Root cause identification method and device for cells with poor network quality

Abstract

The invention discloses a root cause identification method and device for a cell with poor network quality. The method comprises the following steps: firstly, acquiring a network index of a cell with poor sample quality; then, according to the network indexes of the sample poor-quality cells, constructing an initial poor-quality cell root cause identification model; determining a root cause model sequence according to the initial poor-quality cell root cause identification model; and traversing the root cause model sequence to identify the poor-quality root cause of the poor-sample-quality cell, and obtaining the poor-quality root cause of the sample cell. Judging according to the poor-quality root cause of the sample cell to obtain a target root cause recognition model sequence; and finally, according to the target root cause identification model sequence, carrying out root cause identification on the target cell with poor network quality. The method can improve the precision and accuracy of root cause identification of cells with poor network quality.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method and device for identifying a root cause of a cell with poor network quality. Background Art

[0002] With the advent of the 5G era, 3 / 4 / 5G hybrid networking has become the mainstream mode of wireless network operation for current operators. However, the expansion of network scale has also brought about the emergence of poor quality cells. These poor quality cells lead to the decline of wireless network service quality, and in serious cases, even make users unable to communicate normally, which in turn triggers user complaints, brings inconvenience to daily life and work, and may even lead to serious consequences and damage the network reputation of operators.

[0003] Traditional network optimization methods usually rely on manual analysis of network indicators and subjective judgment based on the experience of professionals to locate the root cause of poor quality cells. This method is not only inefficient, but also consumes a lot of manpower and time. In order to solve this problem, new methods that use machine learning technology to locate the root cause of poor quality cells have gradually emerged in recent years.

[0004] For example, the existing patent technology CN116156538A proposes a method for locating the root cause of poor quality cells based on the SMOTE-ReliefF-XGBoost algorithm. The main steps of this method include: establishing a cell base station information matrix and labeling the root cause of poor quality of the base station based on historical experience; using appropriate data preprocessing methods for different features; using SMOTE oversampling technology to balance the amount of data of different categories; due to the redundancy of data features, first use the ReliefF algorithm to calculate the weight of each feature, and then use the sequential forward selection method to select important features; finally, input the processed data into four independent XGBoost classification models for training, and use genetic algorithms for parameter tuning.

[0005] Although this solution provides some solutions, it still has the following shortcomings:

[0006] (1) This technology uses the causes of poor base station quality (such as coverage problems, interference problems, capacity problems, and antenna problems) as features of training samples and relies on historical experience to label the root causes of poor base station quality. This approach may not be accurate enough.

[0007] (2) This technology relies on only a single XGBoost model to identify each quality problem, which may lead to insufficient recognition accuracy.

[0008] (3) This technology does not take into account that the models required for poor quality root cause identification in different scenarios may be different, which may further affect the recognition accuracy. Summary of the invention

[0009] The technical problem to be solved by the present invention is to address the above-mentioned deficiencies in the prior art and to propose a method and device for identifying the root cause of a cell with poor network quality. The method can improve the precision and accuracy of identifying the root cause of a cell with poor network quality.

[0010] In a first aspect, the present invention provides a method for identifying a root cause of a cell with poor network quality, the method comprising the following steps:

[0011] Step S1: Obtain network indicators of sample poor quality cells;

[0012] Step S2: constructing an initial root cause identification model for poor quality cells according to the network indicators of the sample poor quality cells;

[0013] Step S3: determining a root cause model sequence according to the initial poor quality cell root cause identification model; and traversing the root cause model sequence to identify the poor quality root causes of the sample poor quality cells to obtain the poor quality root causes of the sample cells;

[0014] Step S4: Determine whether the accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement:

[0015] If the identification accuracy of the root cause of poor quality of the sample cell does not reach the preset accuracy threshold, steps S2 to S4 are repeatedly performed until any of the following conditions are met: if the maximum preset number of iterations is reached, or the identification accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement;

[0016] If the maximum number of iterations is reached but the accuracy requirement is still not met, the last determined root cause model sequence is set as the target root cause identification model sequence; if the identification accuracy of the root cause of poor quality of the sample cell meets the threshold requirement at a certain moment when the maximum preset number of iterations is not reached, the root cause model sequence at that time is set as the target root cause identification model sequence;

[0017] Step S5: Acquire a target cell with poor network quality; and identify the root cause of the target cell with poor network quality according to a target root cause identification model sequence.

[0018] Furthermore, the step S1 specifically includes the following steps:

[0019] Step S11: Collect all network indicators of sample cells within the target range;

[0020] The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators;

[0021] Step S12: comparing the network indicator with a preset threshold;

[0022] Among them, if a certain network indicator exceeds a preset threshold, the sample cell is determined to be a poor quality cell;

[0023] Step S13: extracting the network indicators of the sample poor-quality cells from all network indicators of the sample cells within the target range.

[0024] Furthermore, the step S2 specifically includes the following steps:

[0025] Step S21: classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells;

[0026] Among them, each pre-defined root cause category has its corresponding initial learning model;

[0027] Step S22: construct a corresponding training data set for each root cause category;

[0028] Step S23: training the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

[0029] Further, in step S3, a sample root cause model sequence is determined according to the initial poor quality cell root cause identification model, specifically:

[0030] Determine a root cause model sequence according to the initial poor quality cell root cause identification model and a multi-armed bandit algorithm;

[0031] The method of determining a sample root cause model sequence according to the initial poor quality cell root cause identification model and using a multi-armed bandit algorithm specifically includes the following steps:

[0032] Step S31: according to the attribute data or scene data of the sample poor quality cells, the sample poor quality cells are divided into different attribute categories or scene categories by a predefined standard or a clustering algorithm;

[0033] Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells;

[0034] Step S32: According to the attribute category or scenario category, a multi-armed bandit algorithm is used to select a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model;

[0035] The formula for selecting the root cause model sequence is as follows:

[0036] a=argmax a∈A (Q(a)+c*U(a));

[0037] Among them, argmax() represents the maximum value function;

[0038] a represents the selected root cause model sequence;

[0039] A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model;

[0040] Q(a) represents the average reward obtained by all a;

[0041] c is a constant, and its value range is (0,1);

[0042] U(a) represents the uncertainty measure;

[0043] The uncertainty measure is used to determine the upper limit of the reward, and its calculation formula is as follows:

[0044]

[0045] Where N(a) represents the number of times a is selected;

[0046] logt represents the logarithm of t, where t is the current iteration number.

[0047] Furthermore, in step S4, judging whether the accuracy of the sample cell meets the accuracy threshold requirement specifically includes the following steps:

[0048] For each attribute category or scene category, use sequence a for the tth time t The root cause of the poor quality of the sample cell identified is verified with the real root cause of the sample cell. If the verification result is greater than the threshold D, a reward r is given. t =1, otherwise r t =0;

[0049] Make the root cause model sequence a t The counter N(a t )=N(a t-1 )+1;

[0050] Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0;

[0051] Make the root cause model sequence a t The average reward value

[0052] Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0;

[0053] Compare the root cause model sequence a t The average reward value and the preset average reward value:

[0054] If the root cause model sequence a t If the average reward value is lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell does not meet the accuracy threshold requirement; otherwise, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

[0055] In a second aspect, the present invention provides a root cause identification device for a cell with poor network quality, the device comprising:

[0056] An acquisition unit, used to acquire network indicators of cells with poor sample quality;

[0057] A construction unit, connected to the acquisition unit, configured to construct an initial poor-quality cell root cause identification model according to the network indicators of the sample poor-quality cells;

[0058] A first processing unit is connected to the construction unit, and is used to determine a root cause model sequence according to the initial poor quality cell root cause identification model; and is used to traverse the root cause model sequence to identify the poor quality root cause of the sample poor quality cell, and obtain the poor quality root cause of the sample cell;

[0059] A determination unit, connected to the first processing unit, configured to determine whether the accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement;

[0060] The construction unit is also connected to the determination unit, and is used to re-construct an initial poor quality cell root cause identification model according to the network indicators of the sample poor quality cell when the determination unit determines that the identification accuracy of the sample cell poor quality root cause does not reach the preset accuracy threshold, until any of the following conditions is met: a maximum preset number of iterations is reached, or the identification accuracy of the sample cell poor quality root cause meets the accuracy threshold requirement;

[0061] a setting unit connected to the determination unit, configured to set the root cause model sequence finally determined as the target root cause identification model sequence when the determination unit determines that the maximum number of iterations has been reached but the accuracy requirement has not been met; and, configured to set the root cause model sequence at that time as the target root cause identification model sequence when the determination unit determines that the identification accuracy of the root cause of poor quality of the sample cell at a certain moment when the maximum preset number of iterations has not been reached meets the threshold requirement;

[0062] The second processing unit is connected to the setting unit, and is used to obtain a target network poor quality cell; and to identify the root cause of the target network poor quality cell according to a target root cause identification model sequence.

[0063] Furthermore, the acquisition unit includes:

[0064] A collection module is used to collect all network indicators of sample cells within the target range;

[0065] The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators;

[0066] A comparison module, connected to the acquisition module, for comparing the network indicator with a preset threshold;

[0067] A determination module, connected to the comparison module, for determining that a certain network indicator exceeds a preset threshold when the comparison module determines that the sample cell is a poor quality cell;

[0068] The extraction module is connected to the determination module and is used to extract the network indicators of the sample cells with poor quality determined by the determination module from all the network indicators of the sample cells within the target range.

[0069] Furthermore, the construction unit comprises:

[0070] A classification module, used for classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells;

[0071] Among them, each pre-defined root cause category has its corresponding initial learning model;

[0072] A construction module, connected to the classification module, for constructing a corresponding training data set for each root cause category;

[0073] The training module is connected to the construction module and is used to train the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

[0074] Furthermore, the first processing unit includes:

[0075] A division module, used to divide the sample quality cells into different attribute categories or scene categories according to the attribute data or scene data of the sample quality cells through a predefined standard or a clustering algorithm;

[0076] Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells;

[0077] A selection module, connected to the division module, for selecting a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model using a multi-armed bandit algorithm according to the attribute category or the scene category;

[0078] The selection module stores the following selection root cause model sequence formula:

[0079] a=argmax a∈A (Q(a)+c*U(a));

[0080] Among them, argmax() represents the maximum value function;

[0081] a represents the selected root cause model sequence;

[0082] A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model;

[0083] Q(a) represents the average reward obtained by all a;

[0084] c is a constant, and its value range is (0,1);

[0085] U(a) represents the uncertainty measure;

[0086] The uncertainty metric is used to determine the upper limit of the reward, and its calculation formula is stored in the selection module. The specific formula is as follows:

[0087]

[0088] Where N(a) represents the number of times a is selected;

[0089] logt represents the logarithm of t, where t is the current iteration number.

[0090] Furthermore, the determination unit includes:

[0091] Verification module, used for each attribute category or scene category to use sequence a for the tth time t The root causes of poor quality of the identified sample plots are verified with the real root causes of the sample plots;

[0092] A reward module is connected to the verification module and is used to store a reward r when the verification result of the verification module is greater than a threshold D. t =1; also used for storing reward r when the verification result of the verification module is not greater than the threshold D t =0;

[0093] A calculation module, connected to the reward module, for calculating the root cause model sequence a t The average reward value of

[0094] The calculation module stores a root cause model sequence a t The counter N(a t )=N(a t-1 )+1;

[0095] Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0;

[0096] The calculation module stores a root cause model sequence at The average reward value

[0097]

[0098] Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0;

[0099] A comparison module, connected to the calculation module, for comparing the root cause model sequence a t The average reward value and the preset average reward value;

[0100] A determination module connected to the comparison module is used to compare the root cause model sequence a in the comparison module. t When the average reward value of is lower than the preset average reward value, it is determined that the accuracy of the root cause of the poor quality of the sample cell does not meet the accuracy threshold requirement; and it is also used to compare the root cause model sequence a in the comparison module. t When the average reward value is not lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

[0101] Beneficial effects of the present invention:

[0102] The present invention can improve the precision and accuracy of identifying the root cause of poor network quality cells, and specifically includes the following contents:

[0103] 1. Accurate root cause identification:

[0104] The present invention can provide more accurate root cause identification by integrating multiple models, combining multiple machine learning models such as XGBoost, random forest, neural network, etc. This is because the advantages of different models complement each other, improving the overall recognition performance and reducing the recognition bias caused by a single model.

[0105] 2. Improved the flexibility and adaptability of the scene:

[0106] The present invention establishes corresponding identification models for different network scenarios, thereby improving the applicability of the model in various environments and ensuring that the accuracy of root cause identification is closely related to the actual situation.

[0107] 3. Real-time dynamic updates:

[0108] The present invention can continuously adjust and optimize the recognition model based on the latest collected data, so that the recognition algorithm can adapt to changes in network status. This real-time improvement effectively improves the prediction accuracy, ensuring that the model can respond quickly and maintain a stable recognition effect when the environment changes, user behavior changes, or the network structure is updated.

[0109] 4. Intelligent optimization and learning:

[0110] The present invention introduces a multi-armed bandit algorithm, which enables the system to intelligently explore and optimize in different models and parameter selections. The multi-armed bandit algorithm can help automatically select the optimal model configuration, thereby continuously improving recognition performance in practical applications and reducing reliance on manual intervention.

[0111] 5. Enhanced system robustness:

[0112] The present invention integrates multiple algorithms and dynamic adaptive characteristics, so that the identification method can still maintain strong robustness when facing uncertain and changing network environments, thus ensuring the stability and reliability of the root cause identification of poor quality cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 Schematic diagram of a method for identifying root causes of cells with poor network quality in an embodiment of the present invention;

[0114] Figure 2 Schematic diagram of the root cause identification process of a cell with poor network quality in an embodiment of the present invention;

[0115] Figure 3 Schematic diagram of a root cause identification framework for a cell with poor network quality in an embodiment of the present invention;

[0116] Figure 4 Schematic diagram of a root cause identification device for a cell with poor network quality in an embodiment of the present invention;

[0117] Among them, the figure numerals are: 10, acquisition unit, 20, construction unit, 30, first processing unit, 40, determination unit, 50, setting unit, 60, second processing unit. DETAILED DESCRIPTION

[0118] In order 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.

[0119] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0120] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0121] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0122] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0123] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0124] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0125] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0126] Embodiment 1:

[0127] like Figure 1 and Figure 2 As shown, this embodiment provides a method for identifying the root cause of a cell with poor network quality, and the method comprises the following steps:

[0128] Step S1: Obtain network indicators of cells with poor sample quality.

[0129] Specifically, step S1 includes the following steps:

[0130] Step S11: Collect all network indicators of sample cells within the target range;

[0131] The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators;

[0132] Step S12: comparing the network indicator with a preset threshold;

[0133] Among them, if a certain network indicator exceeds a preset threshold, the sample cell is determined to be a poor quality cell;

[0134] Step S13: extracting the network indicators of the sample poor-quality cells from all network indicators of the sample cells within the target range.

[0135] Assume that a group of base stations are selected for analysis in a certain urban area. The network indicators of these base stations may include:

[0136] Performance indicators: such as data transmission rate (e.g. download rate and upload rate), network delay, packet loss rate, etc.

[0137] Configuration parameters: such as base station transmission power, frequency, antenna type, coverage radius, etc.

[0138] Measurement report: signal strength, signal-to-noise ratio, and number of users of the user terminal in the cell.

[0139] Alarm indicators: such as equipment failure records, alarm information when flow exceeds preset limits, etc.

[0140] The network indicator is then compared with a preset threshold:

[0141] Assume that some thresholds are set to ensure network quality. For example:

[0142] Data transfer rates below 5Mbps are considered low performance.

[0143] Network latency over 100ms is considered high latency.

[0144] Packet loss rates above 2% are considered an indicator of concern.

[0145] When performing the comparison:

[0146] If a base station records a download rate of 3Mbps (less than 5Mbps), the indicator exceeds the preset threshold.

[0147] Likewise, if the network latency is 120ms, it also exceeds the threshold.

[0148] Therefore, base stations that meet the above conditions will be judged as poor quality cells.

[0149] Finally, the network indicators of the sample cells with poor quality are extracted from all the network indicators of the sample cells within the target range:

[0150] For base stations that have been determined to be of poor quality (for example, base stations with a download rate of 3Mbps and a latency of 120ms), extract the relevant network indicators of these base stations for subsequent analysis:

[0151] Performance indicators: download rate (3Mbps), upload rate (maybe 2Mbps), network latency (120ms), packet loss rate (for example, 3%).

[0152] Configuration parameters: transmit power (e.g., 30 dBm), frequency (e.g., 1800 MHz), antenna type.

[0153] Measurement report: recorded signal strength (perhaps -95dBm), number of users (e.g., 50 simultaneously connected users).

[0154] Alarm indicators: recent equipment failure records and flow alarm information.

[0155] During implementation, this also includes preprocessing the data, that is, cleaning and standardizing the collected network indicators to ensure the accuracy of subsequent analysis. Specific operations include:

[0156] Remove missing values: For processing missing network indicators, you can choose to fill them (such as using the mean, median, etc.), or directly delete samples containing missing values.

[0157] Remove outliers: Identify and remove metric values ​​that are obviously unreasonable, such as extremely high or low transfer rates.

[0158] Step S2: constructing an initial poor quality cell root cause identification model based on the network indicators of the sample poor quality cells.

[0159] Specifically, the step S2 comprises the following steps:

[0160] Step S21: classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells;

[0161] Each predefined root cause category has its corresponding initial learning model; the initial learning model is randomly selected from a supervised learning model or an unsupervised learning model. Supervised learning models: such as decision trees, random forests, support vector machines, neural networks, etc. Unsupervised learning models: such as clustering models (K-means, DBSCAN, etc.) are used for root cause analysis.

[0162] Step S22: construct a corresponding training data set for each root cause category;

[0163] Step S23: training the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

[0164] Step S3: determining a root cause model sequence according to the initial poor quality cell root cause identification model; and traversing the root cause model sequence to identify the poor quality root causes of the sample poor quality cells to obtain the poor quality root causes of the sample cells.

[0165] Specifically, in step S3, according to the initial poor quality cell root cause identification model, a sample root cause model sequence is determined, specifically:

[0166] Determine a root cause model sequence according to the initial poor quality cell root cause identification model and a multi-armed bandit algorithm;

[0167] The method of determining a sample root cause model sequence according to the initial poor quality cell root cause identification model and using a multi-armed bandit algorithm specifically includes the following steps:

[0168] Step S31: according to the attribute data or scene data of the sample poor quality cells, the sample poor quality cells are divided into different attribute categories or scene categories by a predefined standard or a clustering algorithm;

[0169] Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells;

[0170] Step S32: According to the attribute category or scenario category, a multi-armed bandit algorithm is used to select a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model;

[0171] The formula for selecting the root cause model sequence is as follows:

[0172] a=argmax a∈A (Q(a)+c*U(a));

[0173] Among them, argmax() represents the maximum value function;

[0174] a represents the selected root cause model sequence;

[0175] A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model;

[0176] Q(a) represents the average reward obtained by all a;

[0177] c is a constant, and its value range is (0,1);

[0178] U(a) represents the uncertainty measure;

[0179] The uncertainty measure is used to determine the upper limit of the reward, and its calculation formula is as follows:

[0180]

[0181] Where N(a) represents the number of times a is selected;

[0182] logt represents the logarithm of t, where t is the current iteration number.

[0183] Step S4: Determine whether the accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement:

[0184] If the identification accuracy of the root cause of poor quality of the sample cell does not reach the preset accuracy threshold, steps S2 to S4 are repeatedly performed until any of the following conditions are met: if the maximum preset number of iterations is reached, or the identification accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement;

[0185] If the maximum number of iterations is reached but the accuracy requirement is still not met, the finally determined root cause model sequence is set as the target root cause identification model sequence; if the identification accuracy of the root cause of poor quality of the sample cell meets the threshold requirement at a certain moment when the maximum preset number of iterations is not reached, the root cause model sequence at that time is set as the target root cause identification model sequence.

[0186] Specifically, in step S4, judging whether the accuracy of the sample cell meets the accuracy threshold requirement specifically includes the following steps:

[0187] For each attribute category or scene category, use sequence a for the tth time t The root cause of the poor quality of the sample cell identified is verified with the real root cause of the sample cell. If the verification result is greater than the threshold D, a reward r is given. t =1, otherwise r t =0;

[0188] Make the root cause model sequence a t The counter N(a t )=N(a t-1 )+1;

[0189] Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0;

[0190] Make the root cause model sequence a t The average reward value

[0191] Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0;

[0192] Compare the root cause model sequence a t The average reward value and the preset average reward value:

[0193] If the root cause model sequence a t If the average reward value is lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell does not meet the accuracy threshold requirement; otherwise, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

[0194] Step S5: Acquire a target cell with poor network quality; and identify the root cause of the target cell with poor network quality according to a target root cause identification model sequence.

[0195] Specifically, this step includes the following specific steps:

[0196] (1) Obtaining the target cell with poor network quality

[0197] Step (1.1): Data collection

[0198] Assume that in a certain city area, the operator monitors the performance indicators of multiple base stations. The following is the network indicator data of a sample base station:

[0199] Base station A: download rate 3Mbps, upload rate 1Mbps, network delay 150ms, packet loss rate 5%

[0200] Base station B: download rate 15Mbps, upload rate 10Mbps, network delay 50ms, packet loss rate 1%

[0201] Base station C: download rate 4Mbps, upload rate 2Mbps, network delay 120ms, packet loss rate 2%

[0202] Base station D: download rate 12Mbps, upload rate 8Mbps, network delay 30ms, packet loss rate 0.5%

[0203] Step (1.2): Indicator comparison and judgment

[0204] Compare based on preset thresholds:

[0205] Download rate threshold: below 5Mbps is considered poor quality

[0206] Network delay threshold: more than 100ms is considered poor quality

[0207] Packet loss rate threshold: more than 2% is considered poor quality

[0208] After comparison:

[0209] Base station A: download rate 3Mbps (poor quality), delay 150ms (poor quality), packet loss rate 5% (poor quality)

[0210] Base station B: poor quality condition not triggered

[0211] Base station C: download rate 4Mbps (poor quality), delay 120ms (poor quality), packet loss rate 2% (not exceeded)

[0212] Base station D: poor quality condition not triggered

[0213] Base station A and base station C are determined as target cells with poor network quality.

[0214] (2) Root cause identification model sequence

[0215] Step (2.1): Select the root cause identification model sequence

[0216] Assume that three different root cause identification models have been established, named:

[0217] Model 1: Based on XGBoost, it mainly predicts issues that affect coverage.

[0218] Model 2: Based on random forest, it mainly predicts the impact of interference sources.

[0219] Model 3: Based on neural network, mainly predicts capacity issues.

[0220] Step (2.2): Input poor quality cell data

[0221] For the determined target cells (base station A and base station C), their relevant network indicators are input into the model sequence:

[0222] Base station A indicators: download rate 3Mbps, upload rate 1Mbps, network delay 150ms, packet loss rate 5%.

[0223] Base station C indicators: download rate 4Mbps, upload rate 2Mbps, network delay 120ms, packet loss rate 2%.

[0224] Step (2.3): Root cause identification

[0225] Input the indicators of each base station into each model to identify the root cause and obtain the following results:

[0226] Base station A: The XGBoost model identifies coverage issues as the root cause of poor quality.

[0227] Base station C: The neural network confirmed that insufficient capacity was the main cause of poor quality.

[0228] Embodiment 2:

[0229] like Figure 1 and Figure 3 As shown, this embodiment provides a method for identifying the root cause of poor quality cells, and the specific implementation steps of the method are as follows:

[0230] Step K1: Discover poor quality cells and collect cell network indicators;

[0231] Specifically, a poor quality cell refers to a cell where the operating indicators of a wireless network base station are abnormally high or low, which affects the cell's ability to serve users and can trigger user complaints when the problem is serious;

[0232] Collecting cell network indicators includes but is not limited to cell key performance indicators PM, cell configuration parameters CM, cell measurement reports MR and cell alarm indicators FM;

[0233] PM, CM, MR and FM are combined according to the cell ID. One cell data consists of PM, CM, MR and FM. The final data format is (cell ID, PM data, CM data, MR data, FM data).

[0234] Step K2: Establishing a root cause identification model for poor quality cells based on cell network indicators;

[0235] Specifically, this embodiment assumes that all root causes of poor quality have been found in advance and no new root causes will be added, and the abnormal root causes of poor quality cells are completely classified manually, such as fault type, coverage type, interference type, load type, neighboring cell abnormal type and parameter configuration abnormal type;

[0236] Each poor-quality cell data is divided into different classes according to the root cause category, and a root cause identification model training data set is established on each class. The data set uses the cell network indicators as features and whether it is the root cause of this type as a label. The data set includes the data of cells with this type of root cause and cells without this type of root cause;

[0237] The fault root cause identification data set is Dgz, which includes fault root cause cell data and non-fault root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, service cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the cell quality root cause is a non-fault root cause, and 1 indicates that the cell quality root cause is a fault root cause. The sample format is (PM data, CM data, MR data, FM data, label). Two sample examples of the fault root cause identification dataset Dgz are as follows:

[0238]

[0239] The coverage root cause identification data set is Dfg, which includes coverage root cause cell data and non-coverage root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, service cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the cell quality root cause is a non-coverage root cause, and 1 indicates that the cell quality root cause is a coverage root cause. The sample format is (PM data, CM data, MR data, FM data, label), and two sample examples covering the root cause identification dataset Dfg are as follows:

[0240]

[0241] The interference root cause identification data set is Dgr, which includes interference root cause cell data and non-interference root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, service cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the cell quality root cause is a non-interference root cause, and 1 indicates that the cell quality root cause is an interference root cause. The sample format is (PM data, CM data, MR data, FM data, label). Two sample examples of the interference root cause identification dataset Dgr are as follows:

[0242]

[0243] The load root cause identification data set is Dfh, which includes load root cause cell data and non-load root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, serving cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the cell quality root cause is a non-load root cause, and 1 indicates that the cell quality root cause is a load root cause. The sample format is (PM data, CM data, MR data, FM data, label). Two sample examples of the load root cause identification dataset Dfh are as follows:

[0244]

[0245]

[0246] The neighboring cell root cause identification data set is Dlq, which includes neighboring cell root cause cell data and non-neighboring cell root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, serving cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the root cause of the poor cell quality is a non-neighboring cell root cause, and 1 indicates that the root cause of the poor cell quality is a neighboring cell root cause. The sample format is (PM data, CM data, MR data, FM data, label). Two sample examples of the neighborhood class root cause identification dataset Dlq are as follows:

[0247]

[0248] The parameter root cause identification data set is Dcs, which includes parameter root cause cell data and non-parameter root cause cell data. The data set consists of sample features and corresponding sample labels. The sample features include but are not limited to RRC establishment success rate, ERAB establishment success rate, call drop rate, same-frequency switching success rate, dual-stream ratio, CQI greater than 6 ratio, inter-frequency switching success rate, RRC reconstruction success rate, cell offset, serving cell bias, switching amplitude hysteresis, switching event RSRP trigger threshold, measurement report reporting interval, LteScRSRP, LteScTadv, LteScSinrUL, antenna equipment maintenance link abnormality alarm, radio frequency unit standing wave alarm, X2 interface fault alarm, radio frequency unit maintenance link abnormality alarm, radio frequency unit CPRI interface abnormality alarm, transmission optical interface abnormality alarm, S1 interface fault alarm, cell service capability degradation alarm, and the label corresponding to the sample feature is a binary variable 0 or 1. 0 indicates that the cell quality root cause is a non-parameter root cause, and 1 indicates that the cell quality root cause is a parameter root cause. The sample format is (PM data, CM data, MR data, FM data, label). Two sample examples of the parameter class root cause identification dataset Dcs are as follows:

[0249]

[0250] After data preprocessing, exploratory analysis, feature engineering, model selection and hyperparameter tuning for the six types of poor quality root cause data sets Dgz, Dfg, Dgr, Dfh, Dlq and Dcs, a binary classification root cause identification model is established, that is, the network indicators of each cell are used to determine whether the cell quality is caused by this type of root cause; the model here includes but is not limited to KNN, support vector machine, random forest, gradient boosting tree, integrated model or deep learning model;

[0251] Each root cause identification model uses different indicator features for modeling, and the method of selecting features is not limited in this embodiment;

[0252] At this point, corresponding to the above-mentioned root causes, we can obtain the fault root cause identification model Mgz, coverage root cause identification model Mfg, interference root cause identification model Mgr, load root cause identification model Mfh, neighboring root cause identification model Mlq and parameter configuration root cause identification model Mcs, and add these root cause identification models to the root cause identification model set;

[0253] The six models in the root cause identification model set of this embodiment can form K=6! =6*5*4*3*2 different root cause model sequence sets Α={a1, a2, ..., a K}, a k ∈Α represents any root cause model sequence, each sequence a k The length of a is 6, kThe models in the sequence are arranged from front to back according to the recognition order. The models in the front are used for recognition first, and the models in the back are used for recognition later. Each model appears in a sequence and only appears once. For example, the root cause model sequences of two arrangement orders are [Mgz, Mcs, Mfg, Mfh, Mlq, Mgr] and [Mfh, Mfg, Mgr, Mlq, Mgz, Mcs].

[0254] Step K3: determining a root cause model sequence according to a cell quality poor root cause identification model sequence selection algorithm, and traversing the root cause model sequence to identify the root cause of the cell quality poor;

[0255] Specifically, the cell quality root cause identification model sequence selection algorithm refers to selecting a root cause model sequence according to a certain rule, and using the root cause model sequence to identify the root cause of the quality root cause of the cell to be analyzed;

[0256] To improve the accuracy of identification, the cells can be divided into different scenarios to apply the root cause model sequence to identify the root causes of poor quality cells;

[0257] There is no limit to the classification criteria of scenes. It can be geographical scenes including general urban scenes, core business district scenes, high-rise residential scenes, large venue scenes, subway scenes, indoor scenes, etc. It can also be a clustering scene obtained by using KMeans, DBSCAN and other clustering algorithms based on specific community data. Each community belongs to only one scene, and one scene corresponds to multiple communities.

[0258] The root cause identification models of cells with the same root cause category in different scenarios are the same, so the root cause model sequences with the same identification order in different scenarios are also the same;

[0259] Assume that the scenario is divided into three scenarios S1-S3, and the root cause model of each scenario is the same.

[0260] Root cause model of scenario S1: Mgz, Mfg, Mgr, Mfh, Mlq and Mcs,

[0261] Root cause model of scenario S2: Mgz, Mfg, Mgr, Mfh, Mlq and Mcs,

[0262] Root cause model of scenario S2: Mgz, Mfg, Mgr, Mfh, Mlq and Mcs,

[0263] The steps of the cell quality root cause identification model sequence selection algorithm in each scenario of this embodiment are as follows:

[0264] For each scenario, for any root cause model sequence a∈Α, A is the root cause model sequence set obtained in step S102, initialize the counter N(a)=0 and the average reward value Q(a)=0, N(a) represents the number of times sequence a has appeared in the past, and Q(a) represents the average reward r obtained by sequence a in the past;

[0265] Compare the accuracy of the recognition result of sequence a with the predetermined threshold D to obtain the reward r∈{0,1} of the sequence. If the sequence recognition accuracy is higher than the threshold, the reward r=1, otherwise r=0. In this embodiment, the recognition accuracy is taken as the correct rate (the number of correctly recognized samples / the total number of samples);

[0266] For the t-th recognition task, the root cause recognition sequence a is selected according to the following formula: t , where the coefficient c∈(0,1),

[0267]

[0268] The a on the left side of the equal sign represents the root cause model sequence in the sequence set A used in this recognition task. The right side of the equal sign represents the selection of sequence a from set A that maximizes Q(a)+c·U(a), where Q(a) represents the average reward r obtained by sequence a in the past, c is a decimal between 0 and 1, and U(a) represents the uncertainty measure used to determine the upper bound of the reward. The fewer times sequence a is used, the higher the uncertainty of its reward. The uncertainty-based strategy comprehensively considers the existing average reward and uncertainty. argmax is the parameter a that takes the maximum value.

[0269] Take out the root cause model sequence a t The root cause models in the sequence identify the poor quality cells in order. If the current model determines that it is this type of root cause, stop traversing the model sequence and return the root cause type. If it is not this type of root cause, continue to use the next model in the sequence for identification and traverse all models in the sequence until the root cause is found. If no root cause is found after traversing all models in the sequence, return a null value.

[0270] The following analysis is based on scenario S1. Other scenarios are analyzed as in S1. The models in the model sequence of each scenario are the same. The model sequences of different scenarios in the same round may be different or the same.

[0271] For scene S1, during the first recognition, the root cause sequence a = [Mgz, Mcs, Mfg, Mfh, Mlq, Mgr], counter N(a) = 0 and average reward value Q(a) = 0 are initialized. A sample data d is recognized. First, the recognition result of model Mgz is 0, then the recognition result of model Mcs is 0, and then the recognition result of model Mfg is 1. The root cause of the poor quality of sample d is determined to be the coverage class, and the traversal is stopped and the root cause type is returned. The results are as follows:

[0272]

[0273] At the 36th recognition, c=0.5, the average reward Q(a), counter N(a), and uncertainty measure U(a) of each sequence a in the root cause model sequence set A are distributed as follows:

[0274]

[0275] The best sequence a calculated from these distributions is [Mfh, Mgr, Mcs, Mgz, Mfg, Mlq], as shown in the following table:

[0276]

[0277] Use [Mfh, Mgr, Mcs, Mgz, Mfg, Mlq] to identify the poor quality cells in scene S1 for the 36th time and return the identification result of each sample.

[0278] Step K4: Check the root cause identification results of poor cell quality, and update the sequence selection algorithm parameters and root cause identification model.

[0279] Specifically, for each scene, the sequence a is used for the tth time t The root cause of the poor cell quality identified is checked against the real root cause, and the root cause identification accuracy is calculated. If it is greater than the threshold D, a reward r is given. t =1, otherwise r t =0;

[0280] Update sequence a t Counter N(a t )=N(a t-1 )+1;

[0281] Update sequence a t Average reward value

[0282] Add the cell data samples that were misidentified to the data set of the actual root cause class to continue training and update the root cause model of the actual type;

[0283] Return to step K3 and select the root cause model sequence a for the t+1th recognition task based on the updated Q and U t+1 , according to sequence a t+1 Execute this step repeatedly. Q is the average reward of model sequence a, and U is the uncertainty measure.

[0284] The recognition results of all samples identified for the first time are checked against the real root cause, and the recognition accuracy is calculated to be 0.92. Compared with the threshold D = 0.9, it is greater than the threshold, and the reward r = 1 is given. The counter N(a) = 1 and the average reward value Q(a) = 1 are updated. At the same time, the wrong cell samples are added to the actual root cause class data set for continued training, and the root cause model is updated.

[0285] The 36th recognition result is checked against the real root cause, and the recognition accuracy is calculated to be 0.85. Compared with the threshold D = 0.9, it is less than the threshold, and the reward r = 0. For a = [Mfh, Mgr, Mcs, Mgz, Mfg, Mlq], the counter N(a) = 37 is updated, and the average reward value Q(a) = 3.84 + (0-3.84) / 37 = 3.74. At the same time, the wrong cell samples are added to the actual root cause class data set for continued training, and the root cause model is updated.

[0286] As another implementation method of this embodiment, in order to improve the accuracy of identification, the cell can be divided into different scenes to apply the root cause model sequence to identify the root cause of poor quality cells; the scene division standard is not limited, and it can be a geographical scene including a general urban scene, a core business district scene, a high-rise residential community scene, a large venue scene, a subway scene, a room scene, etc., or it can be a clustering scene obtained by using a clustering algorithm such as KMeans and DBSCAN through specific cell data; each cell belongs to only one scene, and one scene corresponds to multiple cells;

[0287] This embodiment assumes that all root causes of poor quality have been found in advance and no new root causes will be added. The abnormal root causes of poor quality cells are completely classified manually, such as fault type, coverage type, interference type, load type, neighboring cell abnormal type and parameter configuration abnormal type.

[0288] Different from the previous implementation method, according to the scenario to which the cell belongs, the cell data of each root cause class is established in each scenario, each poor quality cell is divided into each class according to the root cause category, and a root cause identification model training data set is established on each class. The data set uses the cell network indicator as a feature and whether it is the root cause of this class as a label. The data set includes the data of the cell with the root cause of this class and the cell without the root cause of this class;

[0289] Different from the previous implementation of this embodiment, each scene contains data sets of all quality difference types of the cell in the scene, and each scene is independently modeled using its own data set;

[0290] Assume that three scenarios are divided into S1-S3, and six root cause class datasets are established for each scenario:

[0291] Scene S1 datasets S1_Dgz, S1_Dfg, S1_Dgr, S1_Dfh, S1_Dlq, and S1_Dcs;

[0292] Scene S2 datasets S2_Dgz, S2_Dfg, S2_Dgr, S2_Dfh, S2_Dlq, and S2_Dcs;

[0293] Scene S3 datasets S3_Dgz, S3_Dfg, S3_Dgr, S3_Dfh, S3_Dlq, and S3_Dcs;

[0294] For each scenario, six types of root cause data sets Dgz, Dfg, Dgr, Dfh, Dlq and Dcs are subjected to data preprocessing, exploratory analysis, feature engineering, model selection and hyperparameter tuning to establish a binary classification root cause identification model, that is, to determine whether the root cause of the poor quality of each cell is this type of root cause based on the network indicators of each cell; the models here include but are not limited to KNN, support vector machine, random forest, gradient boosting tree, integrated model or deep learning model;

[0295] Each root cause identification model uses different indicator features for modeling, and the method of selecting features is not limited in this embodiment;

[0296] At this point, each scenario corresponds to the above-mentioned root causes, and the fault root cause identification model Mgz, coverage root cause identification model Mfg, interference root cause identification model Mgr, load root cause identification model Mfh, neighboring root cause identification model Mlq and parameter configuration root cause identification model Mcs of each scenario can be obtained. These root cause identification models are added to the root cause model set of each scenario;

[0297] Scenario 1 root cause model set S1_Mgz, S1_Mfg, S1_Mgr, S1_Mfh, S1_Mlq and S1_Mcs;

[0298] Scenario 2 root cause model set S2_Mgz, S2_Mfg, S2_Mgr, S2_Mfh, S2_Mlq and S2_Mcs;

[0299] Scenario 3 root cause model set S3_Mgz, S3_Mfg, S3_Mgr, S3_Mfh, S3_Mlq and S3_Mcs;

[0300] In this embodiment, the six models in the root cause identification model set of each scene can form K=6! =6*5*4*3*2 different root cause model sequence sets Α={a1, a2, ..., a K}, a k ∈Α represents any root cause model sequence, each sequence a k The length of a is the number of elements in the model set, 6. kThe models in the sequence are arranged from front to back according to the recognition order. The models in the front are used for recognition first, and the models in the back are used for recognition later. Each model appears in a sequence and only appears once. For example, in scenario S1, the root cause model sequences of two arrangement orders are [S1_Mgz, S1_Mcs, S1_Mfg, S1_Mfh, S1_Mlq, S1_Mgr] and [S1_Mfh, S1_Mfg, S1_Mgr, S1_Mlq, S1_Mgz, S1_Mcs], and the same applies to other scenarios.

[0301] Different from the previous implementation, the root cause models of cells of the same root cause category in different scenarios are different, so the root cause model sequences of the same identification order in different scenarios are also different;

[0302] The steps of the cell quality root cause identification model sequence selection algorithm in each scenario of this implementation method are as follows:

[0303] In each scenario, for any recognition sequence a∈Α, initialize the counter N(a)=0 and the average reward value Q(a)=0, where N(a) represents the number of times sequence a has appeared in the past, and Q(a) represents the average reward r obtained by sequence a in the past;

[0304] Compare the accuracy of the recognition result of sequence a with the predetermined threshold D to obtain the reward r∈{0,1} of the sequence. If the sequence recognition accuracy is higher than the threshold, the reward r=1, otherwise r=0. In this embodiment, the recognition accuracy is taken as the correct rate (the number of correctly recognized samples / the total number of samples);

[0305] For the t-th recognition task, the root cause recognition sequence a is selected according to the following formula: t , where the coefficient c∈(0,1),

[0306]

[0307] Take out the recognition sequence a t The root cause models in the sequence identify the poor quality cells in order. If the current model determines that it is this type of root cause, stop traversing the model sequence and return the root cause type. If it is not this type of root cause, continue to use the next model in the sequence for identification and traverse all models in the sequence until the root cause is found. If no root cause is found after traversing all models in the sequence, return a null value.

[0308] Each scene is updated according to its own parameters.

[0309] This embodiment corresponds to Embodiment 1.

[0310] Embodiment 3:

[0311] like Figure 4As shown, this embodiment provides a root cause identification device for a cell with poor network quality, the device comprising:

[0312] An acquisition unit 10 is used to acquire network indicators of sample poor quality cells;

[0313] A construction unit 20, connected to the acquisition unit 10, is used to construct an initial poor quality cell root cause identification model according to the network indicators of the sample poor quality cells;

[0314] The first processing unit 30 is connected to the construction unit 20, and is used to determine a root cause model sequence according to the initial poor quality cell root cause identification model; and is used to traverse the root cause model sequence to identify the poor quality root cause of the sample poor quality cell, and obtain the poor quality root cause of the sample cell;

[0315] The determination unit 40 is connected to the first processing unit 30 and is used to determine whether the accuracy of the root cause of the poor quality of the sample cell meets the accuracy threshold requirement;

[0316] The construction unit 20 is also connected to the determination unit 40, and is used to re-construct an initial poor quality cell root cause identification model according to the network indicators of the sample poor quality cell when the determination unit 40 determines that the identification accuracy of the sample cell poor quality root cause does not reach the preset accuracy threshold, until any of the following conditions is met: the maximum preset number of iterations is reached, or the identification accuracy of the sample cell poor quality root cause meets the accuracy threshold requirement;

[0317] The setting unit 50 is connected to the determination unit 40 and is used to set the root cause model sequence finally determined as the target root cause identification model sequence when the determination unit determines that the maximum number of iterations has been reached but the accuracy requirement has not been met; and when the determination unit determines that the identification accuracy of the root cause of poor quality of the sample cell at a certain moment when the maximum number of preset iterations has not been reached, the root cause model sequence at that time is set as the target root cause identification model sequence;

[0318] The second processing unit 60 is connected to the setting unit 50, and is used to obtain a target network poor quality cell; and to identify the root cause of the target network poor quality cell according to the target root cause identification model sequence.

[0319] As a specific implementation, the acquisition unit 10 includes:

[0320] A collection module is used to collect all network indicators of sample cells within the target range;

[0321] The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators;

[0322] A comparison module, connected to the acquisition module, for comparing the network indicator with a preset threshold;

[0323] A determination module, connected to the comparison module, for determining that a certain network indicator exceeds a preset threshold when the comparison module determines that the sample cell is a poor quality cell;

[0324] The extraction module is connected to the determination module and is used to extract the network indicators of the sample cells with poor quality determined by the determination module from all the network indicators of the sample cells within the target range.

[0325] As a specific implementation, the construction unit 20 includes:

[0326] A classification module, used for classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells;

[0327] Among them, each pre-defined root cause category has its corresponding initial learning model;

[0328] A construction module, connected to the classification module, for constructing a corresponding training data set for each root cause category;

[0329] The training module is connected to the construction module and is used to train the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

[0330] As a specific implementation, the first processing unit 30 includes:

[0331] A division module, used to divide the sample quality cells into different attribute categories or scene categories according to the attribute data or scene data of the sample quality cells through a predefined standard or a clustering algorithm;

[0332] Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells;

[0333] A selection module, connected to the division module, for selecting a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model using a multi-armed bandit algorithm according to the attribute category or the scene category;

[0334] The selection module stores the following selection root cause model sequence formula:

[0335] a=argmax a∈A (Q(a)+c*U(a));

[0336] Among them, argmax() represents the maximum value function;

[0337] a represents the selected root cause model sequence;

[0338] A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model;

[0339] Q(a) represents the average reward obtained by all a;

[0340] c is a constant, and its value range is (0,1);

[0341] U(a) represents the uncertainty measure;

[0342] The uncertainty metric is used to determine the upper limit of the reward, and its calculation formula is stored in the selection module. The specific formula is as follows:

[0343]

[0344] Where N(a) represents the number of times a is selected;

[0345] logt represents the logarithm of t, where t is the current iteration number.

[0346] As a specific implementation, the determination unit 40 includes:

[0347] Verification module, used for each attribute category or scene category to use sequence a for the tth time t The root causes of poor quality of the identified sample plots are verified with the real root causes of the sample plots;

[0348] A reward module is connected to the verification module and is used to store a reward r when the verification result of the verification module is greater than a threshold D. t =1; also used for storing reward r when the verification result of the verification module is not greater than the threshold D t =0;

[0349] A calculation module, connected to the reward module, for calculating the root cause model sequence a t The average reward value of

[0350] The calculation module stores a root cause model sequence a t The counter N(a t )=N(a t-1 )+1;

[0351] Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0;

[0352] The calculation module stores a root cause model sequence a t The average reward value

[0353]

[0354] Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0;

[0355] A comparison module, connected to the calculation module, for comparing the root cause model sequence a t The average reward value and the preset average reward value;

[0356] A determination module connected to the comparison module is used to compare the root cause model sequence a in the comparison module. t When the average reward value of is lower than the preset average reward value, it is determined that the accuracy of the root cause of the poor quality of the sample cell does not meet the accuracy threshold requirement; and it is also used to compare the root cause model sequence a in the comparison module. t When the average reward value is not lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

[0357] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles 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 to be within the scope of protection of the present invention.

Claims

1. A method for identifying the root cause of a cell with poor network quality, characterized in that: The method comprises the following steps: Step S1: Obtain network indicators of sample poor quality cells; Step S2: constructing an initial root cause identification model for poor quality cells according to the network indicators of the sample poor quality cells; Step S3: determining a root cause model sequence according to the initial poor quality cell root cause identification model; and traversing the root cause model sequence to identify the poor quality root causes of the sample poor quality cells to obtain the poor quality root causes of the sample cells; Step S4: Determine whether the accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement: If the identification accuracy of the root cause of poor quality of the sample cell does not reach the preset accuracy threshold, steps S2 to S4 are repeatedly performed until any of the following conditions are met: if the maximum preset number of iterations is reached, or the identification accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement; If the maximum number of iterations is reached but the accuracy requirement is still not met, the last determined root cause model sequence is set as the target root cause identification model sequence; if the identification accuracy of the root cause of poor quality of the sample cell meets the threshold requirement at a certain moment when the maximum preset number of iterations is not reached, the root cause model sequence at that time is set as the target root cause identification model sequence; Step S5: Acquire a target cell with poor network quality; and identify the root cause of the target cell with poor network quality according to a target root cause identification model sequence.

2. The method for identifying the root cause of a poor network quality cell according to claim 1, characterized in that: The step S1 specifically includes the following steps: Step S11: Collect all network indicators of sample cells within the target range; The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators; Step S12: comparing the network indicator with a preset threshold; Among them, if a certain network indicator exceeds a preset threshold, the sample cell is determined to be a poor quality cell; Step S13: extracting the network indicators of the sample poor-quality cells from all network indicators of the sample cells within the target range.

3. The method for identifying the root cause of a cell with poor network quality according to claim 1, characterized in that: The step S2 specifically includes the following steps: Step S21: classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells; Among them, each pre-defined root cause category has its corresponding initial learning model; Step S22: construct a corresponding training data set for each root cause category; Step S23: training the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

4. The method for identifying the root cause of poor network quality cells according to claim 1, characterized in that: In step S3, a sample root cause model sequence is determined according to the initial poor quality cell root cause identification model, specifically: Determine a root cause model sequence according to the initial poor quality cell root cause identification model and a multi-armed bandit algorithm; The method of determining a sample root cause model sequence according to the initial poor quality cell root cause identification model and using a multi-armed bandit algorithm specifically includes the following steps: Step S31: according to the attribute data or scene data of the sample poor quality cells, the sample poor quality cells are divided into different attribute categories or scene categories by a predefined standard or a clustering algorithm; Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells; Step S32: According to the attribute category or scenario category, a multi-armed bandit algorithm is used to select a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model; The formula for selecting the root cause model sequence is as follows: a=argmax a∈A (Q(a)+c*U(a)); Among them, argmax() represents the maximum value function; a represents the selected root cause model sequence; A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model; Q(a) represents the average reward obtained by all a; c is a constant, and its value range is (0,1); U(a) represents the uncertainty measure; The uncertainty measure is used to determine the upper limit of the reward, and its calculation formula is as follows: Where N(a) represents the number of times a is selected; logt represents the logarithm of t, where t is the current iteration number.

5. The method for identifying the root cause of a cell with poor network quality according to any one of claims 1 to 4, characterized in that: In step S4, judging whether the accuracy of the sample cell meets the accuracy threshold requirement specifically includes the following steps: For each attribute category or scene category, use sequence a for the tth time t The root cause of the poor quality of the sample cell identified is verified with the real root cause of the sample cell. If the verification result is greater than the threshold D, a reward r is given. t =1, otherwise r t =0; Make the root cause model sequence a t The counter N(a t )=N(a t-1 )+1; Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0; Make the root cause model sequence a t The average reward value Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0; Compare the root cause model sequence a t The average reward value and the preset average reward value: If the root cause model sequence a t If the average reward value is lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell does not meet the accuracy threshold requirement; otherwise, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

6. A root cause identification device for a cell with poor network quality, characterized in that: The device comprises: An acquisition unit, used to acquire network indicators of cells with poor sample quality; A construction unit, connected to the acquisition unit, configured to construct an initial poor-quality cell root cause identification model according to the network indicators of the sample poor-quality cells; A first processing unit is connected to the construction unit, and is used to determine a root cause model sequence according to the initial poor quality cell root cause identification model; and is used to traverse the root cause model sequence to identify the poor quality root cause of the sample poor quality cell, and obtain the poor quality root cause of the sample cell; A determination unit, connected to the first processing unit, configured to determine whether the accuracy of the root cause of poor quality of the sample cell meets the accuracy threshold requirement; The construction unit is also connected to the determination unit, and is used to re-construct an initial poor quality cell root cause identification model according to the network indicators of the sample poor quality cell when the determination unit determines that the identification accuracy of the sample cell poor quality root cause does not reach the preset accuracy threshold, until any of the following conditions is met: a maximum preset number of iterations is reached, or the identification accuracy of the sample cell poor quality root cause meets the accuracy threshold requirement; a setting unit connected to the determination unit, configured to set the root cause model sequence finally determined as the target root cause identification model sequence when the determination unit determines that the maximum number of iterations has been reached but the accuracy requirement has not been met; and, configured to set the root cause model sequence at that time as the target root cause identification model sequence when the determination unit determines that the identification accuracy of the root cause of poor quality of the sample cell at a certain moment when the maximum preset number of iterations has not been reached meets the threshold requirement; The second processing unit is connected to the setting unit and is used to obtain a target network poor quality cell; and to identify the root cause of the target network poor quality cell according to a target root cause identification model sequence.

7. The root cause identification device for poor network quality cells according to claim 6, characterized in that: The acquisition unit comprises: A collection module is used to collect all network indicators of sample cells within the target range; The network indicators include performance indicators, configuration parameters, measurement reports, and alarm indicators; A comparison module, connected to the acquisition module, for comparing the network indicator with a preset threshold; A determination module, connected to the comparison module, for determining that a certain network indicator exceeds a preset threshold when the comparison module determines that the sample cell is a poor quality cell; The extraction module is connected to the determination module and is used to extract the network indicators of the sample cells with poor quality determined by the determination module from all the network indicators of the sample cells within the target range.

8. The root cause identification device for poor network quality cells according to claim 6, characterized in that: The building block comprises: A classification module, used for classifying the sample poor quality cells into predefined root cause categories according to the network indicators of the sample poor quality cells; Among them, each pre-defined root cause category has its corresponding initial learning model; A construction module, connected to the classification module, for constructing a corresponding training data set for each root cause category; The training module is connected to the construction module and is used to train the initial learning model according to the training data set to obtain an initial poor quality cell root cause identification model.

9. The root cause identification device of poor network quality cells according to claim 6, characterized in that: The first processing unit comprises: A division module, used to divide the sample quality cells into different attribute categories or scene categories according to the attribute data or scene data of the sample quality cells through a predefined standard or a clustering algorithm; Among them, if the sample quality cells are divided into different scene categories, each sample cell belongs to one scene category, and one scene category corresponds to multiple sample cells; A selection module, connected to the division module, for selecting a root cause model sequence from a root cause model sequence set composed of an initial poor quality cell root cause identification model using a multi-armed bandit algorithm according to the attribute category or the scene category; The selection module stores the following selection root cause model sequence formula: a=argmax a∈A (Q(a)+c*U(a)); Among them, argmax() represents the maximum value function; a represents the selected root cause model sequence; A represents the root cause model sequence set composed of the initial poor quality cell root cause identification model; Q(a) represents the average reward obtained by all a; c is a constant, and its value range is (0,1); U(a) represents the uncertainty measure; The uncertainty metric is used to determine the upper limit of the reward, and its calculation formula is stored in the selection module. The specific formula is as follows: Where N(a) represents the number of times a is selected; logt represents the logarithm of t, where t is the current iteration number.

10. The root cause identification device for a cell with poor network quality according to any one of claims 6 to 9, characterized in that: The determination unit comprises: Verification module, used for each attribute category or scene category to use sequence a for the tth time t The root causes of poor quality of the identified sample plots are verified with the real root causes of the sample plots; A reward module is connected to the verification module and is used to store a reward r when the verification result of the verification module is greater than a threshold D. t =1; also used for storing reward r when the verification result of the verification module is not greater than the threshold D t =0; A calculation module, connected to the reward module, for calculating the root cause model sequence a t The average reward value of The calculation module stores a root cause model sequence a t The counter N(a t )=N(a t-1 )+1; Among them, N(a t-1 ) is the root cause model sequence a t-1 The counter of the initial root cause model sequence a0 is N(0)=0; The calculation module stores a root cause model sequence a t The average reward value Among them, Q(a t-1 ) is the root cause model sequence a t-1 The average reward value of the initial root cause model sequence a0 is Q(a0) = 0; A comparison module, connected to the calculation module, for comparing the root cause model sequence a t The average reward value and the preset average reward value; A determination module connected to the comparison module is used to compare the root cause model sequence a in the comparison module. t When the average reward value of is lower than the preset average reward value, it is determined that the accuracy of the root cause of the poor quality of the sample cell does not meet the accuracy threshold requirement; and it is also used to compare the root cause model sequence a in the comparison module. t When the average reward value is not lower than the preset average reward value, the accuracy of determining the root cause of poor quality of the sample cell meets the accuracy threshold requirement.

Images