Communication system fault determination method based on cost-sensitive support vector machine
By introducing cost-sensitive mechanisms and grid search methods in support vector machines, the severity of failures in wireless communication networks is classified, and the problem of low accuracy of fault analysis in the prior art is solved, and a higher root cause of faults is achieved. Analytical accuracy.
Patent Information
- Application Number
- CN202211011375.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-23
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In wireless communication networks, it is difficult for the prior art to effectively analyze the root cause of failures, especially because the severity information of different fault samples is not fully utilized, resulting in low accuracy of fault analysis.
Using the cost-sensitive support vector machine method, faults are divided into three categories: severe, general and mild, and different loss weights are given to samples of different severity. The optimal cost combination of classification errors is found through grid search to improve the fault root Due to the accuracy of the analysis.
Through the use of cost-sensitive support vector machines, misjudgment of faults of various severity is reduced, the accuracy of moderate and minor faults is significantly improved, and the accuracy of root cause analysis of wireless communication system failures is improved.
Smart Images

Figure CN115442835B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of wireless communications, and in particular to a communication system fault root cause analysis method based on a cost-sensitive support vector machine. Background Art
[0002] In the operation and maintenance process of wireless communication networks, the problem of deterioration of key performance indicators (KPI) is inevitable. With the heterogeneity and complexity of the network, technicians need to find the reasons for the degradation of communication system performance among a large number of observed variables, making it difficult to manually analyze the root causes of network failures.
[0003] At present, the existing method for network fault root cause analysis is to use machine learning methods to solve the difficulty of manually performing root cause analysis on the network. As one of the most powerful machine learning methods for classification tasks, support vector machine has become an important method for solving root cause analysis problems. In the methods of applying support vector machine to root cause analysis problems, cost-sensitive learning is rarely used, so that all samples have the same weight, which is obviously unreasonable because the degree of damage to the network by each fault sample is different. Therefore, in actual communication systems, the different severity information of the fault itself cannot be ignored, and this information can be used to improve the performance of network fault root cause analysis.
[0004] Therefore, how to efficiently find the cause of the performance degradation of the communication system has become a technical problem that needs to be solved urgently, and is also the task of root cause analysis of wireless communication system failures. Summary of the invention
[0005] Technical problem: The purpose of the present invention is to provide a communication system fault determination method based on a cost-sensitive support vector machine, using a grid search method to find the optimal value of the classification error cost combination corresponding to different fault severities, so as to further improve the accuracy of root cause analysis of wireless communication systems.
[0006] Technical solution: In a communication system fault determination method based on a cost-sensitive support vector machine of the present invention, the cost sensitivity classifies the fault into severe faults, general faults and minor faults according to the different severity of the fault, and then gives different loss weight factors to samples of different severity in the cost-sensitive machine learning. The present invention uses a cost-sensitive support vector machine to perform fault root cause analysis, and assigns different error root cause classification costs to fault samples of different severity. The higher the severity of the fault, the greater the error cost of the root cause analysis classification, so that the algorithm focuses more reasonably on the classification of fault root causes of different severity, thereby reducing the misjudgment of root cause analysis of faults of various severity.
[0007] The network key performance indicators used in the present invention include Signal-to-interference-plus-noise Ratios (SINR) and Reference Signal Received Power (RSRP).
[0008] This method considers the following network failure types that may occur in the network: excessive uptilt (EU), excessive downtilt (ED), and excessive reduced power (ERP).
[0009] The method specifically comprises the following steps:
[0010] Step 1: Generate network KPIs, determine fault severity, and build a support vector machine training set.
[0011] Generate M∈N through simulation + (where N + is a positive integer set), when various failures occur in the group network, I∈N + RSRP and SINR of each user, and define the set of each group of data as a network KPI vector, where the m∈{1,2,...,M}th network KPI 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; when constructing the training set of the support vector machine, x m As a feature vector, it contains 2I features, and the fault type corresponding to the mth KPI vector is used as a label to form a set of training samples; the total sample space of the training set is
[0012] X={(x1,actual1),(x2,actual2),...,(x m ,actual m ),...,(x M ,actual M )},in,
[0013] actual m∈{1,2,3} is the fault type corresponding to the mth KPI vector, which is divided into severe fault, moderate fault and minor fault;
[0014] Step 2: Support vector machine training:
[0015] S1. 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}.
[0016] When k=1, it means actual m =1 and serious fault;
[0017] When k = 2, it means actual m =1 and it is a medium fault;
[0018] When k=3, it means actual m =1 and it is a minor fault;
[0019] When k = 4, it means actual m =2 and serious fault;
[0020] When k = 5, it means actual m =2 and it is a medium fault;
[0021] When k = 6, it means actual m =2 and it is a minor fault;
[0022] When k=7, it means actual m =3 and serious fault;
[0023] When k = 8, it means actual m =3 and medium fault;
[0024] When k=9, it means actual m =3 and it is a minor fault;
[0025] The basic idea of multi-classification problem is to decompose the multi-classification task into several binary classification problems. Based on the existing methods to achieve this goal, a one-to-rest strategy is used, that is, one of the nine categories is used as a positive example of the classifier, and the rest are set as negative examples.
[0026] S2. Train the support vector machine based on the training set; record w k and b k As the linear classifier and the bias coefficient for the kth class, the binary classification optimization problem is expressed as follows:
[0027]
[0028] where y m Represents x m If x m belongs to the kth category, then y m =1, otherwise y m =-1;ξ m =1-y m (w k ·x m +b k ) represents the sample x m The distance to the boundary of the kth class, ξ = [ξ1, ξ2, ..., ξ N ],V k is the k-th class soft margin penalty parameter, the optimal value is determined by the ten-fold cross validation method, and N is the number of training samples belonging to the k-th class in the total sample space X;
[0029] According to the Lagrange multiplier method and Karush–Kuhn–Tucker conditions, the above two-category optimization problem is converted into the following optimization problem:
[0030]
[0031] where a m , β m is the Lagrange multiplier, α=[α1,α2,…,α N ], β=[β1, β2,...,β N ];
[0032] To solve L(w k , b k ,ξ,α,β) respectively for w k , b k , find the partial derivative of ξ and set it to 0, and we can get:
[0033]
[0034] Substituting formula (3) into formula (2), it can be simplified to m The single variable optimization problem is solved by the existing Sequential minimal optimization method. m * ; The obtained α m * Substitute into formula (3) to obtain w k * and b k * :
[0035]
[0036] Among them, xt is a value in the total sample space X of the training set such that t * >0 samples, y t For x t The corresponding label;
[0037] Step 3: Optimizing the cost parameters of the cost-sensitive support vector machine:
[0038] In the cost-sensitive support vector machine, the cost matrix Cost is defined, and the error cost of severe failure is defined as The medium fault error cost is The cost of a minor fault error is is a set of real numbers. Due to the influence of fault severity on the cost, it can be seen that the error cost of severe faults is greater than the error cost of moderate faults. Similarly, the error cost of moderate faults is greater than the error cost of minor faults, that is, c1≥ c2≥ c3. Therefore, the following cost matrix is introduced:
[0039]
[0040] 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}, c1,c2,c3 are selected by the grid search method according to the system performance. In this process, the F score is used as an indicator to evaluate the performance of the classification problem, which is defined as:
[0041]
[0042]
[0043]
[0044] Where n actual,predicted Indicates the number of samples that the cost-sensitive support vector machine judges the actual class to be the predicted class;
[0045] Step 4: Classification decision of cost-sensitive support vector machine
[0046] Step 4.1, collect network KPIs for root cause analysis;
[0047] x new =[RSRP 1,new ,...,RSRP i,new ,...,RSRP I,new ,SINR 1,new ,...,SINR i,new ,...,SINR I,new],i∈{1,2,...,I}; where RSRP i,new and SINR i,new is the RSRP and SINR of the i-th user in the network to be root-caused;
[0048] Step 4.2, x new Input the trained support vector machine and get the output
[0049] P(predictedclass=j|x new ),
[0050] Where P(predictedclass=j|x new ) means when x new When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j × new +b j Mapping to probability value;
[0051] Step 4.3, conduct root cause analysis if and only if:
[0052]
[0053] The classifier will judge as class predicted class; if class predicted ∈{1,2,3}, it is finally judged as EU failure; if class predicted ∈{4,5,6}, it is finally judged as ED fault; if class predicted ∈{7,8,9}, it is finally determined to be an ERP failure.
[0054] The fault type is: when the actual m =1, this training sample is an EU fault with excessive uptilt; when the actual m =2, the training sample is an ED fault (Excessive downtilt, ED); otherwise, the training sample is an ERP fault (Excessive reduced power, ERP).
[0055] The fault severity is as follows: a severe fault is defined as a fault that has a serious impact on network performance, defined as a fault in which the SINR is lower than 0dB or the RSRP is lower than -110dBm in more than 60% of the network coverage area; a moderate fault is defined as a fault that has a moderate impact on network performance, defined as a fault in which the SINR is lower than 0dB or the RSRP is lower than -110dBm in 40%-60% of the network coverage area; a minor fault is defined as a fault that has a minor impact on network performance, defined as a fault in which the SINR is lower than 0dB or the RSRP is lower than -110dBm in 20%-40% of the network coverage area; the fault severity is judged for the training sample KPI in the total sample space X of the training set according to the above definitions.
[0056] The specific steps of optimizing the cost parameters of the cost-sensitive support vector machine are as follows:
[0057] Step 1, initialize m=1, n actual,predicted =0, The search space of the grid search method is:
[0058] c1∈{c 1min ,c 1min +Δc,c 1min +2Δc,...,c 1max},
[0059] c2∈{c 2min ,c 2min +Δc,c 2min +2Δc,...,c 2max},
[0060] c3∈{c 3min ,c 3min +Δc,c 3min +2Δc,...,c 3max},
[0061] At the same time, considering the constraint condition c1≥c2≥c3>0, initialize c1=c 1min , c2=c 2min , c3=c 3min , where c 1min , c 2min , c 3min Both are set to 1 to ensure that the grid search can cover the search coverage of the classic support vector machine, c 1max , c 2max , c 3max The value should not exceed 15 to prevent oversensitivity to severe errors. 1max >c 2max >c 3maxTo ensure the effectiveness of cost-sensitive search; Δc is the search step size, which should be smaller than the minimum range of the search interval to ensure the effectiveness of the search, that is, Δc <min(c 1max -c 1min ,c 2max -c 2min ,c 3max -c 3min ), and should ensure c 1max -c 1m i n , c 2max -c 2min , c 3max -c 3min Can divide Δc to ensure the divisibility of the grid; c 1min ≥c 2min ≥c 3min And c 1max >c 1min , c 2max >c 2min , c 3max >c 3min ;
[0062] Step 2: Use cost-sensitive support vector machine to train the sequence x in the sample set m Classification is performed if and only if:
[0063]
[0064] The classifier will judge as class predicted class; where class predicted ∈{1,2,3,4,5,6,7,8,9}, Classnum=9 indicates the total number of categories, P(predictedclass=j|x m ) means when x m When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j ·x m +b j Mapping is done by using the existing Platt Scaling method. predicted ∈{1,2,3}, then it is finally judged as EU fault, class predicted =1; if class predicted ∈{4,5,6}, it is finally determined to be an ED fault, class predicted =2; if class predicted ∈{7,8,9}, it is finally judged as ERP failure, class predicted =3;
[0065] Step 3, if classpredicted =actual m , otherwise, m=m+1;
[0066] Step 4: If m≤M, repeat steps 2 and 3 to complete the classification decision for all training sets. Otherwise, calculate the F score under the current c1, c2, c3 assignment to evaluate the classification problem performance and reinitialize n. actual,predicted =0,
[0067] Step 5: Update the values of c1, c2, and c3 using the following method with a step size of Δc, whose value is determined by the search space:
[0068] If c3+Δc≤min(c 3max ,c2), c1=c1, c2=c2, c3=c3+Δc; otherwise, if c2+Δc≤min(c 2max ,c1),c2=c2+Δc,c3=c 3min ; Otherwise, if c1+Δc≤c 1max , c1=c1+Δc, c2=c 2min , c3=c 3min ; Otherwise, all traversals have been completed, jump to step 6;
[0069] Step 6, repeat step 2, step 3, step 4, and step 5 until all value combinations of c1, c2, and c3 in the search interval are traversed;
[0070] Step 7: Find the maximum F score and apply the corresponding optimal c1, c2, c3 values to the cost matrix Cost.
[0071] Beneficial effects: The communication system fault determination method based on cost-sensitive support vector machine of the present invention has the following technical effects:
[0072] The present invention uses a cost-sensitive support vector machine to perform fault root cause analysis, and assigns different error root cause classification costs to fault samples of different severity. The higher the fault severity, the greater the error cost of the root cause analysis classification, so that the algorithm can more reasonably focus on the classification of fault root causes of different severity, thereby reducing the misjudgment of root cause analysis of faults of various severity. The grid search method is used to find the optimal value of the combination of classification error costs corresponding to different fault severities, so as to further improve the accuracy of fault root cause analysis of wireless communication systems, especially the accuracy of faults of medium severity and slight severity is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1is a schematic diagram of a wireless communication system of the present invention;
[0074] Figure 2 It is a flow chart of the method for determining a fault in a wireless communication system of the present invention. DETAILED DESCRIPTION
[0075] See also Figure 1-Figure 2 The specific implementation steps of the communication system fault determination method based on cost-sensitive support vector machine of the present invention are as follows:
[0076] Consider the downlink in a millimeter wave network, with an operating frequency of 80 GHz, each base station has a coverage radius of 25 m, and is equipped with three uniform rectangular arrays (URA), each URA spanning 120°, evenly dividing the horizontal plane. Specifically, the size of each URA 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, the horizontal half-power beam width of the antenna is 65°, the vertical half-power beam width is 35°, the horizontal sidelobe level is 25dB, and the vertical sidelobe level is 20dB. The base station is located at a height of h BS =1.5m, and users are evenly distributed at height h U = 8m hexagonal coverage area of the base station. To simplify the analysis, it is assumed that the user is equipped with a single antenna and is served by the nearest base station among the base stations. i ,y i ) is served by the nearest base station and is interfered by the second and third closest base stations. The influence of shadow fading is also considered, and the standard deviation of shadow fading is 4 dB.
[0077] When the network is operating normally, the base station transmit power is 1W and the antenna tilt angle is 40°. In case of a minor EU fault, the base station transmit power is 1W and the antenna tilt angle is 35°; in case of a moderate EU fault, the base station transmit power is 1W and the antenna tilt angle is 25°; in case of a severe EU fault, the base station transmit power is 1W and the antenna tilt angle is 5°. In case of a minor ED fault, the base station transmit power is 1W and the antenna tilt angle is 45°; in case of a moderate ED fault, the base station transmit power is 1W and the antenna tilt angle is 50°; in case of a severe ED fault, the base station transmit power is 1W and the antenna tilt angle is 55°. In case of a minor ERP fault, the base station transmit power is 0.5W and the antenna tilt angle is 40°; in case of a moderate ERP fault, the base station transmit power is 0.2W and the antenna tilt angle is 40°; in case of a severe ERP fault, the base station transmit power is 0.07W and the antenna tilt angle is 40°.
[0078] Step 1: Collect network KPIs and build support vector machine training sets.
[0079] Through simulation, 900 sets of RSRP and SINR of 144 users in the network when various network failures occur are generated, and the set consisting of each set of data is defined as a network KPI vector, where the m∈{1,2,...,900}th network KPI vector is x m =[RSRP1,SINR1,...,RSRP i ,SINR i ,...,RSRP 144 ,SINR 144 ], i∈{1, 2, ..., 144}, where RSRP i and SINR i is the RSRP and SINR of the ith user. When constructing the training set of the support vector machine, x m As the feature vector, and the fault type corresponding to the mth KPI vector as the label, a set of training samples is formed. The total sample space of the training set is X = {(x1, actual1), (x2, actual2), ..., (x m ,actual m ),...,(x 900 ,actual 900 )}, where actual m ∈{1,2,3} is the fault type corresponding to the mth KPI vector. m = 1, this training sample is an EU fault; when actual m =2, this training sample is an ED fault; otherwise, this training sample is an ERP fault. Then, the human can judge the fault severity of the training sample KPI in the training set X according to the definition.
[0080] Step 2: Support vector machine training. This step includes the following process:
[0081] 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, indicating a serious fault; when k=5, indicating an actual m =2 and it is a moderate fault; when k=6, it means actualm =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 moderate fault; when k=9, it means actual m =2 and it is a minor fault. The basic idea of the multi-classification problem is to decompose the multi-classification task into several binary classification problems. This goal can be achieved based on a variety of existing methods, such as the one-to-rest strategy, which uses one of the nine categories as a positive example for the classifier and sets the rest as negative examples.
[0082] Use support vector machine to train the training set and introduce w k and b k As the linear classifier and the bias coefficient for the kth class, the binary classification optimization problem can be formulated as follows:
[0083]
[0084] where y m Represents x m If x m belongs to the kth category, then y m =1, otherwise y m =-1.ξ m =1-y m (w k ·x m +b k ) represents the sample x m The distance to the boundary of the kth class, ξ = [ξ1, ξ2, ..., ξ N ],V k is the soft margin penalty parameter of the kth class, and the optimal optimization value can be determined by the ten-fold cross validation method. N is the number of training samples belonging to the kth class in the total sample space X.
[0085] According to the Lagrange multiplier method and the Karush–Kuhn–Tucker condition, the above optimization problem can be transformed into the following optimization problem:
[0086]
[0087] where α m , β m is the Lagrange multiplier, α=[α1,α2,…,α N ], β=[β1, β2,...,β N ].
[0088] To solve L(w k , b k ,ξ,α,β) respectively for wk , b k , find the partial derivative of ξ and set it to 0, and we can get:
[0089]
[0090] Substituting (3) into (2), it can be simplified to m The single variable optimization problem can be solved by the existing solution α m * , such as Sequential minimal optimization method. m * Substitute into (3) to obtain w k * and b k * :
[0091]
[0092] Among them, x t is a value in the training set X such that t * >0 samples, y t For x t The corresponding label.
[0093] Step 3: Classification decision and cost parameter optimization of cost-sensitive support vector machine. This step includes the following process:
[0094] In the cost-sensitive support vector machine, the cost matrix Cost is defined, and the faults are divided into three severity levels. The error cost of severe faults is defined as The medium fault error cost is The cost of a minor fault error is Due to the impact of fault severity on the cost, it can be seen that the error cost of a severe fault is greater than the error cost of a moderate fault. Similarly, the error cost of a moderate fault is greater than the error cost of a minor fault, that is, c1≥ c2≥ c3. Therefore, the following cost matrix is introduced:
[0095]
[0096] Where Cost(k,j) represents the cost of classifying the kth class into the jth class. The grid search method selects the optimal value combination of c1, c2, and c3 according to the system performance. In this process, the F score is used as an indicator to evaluate the performance of the classification problem, which is defined as:
[0097]
[0098]
[0099]
[0100] Where n actual,predicted Indicates the number of samples for which the cost-sensitive support vector machine classifies the actual class into the predicted class.
[0101] (1) Initialize m=1, n actual,predicted =0, In this example, Δc = 0.5, and the grid search interval is set to c1∈{1,1.5,2,...15}, c2∈{1,1.5,2,...,10}, c3∈{1,1.5,2,...,5}, and the condition c1≥c2≥c3 is considered. Initialize c1=1, c2=1, c3=1.
[0102] (2) Use cost-sensitive support vector machine to train the sequence x in the sample set m Classification is performed if and only if:
[0103]
[0104] The classifier will judge as class predicted Class. predicted ∈{1,2,3,4,5,6,7,8,9}, Classnum=9 indicates the total number of categories, P(predictedclass=j|x m ) means when x m When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j ·x m +b j Mapping to probability values, in this example, Platt Scaling is used for mapping. Output the final root cause analysis results. predicted ∈{1,2,3}, then it is finally judged as EU fault, class predicted =1; if class predicted ∈{4,5,6}, it is finally determined to be an ED fault, class predicted =2; if class predicted ∈{7,8,9}, it is finally judged as ERP failure, class predicted =3.
[0105] (3) If otherwise,
[0106]
[0107] (4) If m≤M, repeat (2) and (3) to complete the classification decision for all training sets. Otherwise, calculate the F score under the current c1, c2, c3 assignments to evaluate the classification problem performance and reinitialize n actual,predicted =0,
[0108] (5) Update the values of c1, c2, and c3 using the following method:
[0109] If c3+0.5≤min(5,c2), c1=c1, c2=c2, c3=c3+0.5; otherwise, if c2+0.5≤min(10,c1), c2=c2+0.5, c3=1; otherwise, if c1+0.5≤15, c1=c1+0.5, c2=1, c3=1; otherwise, all traversals have been completed, jump to (6).
[0110] (6) Repeat (2), (3), (4), and (5) until all combinations of c1, c2, and c3 are traversed.
[0111] (7) Find the maximum F score and apply the corresponding optimal c1, c2, c3 values to the cost matrix Cost. In this example, (c1, c2, c3) = (9, 5, 3.5) to obtain the maximum F score, and apply the c1, c2, c3 values to the cost matrix Cost.
[0112]
[0113] Step 4: Classification decision of cost-sensitive support vector machine
[0114] (1) Collect the network KPI for root cause analysis as x new =[RSRP 1,new ,...,RSRP i,new ,...,RSRP 144,new ,SINR 1,new ,...,SINR i,new ,...,SINR 144,new ], i∈{1, 2, ..., 144}. Among them, RSRP i,new and SINR i,new are the RSRP and SINR of the i-th user in the coverage area of the base station in the network to be rooted.
[0115] (2) Input the trained support vector machine and get the output P(predictedclass=j|x new ),
[0116] Where P(predictedclass=j|x new) means when x new When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j ·x new +b j Mapping to probability values.
[0117] (3) A root cause analysis is conducted if and only if:
[0118]
[0119] The classifier will judge as class predicted Class. If class predicted ∈{1,2,3}, it is finally judged as EU failure; if class predicted ∈{4,5,6}, it is finally judged as ED fault; if class predicted ∈{7,8,9}, it is finally determined to be an ERP failure.
[0120] The present invention provides a communication system fault root cause analysis method based on a cost-sensitive support vector machine, wherein the cost sensitivity is targeted at different fault severities, and the faults are divided into serious faults, general faults and minor faults, and then different loss weight factors are given to samples of different severities in the cost-sensitive machine learning. The present invention uses a cost-sensitive support vector machine to perform fault root cause analysis, and assigns different error root cause classification costs to fault samples of different severities. The higher the fault severity, the greater the root cause analysis classification error cost, so that the algorithm is more reasonably focused on the classification of fault root causes of different severities, thereby reducing the misjudgment of root cause analysis of faults of various severities. A grid search method is used to find the optimal value of the classification error cost combination corresponding to different fault severities, so as to further improve the accuracy of fault root cause analysis of the wireless communication system.
[0121] The specific embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.
Claims
1. A communication system fault determination method based on cost-sensitive support vector machine, It is characterized in that The method specifically comprises the following steps: Step 1: Generate network KPIs, determine fault severity, and build a support vector machine training set. Generate M∈N through simulation + , where N + is a set of positive integers. When various failures occur in the group network, I∈N + RSRP and SINR of each user, and define the set of each group of data as a network KPI vector, where the m∈{1,2,...,M}th network KPI vector is x m =[RSRP 1 ,SINR 1 ,...,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; when constructing the training set of the support vector machine, x m As a feature vector, it contains 2I features, and the fault type corresponding to the mth KPI vector is used as a label to form a set of training samples; the total sample space of the training set is X = {(x 1 , actual 1 ), (x 2 , actual 2 ),..., (x m , actual m ),..., (x M , actual M )}, where, actual m ∈{1,2,3} is the fault type corresponding to the mth KPI vector, which is divided into severe fault, moderate fault and minor fault; Step 2: Support vector machine training: S1. 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 medium 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 =3 and medium fault; When k=9, it means actual m =3 and it is a minor fault; The multi-classification problem is to decompose the multi-classification task into several binary classification problems, using a one-to-rest strategy, that is, one of the nine categories is used as a positive example of the classifier, and the rest are set as negative examples; S2. Train the support vector machine based on the training set; record w k and b k As the linear classifier and the bias coefficient for the kth class, the binary classification optimization problem is expressed as follows: where y m Represents x m If x m belongs to the kth category, then y m =1, otherwise y m =-1;ξ m =1-y m (w k ·x m +b k ) represents the sample x m The distance to the boundary of the kth class, ξ=[ξ 1 ,ξ 2 , ..., ξ N ],V k is the k-th class soft margin penalty parameter, the optimal value is determined by the ten-fold cross validation method, and N is the number of training samples belonging to the k-th class in the total sample space X; According to the Lagrange multiplier method and the Karush–Kuhn–Tucker condition, the above two-category optimization problem is converted into the following optimization problem: where α m , β m is the Lagrange multiplier, α=[α 1 , α 2 , ..., α N ],β=[β 1 , β 2 , ..., β N ]; To solve L(w k , b k , ξ, α, β) respectively for w k , b k , find the partial derivative of ξ and set it to 0, and we can get: Substituting formula (3) into formula (2), it can be simplified to m The single variable optimization problem is solved by the existing Sequential minimal optimization method. m * ; The obtained α m * Substitute into formula (3) to obtain w k * and b k * : Among them, x t is a value in the total sample space X of the training set such that t * >0 samples, y t For x t The corresponding label; Step 3: Optimizing the cost parameters of the cost-sensitive support vector machine: In the cost-sensitive support vector machine, the cost matrix Cost is defined, and the error cost of severe failure is defined as The medium fault error cost is The cost of a minor fault error is is a real number set. Due to the influence of fault severity on the cost, it can be seen that the error cost of severe faults is greater than that of moderate faults. Similarly, the error cost of moderate faults is greater than that of minor faults, that is, c 1 ≥c 2 ≥c 3 , so the following cost matrix is introduced: 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}, c 1 ,c 2 ,c 3 The grid search method selects the optimal value combination based on the system performance; in this process, the F score is used as an indicator to evaluate the performance of the classification problem, which is defined as: Where n actual,predicted Indicates the number of samples that the cost-sensitive support vector machine judges the actual class to be the predicted class; Step 4: Classification decision of cost-sensitive support vector machine Step 4.1, collect network KPIs for root cause analysis; x new =[RSRP 1,new ,...,RSRP i,new ,...,RSRP I,new ,SINR 1,new ,...,SINR i,new ,...,SINR I,new ], i∈{1, 2, ..., I}; where RSRP i,new and SINR i,new is the RSRP and SINR of the i-th user in the network to be root-caused; Step 4.2, x new Input the trained support vector machine and get the output Where P(predictedclass=j|x new ) means when x new When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j ·x new +b j Mapping to probability value; Step 4.3, conduct root cause analysis if and only if: The classifier will judge as class predicted class; if class predicted ∈{1,2,3}, it is finally judged as EU failure; if class predicted ∈{4,5,6}, it is finally judged as ED fault; if class predicted ∈{7,8,9}, it is finally determined to be an ERP failure; The specific steps of optimizing the cost parameters of the cost-sensitive support vector machine are as follows: Step 1, initialize m=1, n actual,predicted =0, The search space of the grid search method is: c 1 ∈{c 1min ,c 1min +Δc,c 1min +2Δc,...,c 1max }, c 2 ∈{c 2min ,c 2min +Δc,c 2min +2Δc,...,c 2max }, c 3 ∈{c 3min ,c 3min +Δc,c 3min +2Δc,...,c 3max }, Considering the constraint condition c 1 ≥c 2 ≥c 3 >0, initialize c 1 =c 1min , c 2 =c 2min , c 3 =c 3min , where c 1min , c 2min , c 3min Both are set to 1 to ensure that the grid search can cover the search coverage of the classic support vector machine, c 1max , c 2max , c 3max The value should not exceed 15 to prevent oversensitivity to severe errors. 1max >c 2max >c 3max To ensure the effectiveness of cost-sensitive search; Δc is the search step size, which should be smaller than the minimum range of the search interval to ensure the effectiveness of the search, that is, Δc <min(c 1max -c 1min ,c 2max -c 2min ,c 3max -c 3min ), and should ensure c 1max -c 1min , c 2max -c 2min , c 3max -c 3min Can divide Δc to ensure the divisibility of the grid; c 1min ≥c 2min ≥c 3min And c 1max >c 1min , c 2max >c 2min , c 3max >c 3min ; Step 2: Use cost-sensitive support vector machine to train the sequence x in the sample set m Classification is performed if and only if: The classifier will judge as class predicted class; where class predicted ∈{1,2,3,4,5,6,7,8,9}, Classnum=9 indicates the total number of categories, P(predictedclass=j|x m ) means when x m When making classification judgments, the probability of being judged as the jth class is output by the trained support vector machine, w j ·x m +b j Mapping is done to probability values, using the existing method PlattScaling for mapping; if class predicted ∈{1,2,3}, then it is finally judged as EU fault, class predicted =1; if class predicted ∈{4,5,6}, it is finally determined to be an ED fault, class predicted =2; if class predicted ∈{7,8,9}, it is finally judged as ERP failure, class predicted =3; Step 3, if class predicted =actual m , otherwise, m=m+1; Step 4: If m≤M, repeat steps 2 and 3 to complete the classification decision for all training sets. Otherwise, calculate the current c 1 ,c 2 ,c 3 Assign the F score under the value to evaluate the classification problem performance and reinitialize n actual,predicted =0, Step 5: Update c with a step size of Δc, whose value is determined by the search space, using the following method: 1 ,c 2 ,c 3 Values: If c 3 +Δc≤min(c 3max ,c 2 ), c 1 =c 1 , c 2 =c 2 , c 3 =c 3 +Δc; otherwise, if c 2 +Δc≤min(c 2max ,c 1 ), c 2 =c 2 +Δc,c 3 =c 3min ; Otherwise, if c 1 +Δc≤c 1max , c 1 =c 1 +Δc,c 2 =c 2min , c 3 =c 3min ; Otherwise, all traversals have been completed, jump to step 6; Step 6: Repeat steps 2, 3, 4, and 5 until all c in the search interval are traversed. 1 ,c 2 ,c 3 All value combinations; Step 7: Find the maximum F score and its corresponding optimal c 1 ,c 2 ,c 3 The value of is applied to the cost matrix Cost.
2. The communication system fault determination method based on cost-sensitive support vector machine according to claim 1, It is characterized in that The fault type is: when the actual m =1, this training sample is EU antenna tilt angle too small EU fault; when actual m =2, this training sample is an ED fault due to excessive ED antenna tilt angle; otherwise, this training sample is an ERP fault due to excessive ERP power reduction.
3. The communication system fault determination method based on cost-sensitive support vector machine according to claim 1, It is characterized in that The fault severity levels are: a severe fault is defined as a fault that is severe to network performance, defined as a fault where the SINR is less than 0dB or the RSRP is less than -110dBm in more than 60% of the network coverage area; Moderate faults are defined as faults with moderate impact on network performance, defined as faults where SINR is less than 0dB or RSRP is less than -110dBm in 40%-60% of the network coverage area; A minor fault is defined as a fault that has a slight impact on network performance, defined as a fault in which the SINR is lower than 0dB or the RSRP is lower than -110dBm in 20%-40% of the network coverage area. The severity of the fault is judged based on the above definition for the training sample KPI in the total sample space X of the training set.
Citation Information
Patent Citations
Industrial big data early failure detection method based on image semi-supervising cost-sensitive
CN107918379A
Method for fault detection of ad hoc network based on cost sensitivity
CN108768772A
