and perform feature extraction to form a training set, including: For any vector p=(p1,p2,...,p l ) T and q=(q1,q2,...,q l ) T , and its Mahalanobis distance is defined as follows: cov(p,q)=E(p T q)-E(p)E(q) where d M (·) is the Mahalanobis distance, E(·) and (·) T They represent expectation and transposition respectively, cov(p,q) is the covariance of p and q; suppose the base station coverage area contains N section sectors, compare the user data between sectors, and extract the sector data most relevant to fault confirmation as the machine learning training sequence; at the same time, to cope with the challenge of user sparsity, each sector is evenly divided into partitions, and the features of users in the partitions are extracted, so that the number of users required for feature extraction is smaller; the base station coverage area needs to be divided into N setnum Partition and extract features, each sector needs to be divided into at least two, that is, N setnum 32·N section , considering that dividing too many regions will lead to too many features being extracted and the extraction efficiency is too low, we take N setnum =2·N section ; When constructing a training set for machine learning, x m_1 As the feature vector, and the fault type corresponding to the m-th KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is X MD ={(x 1_1 ,actual1),(x 2_1 ,actual1),...,(x M_1 ,actual1)} Among them actual m ∈{1,2,3} is the fault type corresponding to the mth KPIs vector. m = 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault; Network fault types include: antenna tilt angle is too small EU, antenna tilt angle is too large ED, and excessive power reduction ERP; Step 3: Principal component analysis feature extraction to form a training set, including the following process: Principal component analysis (PCA) is the most commonly used data dimensionality reduction method in conjunction with support vector machines. Set all x m Composition matrix X: Decentralize the matrix X, that is, subtract the mean of each column from the mean of the column to perform zero mean processing; Calculate the covariance matrix R is a set of real numbers, and C is decomposed to obtain the eigenvalue λ of the covariance matrix c , and its corresponding eigenvector v c ; The obtained λ c Arrange them in descending order and extract n c eigenvalues; for n c The value of follows the following principles: first determine a threshold thre l ∈(0,1), which reflects the proportion of the extracted features in the total features. Considering that we want to reduce the data dimension and retain the principal components, we choose thre l =0.95, n λ is the number of eigenvalues ​​of C, n c The value is the smallest integer that satisfies the following formula: Combine the eigenvectors corresponding to the extracted eigenvalues ​​into a matrix The data after feature extraction is Thus, principal component analysis feature extraction is achieved; When constructing a training set for machine learning, divide X' into M x rows m_2 As the feature vector, and the fault type corresponding to the m-th KPIs vector as the label, a set of training samples is formed. The total sample space of the training set is X PCA ={(x 1_2 ,actual1),(x 2_2 ,actual1),...,(x M_2 ,actual1)} Among them actual m ∈{1,2,3} is the fault type corresponding to the mth KPIs vector. m = 1, the training sample is EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault; Step 4: Training of cost-sensitive support vector machine, including the following process: The cost-sensitive support vector machine used assigns different classification error costs to faults of different severity, considering three fault severity levels: severe faults, moderate faults, and minor faults. Humans can manually judge the severity of faults based on the definition of raw network key indicators, i.e., key network indicators without feature extraction. In the second and third steps, two different training sets are obtained, namely X MD and X PCA , respectively train two cost-sensitive support vector machines. Due to their similarity, X PCA Just repeat X MD step; In the cost-sensitive support vector machine, the cost matrix Cost is introduced to give different costs to each sample misjudgment; different error classification costs are set for three fault samples with different severity levels, and c1∈R is defined as the error classification cost of severe faults; c2∈R is the error classification cost of moderate faults; and c3∈R is the error classification cost of minor faults. Since the greater the severity of the fault, the greater the negative impact on the network, a larger error classification cost is required to more reasonably deploy cost-sensitive machine learning, that is, c13c23c3 Where Cost(k,j) represents the cost of classifying the kth class into the jth class, k,j∈{1,2,3,4,5,6,7,8,9}; a variety of existing hyperparameter optimization methods can be used to optimize the parameters of c1,c2,c3. Bayesian optimization predicts the most likely optimal c1,c2,c3 as the value of the next calculation, and iterates N times. BY times, and finally find the optimal c1 * ,c2 * ,c3 * ; Usually the default iteration is 30 times, which does not take too much time and can improve the optimization effect, that is, N BY =30; c1 * ,c2 * ,c3 * Substitute the above formula to obtain the optimal cost matrix Cost * , making the classifier performance optimal; Find w k * 、b k * and Cost * Then, the cost-sensitive support vector machine is used to train the sample set X MD For each sequence x m_1 For classification, x m_1 Determined as class predicted ∈{1,2,3,4,5,6,7,8,9}, if and only if: Where P(predictedclass=j|x m_1 ) means when x m_1 When making classification judgments, the probability of judging as class j needs to be the output value w of the support vector machine j * ·x m_1 +b j * Mapping to probability value; mapping the nine-category problem to a three-category problem, if class predicted ∈{1,2,3}, then it is finally determined to be an EU fault, class predicted =1; if class predicted ∈{4,5,6}, then it is finally determined to be an ED fault, class predicted =2; if class predicted ∈{7,8,9}, then it is finally determined to be an ERP failure, class predicted =3, Use F actual -score as a performance indicator, F actual The larger the -score, the better the system performance. It is defined as follows: where n actual,predicted represents the number of samples that the cost-sensitive support vector machine classifies from actual∈{1,2,3} to predicted∈{1,2,3}, n actual,actual Indicates the number of samples that the cost-sensitive support vector machine judges the actual class to be the actual class; Step 5: Classification decision based on ensemble learning Training set X MD There is random sampling in the generation process, so the stability is low, but because the number of users in the area is low, the performance is good when facing uneven distribution of network users; training set X PCA There is no random sampling in the process, so the stability is high, but a large number of users are required. When faced with an uneven distribution of network users, the performance is poor. Therefore, ensemble learning is used to obtain a stable and efficient classifier. The specific steps are as follows: (1) Collect the key network indicators (KPIs) x of the network to be confirmed as faulty. new =[RSRP 1_new ,SINR 1_new ,...,RSRP i_new ,SINR i_new ,...,RSRP I_new ,SINR I_new ],i∈{1,2,...,I}; where RSRP i_new and SINR i_new are the RSRP and SINR of the i-th user in the network to be confirmed for fault; (2) Change x new Substitute into the second and third steps respectively, perform feature extraction and obtain x new_1 、x new_2 ; (3) Substitute the cost-sensitive support vector machine solved in step 4 and output P(predictedclass=j|x new_1 ) and P(predictedclass=j|x new_2 ); (4). Using ensemble learning, new Perform classification judgment, x m_1 Determined as class predicted_new ∈{1,2,3,4,5,6,7,8,9}, if and only if: If class predicted_new ∈{1,2,3}, it is finally determined to be an EU fault; if class predicted_new ∈{4,5,6}, it is finally determined to be an ED fault; if class predicted_new ∈{7,8,9}, it is finally determined to be an ERP failure.

3. A fault confirmation method for a user-sparse communication system according to claim 2, characterized in that: The sectors are specifically divided as follows: In each sector of the base station, the bisector of the central angle of the sector is used as the dividing line to evenly divide it into two partitions, and the total coverage area is divided into N setnum Partitions, in each partition, use Mahalanobis distance to calculate the similarity between the KPIs of the faulty network and the normal network, and divide I users into regions, where there are I setnum The users are distributed in the setnum∈{1,...,N setnum } partition, whose set is According to UE setnum , for each eigenvector x m Divided into N setnum Partition feature vector The same division is also performed on the network KPIs when the network is operating normally, and the eigenvectors for each feature vector are obtained. Divided into N setnum Partition feature vector Calculate the Mahalanobis distance between the partition feature vectors of all M_Normal groups during normal operation and the sample partition feature vectors, and sum them as x m Index of the partition in setnum 4. A fault confirmation method for a user-sparse communication system according to claim 3, characterized in that: The larger the index of the partition, the lower the similarity between the partition and the normal operation of the network, the more it can represent the characteristics of the fault, and the greater the impact on the fault confirmation result. In order to deal with the problem of uneven distribution of network users, only some users' RSRP and SINR are extracted from some areas that can represent the fault characteristics as the training sequence for subsequent machine learning. Consider extracting N setnum The data of three partitions in the partitions can reflect the fault characteristics without introducing too much redundant data; take N setnum x in each partition m The maximum value of the Mahalanobis distance index is the lowest similarity, and the corresponding partition is setnum_1; the second largest Mahalanobis distance index corresponds to the partition setnum_2, and the third largest Mahalanobis distance index corresponds to the partition setnum_3; in the user set UE corresponding to setnum_1 setnum_1 Randomly select N1 users to form a set Because users in the same area are close to each other, their RSRP and SINR have statistical similarities. No more than 8 features and no less than 2 features are extracted, that is, 2≤N1<8; in the user set UE corresponding to setnum_2 setnum_2 Randomly select N2 users to form a set Because the similarity of the user set partition is higher than setnum_1, the higher the similarity, the smaller the impact of the partition on the confirmation result, and the number of extracted users should be less than or equal to N1, that is, 2≤N2≤N1; in the user set UE corresponding to setnum_3 setnum_3 Randomly select N3 users to form a set The same principle as N2, N3 should satisfy 2≤N3≤N2. The feature sequence after feature extraction is as follows:

5. A fault confirmation method for a user-sparse communication system according to claim 2, characterized in that: The fault severity is defined as follows: A severe fault is defined as a fault that has a serious impact on network performance, specifically defined as a fault in which the SINR of more than 60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a moderate fault is defined as a fault that has a moderate negative impact on network performance, specifically defined as a fault in which, among the original network key indicators, the SINR of 40%-60% of users in the network coverage area is lower than 0dB or the RSRP is lower than -110dBm; a minor fault is defined as a fault that has a slight negative impact on network performance, specifically defined as a fault in which the SINR of 20%-40% of users in the network coverage area is lower than 0dBm or the RSRP is lower than -110dBm.

6. A fault confirmation method for a user-sparse communication system according to claim 2, characterized in that: According to the fault severity and fault type of the training sequence, the training sequence is divided into nine categories and marked with k∈{1,2,3,4,5,6,7,8,9}. When k=1, it means actual m =1 and serious fault; when k=2, it means actual m =1 and it is a moderate fault; when k=3, it means actual m =1 and it is a minor fault; when k=4, it means actual m =2 and serious fault; when k=5, it means actual m =2 and it is a medium fault; when k=6, it means actual m =2 and it is a minor fault; when k=7, it means actual m =3 and serious fault; when k=8, it means actual m =2 and it is a medium fault; when k=9, it means actual m =2 and it is a minor fault.

7. A fault confirmation method for a user-sparse communication system according to claim 2, characterized in that: In the fourth step, two cost-sensitive support vector machines X are respectively MD and X PCA Conduct training, Convert the multi-classification problem into multiple binary classification problems. For the k-th classifier, let y m Represents x m_1 The binary classification label, if x m_1 Belongs to the kth category, then y m =1, otherwise y m = -1, N is the total sample space X MD The number of training samples belonging to the kth class in ; Based on the training set X MD The cost-sensitive support vector machine training is defined as the hyperplane normal vector and intercept of the k-th linear classifier as the classification boundary as w k and b k , then the two-classification problem is expressed as follows: st y m (w k ·x m_1 +b k )≥1-ξ m Among them, ||·|| 2 Indicates the modulus of the vector, ξ m =1-(w k ·x m_1 +b k ) represents the sample x m_1 The distance to the kth class boundary, ξ=[ξ1,ξ2,…,ξ N ],V k is the k-th soft margin penalty parameter, and the optimal value is determined by the ten-fold cross validation method; Through the Karush–Kuhn–Tucker condition and the Lagrange multiplier method, the above problem is transformed into the following problem: where α m , b m is the Lagrange multiplier, α=[α1,α2,…,α N ], β=[β1, β2,...,β N ]; Use partial derivatives to solve the above optimization problem and calculate L(w k , b k ,ξ,α,β) respectively for w k , b k , and set the partial derivative of ξ to 0, we can get: Substituting Equation (8) into Equation (7), the multivariable optimization problem can be simplified to m The single variable optimization problem is solved by using the Sequential minimal optimization algorithm, and the optimal solution is α m * , substituting into formula (8) we can get w k and b k The optimal solution, that is, w k * and b k * : Among them, x t_1 For the training set X MD There exists such a t * >0 for a sample, y t is x t_1 The corresponding label.

8. A fault confirmation method for a user-sparse communication system according to claim 2, characterized in that: The X MD and X PCA The corresponding cost-sensitive support vector machine acquisition steps are as follows: Initialize m=1, n BY =1; (1) Use cost-sensitive support vector machine to train the sample set X MD mid-sequence x m_1 Make classification judgments; (2). If class predicted =actual m , otherwise, m=m+1; (3) If m≤M, repeat (2) and (3) to complete the classification decision for all training sequences. Otherwise, calculate the performance index F of the classifier under the current c1, c2, c3 actual -score, and use Bayesian optimization to predict the optimal value c1 of c1, c2, c3 * ,c2 * ,c3 * , update, n BY =n BY +1, c1=c1 * , c2=c2 * , c3=c3 * ; (4). Reinitialize (5). If n BY ≤N BY , repeat (2), (3), (4), (5), otherwise, find F in all combinations of c1, c2, c3 actual -score, and apply its corresponding c1, c2, c3 values ​​to the cost matrix Cost. After the above steps, we can get the training set X MD The corresponding cost-sensitive support vector machine, X MD Change to X PCA , repeat the above steps to obtain the training set X PCA The corresponding cost-sensitive support vector machine.

Images