Root cause analysis method for mobile base station complaints based on Bayesian network

Abstract

The application discloses a mobile base station complaint root cause analysis method based on a Bayesian network, takes operation and maintenance variables and user complaint variables as nodes, constructs a Bayesian network according to association rules describing business knowledge, then generates embedded vectors of each variable in the Bayesian network, carries out probability reasoning based on the variable embedded vectors, obtains conditional probabilities of each variable occurring failure, finally calculates contribution degrees of each operation and maintenance variable to the failure, and takes operation and maintenance variables with larger contribution degrees as a mobile base station complaint root cause set. The application takes operation and maintenance variables and user complaint variables as nodes to construct a Bayesian network, infers contribution degrees of each operation and maintenance variable to the failure based on the Bayesian network, and further obtains a mobile base station complaint root cause set with high accuracy and good interpretability.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile base stations, and more specifically, relates to a mobile base station complaint root cause analysis method based on a Bayesian network. Background Art

[0002] The root cause analysis of mobile base station complaints aims to discover the causes of user complaints, locate and solve problems in mobile base station operation and maintenance, and provide targeted traceability results for mobile base station operation and maintenance personnel. However, base station operation and maintenance data involves many variables (called operation and maintenance variables), and there are dependencies between variables. Only a few variables are the root cause of the fault. Traditional base station fault location mainly relies on the work experience and professional knowledge of operation and maintenance personnel. When faced with massive and complex operation and maintenance data, its accuracy, operation and maintenance efficiency and labor costs can no longer meet the needs of the rapid advancement of new infrastructure, nor can it meet the increasingly urgent needs of people for accurate diagnosis and rapid fault handling of base station faults. How to use computer technology, data analysis and knowledge reasoning methods to perform root cause analysis on large-scale base station operation and maintenance data with dependencies between variables has become an urgent problem that needs to be solved. It has important research value and practical significance.

[0003] Among the well-known root cause analysis methods, the most representative one is the neural network method. However, this method does not take into account the dependencies between operation and maintenance variables, which makes the root cause analysis results lack robustness. For example, Liu Fagui et al. (<Patent 202210077012.9>, 2022) preprocessed and divided the time series indicator data in the cloud server, trained an unsupervised anomaly detection model on the divided data set, performed anomaly detection, and then obtained the anomaly probability, determined the abnormal data points, marked the abnormal intervals, and calculated the weights of multiple variables in the abnormal intervals through feature selection methods. According to the weight size, the indicators affecting the anomaly were obtained. Wu Meng et al. (<Patent 202111440027.9>, 2021) extracted features from the detection data, input the features into supervised anomaly detectors, semi-supervised anomaly detectors, and unsupervised anomaly detectors. After the equipment anomaly was found, the integrated anomaly detector was combined with the root cause analysis to provide a method for locating abnormal data and obtain root cause analysis results, providing a basis for decision-making and management.

[0004] A probabilistic graphical model is a graphical model that can express uncertain dependencies. A Bayesian Network (BN), an important probabilistic graphical model that supports uncertain knowledge representation and reasoning, is a directed acyclic graph (DAG) with random variables as nodes, and each node has a conditional probability table (CPT) that can quantitatively describe the degree of mutual influence between variables. Intuitively, the interdependencies between operation and maintenance variables can be qualitatively and quantitatively described using BN, and the existing dependencies can be inferred and analyzed using probabilistic reasoning algorithms. Well-known BN-based operation and maintenance data mining and root cause analysis methods can well express the dependencies between operation and maintenance variables and their uncertainties. For example, Xiang Chaodeng (<Patent 202011435334.3>, 2020) uses alarm data to construct a training set and combines it with the Spark engine to construct an alarm BN. Based on this, they identify the correlation between real-time alarms to locate the root cause of the alarm. Yang Rui et al. (Information Technology, 2022) used relational modeling to generate a BN for industrial alarm data. They then used a forward-iterative path inference algorithm based on abnormal alarm information to infer a set of abnormal paths. This set of abnormal paths, combined with a set of evidence, was then used to infer the root cause and path of the abnormality. However, when dealing with large-scale operational and maintenance variable sets, these known methods suffer from repeated calculations and low efficiency.

[0005] Mobile base station operation and maintenance variables not only have interdependencies but also potential causal relationships. Compared to dependency, causality strictly distinguishes between cause and effect variables, further deepening the dependency relationship. However, the commonly used BN root cause analysis method for operation and maintenance data does not consider the potential causal relationships between variables, resulting in a lack of accuracy and interpretability in the root cause analysis results. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a mobile base station complaint root cause analysis method based on Bayesian network. Operation and maintenance variables and user complaint variables are used as nodes to construct a Bayesian network. Based on the Bayesian network, the contribution of each operation and maintenance variable to the fault is inferred, and then a set of mobile base station complaint root causes with high accuracy and good interpretability is obtained.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides a method for analyzing the root cause of mobile base station complaints based on a Bayesian network, comprising the following steps:

[0008] S1: Construct a Bayesian network based on the association rules that describe business knowledge. The specific method is as follows:

[0009] S1.1: Based on the current user complaint information, locate the mobile base station corresponding to the user's area and obtain N historical complaint root cause analysis records of the mobile base station with the same problem as the current user's complaint. The historical complaint root cause analysis records include the complaint problem, the abnormal operation and maintenance variables and values ​​of the mobile base station in the area where the complaint occurred; based on the historical complaint root cause analysis records, establish a variable set Q = (x1,…,x i ,…,x N ,x N+1 ), where x n Represents the nth operation and maintenance variable, and the number of possible values ​​corresponding to it is recorded as M n , n=1,2,…,N,x N+1 Represents the user complaint variable, which has the value M N+1 =2, respectively 0 and 1, x N+1 =1 indicates user complaints, x N+1 =0 means the user has not complained;

[0010] S1.2: Use the following formula to calculate any two variables x in the variable set Q: i and x j The support between sup(x i ,x j ):

[0011]

[0012] Where i, j = 1, 2, ..., N and i ≠ j, #(x i ,x j ) represents the variable x i and x j The number of times they appear together in historical complaint root cause analysis records;

[0013] Set the minimum support sup min ,0≤sup min ≤1, find all the values ​​that satisfy sup(x i ,x j )≥sup min The variable pairs are recorded as W, and the frequent binomial set V=(v1,…,v w ,…,v W ), where v w Indicates the w-th variable pair selected based on support Represent the two variables in the w-th variable pair, w=1,2,…,W,1≤i w ,j w ≤N and i w ≠j w ;

[0014] S1.3: Set the minimum confidence level c min , 0≤c min ≤1, the following formula is used to calculate the frequent binomial set V for each variable pair Confidence

[0015]

[0016] in, Representing variables and The number of times it appears simultaneously in historical complaint root cause analysis records, Represents variables in historical complaint root cause analysis records Number of occurrences;

[0017] Then find all the Variable pairs, and arrange the selected variable pairs from large to small according to the confidence level. The number of selected variable pairs is T, and the tth variable pair is Represent the two variables in the t-th variable pair, r=1,2,…,T,1≤i t ,j t ≤N and i t ≠j t ; According to the variable Generate strong association rules for variables Construct variable strong association rule set R=(r1,…,r t ,…,r T ), where r t Represents strong association rules of variables

[0018] S1.4: Generate a directed acyclic graph of the Bayesian network based on the strong association rules in the variable strong association rule set R. The specific steps are: use the variables in the variable set Q as nodes of the Bayesian network, and take out each variable strong association rule in the variable strong association rule set R in turn. If the variables in the Bayesian network and If there is no directed edge between them, a variable is generated. Pointer to variable ’s directed edges, thus obtaining the directed acyclic graph of the Bayesian network;

[0019] Then, based on the directed acyclic graph of the Bayesian network and the mobile base station operation and maintenance data, the maximum likelihood estimation is used to obtain the value of each variable x in the directed acyclic graph. i The conditional probability parameter The calculation formula is as follows:

[0020]

[0021] Among them, q i,h is the variable x i The parent node set π(x i ) the hth value combination of the nodes, h=1,2,…,H i , H i For the parent node set π(x i ), is the operation and maintenance variable x i mth i possible values, m i =0,1,…,M i -1, #() is the number of operation and maintenance data records that meet the search conditions;

[0022] S2: Generate the embedding vectors of each variable in the Bayesian network. The specific method is as follows:

[0023] S2.1: Perform S sampling on the variable set Q of the directed acyclic graph to generate S samples, forming a sample set D. c =(D c,1 ,…,D c,s ,…,D c,S ), D c,s Denotes the sth sample, s = 1, 2, ..., S. The specific process of generating samples each time is as follows:

[0024] With the variable set Q=(x1,…,x i ,…,x N ,x N+1 ) is sampled one by one in the order of variables as the sampling order. According to the forward sampling algorithm, for the variable x i If it is the root node in the Bayesian network directed acyclic graph, the sampling value is P(x i ) is sampled with probability; if the variable x i If it is not the root node, the sampling value is P(x i |π(x i )=q i ) is sampled with probability, and variable x is recorded i The value of The probability of Generate a random number ram, 0<ram<1, if where p i,-1 =0, then the variable x i Assign value to mth i possible values;

[0025] S2.2: Based on the sample set D c , the following formula is used to calculate the point mutual information value of different operation and maintenance variables for different value combinations

[0026]

[0027] in, Represents the variable x in the sample set D i The value of The number of samples, Represents the variable x in the sample set D j The value of The number of samples, m j =0,1,…,M j -1,M j Represents the variable x j The number of representable values ​​of Indicates that the sample set D satisfies the variable x at the same time i The value of And the operation variable x j The value of The number of samples;

[0028] Then according to Construct a point mutual information matrix W of size G×G, where

[0029] S2.3: Calculate the variable correlation matrix M using the following formula OBN :

[0030] M OBN =W-logβ

[0031] Where β represents the number of non-adjacent node pairs in the Bayesian network;

[0032] S2.4: Variable correlation matrix M OBN Low-rank decomposition is used for dimensionality reduction, and the resulting matrix of size G×d is used as the embedding matrix Y, where each row is the embedding vector of the corresponding variable value;

[0033] S3: Perform probabilistic reasoning based on the variable embedding vector to obtain the conditional probability of each variable failing. The specific method is as follows:

[0034] S3.1: According to the embedding matrix Y, get the variable x i The value of The embedding vector and the variable x j The value of The embedding vector Calculate the embedding vector and embedding vector The similarity between

[0035] S3.2: Compute the similarity between the embedding vectors of a variable x and its neighbors i The value is Probability The calculation method is as follows:

[0036] Note the variable x i The set of neighbor variables is Ne i , for the neighbor variable set Ne i For each variable x j′ , remember the embedding vector and embedding vector The similarity between x j′ ∈Ne i , m j′ =0,1,…,M j′ -1,M j′ Represents the variable x j′ The number of possible values ​​of ; then calculate the embedding vector and the neighbor variable set Ne i All variables x j′ The average value of embedding vector similarity at different values Then use the following formula to calculate the variable x i The value is Probability

[0037]

[0038] S3.3: User complaint variable x N+1 As an evidence node, traverse each operation and maintenance variable x n , perform Bayesian network approximate reasoning according to the following steps to calculate the operation and maintenance variable x n In different values The conditional probability of failure The specific steps are as follows:

[0039] (1) Initialize each operation and maintenance variable x n The value is Number

[0040] (2) Let the user complain about variable x N+1 The value of is 1, and then the value of each operation and maintenance variable in the variable set Q is randomly generated, and the operation and maintenance reasoning sample D is initialized. o,1 , let the operation and maintenance reasoning sample number z = 1;

[0041] (3) For each operation and maintenance variable x n , if the operation and maintenance reasoning sample D o,zMid-O&M variables Then order Otherwise, no action will be taken;

[0042] (4) Based on operation and maintenance reasoning sample D o,z Generate new operation and maintenance reasoning sample D o,z+1 The specific method is: resample the operation and maintenance variables according to the sequence number of each operation and maintenance variable in the variable set Q, and for the operation and maintenance variable x n Generate random number ram′ n ,if Where P(x n =k n,-1 )=0, then the operation and maintenance variable x n The value of is updated to its mth n possible values;

[0043] (5) Determine whether z < Z, where Z is the preset number of operation and maintenance reasoning samples. If so, set z = z + 1 and return to step (3); otherwise, proceed to step (6);

[0044] (6) Based on the generated Z operation and maintenance reasoning samples D o,z , the following formula is used to calculate the conditional probability of each operation and maintenance variable failing when a known base station fails under user complaints:

[0045]

[0046] S4: Based on the conditional probability of failure of the operation and maintenance variable under the user complaint situation, the operation and maintenance variable x is calculated according to the following formula n The value is and The absolute value of the conditional probability difference under

[0047]

[0048] Among them, m n ,m′ n =0,1,…,M n -1,m n ≠m′ n ;

[0049] The operation and maintenance variable x is calculated using the following formula n In the value Contribution to base station failure

[0050]

[0051] Among them, vd max (x nrepresents an operation and maintenance variable x n The maximum value in the conditional probability difference of the value pair, represents rounding up;

[0052] The contribution degree db(x n ) of the operation and maintenance variable x n ) is calculated by the following formula:

[0053]

[0054] The operation and maintenance variables are sorted according to the contribution degree from large to small, and the first γ operation and maintenance variables constitute the mobile base station complaint root cause set. The size of γ is set according to actual needs.

[0055] In order to achieve the above-mentioned purposes, the mobile base station complaint root cause analysis method based on the Bayesian network of the present application takes the operation and maintenance variable and the user complaint variable as a node, constructs the Bayesian network according to the association rule describing the business knowledge, then generates the embedding vector of each variable in the Bayesian network, carries out the probability reasoning based on the variable embedding vector, obtains the conditional probability of each variable failure, and finally calculates the contribution degree of each operation and maintenance variable to the failure, and takes the operation and maintenance variable with a larger contribution degree as the mobile base station complaint root cause set.

[0056] The present application has the following beneficial effects:

[0057] 1) The present application represents the operation and maintenance variable and the user complaint variable as a unified, low-dimensional and dense node embedding vector through the construction and embedding of the Bayesian network, fully mines the causal relationship between the operation and maintenance variables, and uses the approximate probability reasoning method based on the embedding vector to obtain the conditional probability of each operation and maintenance variable failure with the user complaint variable as the evidence node. Through the embedding of the Bayesian network, the problem of low reasoning efficiency caused by multiple repeated calculations can be solved, and the failure analysis efficiency is improved.

[0058] 2) The present application combines the association rule mining, maximum likelihood estimation method, graph embedding technology, Bayesian network reasoning algorithm and causal model, efficiently and accurately analyzes the causes of base station failure, and provides a scientific basis for the operation and maintenance personnel to eliminate base station failure. DETAILED DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is the specific implementation mode flowchart of the mobile base station complaint root cause analysis method based on the Bayesian network of the present application;

[0060] Figure 2 is the flowchart of constructing the Bayesian network in the present application;

[0061] Figure 3 is the flowchart of generating the embedding vector of the operation and maintenance variable in the Bayesian network in the present application;

[0062] Figure 4 It is a flowchart of probabilistic reasoning based on variable embedding vectors in the present invention;

[0063] Figure 5 Schematic diagram of the directed acyclic graph of the Bayesian network in this embodiment. DETAILED DESCRIPTION

[0064] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0065] Example

[0066] Figure 1 This is a flow chart of a specific implementation of the mobile base station complaint root cause analysis method based on the Bayesian network of the present invention. Figure 1 As shown, the specific steps of the mobile base station complaint root cause analysis method based on the Bayesian network of the present invention include:

[0067] S101: Bayesian network construction:

[0068] A directed acyclic graph of the Bayesian network is constructed based on the association rules that describe business knowledge. Based on the operation and maintenance data of mobile base stations, the maximum likelihood estimation (MLE) is used to learn the probability parameters of the Bayesian network. Figure 2 This is a flow chart of constructing the Bayesian network in the present invention. Figure 2 As shown, the specific steps of constructing the Bayesian network in the present invention include:

[0069] S201: Construct variable sets based on historical complaint root cause analysis records:

[0070] When a mobile network failure occurs, when a user files a complaint against the mobile network, corresponding user complaint information will be generated, including a description of the user's complaint problem and the location information of the user's area. Therefore, in the present invention, for the current user complaint information, the mobile base station corresponding to the user's area is located, and N historical complaint root cause analysis records of the mobile base station that are the same as the current user's complaint problem are obtained. The historical complaint root cause analysis records include the complaint problem, the abnormal operation and maintenance variables and values ​​of the mobile base station in the area where the complaint occurred. Based on the historical complaint root cause analysis records, a variable set Q = (x1,…,x i ,…,x N ,x N+1 ), where x n Represents the nth operation and maintenance variable, and the number of possible values ​​corresponding to it is recorded as M n, n=1,2,…,N,x N+1 Represents the user complaint variable, which has the value M N+1 =2, respectively 0 and 1, x N+1 =1 indicates user complaints, x N+1 =0 means the user has not filed a complaint.

[0071] S202: Screening frequent binomial sets:

[0072] Use the following formula to calculate any two variables x in the variable set Q i and x j The support between sup(x i ,x j ):

[0073]

[0074] Where i, j = 1, 2, ..., N and i ≠ j, #(x i ,x j ) represents the variable x i and x j The number of times it appears in the historical complaint root cause analysis records. Obviously, the support degree sup(x i ,x j ) reflects the relationship between any two variables x i and x j The probability of simultaneous occurrence in historical complaint root cause analysis records.

[0075] Set the minimum support sup min ,0≤sup min ≤1, find all the values ​​that satisfy sup(x i ,x j )≥sup min The variable pairs are recorded as W, and the frequent binomial set V=(v1,…,v w ,…,v W ), where v w Indicates the w-th variable pair selected based on support Represent the two variables in the w-th variable pair, w=1,2,…,W,1≤i w ,j w ≤N and i w ≠j w .

[0076] S203: Generate variable strong association rules:

[0077] Set the minimum confidence level c min , 0≤c min≤1, the following formula is used to calculate the frequent binomial set V for each variable pair Confidence

[0078]

[0079] in, Representing variables and The number of times it appears simultaneously in historical complaint root cause analysis records, Represents variables in historical complaint root cause analysis records Number of occurrences.

[0080] Then find all the Variable pairs, and arrange the selected variable pairs from large to small according to the confidence level. The number of selected variable pairs is T, and the tth variable pair is Represent the two variables in the t-th variable pair, r=1,2,…,T,1≤i t ,j t ≤N and i t ≠j t According to the variable Generate strong association rules for variables Construct variable strong association rule set R=(r1,…,r t ,…,r T ), where r t Represents strong association rules of variables Strong association rules for variables Used to describe the dependency relationship between variables.

[0081] S204: Constructing a Bayesian network:

[0082] Generate a directed acyclic graph of the Bayesian network based on the strong association rules in the variable strong association rule set R. The specific steps are: use the variables in the variable set Q as the nodes of the Bayesian network, and take out each variable strong association rule in the variable strong association rule set R in turn. If the variables in the Bayesian network and If there is no directed edge between them, a variable is generated. Pointer to variable , thus obtaining the directed acyclic graph of the Bayesian network.

[0083] Then, based on the directed acyclic graph of the Bayesian network and the mobile base station operation and maintenance data, the maximum likelihood estimation is used to obtain the value of each variable x in the directed acyclic graph. i The conditional probability parameter The calculation formula is as follows:

[0084]

[0085] Among them, q i,h is the variable x i The parent node set π(x i ) the hth value combination of the nodes, h=1,2,…,H i , H i For the parent node set π(x i ), is the operation and maintenance variable x i mth i possible values, m i =0,1,…,M i -1, #() is the number of operation and maintenance data records that meet the search conditions.

[0086] S102: Embedding processing of Bayesian network:

[0087] Point Mutual Information (PMI) is a correlation metric used in information theory and statistics, typically used to measure the correlation between two entities. In this paper, the dependencies between operational variables in a Bayesian network are transformed into a point mutual information matrix, which is then used to generate an operational variable association matrix. This matrix is ​​then reduced in dimension using singular value decomposition (SVD) to generate embedding vectors for each variable in the Bayesian network. This serves as the basis for subsequent probabilistic reasoning of operational variables and root cause analysis of mobile base station complaints. Figure 3 The flowchart of the generation of embedding vectors of operational variables in the Bayesian network of the present invention. Figure 3 As shown, the specific steps of generating the embedding vector of the operation and maintenance variables in the Bayesian network in the present invention include:

[0088] S301: Generate a sampling sample set:

[0089] According to the variable set Q of the directed acyclic graph, S samples are sampled to generate S samples, forming the sampling sample set D c =(D c,1 ,…,D c,s ,…,D c,S ), D c,s Denotes the sth sample, s = 1, 2, ..., S. The specific process of generating samples each time is as follows:

[0090] With the variable set Q=(x1,…,x i ,…,x N ,x N+1 ) is sampled one by one in the order of variables as the sampling order. According to the forward sampling algorithm, for the variable xi If it is the root node in the Bayesian network directed acyclic graph, the sampling value is P(x i ) is sampled with probability; if the variable x i If it is not the root node, the sampling value is P(x i |π(x i )=q i ) is sampled with probability, and variable x is recorded i The value of The probability of Obviously, P(x i ) and P(x i |π(x i )=q i ) can be directly accessed through the variable x i The conditional probability table is obtained. Generate a random number ram, 0<ram<1, if where p i,-1 =0, then the variable x i The value assigned is the mth possible value. After sampling the variables in sequence, a sample is generated.

[0091] In this embodiment, the number of sample samples S is determined by the following method: an initial value S0 of the number of sample samples is set, and then the number of sample samples is gradually increased. For each number of sample samples, each conditional probability parameter obtained according to the forward sampling algorithm is statistically analyzed. Corresponding conditional probability parameters in the conditional probability table based on the Bayesian network If the average value of the difference is less than a preset threshold (0.01 in this embodiment), the current number of sampling samples is used as the final number of sampling samples S.

[0092] S302: Constructing point mutual information matrix:

[0093] According to the sample set D c , the following formula is used to calculate the point mutual information value of different operation and maintenance variables for different value combinations

[0094]

[0095] in, Represents the variable x in the sample set D i The value of The number of samples, Represents the variable x in the sample set D j The value of The number of samples, m j =0,1,…,M j -1,M j Represents the variable xj The number of values ​​that can be represented. Indicates that the sample set D satisfies the variable x at the same time i The value of And the operation variable x j The value of The number of samples.

[0096] Then according to Construct a point mutual information matrix W of size G×G, where

[0097] S303: Calculate the operation and maintenance variable correlation matrix:

[0098] In order to make the interrelated operation and maintenance variables in OBN have similar embedding, the present invention further expresses the point mutual information matrix W as the variable association matrix M OBN , so that the positive correlation of the operation and maintenance variable pairs has a more similar embedding than the negative correlation. The variable correlation matrix M OBN The calculation formula is:

[0099] M OBN =W-logβ

[0100] Among them, β represents the number of non-adjacent node pairs in the Bayesian network (i.e., the number of non-neighbor node pairs)

[0101] S304: Get the embedding matrix:

[0102] Since the variable correlation matrix M OBN The associations between some operation and maintenance variables are redundant, so the variable association matrix M in the present invention is OBN Low-rank decomposition is used for dimensionality reduction, and the resulting matrix of size G×d is used as the embedding matrix Y, where each row is the embedding vector of the corresponding variable value.

[0103] In this embodiment, the operation and maintenance variable correlation matrix M is calculated based on the singular value decomposition method. OBN In the conventional singular value decomposition method, low-rank decomposition is achieved by minimizing the following loss function J:

[0104]

[0105] in, and M OBN The two low-dimensional orthogonal matrices obtained by singular value decomposition, (V d ) T Indicates V d The transposed matrix of is a diagonal matrix, Σ=(σ1,…,σ d ), ση are singular values, η = 1, 2, …, d, and σ1≥…≥σ d > 0, σ η The size represents the correlation between the value pairs of the variable pairs, d is the dimension of the low-rank space, and the subscript F represents the Euclidean norm (i.e., the distance norm between vectors).

[0106] In this embodiment, in order to improve the operation and maintenance variable correlation matrix M OBN The efficiency of low-rank vector decomposition, according to the matrix perturbation theory, the F-norm of matrix A can be expressed as Wherein, tr is the trace of matrix A T λ l is the eigenvalue of matrix A T A, D N is the dimension of matrix A T A. Therefore, the loss function J can be rewritten as:

[0107]

[0108] Wherein, λ g is the eigenvalue of matrix M OBN -U d Σ d (V d ) T .

[0109] Minimizing the above loss function J, the variable correlation matrix M OBN can be approximately represented by the first d-dimensional matrix of , and the singular value decomposition method is used to decompose as follows:

[0110]

[0111] Obtaining two low-dimensional matrices and and The low-dimensional matrix U is used as the embedding matrix Y of the variable correlation matrix M OBN .

[0112] S103: Perform probabilistic reasoning based on variable embedding vectors:

[0113] Next, the correlation between the variable embedding vectors obtained in step S102 is calculated, and the probability of different values of the variable is calculated accordingly, which is the basis for generating operation and maintenance reasoning samples. Further, using the user complaint variable as evidence, using the probabilistic reasoning method based on the variable embedding vector, the conditional probability of each variable failure is obtained, and the size of the conditional probability reflects the possibility of the operation and maintenance variable failure under the user complaint. Figure 4 is the flowchart of the probabilistic reasoning based on the variable embedding vector in the present application. As shown in Figure 4 As shown, the specific steps of performing probabilistic reasoning based on variable embedding vectors in the present invention include:

[0114] S401: Calculate the similarity of variable embedding vectors:

[0115] According to the embedding matrix Y, we get the variable x i The value of The embedding vector and the variable x j The value of The embedding vector Calculate the embedding vector and embedding vector The similarity between In this embodiment, the similarity between the embedding vectors adopts the Euclidean distance. Generally speaking, the larger the point mutual information value between variables, the shorter the distance between the variable embedding vectors, and the stronger the correlation between them.

[0116] S402: Calculate the probability of different values ​​of the variable:

[0117] variable x i The value is Probability Reflects the variable x i The value is In the present invention, the variable x is calculated based on the similarity between the embedding vectors of the variable and its neighbors. i The value is Probability The calculation method is as follows:

[0118] Note the variable x i The set of neighbor variables is Ne i , for the neighbor variable set Ne i For each variable x j′ , remember the embedding vector and embedding vector The similarity between x j′ ∈Ne i , m j′ =0,1,…,M j′ -1,M j′ Represents the variable x j′ The number of possible values ​​of . Then calculate the embedding vector and the neighbor variable set Ne i All variables x j′ The average value of embedding vector similarity at different values Then use the following formula to calculate the variable x i The value is Probability

[0119]

[0120] S403: Calculate the probability of failure of the operation and maintenance variable:

[0121] Take the user complaint variable x N+1 As an evidence node, traverse each operation and maintenance variable x n , perform Bayesian network approximate reasoning according to the following steps to calculate the operation and maintenance variable x n In different values The conditional probability of failure The specific steps are as follows:

[0122] (1) Initialize each operation and maintenance variable x n The value is Number

[0123] (2) Let the user complain about variable x N+1 The value of is 1, and then the value of each operation and maintenance variable in the variable set Q is randomly generated, and the operation and maintenance reasoning sample D is initialized. o,1 , let the operation and maintenance reasoning sample number z = 1.

[0124] (3) For each operation and maintenance variable x n , if the operation and maintenance reasoning sample D o,z Mid-O&M variables Then order Otherwise, no action is taken.

[0125] (4) Based on operation and maintenance reasoning sample D o,z Generate new operation and maintenance reasoning sample D o,z+1 The specific method is: resample the operation and maintenance variables according to the sequence number of each operation and maintenance variable in the variable set Q, and for the operation and maintenance variable x n Generate random number ram′ n ,if in , then the operation and maintenance variable x n The value of is updated to its mth n possible values.

[0126] (5) Determine whether z < Z, where Z is the preset number of operation and maintenance reasoning samples. If so, set z = z + 1 and return to step (3); otherwise, proceed to step (6).

[0127] (6) Based on the generated Z operation and maintenance reasoning samples D o,z , the following formula is used to calculate the conditional probability of each operation and maintenance variable failing when a known base station fails under user complaints:

[0128]

[0129] Among them, m n ,m′ n =0,1,…,M n -1,m n ≠m′ n .

[0130] Obviously, the conditional probability The larger the operation variable x is, the n The value is The greater the possibility of failure.

[0131] S104: Analysis of root causes of mobile base station complaints:

[0132] Based on the conditional probability of each operation and maintenance variable failing, the Blame algorithm in the causal model is used to calculate the contribution of each operation and maintenance variable to the failure, thereby obtaining the root cause set of mobile base station complaints. The Blame algorithm in the causal model can be used to analyze the causal relationship between operation and maintenance variables and calculate the contribution of the cause to the result. In the present invention, based on the conditional probability of each operation and maintenance variable failing under the user complaint scenario, the operation and maintenance variable x is calculated according to the following formula n The value is and The absolute value of the conditional probability difference under

[0133]

[0134] The operation and maintenance variable x is calculated using the following formula n In the value Contribution to base station failure

[0135]

[0136] Among them, vd max (x n ) represents the operation and maintenance variable x n The maximum value among the conditional probability differences of the value pairs, Indicates rounding up.

[0137] The operation and maintenance variable x is calculated using the following formula n Contribution db(x n ):

[0138]

[0139] From the above formula, we can see that the operation and maintenance variable x n Contribution db(x nrepresents the operation and maintenance variable x n Expectation under all different values, i.e. operation and maintenance variable x n The size of the impact of the fault occurrence, db(x n The greater the value is, the more likely it is the root cause variable of the mobile base station fault. Therefore, the operation and maintenance variables are sorted according to the contribution degree from large to small, and the first γ operation and maintenance variables constitute the mobile base station complaint root cause set, and the size of γ is set according to actual needs. It can be seen that the mobile base station complaint root cause set has good interpretability and can provide scientific basis for the operation and maintenance personnel to exclude the base station fault.

[0140] In order to better illustrate the technical effect of the present application, a specific example is used to experimentally verify the present application. In this experimental verification, the user complaint problem is "VoLTE voice call intermittent", and the operation and maintenance data of the mobile base station near the user is analyzed.

[0141] 1: Bayesian network construction:

[0142] According to the user complaint information, the user's area position is located, and the operation and maintenance data and historical complaint root cause analysis records of the mobile base station near the user are obtained (only the records of the "VoLTE voice call intermittent" related historical complaint root cause analysis are needed). Based on the historical complaint root cause analysis records, a variable set Q is established, wherein the user complaint variable is set as x N+1 . According to experience, the minimum support sup min is set to 0.3, and the minimum confidence c min is set to 0.5. The support between any two operation and maintenance variables is calculated, and all variable pairs satisfying the support greater than or equal to the minimum support sup min are found by searching to generate the frequent binomial set V={(“SRVCC handover success rate”, “SRVCC handover request times”), (“RRC link reestablishment request times”, “SRVCC handover request times”), …, (“RRC link reestablishment request times”, “seq high RTP uplink packet loss rate”), (“seq high RTP uplink packet loss”, “seq high RTP uplink packet loss rate”)}.

[0143] Then the confidence of all variable pairs in the frequent binomial set V is calculated, and all variable pairs satisfying the confidence greater than or equal to the minimum confidence c minVariable pairs are sorted from highest to lowest according to confidence, forming a variable strong association rule set R = {"SRVCC handover success rate" → "SRVCC handover request count", ..., "seq high RTP uplink packet loss rate" → "seq high RTP uplink packet loss rate", "RRC link reestablishment request count" → "seq high RTP uplink packet loss rate"}. The variables in the variable set Q are used as nodes in the Bayesian network. Strong association rules in the variable strong association rule set R are sequentially extracted based on confidence levels to generate directed edges between the variables in the Bayesian network, resulting in a directed acyclic graph of the Bayesian network. Figure 5 Schematic diagram of the directed acyclic graph of the Bayesian network in this embodiment.

[0144] Then, based on the operation and maintenance data of the mobile base station and the directed acyclic graph structure of the Bayesian network, the conditional probability table of the variables is obtained. The row title of each variable conditional probability table is the different values ​​corresponding to the variable, the column title is the different value combinations of the parent node, and the value in the table is the corresponding conditional probability distribution value. Table 1 is Figure 5 Conditional probability table of "eSRVCC handover request count" in the example shown. As shown in Table 1, the frequency of the operation and maintenance variable "SRVCC handover request count" in the root cause analysis depends on the frequency of the operation and maintenance variable "SRVCC handover success rate".

[0145]

[0146] Table 1

[0147] 2: Bayesian Network Embedding

[0148] Sampling is performed in the order of variables in the variable set Q. For example, the operation and maintenance variable "seq high load leads to low rate" is used as the initial node. From its conditional probability table, we know that P("seq high load leads to low rate" = "Normal") = 0.7 and P("seq high load leads to low rate" = "High") = 0.3. Generate a random number r = 0.5. Since P("seq high load leads to low rate" = "High") = 0.3 <r<(“seq高负荷导致低速率”=“Normal”)=0.7,则运维变量“seq高负荷导致低速率”值被赋值为“High”。同理,再采样运维变量“seq高RTP上行丢包率”,生成的随机数r=0.7,则运维变量“seq高RTP上行丢包率”的值被赋值为“Normal”。对变量集Q余下的变量采样并赋值,生成采样样本D c,1 In this embodiment, the number of sample sets is determined to be 3000 by calculating the conditional probability parameters of the variables, and finally the sample set D is obtained. c =(D c,1 ,D c,2 ,…,Dc,3000 Table 2 is a partial sample information table of the sample set in this embodiment.

[0149]

[0150] Table 2

[0151] Then construct the mutual information matrix in the sampling sample set D c The number of different values ​​of each operation and maintenance variable and the number of operation and maintenance variable pairs in the statistical variable set Q are used to calculate the PMI value of each operation and maintenance variable pair. For example, 36 samples meet the "seq high RTP uplink packet loss rate" as "High", 53 samples meet the "seq high load leads to low rate" as "High", and 25 samples meet both values ​​at the same time. Then multiplying by 3000 is approximately equal to 1.59

[0152]

[0153] In this embodiment, for convenience of representation, x is used. 1.0 Indicates the first value of the operation and maintenance variable "seq high load leads to low rate" is "Normal", x 1.1 Indicates the second value "High", and so on, the calculation is performed for 18 nodes. Table 3 is a partial data table of the point mutual information matrix obtained in this embodiment.

[0154] <![CDATA[k 1.0 ]]> <![CDATA[k 1.1 ]]> <![CDATA[k 2.0 ]]> <![CDATA[k 2.1 ]]> <![CDATA[k 2.2 ]]> ... k 17.1 ]]> <![CDATA[k 17.2 ]]> <![CDATA[k 18.0 ]]> k 18.1 ]]> <![CDATA[k1 .0 ]]> 0 0 3.35 2.54 0.73 ... 0 0 0 0 k 1.1 ]]> 0 0 1.34 -0.89 1.59 ... 0 0 0 0 <![CDATA[k 2.0 ]]> 0 0 0 0 0 ... 0 0 0 0 <![CDATA[k 2.1 ]]> 0 0 0 0 0 ... 0 0 0 0 k 2.2 ]]> 0 0 0 0 0 ... 0 0 0 0 ... ... ... ... ... ... ... ... ... ... ... <![CDATA[k 17.1 ]]> 0 0 0 0 0 ... 0 0 3.27 -0.36 k 17.2 ]]> 0 0 0 0 0 ... 0 0 1.75 -1.37 <![CDATA[k 18.0 ]]> 0 0 0 0 0 ... 0 0 0 0 <![CDATA[k 18.1 ]]> 0 0 0 0 0 ... 0 0 0 0

[0155] Table 3

[0156] The variable association matrix is ​​then further derived from the point mutual information matrix, and low-rank decomposition is performed on the variable association matrix to reduce its dimensionality and obtain an embedding matrix. This embodiment uses singular value decomposition to achieve low-rank decomposition, obtaining embedding vectors for 18 variables. Table 4 shows a partial data table of the embedding matrix of the Bayesian network in this embodiment.

[0157]

[0158]

[0159] Table 4

[0160] 3: Probabilistic reasoning based on variable embedding vectors:

[0161] First, we calculate the similarity of variables for different value pairs based on the embedding vectors in Table 4, and then calculate the probability of each variable for different values.

[0162] Set the user complaint variable x N+1 As evidence nodes, 5000 operation and maintenance reasoning samples are generated. Table 5 is a partial data table of the operation and maintenance reasoning samples in this embodiment.

[0163] x1 <![CDATA[x2]]> <![CDATA[x3]]> ... <![CDATA[x 17 ]]> x 18 ]]> 1 0 1 1 ... 2 0 2 0 0 1 ... 2 0 3 0 0 0 ... 0 0 ... ... ... ... ... ... ... 4999 1 0 2 ... 1 2 5000 1 0 2 ... 1 1

[0164] Table 5

[0165] Based on the operation and maintenance reasoning sample set, calculate the operation and maintenance variable set Q q Table 6 is a partial data table of the conditional probability of failure of each operation and maintenance variable in this embodiment.

[0166] <![CDATA[k 1.0 ]]> <![CDATA[k 1.0 ]]> <![CDATA[k 2.0 ]]> <![CDATA[k 2.1 ]]> <![CDATA[k 2.2 ]]> ... <![CDATA[k 18.0 ]]> <![CDATA[ k18.1 ]]> <![CDATA[x N+1 ]]> 0.62 0.38 0.56 0.32 0.12 ... 0.46 0.54

[0167] Table 6

[0168] 4: Analysis of the root causes of mobile base station complaints

[0169] Calculate the contribution of each operation and maintenance variable in the variable set Q to the failure under user complaints. For example, for the node set Q q The O&M variable "seq high load leads to low rate" has the values ​​"Normal" and "High", which yields the following:

[0170] P(“high seq load leads to low rate” = “Normal” | q = “VoLTE voice call is intermittent”) = 0.62, dr(“high seq load leads to low rate” = “Normal”) ≈ 0.15,

[0171] P("High seq load leads to low rate" = "High" | q = "VoLTE voice call is intermittent") = 0.38. Using formula 4-3, we can get dr("High seq load leads to low rate" = "High") ≈ 0.10.

[0172] Table 7 is a descending table of the contribution values ​​of the 17 operation and maintenance variables in this embodiment.

[0173] Table 7. Operation and maintenance variable db value descending table

[0174] Serial number variable db value 1 SRVCC handover success rate 0.62 2 Number of RRC link re-establishment requests 0.57 3 Number of SRVCC handover requests 0.52 4 VoLTE downlink packet loss rate 0.46 ... ... ... 15 seq high RTP uplink packet loss rate 0.12 16 Interference level greater than 105dBm tag 0.09 17 VoLTE voice peak data 0.06

[0175] Table 7

[0176] Table 7 shows that when analyzing user complaints about "poor and intermittent call quality" during VoLTE voice calls, the "SRVCC handover success rate," "RRC link reestablishment request count," and "SRVCC handover request count" variables from base station operation and maintenance data contribute significantly. When γ is set to 3, the "SRVCC handover success rate," "RRC link reestablishment request count," and "SRVCC handover request count" constitute the root cause set for mobile base station complaints.

[0177] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for analyzing root causes of mobile base station complaints based on a Bayesian network, characterized in that: The following steps are involved: S1: Construct a Bayesian network based on the association rules that describe business knowledge. The specific method is as follows: S1.1: Based on the current user complaint information, locate the mobile base station corresponding to the user's area and obtain N historical complaint root cause analysis records of the mobile base station with the same problem as the current user's complaint. The historical complaint root cause analysis records include the complaint problem, the abnormal operation and maintenance variables and values ​​of the mobile base station in the area where the complaint occurred; based on the historical complaint root cause analysis records, establish a variable set Q = (x1,…,x i ,…,x N ,x N+1 ), where x n Represents the nth operation and maintenance variable, and the number of possible values ​​corresponding to it is recorded as M n , n=1,2,…,N,x N+1 Represents the user complaint variable, which has the value M N+1 =2, respectively 0 and 1, x N+1 =1 indicates user complaints, x N+1 =0 means the user has not complained; S1.2: Use the following formula to calculate any two variables x in the variable set Q: i and x j The support between sup(x i ,x j ): Where i, j = 1, 2, ..., N and i ≠ j, #(x i ,x j ) represents the variable x i and x j The number of times they appear together in historical complaint root cause analysis records; Set the minimum support sup min ,0≤sup min ≤1, find all the values ​​that satisfy sup(x i ,x j )≥sup min The variable pairs are recorded as W, and the frequent binomial set V=(v1,…,v w ,…,v W ), where v w Indicates the w-th variable pair selected based on support Represent the two variables in the w-th variable pair, w=1,2,…,W,1≤i w ,j w ≤N and i w ≠j w ; S1.3: Set the minimum confidence level c min , 0≤c min ≤1, the following formula is used to calculate the frequent binomial set V for each variable pair Confidence in, Representing variables and The number of times it appears simultaneously in historical complaint root cause analysis records, Represents variables in historical complaint root cause analysis records Number of occurrences; Then find all the Variable pairs, and arrange the selected variable pairs from large to small according to the confidence level. The number of selected variable pairs is T, and the tth variable pair is Represent the two variables in the t-th variable pair, r=1,2,…,T,1≤i t ,j t ≤N and i t ≠j t ; According to the variable Generate strong association rules for variables Construct variable strong association rule set R=(r1,…,r t ,…,r T ), where r t Represents strong association rules of variables S1.4: Generate a directed acyclic graph of the Bayesian network based on the strong association rules in the variable strong association rule set R. The specific steps are: use the variables in the variable set Q as nodes of the Bayesian network, and take out each variable strong association rule in the variable strong association rule set R in turn. If the variables in the Bayesian network and If there is no directed edge between them, a variable is generated. Pointer to variable ’s directed edges, thus obtaining the directed acyclic graph of the Bayesian network; Then, based on the directed acyclic graph of the Bayesian network and the operation and maintenance data of the mobile base station, the maximum likelihood estimation is used to obtain the value of each variable x in the directed acyclic graph. i The conditional probability parameter The calculation formula is as follows: Among them, q i,h is the variable x i The parent node set π(x i ) the hth value combination of the nodes, h=1,2,…,H i , H i For the parent node set π(x i ), is the operation and maintenance variable x i mth i possible values, m i =0,1,…,M i -1, #() is the number of operation and maintenance data records that meet the search conditions; S2: Generate the embedding vectors of each variable in the Bayesian network. The specific method is as follows: S2.1: Perform S sampling on the variable set Q of the directed acyclic graph to generate S samples, forming a sample set D. c =(D c,1 ,…,D c,s ,…,D c,S ), D c,s Denotes the sth sample, s = 1, 2, ..., S. The specific process of generating samples each time is as follows: With the variable set Q=(x1,…,x i ,…,x N ,x N+1 ) is sampled one by one in the order of variables as the sampling order. According to the forward sampling algorithm, for the variable x i If it is the root node in the Bayesian network directed acyclic graph, the sampling value is P(x i ) is sampled with probability; if the variable x i If it is not the root node, the sampling value is P(x i |π(x i )=q i ) is sampled with probability, and variable x is recorded i The value of The probability of Generate a random number ram, 0<ram<1, if where p i,-1 =0, then the variable x i Assign value to mth i possible values; S2.2: Based on the sample set D c , the following formula is used to calculate the point mutual information value of different operation and maintenance variables for different value combinations in, Represents the variable x in the sample set D i The value of The number of samples, Represents the variable x in the sample set D j The value of The number of samples, m j =0,1,…,M j -1,M j Represents the variable x j The number of representable values ​​of Indicates that the sample set D satisfies the variable x at the same time i The value of And the operation variable x j The value of The number of samples; Then according to Construct a point mutual information matrix W of size G×G, where S2.3: Calculate the variable correlation matrix M using the following formula OBN : M OBN =W-logβ Where β represents the number of non-adjacent node pairs in the Bayesian network; S2.4: Variable correlation matrix M OBN Low-rank decomposition is used for dimensionality reduction, and the resulting matrix of size G×d is used as the embedding matrix Y, where each row is the embedding vector of the corresponding variable value; S3: Perform probabilistic reasoning based on the variable embedding vector to obtain the conditional probability of each variable failing. The specific method is as follows: S3.1: According to the embedding matrix Y, get the variable x i The value of The embedding vector and the variable x j The value of The embedding vector Calculate the embedding vector and embedding vector The similarity between S3.2: Compute the similarity between the embedding vectors of a variable x and its neighbors i The value is Probability The calculation method is as follows: Note the variable x i The set of neighbor variables is Ne i , for the neighbor variable set Ne i For each variable x j′ , remember the embedding vector and embedding vector The similarity between x j′ ∈Ne i , m j′ =0,1,…,M j′ -1,M j′ Represents the variable x j′ The number of possible values ​​of ; then calculate the embedding vector and the neighbor variable set Ne i All variables x j′ The average value of embedding vector similarity at different values Then use the following formula to calculate the variable x i The value is Probability S3.3: User complaint variable x N+1 As an evidence node, traverse each operation and maintenance variable x n , perform Bayesian network approximate reasoning according to the following steps to calculate the operation and maintenance variable x n In different values The conditional probability of failure The specific steps are as follows: (1) Initialize each operation and maintenance variable x n The value is Number (2) Let the user complain about variable x N+1 The value of is 1, and then the value of each operation and maintenance variable in the variable set Q is randomly generated, and the operation and maintenance reasoning sample D is initialized. o,1 , let the operation and maintenance reasoning sample number z = 1; (3) For each operation and maintenance variable x n , if the operation and maintenance reasoning sample D o,z Mid-O&M variables Then order Otherwise, no action will be taken; (4) Based on operation and maintenance reasoning sample D o,z Generate new operation and maintenance reasoning sample D o,z+1 The specific method is: resample the operation and maintenance variables according to the sequence number of each operation and maintenance variable in the variable set Q, and for the operation and maintenance variable x n Generate random number ram′ n ,if Where P(x n =k n,-1 )=0, then the operation and maintenance variable x n The value of is updated to its mth n possible values; (5) Determine whether z < Z, where Z is the preset number of operation and maintenance reasoning samples. If so, set z = z + 1 and return to step (3); otherwise, proceed to step (6); (6) Based on the generated Z operation and maintenance reasoning samples D o,z , the following formula is used to calculate the conditional probability of each operation and maintenance variable failing when a known base station fails under user complaints: S4: Based on the conditional probability of failure of the operation and maintenance variable under the user complaint situation, the operation and maintenance variable x is calculated according to the following formula n The value is and The absolute value of the conditional probability difference under Among them, m n ,m′ n =0,1,…,M n -1,m n ≠m′ n ; The operation and maintenance variable x is calculated using the following formula n In the value Contribution to base station failure Among them, vd max (x n ) represents the operation and maintenance variable x n The maximum value among the conditional probability differences of the value pairs, Indicates rounding up; The operation and maintenance variable x is calculated using the following formula n Contribution db(x n ): Sort the operation and maintenance variables from large to small according to their contribution, and use the first γ operation and maintenance variables to form the root cause set of mobile base station complaints. The size of γ is set according to actual needs.

2. The mobile base station complaint root cause analysis method according to claim 1, characterized in that: The number of sample samples S in step S2.1 is determined by setting an initial value S0 for the number of sample samples, then gradually increasing the number of sample samples, and for each number of sample samples, calculating each conditional probability parameter obtained by the forward sampling algorithm. Corresponding conditional probability parameters in the conditional probability table based on the Bayesian network If the average value of the difference is less than the preset threshold, the current number of sampling samples is used as the final number of sampling samples S.

3. The mobile base station complaint root cause analysis method according to claim 1, characterized in that: In step S2.4, the variable correlation matrix M OBN The method for low-rank decomposition is: Minimize the following loss function J, and transform the variable association matrix M OBN Approximately use the matrix of its first d dimensions To express: Among them, λ g is the matrix M OBN -U d Σ d (V d ) T The characteristic value of Then the singular value decomposition method is used to The breakdown is as follows: Get the matrix Two low-dimensional matrices and The low-dimensional matrix U is used as the variable correlation matrix M OBN The embedding matrix Y of .

Images