A method, apparatus and medium for root cause analysis of poor quality wireless network cells

Abstract

This invention provides a method, apparatus, and medium for root cause analysis of poor-quality wireless network cells. The method includes: acquiring multi-dimensional feature index sample data of multiple wireless network cells; evaluating multiple classification models based on the feature index sample data to select the optimal classification model; selecting the top N key feature indicators with the highest impact on identifying poor-quality cells from the feature index sample data based on the optimal classification model; clustering all feature indicators in the feature index sample data to obtain M root cause categories; calculating a comprehensive score for each root cause category based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of poor-quality cells, as well as their importance; and determining the primary and secondary causes of the problem in poor-quality cells based on the ranking of the comprehensive scores. This method, apparatus, and medium can solve the problem that existing technologies generally rely on human experience in root cause analysis of poor-quality wireless network cells.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method, apparatus and medium for root cause analysis of poor-quality cells in wireless networks. Background Technology

[0002] A poor-quality cell refers to a cell with one or more network communication quality issues. These issues encompass various aspects, such as voice call completion rate, call drop rate, packet loss rate, data service disconnection rate, access success rate, and user-perceived speed. Currently, identifying the types of problems in poor-quality cells primarily employs traditional network optimization (or network optimization) methods. Specifically, this involves first collecting various key performance indicators (KPIs) for the cell. Then, network optimization personnel can use their expert knowledge and experience to set thresholds for each KPI. Based on these thresholds and the cell's overall KPIs, they can determine the type of poor-quality issue. Next, network optimization personnel analyze the data, manually troubleshooting various network problems such as coverage, interference, and capacity issues. Leveraging their experience, they conduct root cause analysis and develop optimization solutions to address the network problems.

[0003] Currently, the analysis of root causes of poor wireless network quality in certain cells generally relies on manual experience combined with various software tools, which presents several problems:

[0004] 1. A shortage of personnel with specialized skills; a large number of people with few or no expertise has become a bottleneck for optimization. 2. Network problems are complex, with numerous causes, making problem localization difficult. 3. Problems are repeatedly handled, resulting in poor timeliness and long processing cycles. 4. Network problem discovery has certain limitations; the true pain points are not clearly identified, leading to high manpower costs and low efficiency. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing a method, apparatus and medium for root cause analysis of poor quality cells in wireless networks, so as to solve the problem that the prior art generally relies on human experience in the root cause analysis of poor quality cells in wireless networks.

[0006] In a first aspect, the present invention provides a method for root cause analysis of poor-quality cells in a wireless network, comprising:

[0007] Acquire multi-dimensional feature index sample data from multiple wireless network cells;

[0008] Based on the sample data of the aforementioned feature indicators, multiple classification models are evaluated to select the optimal classification model;

[0009] Based on the optimal classification model, select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data, where 20≤N≤40;

[0010] Clustering is performed on all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N;

[0011] For each poor-quality cell among the plurality of wireless network cells, a comprehensive score for each root cause category is calculated based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance. The primary and secondary causes of the problem in the poor-quality cell are determined based on the ranking of the comprehensive scores.

[0012] Furthermore, the step of evaluating multiple classification models based on the feature index sample data to select the optimal classification model specifically includes:

[0013] A classification training set is constructed by using multi-dimensional feature index sample data of each poor-quality cell in the plurality of wireless network cells as positive samples and multi-dimensional feature index sample data of each non-poor-quality cell in the plurality of wireless network cells as negative samples.

[0014] Using the classification training set, the multiple classification models are trained respectively to obtain each of the trained classification models;

[0015] The final evaluation score of each of the classification models after training is calculated based on the accuracy of each model after training, the balanced data accuracy, the area under the receiver operating characteristic (ROC) curve, the F1 score, and the training time.

[0016] The classification model with the highest final evaluation score is taken as the optimal classification model.

[0017] Further, the final evaluation score of each of the trained classification models is calculated based on the accuracy, balanced data accuracy, area under the receiver operating characteristic (ROC) curve, F1 score, and training duration. This specifically includes:

[0018] The final evaluation score of each of the trained classification models is calculated using the following formula:

[0019] Final_Score=sum(Accuracy+Balance_Accuracy+Roc_Auc_Score

[0020] +F1_Score)-TimeTaken / 1000

[0021] Where Final_Score is the final evaluation score of the classification model, Accuracy is the accuracy of the classification model, Balance_Accuracy is the balanced data accuracy of the classification model, Roc_Auc_Score is the area under the ROC curve of the classification model, F1_Score is the F1 evaluation score of the classification model, and TimeTaken is the training time of the classification model.

[0022] Furthermore, the step of selecting the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data according to the optimal classification model specifically includes:

[0023] Based on the training results of the optimal classification model, the importance of each feature indicator in the feature indicator sample data is determined. The importance of the feature indicator is used to measure the degree of influence of the feature indicator on the prediction target of the optimal classification model.

[0024] The importance of each feature index in the feature index sample data is sorted in descending order, and the top N feature indices after descending order are selected as the top N key feature indices with the highest impact on identifying poor-quality cells.

[0025] Furthermore, the step of clustering all feature indicators in the feature indicator sample data to obtain M root cause classifications specifically includes:

[0026] The affinity propagation model is used to cluster all feature indicators in the feature indicator sample data to obtain the M root cause classifications.

[0027] Furthermore, the root cause classification includes: alarm type, coverage type, capacity type, interference type, structure type, access type, mobility type, persistence type, and sensing type.

[0028] Furthermore, the step of calculating the comprehensive score for each root cause category based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality community, as well as their importance, specifically includes:

[0029] Based on the importance of all key feature indicators in the M root cause categories, the importance ranking of the indicators for each root cause category is obtained.

[0030] Obtain the priority ranking of each root cause category based on expert experience;

[0031] For each of the M root cause categories, calculate the ratio of the number of all feature indicators in the root cause category to the total number of feature indicators in all root cause categories to obtain the first indicator score of the root cause category;

[0032] The second index score for each root cause category is calculated based on the first index score for each root cause category and the number of all key feature indicators in each root cause category.

[0033] The ranking of the index numbers for each root cause category is obtained by scoring the second index number for each root cause category.

[0034] Calculate the comprehensive score for each root cause category based on the ranking of the importance, priority, and number of indicators for each root cause category.

[0035] Furthermore, determining the primary and secondary causes of the poor-quality cells based on the ranking of the comprehensive scores specifically includes:

[0036] The root cause category ranked first in the comprehensive score is identified as the primary cause of the poor-quality community, and the root cause category ranked second in the comprehensive score is identified as the secondary cause of the poor-quality community.

[0037] The characteristic indicators involved in the primary and secondary causes of the poor-quality cells are calibrated by using a preset value threshold range.

[0038] Based on the characteristic indicators involved in the preliminary primary and secondary causes of the poor-quality community after indicator calibration, update the indicator importance ranking of each root cause classification.

[0039] The overall score for each root cause category is updated based on the updated ranking of the importance of the indicators, the ranking of the priority of the indicators, and the ranking of the number of indicators for each root cause category.

[0040] The root cause category ranked first in the updated comprehensive score is determined as the primary cause of the problem in the poor-quality cell, and the root cause category ranked second in the updated comprehensive score is determined as the secondary cause of the problem in the poor-quality cell.

[0041] Furthermore, after determining the primary and secondary causes of the poor-quality cells based on the ranking of the comprehensive score, the method further includes:

[0042] Based on the primary and secondary causes of the poor-quality cell problem and the preset network optimization case library, an optimization scheme for the poor-quality cell is obtained.

[0043] In a second aspect, the present invention provides a root cause analysis device for poor-quality wireless network cells, comprising:

[0044] The acquisition module is used to acquire multi-dimensional feature indicator sample data of multiple wireless network cells.

[0045] An evaluation and screening module, connected to the acquisition module, is used to evaluate multiple classification models based on the feature index sample data in order to select the optimal classification model.

[0046] The selection module, connected to the evaluation and screening module, is used to select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data according to the optimal classification model, wherein 20≤N≤40;

[0047] The clustering module, connected to the selection module, is used to cluster all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N;

[0048] The root cause analysis module, connected to the clustering module, is used to calculate a comprehensive score for each poor-quality cell among the plurality of wireless network cells, based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance, and to determine the primary and secondary causes of the problem of the poor-quality cell based on the ranking of the comprehensive scores.

[0049] Thirdly, the present invention provides a root cause analysis device for poor quality cells in wireless networks, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the root cause analysis method for poor quality cells in wireless networks described in the first aspect.

[0050] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the root cause analysis method for poor wireless network quality cells described in the first aspect.

[0051] The method, device and medium for analyzing the root cause of poor-quality wireless network cells provided by the present invention first obtain multi-dimensional characteristic index sample data of multiple wireless network cells; then evaluate multiple classification models according to the characteristic index sample data to screen out the optimal classification model; then according to the optimal classification model, select the top N key characteristic indexes with the highest influence degree on identifying poor-quality cells from the characteristic index sample data, where 20 ≤ N ≤ 40; then cluster all the characteristic indexes in the characteristic index sample data to obtain M root cause classifications, where M < N; finally, for each poor-quality cell in the multiple wireless network cells, calculate the comprehensive score of each root cause classification according to the number of all characteristic indexes, the number of all key characteristic indexes and the importance in the M root cause classifications of the poor-quality cell, and determine the main problem cause and secondary problem cause of the poor-quality cell according to the ranking of the comprehensive scores. By introducing machine learning feature engineering, feature selection and importance determination ideas, the present invention screens the optimal classification model based on multi-dimensional characteristic index sample data of multiple wireless network cells, extracts key characteristic indexes based on the optimal classification model, uses clustering problem rules to sort out M root cause classifications of poor-quality cells and calculate the comprehensive score of each root cause classification, and determines the main problem cause and secondary problem cause of the poor-quality cell based on the ranking of the comprehensive scores of each root cause classification, so as to accurately locate wireless network problems, finally improve the quality of wireless network services and user perception, and solve the problem that the prior art generally relies on manual experience in analyzing the root cause of poor-quality wireless network cell problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of a method for analyzing the root cause of poor-quality wireless network cells according to Embodiment 1 of the present invention;

[0053] Figure 2 It is a flowchart of another method for analyzing the root cause of poor-quality wireless network cells according to an embodiment of the present invention;

[0054] Figure 3 It is a sorting schematic diagram of an importance score example of characteristic indexes according to an embodiment of the present invention;

[0055] Figure 4 It is a schematic diagram of the clustering process of the affinity propagation model according to an embodiment of the present invention;

[0056] Figure 5 It is a flowchart of outputting a comprehensive score according to an embodiment of the present invention;

[0057] Figure 6 It is a structural schematic diagram of a device for analyzing the root cause of poor-quality wireless network cells according to Embodiment 2 of the present invention;

[0058] Figure 7This is a schematic diagram of the structure of a wireless network poor quality cell root cause analysis device according to Embodiment 3 of the present invention. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0060] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0061] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0062] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0063] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0064] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0065] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0066] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0067] Example 1:

[0068] This embodiment provides a root cause analysis method for poor-quality wireless network cells, such as... Figure 1 As shown, the method includes:

[0069] Step S101: Obtain multi-dimensional feature index sample data of multiple wireless network cells.

[0070] It should be noted that wireless networks can refer to 4G networks or 5G networks, etc., and the multi-dimensional characteristic indicator sample data includes network management performance data, measurement and collection data, fault alarm data, etc.

[0071] Step S102: Evaluate multiple classification models based on the feature index sample data to select the optimal classification model.

[0072] In this embodiment, based on multi-dimensional feature index sample data from multiple wireless network cells, different classification models are used for comparative evaluation, and the best-performing classification model is selected as the model to be used subsequently.

[0073] In one optional embodiment, the step of evaluating multiple classification models based on the feature index sample data to select the optimal classification model specifically includes:

[0074] A classification training set is constructed by using multi-dimensional feature index sample data of each poor-quality cell in the plurality of wireless network cells as positive samples and multi-dimensional feature index sample data of each non-poor-quality cell in the plurality of wireless network cells as negative samples.

[0075] Using the classification training set, the multiple classification models are trained respectively to obtain each of the trained classification models;

[0076] The final evaluation score of each of the classification models after training is calculated based on the accuracy of each model after training, the balanced data accuracy, the area under the receiver operating characteristic (ROC) curve, the F1 score, and the training time.

[0077] The classification model with the highest final evaluation score is taken as the optimal classification model.

[0078] Specifically, multi-dimensional feature index sample data of each poor-quality cell is used as positive samples, and multi-dimensional feature index sample data of each non-poor-quality cell is used as negative samples to construct a classification training set. The classification training set is then used to predict and evaluate different classification models. Each classification model is scored based on its accuracy, balanced data accuracy, area score under the ROC curve (Receiver Operating Characteristic Curve), F1 score, and training time. The classification model with the highest final evaluation score is selected as the optimal classification model.

[0079] Specifically, the final evaluation score of each of the trained classification models is calculated using the following formula:

[0080] Final_Score=sum(Accuracy+Balance_Accuracy+Roc_Auc_Score

[0081] +F1_Score)-TimeTaken / 1000

[0082] Where Final_Score is the final evaluation score of the classification model, Accuracy is the accuracy of the classification model, Balance_Accuracy is the balanced data accuracy of the classification model, Roc_Auc_Score is the area under the ROC curve of the classification model, F1_Score is the F1 evaluation score of the classification model, and TimeTaken is the training time of the classification model.

[0083] Step S103: Based on the optimal classification model, select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data, where 20≤N≤40.

[0084] Specifically, based on the training results of the optimal classification model, the importance of each feature indicator in the feature indicator sample data is determined. The importance of the feature indicator is used to measure the degree of influence of the feature indicator on the prediction objective of the optimal classification model (identifying whether a wireless network cell is a poor-quality cell). The importance of each feature indicator in the feature indicator sample data is sorted in descending order, and the top N feature indicators after descending order are selected as the top N key feature indicators with the highest influence on identifying poor-quality cells, as needed. This serves as the underlying data basis for the root cause analysis output.

[0085] Step S104: Cluster all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N。

[0086] Specifically, the affinity propagation model is used to cluster all feature indicators in the feature indicator sample data based on their feature indicator values ​​to obtain the M root cause classifications.

[0087] Step S105: For each poor-quality cell among the plurality of wireless network cells, calculate the comprehensive score for each root cause category based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance, and determine the primary and secondary causes of the problem for the poor-quality cell based on the ranking of the comprehensive scores.

[0088] It should be noted that the root cause classification includes: alarm type, coverage type, capacity type, interference type, structure type, access type, mobility type, persistence type, and sensing type.

[0089] In an optional embodiment, the step of calculating the comprehensive score for each root cause category based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality community, as well as their importance, specifically includes:

[0090] Based on the importance of all key feature indicators in the M root cause categories, the importance ranking of the indicators for each root cause category is obtained.

[0091] Obtain the priority ranking of each root cause category based on expert experience;

[0092] For each of the M root cause categories, calculate the ratio of the number of all feature indicators in the root cause category to the total number of feature indicators in all root cause categories to obtain the first indicator score of the root cause category;

[0093] The second index score for each root cause category is calculated based on the first index score for each root cause category and the number of all key feature indicators in each root cause category.

[0094] The ranking of the index numbers for each root cause category is obtained by scoring the second index number for each root cause category.

[0095] Calculate the comprehensive score for each root cause category based on the ranking of the importance, priority, and number of indicators for each root cause category.

[0096] Specifically, the importance of all key feature indicators in each of the M root cause categories is summed, and based on the summed importance of all key feature indicators in each root cause category, the importance ranking of the indicators for each root cause category is obtained. The priority ranking of each root cause category is then obtained based on expert optimization experience.

[0097] Specifically, the total number of feature indicators in all root cause categories is calculated and divided by the total number of feature indicators in each of the M root cause categories to obtain the first indicator score for each root cause category. The first indicator score for each root cause category is then multiplied by the total number of key feature indicators in each root cause category to obtain the second indicator score for each root cause category and the indicator ranking for each root cause category.

[0098] Specifically, the comprehensive score for each root cause category is obtained by summing the importance ranking, priority ranking, and number ranking of indicators for each root cause category.

[0099] In one optional embodiment, determining the primary and secondary causes of the poor-quality cell based on the ranking of the comprehensive score specifically includes:

[0100] The root cause category ranked first in the comprehensive score is identified as the primary cause of the poor-quality community, and the root cause category ranked second in the comprehensive score is identified as the secondary cause of the poor-quality community.

[0101] The characteristic indicators involved in the primary and secondary causes of the poor-quality cells are calibrated by using a preset value threshold range.

[0102] Based on the characteristic indicators involved in the preliminary primary and secondary causes of the poor-quality community after indicator calibration, update the indicator importance ranking of each root cause classification.

[0103] The overall score for each root cause category is updated based on the updated ranking of the importance of the indicators, the ranking of the priority of the indicators, and the ranking of the number of indicators for each root cause category.

[0104] The root cause category ranked first in the updated comprehensive score is determined as the primary cause of the problem in the poor-quality cell, and the root cause category ranked second in the updated comprehensive score is determined as the secondary cause of the problem in the poor-quality cell.

[0105] Specifically, the root causes ranked first and second in the comprehensive score are identified as the primary and secondary causes of the poor-quality community. Using statistical methods combined with optimization experience, threshold ranges are set for each characteristic indicator. The characteristic indicators related to the primary and secondary causes of the initial problems are calibrated according to the threshold configuration. The calibrated data are clustered again, and the importance ranking of the indicators is updated. The ranking is then combined with the priority ranking and the number of indicators. The one ranked first is the final primary cause of the problem, and the one ranked second is the final secondary cause of the problem.

[0106] In an optional embodiment, after determining the primary and secondary causes of the poor-quality cells based on the ranking of the comprehensive score, the method further includes:

[0107] Based on the primary and secondary causes of the poor-quality cell problem and the preset network optimization case library, an optimization scheme for the poor-quality cell is obtained.

[0108] Specifically, based on on-site optimization cases and the optimization experience of network experts, a network optimization case library is compiled. By identifying the primary and secondary causes of poor-quality cells, and combining this library with the network optimization case library, understandable optimization solutions are provided.

[0109] In a specific embodiment, taking a 4G wireless network as an example, such as... Figure 2 As shown, the root cause analysis method for poor-quality wireless network cells includes the following steps:

[0110] Step S1: 4G network multi-dimensional data source fusion processing

[0111] Specifically, multi-dimensional data from 4G network wireless cells, including PM (network management performance data), MR (measurement and acquisition data), and FM (fault alarm data), are used. The accessed data is modeled according to business requirements, outputting a data dictionary table to support upper-layer applications. By extracting tables and indicators describing different business types, a multi-dimensional performance table of cell indicators used for machine learning is automatically generated. Details are as follows:

[0112] 4G cell performance hourly granular data, 4G cell performance daily granular data, 4G operating parameters, and MR data are processed to obtain 4G cell characteristic indicators. These indicator features are then classified and fused together at the cell level to form a multi-dimensional data source. A sample table of the 4G wireless cell data structure is shown in Table 1.

[0113] Table 1: Sample Data Structure of 4G Wireless Cells

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] The data samples with poor quality issues from the above communities are merged to distinguish between positive and negative samples. After completion, proceed to step S2.

[0120] Step S2: Optimal Model Selection

[0121] Based on the various indicator values ​​and the quality defect label "problem_flag" obtained from the fusion of multi-dimensional data sources from 4G cells in step S1, different models were used for comparative evaluation, and the best-performing model was selected for subsequent use. The selection of an appropriate classification model requires comprehensive consideration of accuracy, robustness, stability, and performance; therefore, the optimal model was selected based on a comprehensive score. Furthermore, due to the imbalance of the data, some unique evaluation indicators were adopted. The evaluation method assesses model effectiveness across five dimensions, with the total score being the sum of the scores for each sub-item. Using historical data from the same cell, problem_flag == 1 as positive samples and problem_flag == 0 as negative samples, a classification training set was constructed. Different models were used for prediction and evaluation, and the models were scored based on the evaluation. The evaluation items are as follows:

[0122] Accuracy: Accuracy measures the proportion of times a model correctly classifies all predicted samples.

[0123]

[0124] Balanced accuracy score: A commonly used performance metric when dealing with imbalanced datasets. Unlike regular accuracy, balanced accuracy takes into account the imbalance in the number of samples between classes, making it more suitable for evaluating models in imbalanced data scenarios.

[0125]

[0126] Receiver Operating Characteristic Curve (ROC) and Area Under the ROC Curve (AUC): The Receiver Operating Characteristic Curve (ROC) and the Area Under the ROC Curve (AUC) are commonly used performance evaluation tools in binary classification problems to assess the accuracy and reliability of classification models. The ROC curve describes the relationship between the True Positive Rate (TPR, also known as recall) and the False Positive Rate (FPR) at different classification thresholds. True Positive Rate Where TP is the number of True Positives, P is the actual number of positive samples, and the false positive rate is... Where FP represents the number of false positives, N represents the actual number of negative samples, the horizontal axis of the ROC curve is FPR, and the vertical axis is TPR. AUC is the area under the ROC curve, ranging from 0 to 1. The larger the AUC, the better the model's classification performance, and the stronger its ability to distinguish between positive and negative samples. When AUC equals 0.5, it means the model's classification performance is equivalent to random guessing; when AUC is greater than 0.5 and close to 1, it means the model's classification performance is relatively good.

[0127] F1 Score: F1 evaluation is a performance metric that takes into account both precision and recall.

[0128]

[0129] Training time (TimeTaken): Since the training time varies greatly among different models, and the root cause algorithm needs to be calculated for each cell, the time is an important evaluation metric. Too long a calculation time makes the model unusable, such as deep models.

[0130] Final score: A higher score indicates shorter training time and a higher model evaluation. Calculation rules:

[0131] Final_Score=sum(Accuracy+Balance_Accuracy+Roc_Auc_Score

[0132] +F1_Score)-TimeTaken / 1000

[0133] The selection of models employed a large number of classification models, and the final results are shown in Table 2:

[0134] Table 2: Results of each classification model

[0135]

[0136] Through a series of evaluations based on live network data, the following model scores are output, and the optimal model is selected based on these scores: model effectiveness is evaluated across five dimensions: accuracy, balanced data accuracy, area under the ROC curve, F1 score, and training time. Higher scores indicate shorter training time and higher model evaluation. Then, the classification model with the best overall score is selected for feature selection training, and ExtraTreesClassifier is ultimately chosen as the final model. After completion, proceed to step S3.

[0137] Step S3: Extraction of TOP N key features of poor-quality 4G cells

[0138] Specifically, historical data from 4G cells is analyzed and modeled. The model is built using a quality difference label "problem_flag" of 0 or 1 as the target. The model strives to find the data items most likely to influence the outcome when analyzing feature weights. After training the optimal model selected in step S2, the importance of the indicator features for each 4G cell is sorted in descending order. The top 30 features with the highest impact are selected as the underlying data basis for root cause analysis output. The TopN key features utilize the feature importance attribute of the AI ​​model. In supervised models, feature importance refers to the degree of influence of each feature (input variable) on the prediction target (output variable) in machine learning. In extremely randomized trees, the Gini coefficient or information gain is used for evaluation. During the construction of each decision tree, the Gini coefficient or information gain can be calculated at each node split to measure the importance of each feature. Both the Gini coefficient and information gain are used to measure the purity or uncertainty of a sample set. For each feature, its average Gini coefficient or average information gain can be calculated as a measure of feature importance. The Gini coefficient (or Gini Impurity) is an indicator used to measure the purity or uncertainty of a dataset. For binary classification problems, the calculation of the Gini coefficient can be simplified to: Gini = 1 - (p1^2 + p2^2), where p1 is the proportion of samples belonging to class 1 out of the total number of samples, and p2 = 1 - p1 is the proportion of samples belonging to class 2 out of the total number of samples. The Gini coefficient ranges from 0 to 1. The smaller the value, the higher the purity of the dataset, meaning the samples tend to belong to the same class. When the Gini coefficient is 0, it means that all samples in the dataset belong to the same class, and the dataset is pure; when the Gini coefficient is 1, it means that the samples in the dataset are evenly distributed across classes, and the dataset is impure. In the process of constructing a decision tree, selecting features with smaller Gini coefficients for node splitting can result in higher purity for each node, thus constructing a more effective decision tree model. Using an extremely random tree model, after sorting each indicator feature in descending order of Gini factor importance, the top 30 features with the highest influence are selected as the underlying data basis for root cause analysis output. Example of a TOP importance feature index (i.e., the importance of the feature index) score: Figure 3 As shown: After completion, proceed to step S4.

[0139] Step S4: Clustering of Key Features of Poor-Quality 4G Cells

[0140] Specifically, there are numerous indicators related to 4G wireless cells (thousands of which are directly relevant), making the analysis of such a large number of indicators extremely difficult for users. If there were a method that could automatically categorize similar indicators and use one (or a few) representative indicators to represent this group of categories, the barrier to entry would be significantly lowered. This invention utilizes an unsupervised clustering algorithm from AI algorithms to meet this requirement, namely the Affinity Propagation Model. This is a clustering algorithm that determines the cluster centers of data points based on a similarity metric between them. By using this clustering algorithm, the following functions are achieved: dynamically clustering key feature indicators using the Affinity Propagation Model; validating these classification indicators based on network optimization experience; manually adjusting the assignment of unreasonable feature indicators; and allocating weight coefficients according to the priority of classification optimization processing.

[0141] 1. Algorithm Advantages: 1) Automatic Cluster Center Determination: The affinity propagation model can automatically select cluster centers without pre-specifying the number of clusters. It determines which data points serve as cluster centers by propagating affinity between data points, thus adapting to the characteristics of the data. 2) Applicable to Irregular Cluster Shapes: Compared to traditional distance-based clustering algorithms (such as K-means), the affinity propagation model can adapt to irregular cluster shapes. It uses the concepts of affinity and attraction to establish complex relationships between data points, thereby capturing richer cluster structures. 3) Unaffected by Initial Value Selection: The affinity propagation model is relatively insensitive to the selection of initial values. It uses a message-passing approach to update the affinity and attraction between data points, thus adaptively adjusting the clustering results and reducing dependence on initial values.

[0142] 2. Algorithm Disadvantages: 1) High Time and Space Complexity: The affinity propagation model has high computational overhead on large-scale datasets. It needs to calculate the affinity and attraction between each pair of data points, resulting in a time complexity of O(N^2T), where N is the number of data points and T is the number of iterations. Furthermore, it needs to store the affinity and attraction values ​​between each pair of data points, leading to high space complexity. 2) Sensitive to Parameters: The affinity propagation model has some parameters that need to be manually set, such as the damping factor and the number of iterations. The selection of these parameters may affect the clustering results, so parameter tuning is necessary to obtain better results. 3) Susceptible to Noise and Outliers: The affinity propagation model is sensitive to noise and outliers. Since it determines cluster centers based on similarity metrics between data points, a large amount of noise or outliers in the dataset may lead to inaccurate clustering results.

[0143] Affinity models can cluster data into groups based on large datasets, meaning data within the same group represent the same meaning. The clustering process is as follows: Figure 4 As shown in the demonstration: Figure 4 The diagram illustrates the different stages (ITERATION#1 to ITERATION#6) of the affinity model clustering process during iteration, starting from initialization and proceeding through multiple iterations until final convergence. Currently, this process converges into a broad category classification, which involves performing business analysis on classification indicators based on network optimization experience, adjusting the assignment of unreasonable feature indicators, and outputting a feature indicator classification table. The affinity model clusters 4G feature indicators based on their values, assigns cluster numbers, and then performs business analysis on the clusters based on network optimization experience, ultimately resulting in the 4G feature indicator classification configuration table shown in Table 3.

[0144] Table 3: Classification and Configuration Table of 4G Feature Indicators

[0145]

[0146]

[0147]

[0148]

[0149]

[0150] The data shows that a lot of data can indeed be grouped into one category. After that, proceed to step S5.

[0151] Step S5: Root cause identification of poor 4G quality cells

[0152] 1. For example Figure 5 As shown, the process involves three steps: summing the importance scores of the top N key indicators of a cell (i.e., ranking the importance of indicators), scoring the weight coefficient of indicator category proportion (i.e., ranking the number of indicators), and scoring the indicator category priority (i.e., ranking the priority). The final output is a comprehensive score for the 4G network cell feature classification priority (i.e., a comprehensive score for root cause classification).

[0153] 1-1: Aggregate and rank the top N key features (TOP features) output by the indicator importance model according to the root cause classification dimension. Examples are shown in Table 4.

[0154] Table 4: Ranking of Top Indicators by Eigenfactor Classification Based on Indicator Importance

[0155]

[0156] These top indicators were aggregated according to the root cause classification dimension, and the importance of each indicator in the root cause classification of the cell was ranked according to the score. As shown in the table below, the cell's eci = 1461930906, and there are 12 access category items in the root cause classification. The importance of these indicators was accumulated, and the final score was 0.2789, ranking first, as shown in Table 5.

[0157] Table 5: Ranking Results of Indicator Importance Based on Indicator Importance

[0158]

[0159] 1-2: Based on the 4G cell characteristics obtained from expert optimization experience, root cause classification and priority ranking were performed, as shown in Table 6:

[0160] Table 6: Priority and Corresponding Scores for 4G Cell Root Cause Classification Based on Expert Experience

[0161]

[0162]

[0163] Based on expert experience, alarms are given the highest priority in root cause analysis and therefore receive full marks. Coverage, capacity, and interference are ranked second, third, and fourth respectively based on field experience, and so on, to complete the sorting of all root cause categories and the corresponding score allocation.

[0164] 1-3: Based on the existing feature indicators and root cause classification, rank the number of indicators for each root cause category. The root cause classification indicator score (i.e., the first indicator score) = total number of feature indicators / number of root cause classification indicators, as shown in Table 7.

[0165] Table 7: Scoring of Root Cause Classification Indicators

[0166] Root cause classification Number of root cause classification indicators Total number of feature indicators Root cause classification index score network Preservation Class 18 194 10.78 4G Override class 11 194 17.64 4G Perception class 50 194 3.88 4G Interference 12 194 16.17 4G Alarms 6 194 32.33 4G Access Class 49 194 3.96 4G Structure class 4 194 48.50 4G Capacity 15 194 12.93 4G Mobile Class 27 194 7.19 4G

[0167] As can be seen from the table above, the total number of feature indicators is the number of indicators in the "4G Feature Indicator Classification Configuration Table" in step S4. Among them, there are 18 indicators for the maintenance category, 11 indicators for the coverage category, 50 indicators for the sensing category, 12 indicators for the interference category, 6 indicators for the alarm category, 49 indicators for the access category, 4 indicators for the structure category, 15 indicators for the capacity category, and 27 indicators for the mobility category. According to the calculation formula above, the root cause classification indicator score results for different root cause classifications can be calculated.

[0168] Then, the scores for different indicator numbers are completed using the following formula: Indicator number score (i.e., the second indicator number score) = Root cause classification indicator number score * Number of root cause classification indicators in the top N indicators. The indicator number scores and ranking table are shown in Table 8.

[0169] Table 8: Indicator Scoring and Ranking Table

[0170]

[0171]

[0172] Based on the statistical results of the number of root cause category indicators in the top N indicators of different root cause categories in the community eci=146193906 in Table 4, multiplying them by the root cause category indicator score results in Table 7, the indicator score can be obtained.

[0173] Finally, based on the rankings of indicator importance, root cause classification priority, and number of indicators, the comprehensive ranking of 4G cell feature classification priority is obtained by summing the rankings of the three dimensions, as shown in Table 9:

[0174] Table 9: Comprehensive Ranking of 4G Cell Feature Classification Priority

[0175] eci Root cause classification Ranking of indicator importance Root cause classification priority ranking Ranking of Indicators Overall score 146193906 Interference 3 3 1 7 146193906 Override class 4 1 2 7 146193906 Structure class 7 4 3 14 146193906 Access Class 2 5 4 11 146193906 Capacity 5 2 5 12 146193906 Perception class 1 7 6 14 146193906 Mobile Class 6 6 7 19

[0176] 2. Delineating the primary and secondary causes of problems in the community

[0177] Based on the results of the 4G cell feature classification priority comprehensive scoring system, the classification with the higher comprehensive ranking is selected as the boundary for poor-quality cells. The labels are customized for ascending order of the single-cell problem classification, and after being associated with the 4G cell problem list, the primary cause (first problem) and secondary cause (second problem) of the 4G problem cells are output. As shown in Table 10, interference is the primary cause, and coverage is the secondary cause, ranked second.

[0178] Table 10: Sample Results of Primary and Secondary Causes of Problem Cells in 4G

[0179]

[0180]

[0181] 3. Root cause analysis of wireless network indicators in the community

[0182] The characteristic indicators are analyzed, and statistical methods combined with optimization experience are used to set threshold ranges for each characteristic indicator. The indicators related to the primary and secondary causes (i.e., the preliminary primary and secondary causes) of the problem areas identified in the previous step are calibrated according to the threshold configuration to obtain more accurate indicators related to the root causes of the problems. This outputs the primary and secondary causes of the root cause localization, and the values ​​of the indicators related to the primary and secondary causes. The calibrated data is then clustered again based on the problem areas, and ranked jointly with the problem classification priority and the number of root cause-related indicators. The first-ranked indicator is the primary cause (i.e., the final primary cause), and the second-ranked indicator is the secondary cause (i.e., the final secondary cause). For areas where only one root cause is identified, the primary cause is the identified root cause, and the secondary cause is filled with "none". Examples are shown in Table 11.

[0183] Table 11: Examples of Root Cause Analysis for 4G Wireless Cells

[0184]

[0185] After completion, proceed to step S6.

[0186] Step S6: Output suggested solutions to cell problems

[0187] Specifically, based on on-site optimization cases and the optimization experience of network experts, a network optimization case library is compiled as a configuration for optimization suggestions. This library can accumulate case content over time to form the foundation for the next step of the knowledge base. (See Table 12.)

[0188] Table 12: Network Optimization Case Library

[0189]

[0190]

[0191] Based on the identified root causes of the problem and combined with a network optimization case library compiled by experts, a descriptive text is output, providing an understandable optimization solution. Example suggestions for solutions to poor 4G quality cells are shown in Table 13:

[0192] Table 13: Suggestions for Solutions to Poor 4G Cell Quality

[0193]

[0194]

[0195] After completion, proceed to step S7.

[0196] Step S7: End

[0197] It is worth mentioning that, in order to solve the above-mentioned problems, the present invention proposes to incorporate the concept of AI intelligent analysis, establish an AI model for analyzing the root causes of poor-quality cell problems, realize the intelligent mining of the root causes of 4G network problems, and enhance the initiative of problem discovery. By introducing machine learning feature engineering, feature selection and importance determination ideas, correlation analysis is carried out on multi-data sources such as network performance data, MR measurement data, interference data, and alarm data. And extract problem index features, use clustering problem rules to sort out the TOP N reasons, accurately locate 4G network problems, and finally improve the service quality of 4G networks and user perception. The present invention automatically analyzes the hidden root causes of a large amount of data by introducing concepts such as feature engineering and feature importance in AI. Extract problem symptoms, clustering problem rules, sort out the TOP reasons for clustering, focus on TOP factors, simplify the complex. And based on automation, data sources such as network performance data, MR measurement data, interference data, and alarm data are uniformly accessed. Output the main reasons and secondary reasons for the problem cell, accurately locate network problems,挖掘潜在问题风险点,固化平台经验,完善问题库建设,指导优化人员有的放矢的进行网络问题的处理。提高一线人员工作效率,降低网络运维投入成本,切实做到“降本增效”的目标。

[0198] The method for analyzing the root causes of poor-quality cells in a wireless network provided by an embodiment of the present invention first obtains multi-dimensional feature index sample data of multiple wireless network cells; then evaluates multiple classification models according to the feature index sample data to screen out the optimal classification model; then according to the optimal classification model, select the top N key feature indexes with the highest influence on identifying poor-quality cells from the feature index sample data, where 20 ≤ N ≤ 40; then cluster all the feature indexes in the feature index sample data to obtain M root cause classifications, where M < N; finally, for each poor-quality cell in the multiple wireless network cells, calculate the comprehensive score of each root cause classification according to the number of all feature indexes and the number and importance of all key feature indexes in the M root cause classifications of the poor-quality cell, and determine the main problem cause and secondary problem cause of the poor-quality cell according to the ranking of the comprehensive scores. The present invention screens out the optimal classification model based on the multi-dimensional feature index sample data of multiple wireless network cells by introducing machine learning feature engineering, feature selection and importance determination ideas, extracts key feature indexes based on the optimal classification model, uses clustering problem rules to sort out M root cause classifications of poor-quality cells and calculates the comprehensive score of each root cause classification, and determines the main problem cause and secondary problem cause of the poor-quality cell based on the ranking of the comprehensive scores of each root cause classification, so as to accurately locate wireless network problems, and finally improve the service quality of wireless networks and user perception, and solve the problem that the prior art generally relies on manual experience in analyzing the root causes of poor-quality cell problems in wireless networks.

[0199] It should be noted that there is an unclear expression "挖掘潜在问题风险点,固化平台经验,完善问题库建设,指导优化人员有的放矢的进行网络问题的处理。" in the original Chinese text. The above translation tries to make sense of the overall context while keeping this part as intact as possible. You may need to check and correct this part according to the actual situation.Example 2:

[0200] like Figure 6 As shown, this embodiment provides a root cause analysis device for poor-quality wireless network cells, used to perform the above-described root cause analysis method for poor-quality wireless network cells, including:

[0201] The acquisition module 11 is used to acquire multi-dimensional feature indicator sample data of multiple wireless network cells;

[0202] The evaluation and screening module 12, connected to the acquisition module 11, is used to evaluate multiple classification models based on the feature index sample data in order to select the optimal classification model.

[0203] Selection module 13, connected to the evaluation and screening module 12, is used to select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data according to the optimal classification model, wherein 20≤N≤40;

[0204] Clustering module 14, connected to selection module 13, is used to cluster all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N;

[0205] The root cause analysis module 15, connected to the clustering module 14, is used to calculate a comprehensive score for each poor-quality cell among the plurality of wireless network cells, based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance, and to determine the primary and secondary causes of the problem of the poor-quality cell based on the ranking of the comprehensive scores.

[0206] Furthermore, the evaluation and screening module 12 specifically includes:

[0207] The construction unit is used to construct a classification training set by taking the multi-dimensional feature index sample data of each poor-quality cell in the plurality of wireless network cells as positive samples and taking the multi-dimensional feature index sample data of each non-poor-quality cell in the plurality of wireless network cells as negative samples.

[0208] The training unit is used to train the plurality of classification models using the classification training set to obtain each of the trained classification models.

[0209] The first calculation unit is used to calculate the final evaluation score of each of the classification models after training based on the accuracy of each of the training classification models, the balanced data accuracy, the area under the receiver operating characteristic (ROC) curve, the F1 score, and the training time.

[0210] As a unit, it is used to select the training classification model with the highest final evaluation score as the optimal classification model.

[0211] Furthermore, the first computing unit is specifically used for:

[0212] The final evaluation score of each of the trained classification models is calculated using the following formula:

[0213] Final_Score=sum(Accuracy+Balance_Accuracy+Roc_Auc_Score

[0214] +F1_Score)-TimeTaken / 1000

[0215] Where Final_Score is the final evaluation score of the classification model, Accuracy is the accuracy of the classification model, Balance_Accuracy is the balanced data accuracy of the classification model, Roc_Auc_Score is the area under the ROC curve of the classification model, F1_Score is the F1 evaluation score of the classification model, and TimeTaken is the training time of the classification model.

[0216] Furthermore, the selection module 13 specifically includes:

[0217] The first determining unit is used to determine the importance of each feature indicator in the feature indicator sample data based on the training results of the optimal classification model. The importance of the feature indicator is used to measure the degree of influence of the feature indicator on the prediction target of the optimal classification model.

[0218] The sorting and selection unit is used to sort the importance of each feature index in the feature index sample data in descending order, and select the top N feature indices after descending sorting as the top N key feature indices with the highest impact on identifying poor-quality cells.

[0219] Furthermore, the clustering module 14 specifically includes:

[0220] Clustering unit, used to cluster all feature indicators in the feature indicator sample data using the affinity propagation model, to obtain the M root cause classifications.

[0221] Furthermore, the root cause classification includes: alarm type, coverage type, capacity type, interference type, structure type, access type, mobility type, persistence type, and sensing type.

[0222] Furthermore, the root cause analysis module 15 includes:

[0223] The first obtaining unit is used to obtain the ranking of the importance of indicators for each root cause category based on the importance of all key feature indicators in the M root cause categories.

[0224] The acquisition unit is used to obtain the priority ranking of each root cause category based on expert experience;

[0225] The calculation unit is used to calculate the ratio of the number of all feature indicators in the root cause category to the total number of feature indicators in all root cause categories for each of the M root cause categories, and to obtain the first indicator score of the root cause category.

[0226] The second calculation unit is used to calculate the second indicator score for each root cause category based on the first indicator score for each root cause category and the number of all key feature indicators in each root cause category.

[0227] The second obtaining unit is used to score each root cause category based on the number of second indicators to obtain the ranking of the number of indicators for each root cause category.

[0228] The third calculation unit is used to calculate the comprehensive score of each root cause category based on the ranking of the importance, priority, and number of indicators of each root cause category.

[0229] Furthermore, the root cause analysis module 15 also includes:

[0230] The second determining unit is used to classify the root cause ranking first in the comprehensive score as the primary cause of the problem in the poor-quality community, and to classify the root cause ranking second in the comprehensive score as the secondary cause of the problem in the poor-quality community.

[0231] The calibration unit is used to calibrate the characteristic indicators involved in the primary and secondary causes of the initial problems of the poor-quality cell by means of a preset value threshold range.

[0232] The first update unit is used to update the indicator importance ranking of each root cause classification based on the characteristic indicators involved in the preliminary primary and secondary causes of the poor-quality cells after indicator calibration.

[0233] The second update unit is used to update the comprehensive score of each root cause category based on the updated ranking of the importance of indicators, the ranking of the priority of indicators, and the ranking of the number of indicators of each root cause category.

[0234] The third determining unit is used to determine the root cause classification ranked first in the updated comprehensive score as the final primary cause of the problem in the poor-quality cell, and to determine the root cause classification ranked second in the updated comprehensive score as the final secondary cause of the problem in the poor-quality cell.

[0235] Furthermore, the device also includes:

[0236] The module is used to obtain an optimization scheme for the poor-quality cell based on the primary and secondary causes of the problem and a preset network optimization case library.

[0237] Example 3:

[0238] refer to Figure 7 This embodiment provides a root cause analysis device for poor wireless network quality cells, including a memory 21 and a processor 22. The memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the root cause analysis method for poor wireless network quality cells in Embodiment 1.

[0239] The memory 21 is connected to the processor 22. The memory 21 can be a flash memory, a read-only memory or other memory, and the processor 22 can be a central processing unit or a microcontroller.

[0240] Example 4:

[0241] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the root cause analysis method for poor wireless network quality cells in Embodiment 1 above.

[0242] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules, or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.

[0243] In summary, the method, device and medium for analyzing the root cause of poor-quality wireless network cells provided by the embodiments of the present invention first obtain multi-dimensional characteristic index sample data of multiple wireless network cells; then evaluate multiple classification models according to the characteristic index sample data to screen out the optimal classification model; then according to the optimal classification model, select the top N key characteristic indexes with the highest influence on identifying poor-quality cells from the characteristic index sample data, where 20 ≤ N ≤ 40; then cluster all the characteristic indexes in the characteristic index sample data to obtain M root cause classifications, where M < N; finally, for each poor-quality cell in the multiple wireless network cells, calculate the comprehensive score of each root cause classification according to the number and importance of all characteristic indexes and all key characteristic indexes in the M root cause classifications of the poor-quality cell, and determine the main problem cause and secondary problem cause of the poor-quality cell according to the ranking of the comprehensive scores. By introducing the ideas of machine learning feature engineering, feature selection and importance determination, the present invention screens the optimal classification model based on the multi-dimensional characteristic index sample data of multiple wireless network cells, extracts key characteristic indexes based on the optimal classification model, uses clustering problem rules to sort out M root cause classifications of poor-quality cells and calculate the comprehensive score of each root cause classification, and determines the main problem cause and secondary problem cause of poor-quality cells based on the ranking of the comprehensive scores of each root cause classification, so as to accurately locate wireless network problems, ultimately improve the quality of wireless network services and user perception, and solve the problem that the prior art generally relies on manual experience in analyzing the root cause of poor-quality wireless network cell problems.

[0244] It can be understood that the above embodiments are merely exemplary embodiments adopted to illustrate the principle of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered within the protection scope of the present invention.

Claims

1. A method for root cause analysis of poor-quality wireless network cells, characterized in that, The method includes: Acquire multi-dimensional feature index sample data from multiple wireless network cells; Based on the sample data of the aforementioned feature indicators, multiple classification models are evaluated to select the optimal classification model; Based on the optimal classification model, select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data, where 20≤N≤40; Clustering is performed on all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N; For each poor-quality cell among the plurality of wireless network cells, a comprehensive score for each root cause category is calculated based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance. The primary and secondary causes of the problem in the poor-quality cell are determined based on the ranking of the comprehensive scores. The step of evaluating multiple classification models based on the feature index sample data to select the optimal classification model specifically includes: A classification training set is constructed by using multi-dimensional feature index sample data of each poor-quality cell in the plurality of wireless network cells as positive samples and multi-dimensional feature index sample data of each non-poor-quality cell in the plurality of wireless network cells as negative samples. Using the classification training set, the multiple classification models are trained respectively to obtain each of the trained classification models; The final evaluation score for each of the trained classification models is calculated based on the accuracy, balanced data accuracy, area under the receiver operating characteristic (ROC) curve, F1 score, and training duration. The balanced data accuracy is determined by summing the recall of each class and the number of classes; the area under the ROC curve evaluates the model's ability to distinguish between positive and negative samples; and the F1 score is determined based on precision and recall. The classification model with the highest final evaluation score is taken as the optimal classification model.

2. The method according to claim 1, characterized in that, The final evaluation score for each of the trained classification models is calculated based on the accuracy, balanced data accuracy, area under the receiver operating characteristic (ROC) curve, F1 score, and training duration. Specifically, this includes: The final evaluation score of each of the trained classification models is calculated using the following formula: Final_Score = sum(Accuracy + Balance_Accuracy + Roc_Auc_Score + F1_Score)– TimeTaken / 1000 Where Final_Score is the final evaluation score of the classification model, Accuracy is the accuracy of the classification model, Balance_Accuracy is the balanced data accuracy of the classification model, Roc_Auc_Score is the area under the ROC curve of the classification model, F1_Score is the F1 evaluation score of the classification model, and TimeTaken is the training time of the classification model.

3. The method according to claim 1, characterized in that, The step of selecting the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data according to the optimal classification model specifically includes: Based on the training results of the optimal classification model, the importance of each feature indicator in the feature indicator sample data is determined. The importance of the feature indicator is used to measure the degree of influence of the feature indicator on the prediction target of the optimal classification model. The importance of each feature index in the feature index sample data is sorted in descending order, and the top N feature indices after descending order are selected as the top N key feature indices with the highest impact on identifying poor-quality cells.

4. The method according to claim 1, characterized in that, The step of clustering all feature indicators in the feature indicator sample data to obtain M root cause classifications specifically includes: The affinity propagation model is used to cluster all feature indicators in the feature indicator sample data to obtain the M root cause classifications.

5. The method according to claim 1, characterized in that, The root cause classification includes: alarm, coverage, capacity, interference, structure, access, mobility, persistence, and sensing.

6. The method according to claim 1, characterized in that, The calculation of the comprehensive score for each root cause category based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality community, as well as their importance, specifically includes: Based on the importance of all key feature indicators in the M root cause categories, the importance ranking of the indicators for each root cause category is obtained. Obtain the priority ranking of each root cause category based on expert experience; For each of the M root cause categories, calculate the ratio of the number of all feature indicators in the root cause category to the total number of feature indicators in all root cause categories to obtain the first indicator score of the root cause category; The second index score for each root cause category is calculated based on the first index score for each root cause category and the number of all key feature indicators in each root cause category. The ranking of the index numbers for each root cause category is obtained by scoring the second index number for each root cause category. Calculate the comprehensive score for each root cause category based on the ranking of the importance, priority, and number of indicators for each root cause category.

7. The method according to claim 6, characterized in that, The determination of the primary and secondary causes of the poor-quality cells based on the ranking of the comprehensive score specifically includes: The root cause category ranked first in the comprehensive score is identified as the primary cause of the poor-quality community, and the root cause category ranked second in the comprehensive score is identified as the secondary cause of the poor-quality community. The characteristic indicators involved in the primary and secondary causes of the poor-quality cells are calibrated by using a preset value threshold range. Based on the characteristic indicators involved in the preliminary primary and secondary causes of the poor-quality community after indicator calibration, the indicator importance ranking of each root cause classification is updated. The overall score for each root cause category is updated based on the updated ranking of the importance of the indicators, the ranking of the priority of the indicators, and the ranking of the number of indicators for each root cause category. The root cause category ranked first in the updated comprehensive score is determined as the primary cause of the problem in the poor-quality cell, and the root cause category ranked second in the updated comprehensive score is determined as the secondary cause of the problem in the poor-quality cell.

8. The method according to claim 1, characterized in that, After determining the primary and secondary causes of the poor-quality cells based on the ranking of the comprehensive score, the method further includes: Based on the primary and secondary causes of the poor-quality cell problem and the preset network optimization case library, an optimization scheme for the poor-quality cell is obtained.

9. A root cause analysis device for poor-quality wireless network cells, characterized in that, include: The acquisition module is used to acquire multi-dimensional feature indicator sample data of multiple wireless network cells. An evaluation and screening module, connected to the acquisition module, is used to evaluate multiple classification models based on the feature index sample data in order to select the optimal classification model. The selection module, connected to the evaluation and screening module, is used to select the top N key feature indicators with the highest impact on identifying poor-quality communities from the feature indicator sample data according to the optimal classification model, wherein 20≤N≤40; The clustering module, connected to the selection module, is used to cluster all feature indicators in the feature indicator sample data to obtain M root cause classifications, where M <N; The root cause analysis module, connected to the clustering module, is used to calculate a comprehensive score for each poor-quality cell among the plurality of wireless network cells, based on the number of all feature indicators and the number of all key feature indicators in the M root cause categories of the poor-quality cell, as well as their importance, and to determine the primary and secondary causes of the problem of the poor-quality cell based on the ranking of the comprehensive scores. The evaluation and screening module specifically includes: The construction unit is used to construct a classification training set by taking the multi-dimensional feature index sample data of each poor-quality cell in the plurality of wireless network cells as positive samples and taking the multi-dimensional feature index sample data of each non-poor-quality cell in the plurality of wireless network cells as negative samples. The training unit is used to train the plurality of classification models using the classification training set to obtain each of the trained classification models. The first calculation unit is used to calculate the final evaluation score of each of the training classification models based on the accuracy, balanced data accuracy, area under the receiver operating characteristic (ROC) curve, F1 score, and training time of each training classification model. The balanced data accuracy is determined based on the sum of the recall rates of the classes and the number of classes; the area under the ROC curve is used to evaluate the model's ability to distinguish between positive and negative samples; and the F1 score is determined based on precision and recall. As a unit, it is used to select the training classification model with the highest final evaluation score as the optimal classification model.

10. A root cause analysis device for poor-quality wireless network cells, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the root cause analysis method for poor quality cells in wireless networks as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the root cause analysis method for poor-quality wireless network cells as described in any one of claims 1-8.

Images