CN116156538A - Quality difference cell root cause positioning method based on SMOTE-ReliefF-XGBoost algorithm - Google Patents

Abstract

The invention provides a quality difference cell root cause positioning method based on an SMOTE-ReliefF-XGBoost algorithm, which comprises the following steps: establishing a cell base station information matrix, and labeling a base station quality difference root cause according to historical experience; adopting a proper data preprocessing method according to the characteristics of different characteristics of the base station information; equalizing different types of data by using an SMOTE oversampling method; because of redundancy of the data features, firstly calculating weight values of the features by adopting a ReliefF algorithm, and then selecting important features according to a Sequential Forward Selection method; finally, the processed data are sent into 4 independent XGBoost classification models for training, and a genetic algorithm is adopted for parameter tuning; the invention effectively avoids the problem of over-fitting by adopting the SMOTE over-sampling technology, saves the memory resources of a computer and eliminates redundant features by adopting a Relieff feature selection algorithm, improves the performance of the model, and finally reduces the training time of the model by adopting a genetic algorithm to optimize the XGBoost model.

Description

Root Cause Location of Poor Quality Cells Based on SMOTE-ReliefF-XGBoost Algorithm method

technical field

The invention belongs to the technical field of communication, and in particular relates to a method for locating the root cause of a cell with poor quality based on the SMOTE-ReliefF-XGBoost algorithm.

Background technique

With the continuous and rapid development of mobile communication technology, wireless networks have become an indispensable part of life and work. The current mobile communication network is getting larger and larger, and the network structure is becoming more and more complex. Network problems also cropped up.

The wireless network transmits information through base stations, and each base station contains its local area network data information, mainly including bandwidth, frequency band, macro station power, air interface up/downlink traffic, busy up/downlink PRB average utilization rate, and dual-stream ratio, etc. characteristic information.

The poor wireless network of the base station is affected by many factors, such as the number of users and the power of the base station; if you want to optimize the wireless network of the base station, you first need to know the root cause of the poor quality of the base station, so as to prescribe the right medicine. According to the data analysis of poor-quality base stations, the cause of poor quality is a problem that must be faced in wireless network optimization. If poor-quality base stations are not optimized in time, it will not only affect user experience, but may even lead to user loss.

At present, the root cause of the poor quality of the base station wireless network is mainly analyzed by manpower, which is time-consuming and laborious; there are few algorithms to realize the automatic analysis of the root cause of the poor quality of the base station wireless network. The application is exhausted, and the current automatic poor quality root cause location algorithm is simple, which cannot accurately locate the root cause of poor quality.

Therefore, in view of the above technical problems, a root cause location method for poor quality cells based on the SMOTE-ReliefF-XGBoost algorithm is proposed to realize automatic and reliable location of the root cause of poor quality.

Contents of the invention

The purpose of the present invention is to improve the reliability of intelligent location of the root cause of poor quality based on the deficiencies of the prior art, and provide a method for location of the root cause of poor quality cells based on the SMOTE-ReliefF-XGBoost algorithm.

For above-mentioned purpose, the present invention realizes based on following steps, mainly comprises:

S1: Establish a cell base station information matrix, and mark the reasons for the poor quality of each base station according to different status information according to the bandwidth, scene, average number of users, and total traffic of the cell base station; the reasons for the poor quality of the base station are: coverage problems, Interference issues, capacity issues and antenna feeder issues correspond to labels 1 to 4 respectively;

S2: Label encoding, marking the degree of the four labels in S1 corresponding to the question. Among them, the coverage problem is divided into weak coverage, over-coverage, and overlapping coverage, corresponding to labels 0-2; the interference problem is divided into external interference and system interference, corresponding to labels 0-1; the capacity problem is divided into high load and license-restricted sectors Unbalanced traffic load and high load, other reasons, correspond to labels 0-1; antenna feeder problems are divided into dual-stream ratio problems and antenna feeder disconnection problems, corresponding to labels 0-1; Belonging to any of the above questions, it is called a blank set, followed by the codes of the sub-questions under each question, for example, the code of the blank set in the covering question is 3;

S3: data preprocessing, which is processed separately according to the characteristics of different information of the cell base station. For discrete information such as the number (times) of RRC connection establishment failures caused by frequency band, area, bandwidth, and resource allocation failures, one-hot encoding is used for processing; for weekly downlink cell self-busy average traffic (GByte) and air interface uplink services For information with large units such as traffic (GByte), use the Z-Score method for standardized processing;

S4: Data equalization, using the SMOTE method to balance the sub-problem data under different causes of poor quality, so that the amount of base station data corresponding to the sub-problem labels under each problem label is almost the same;

S5: Feature selection, calculate the importance score of each feature through the ReliefF algorithm for the balanced data under each label question in S4, and then select the important features corresponding to each label question through the Sequential Forward Selection method, and eliminate irrelevant features;

S6: Model training. The features selected in S5 for each label problem and the matrix composed of sub-problem labels are sent to four XGBoost classification models for independent training, and the genetic optimization algorithm is used to adjust the loss function (cross entropy function) to obtain an extreme value. Small hyperparameters, test model performance by cross-validation;

S7: Locating the root cause of poor quality, applying the optimal hyperparameters debugged in S6 to the model, and then inputting the cell base station information matrix into the model to obtain the root cause of poor quality of the base station, thereby optimizing the base station layout and related configuration according to the root cause of poor quality.

The steps S1-S3 can be specifically described as:

1≤k≤n,1≤i≤m。Assume that cell base station information is represented by F={X ₁ , X ₂ ,...,X _n } ^T ={F ₁ ,F ₂ ,...,F _m }, where

1≤k≤n, 1≤i≤m.

1≤j≤4。The cell base station label is represented by L={L ¹ , L ² , L ³ , L ⁴ }, where

1≤j≤4.

公式归一化其特征值，其中μ_i是特征均值，σ_i是特征标准差。For features with larger values in cell base station information F, according to

The formula normalizes its eigenvalues, where μ _i is the feature mean and σ _i is the feature standard deviation.

For the discrete value features in cell base station information F, one-hot coding is adopted. Finally, the data is divided into training set and test set.

The S4 can be specifically described as:

计算其与剩余样本的距离，并随机从k(k＝3)个最近邻样本中选取一个。First randomly select a minority sample X _t from the training set, 1≤t≤n, and then according to the Euclidean distance formula

Calculate the distance between it and the remaining samples, and randomly select one of the k (k=3) nearest neighbor samples.

Finally, randomly select a point on the mapping path between the minority sample X _t and the randomly selected sample as the new sampled sample. Sampling is repeated many times until the data volume of various samples reaches a balance.

The ReliefF algorithm in S5 first randomly selects a sample R from the training set, calculates the distance between it and its similar and heterogeneous samples according to the Euclidean distance formula, and selects its k nearest neighbors of the same and heterogeneous samples respectively.

Each feature weight is iteratively calculated according to the following formula:

中的第j个最近邻样本，m表示迭代次数，k表示选取的最近邻样本个数。Among them, diff(F _i ,R,H _j ) represents the difference between sample R and H _j on feature F _i , H _j represents the jth nearest neighbor sample in category C∈class(R), M _j (C) Indicates category

The jth nearest neighbor sample in , m represents the number of iterations, and k represents the number of selected nearest neighbor samples.

The calculation of diff(A,R ₁ ,R ₂ ) is as follows:

根据ReliefF算法计算的特征权重信息迭代向S中逐个添加特征，直到找出使模型性能达到最佳的特征子集时停止。The initial feature subset S selected by the Sequential Forward Selection algorithm in S5 is

According to the feature weight information calculated by the ReliefF algorithm, iteratively add features to S one by one, until the feature subset that makes the model performance optimal is found and stops.

The essence of the XGBoost classification algorithm in S6 is to assemble several base classifier models into a base classifier-based model set.

The essence of the genetic optimization algorithm in S6 is to realize parameter tuning through computer simulation of the process of biological and human evolution.

Beneficial effects of the present invention:

1. The present invention uses the ReliefF algorithm to calculate the correlation weight between features and labels, and the supervised feature selection method can improve the effectiveness of the selected features;

2. Aiming at the characteristics of unbalanced data types, the present invention adopts the SMOTE oversampling algorithm to balance the amount of data of each type, effectively avoiding the influence of over-fitting and weight deviation of the model;

3. The present invention adopts a divide-and-conquer method for each major root cause, respectively applying four independent XGBoost classification models to locate the root cause of poor base station quality, effectively improving the reliability of locating the root cause;

4. The present invention uses a genetic algorithm to tune the hyperparameters of the XGBoost classification model, which can effectively reduce the training time of the model and improve the performance of the model.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the embodiments of the present invention or in the description of the prior art. Obviously, the accompanying drawings described below are only illustrations of the present invention For some embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without creative effort.

Fig. 1 is an overall system flow chart of a method for locating the root cause of a poor quality cell based on the SMOTE-ReliefF-XGBoost algorithm of the present invention.

FIG. 2 is a block diagram of an overall system algorithm of a method for locating the root cause of a poor-quality cell based on the SMOTE-ReliefF-XGBoost algorithm of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the implementation of the flow chart shown in the accompanying drawings. However, this embodiment does not limit the present invention. Based on the embodiments of the present invention, any structural, method or functional changes made by those skilled in the art according to the method are included in the protection scope of the present invention.

The present invention provides a method for locating the root cause of poor quality cells based on the SMOTE-ReliefF-XGBoost algorithm, as shown in Figure 1-2, the method includes steps:

S1: Establish a cell base station information matrix, and mark the reasons for the poor quality of each base station according to different status information according to the bandwidth, scene, average number of users, and total traffic of the cell base station; the reasons for the poor quality of the base station are: coverage problems, Interference issues, capacity issues and antenna feeder issues correspond to labels 1 to 4 respectively;

S2: Label encoding, marking the degree of the four labels in S1 corresponding to the question. Among them, the coverage problem is divided into weak coverage, over-coverage, and overlapping coverage, corresponding to labels 0-2; the interference problem is divided into external interference and system interference, corresponding to labels 0-1; the capacity problem is divided into high load and license-restricted sectors Unbalanced traffic load and high load, other reasons, correspond to labels 0-1; antenna feeder problems are divided into dual-stream ratio problems and antenna feeder disconnection problems, corresponding to labels 0-1; Belonging to any of the above questions, it is called a blank set, followed by the codes of the sub-questions under each question, for example, the code of the blank set in the covering question is 3;

S3: data preprocessing, which is processed separately according to the characteristics of different information of the cell base station. For discrete information such as the number (times) of RRC connection establishment failures caused by frequency band, area, bandwidth, and resource allocation failures, one-hot encoding is used for processing; for weekly downlink cell self-busy average traffic (GByte) and air interface uplink services For information with large units such as traffic (GByte), use the Z-Score method for standardized processing;

S4: Data equalization, using the SMOTE method to balance the sub-problem data under different causes of poor quality, so that the amount of base station data corresponding to the sub-problem labels under each problem label is almost the same;

S5: Feature selection, calculate the importance score of each feature through the ReliefF algorithm on the balanced data under each label question in S4, and then select the important features corresponding to each label question through the Sequential Forward Selection method, and eliminate irrelevant features;

S6: Model training. The features selected in S5 for each label problem and the matrix composed of sub-problem labels are sent to four XGBoost classification models for independent training, and the genetic optimization algorithm is used to adjust the loss function (cross entropy function) to obtain an extreme value. Small hyperparameters, test model performance by cross-validation;

S7: Locating the root cause of poor quality, applying the optimal hyperparameters debugged in S6 to the model, and then inputting the cell base station information matrix into the model to obtain the root cause of poor quality of the base station, thereby optimizing the base station layout and related configuration according to the root cause of poor quality.

Wherein, the steps S1-S3 can be specifically described as:

1≤k≤n,1≤i≤m。Assume that cell base station information is represented by F={X ₁ ,X ₂ ,...,X _n } ^T ={F ₁ ,F ₂ ,...,F _m }, where

1≤k≤n, 1≤i≤m.

1≤j≤4。The cell base station label is represented by L={L ¹ , L ² , L ³ , L ⁴ }, where

1≤j≤4.

公式归一化其特征值，其中μ_i是特征均值，σ_i是特征标准差。For features with larger values in cell base station information F, according to

The formula normalizes its eigenvalues, where μ _i is the feature mean and σ _i is the feature standard deviation.

For the discrete value features in cell base station information F, one-hot coding is adopted. Finally, the data is divided into training set and test set.

Specifically, the S4 can be specifically described as:

计算其与剩余样本的距离，并随机从k(k＝3)个最近邻样本中选取一个。First randomly select a minority sample X _t from the training set, 1≤t≤n, and then according to the Euclidean distance formula

Calculate the distance between it and the remaining samples, and randomly select one of the k (k=3) nearest neighbor samples.

Finally, randomly select a point on the mapping path between the minority sample X _t and the randomly selected sample as the new sampled sample. Sampling is repeated many times until the data volume of various samples reaches a balance.

Specifically, as shown in Figure 2, the ReliefF algorithm in S5 first randomly selects a sample R from the training set, calculates the distances between its similar and heterogeneous samples according to the Euclidean distance formula, and selects its k nearest neighbors Homogeneous and heterogeneous samples.

Each feature weight is iteratively calculated according to the following formula:

中的第j个最近邻样本，m表示迭代次数，k表示选取的最近邻样本个数。Among them, diff(F _i ,R,H _j ) represents the difference between sample R and H _j on feature F _i , H _j represents the jth nearest neighbor sample in category C∈class(R), M _j (C) Indicates category

The jth nearest neighbor sample in , m represents the number of iterations, and k represents the number of selected nearest neighbor samples.

The calculation of diff(A,R ₁ ,R ₂ ) is as follows:

根据所述S4中ReliefF算法计算的特征权重信息迭代向特征子集S中逐个添加特征，直到找出使模型性能达到最佳的特征子集时停止。Specifically, the initial feature subset S selected by the Sequential Forward Selection algorithm in S5 is

According to the feature weight information calculated by the ReliefF algorithm in S4, iteratively add features to the feature subset S one by one, and stop until the feature subset that makes the model performance optimal is found.

As shown in Figure 2, the essence of the XGBoost classification algorithm in S6 is to assemble several base classifier models into a base classifier-based model set. The base classifier generally uses a decision tree. The principle of the XGBoost classification algorithm is to The basis of the classifier continues to increase until the accuracy of the algorithm no longer increases significantly.

Specifically, assuming that there are K base classifiers f _k in XGBoost, 1≤k≤K, for the ith (1≤i≤n) sample X _i , the jth (1≤j≤4) XGBoost model is The output of i samples is:

Build the following objective function:

The training of the XGBoost model is achieved by minimizing the objective function argmin Obj ^j .

Specifically, the above-mentioned optimization process is implemented according to the genetic optimization algorithm in S6, and its essence is to realize parameter tuning by simulating the process of biological and human evolution by computer.

The above are only preferred embodiments of the present invention, and only specifically describe the technical principle of the present invention. These descriptions are only for explaining the principle of the present invention, and cannot be interpreted as limiting the protection scope of the present invention in any way. Based on the explanations here, any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, and those skilled in the art who can think of other specific implementations of the present invention without creative work are all Should be included within the protection scope of the present invention.

Claims (7)

1. a kind of quality poor cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm, it is characterized in that, described method comprises the following steps: S1: Establish a cell base station information matrix, and mark the reasons for the poor quality of each base station according to different status information according to the bandwidth, scene, average number of users, and total traffic of the cell base station; the reasons for the poor quality of the base station are: coverage problems, Interference issues, capacity issues and antenna feeder issues correspond to labels 1 to 4 respectively; S2: Label coding, marking the degree of the four labels in S1 corresponding to the problem; among them, the coverage problem is divided into weak coverage, over-coverage, and overlapping coverage, corresponding to labels 0-2; the interference problem is divided into external interference and internal system interference, Corresponds to labels 0-1; capacity problems are divided into high load, unbalanced traffic load and high load among license-restricted sectors, and other reasons, corresponding to labels 0-1; antenna feeder problems are divided into dual-flow ratio problems and antenna feeder disconnection Questions, corresponding to labels 0-1; In addition, there are some reasons for the poor quality of base stations that do not belong to any of the above problems, which are called blank sets, which are closely followed by the codes of sub-problems under each question. For example, the blank set code in the coverage problem is 3; S3: Data preprocessing, which is processed separately according to the characteristics of different information of the cell base station; for discrete information such as the number of RRC connection establishment failures caused by frequency band, area, bandwidth, and resource allocation failures, one-hot encoding is used for processing ;Aiming at the average traffic (GByte) and air interface uplink business traffic (GByte) when the weekly downlink cell is busy, the Z-Score method is used to standardize the processing; S4: Data equalization, using the SMOTE method to balance the sub-problem data under different causes of poor quality, so that the amount of base station data corresponding to the sub-problem labels under each problem label is almost the same; S5: Feature selection, calculate the importance score of each feature through the ReliefF algorithm for the balanced data under each label question in S4, and then select the important features corresponding to each label question through the Sequential Forward Selection method, and eliminate irrelevant features; S6: Model training. The features selected in S5 for each label problem and the matrix composed of sub-problem labels are sent to four XGBoost classification models for independent training, and the genetic optimization algorithm is used to adjust the loss function (cross entropy function) to obtain an extreme value. Small hyperparameters, test model performance by cross-validation; S7: Locating the root cause of poor quality, applying the optimal hyperparameters debugged in S6 to the model, and then inputting the cell base station information matrix into the model to obtain the root cause of poor quality of the base station, thereby optimizing the base station layout and related configuration according to the root cause of poor quality. 2. a kind of poor-quality cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm according to claim 1, is characterized in that, described steps S1-S3 can be specifically described as: Assume that cell base station information is represented by F={X ₁ ,X ₂ ,…,X _n } ^T ={F ₁ ,F ₂ ,…,F _m }, where

The cell base station label is represented by L={L ¹ , L ² , L ³ , L ⁴ }, where />

For features with larger values in cell base station information F, according to

The formula normalizes its eigenvalues, where μ _i is the feature mean and σ _i is the feature standard deviation; for the discrete value features in the cell base station information F, one-hot encoding is used; finally, the data is divided into training set and test set. 3. a kind of quality poor cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm according to claim 1, it is characterized in that, among the described S4, SMOTE algorithm can specifically be described as: First randomly select a minority sample X _t from the training set, 1≤t≤n, and then according to the Euclidean distance formula

Calculate the distance between it and the remaining samples, and randomly select one of the k (k=3) nearest neighbor samples, and finally randomly select a point on the mapping path between the minority sample X _t and the randomly selected sample as the newly sampled Sample; repeated sampling multiple times until the data volume of various samples reaches a balance. 4. a kind of quality poor cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm according to claim 1, it is characterized in that, ReliefF algorithm first randomly selects a sample R from training set in the described S5, according to Euclidean distance The formula calculates the distance between its similar and heterogeneous samples, and selects its k nearest neighbors of the same and heterogeneous samples; iteratively calculates the weight of each feature according to the following formula:

Among them, diff(F _i ,R,H _j ) represents the difference between sample R and H _j on feature F _i , H _j represents the jth nearest neighbor sample in category C∈class(R), M _j (C) Indicates category

The jth nearest neighbor sample in , m represents the number of iterations, and k represents the number of selected nearest neighbor samples; the calculation of diff(A,R ₁ ,R ₂ ) is as follows:

5. a kind of quality poor cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm according to claim 1, it is characterized in that, among the described S5, the selected initial feature subset S of Sequential Forward Selection algorithm is

According to the feature weight information calculated by the ReliefF algorithm, it iteratively adds features to S one by one, and stops until the model performance is found to be optimal. 6. a kind of quality poor cell root cause localization method based on SMOTE-ReliefF-XGBoost algorithm according to claim 1, it is characterized in that, the essence of XGBoost classification algorithm among the described S6 is that several base classifier models are assembled into one An ensemble of models based on base classifiers. 7. a kind of method for localizing the root cause of poor-quality residential areas based on the SMOTE-ReliefF-XGBoost algorithm according to claim 1, is characterized in that, the essence of the genetic optimization algorithm among the described S6 is to simulate the process of biological and human evolution by computer Realize parameter tuning.

Images