Network poor quality analysis model training method and device, equipment, medium and product

Abstract

The invention provides a training method and device of a network poor quality analysis model, equipment, a medium and a product, and relates to the technical field of wireless communication. By training the time network and obtaining the backbone model, intelligent analysis and prediction of the operation indexes and the high-advance data can be realized, and improvement of the analysis efficiency of the poor-network-quality work order is facilitated. Through supervised training, fine tuning of an output layer of a backbone model, an anomaly detection branch, a root cause positioning branch and an index prediction branch, prediction of abnormal poor quality, network poor quality reasons and adjusted key operation indexes is further realized, so that intelligent analysis and processing of a network poor quality work order and evaluation of a processing result are realized; and the efficiency of processing the fault in the wireless poor-quality work order is improved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a training method, apparatus, device, medium and product for a network quality poor analysis model. Background Technology

[0002] With the development of the communications industry, mobile communication networks provide network connectivity services to users. The endpoint of a mobile communication network is the wireless access side, and each endpoint corresponds to different wireless receiving and transmitting devices. Different logical cells can be defined based on these devices. A region often has thousands or even tens of thousands of wireless cells. With such a large number of cells, network quality can easily deteriorate when the internal and external environment changes.

[0003] There is a relatively mature operation and maintenance process for handling network quality problems. When a poor quality work order occurs in a certain cell, the operation and maintenance personnel often need to comprehensively analyze the various operating indicators of the cell to determine whether the cell has occasional fluctuations or a real quality problem. For the latter, it is necessary to further determine the cause of the quality problem, optimize network parameters according to the guidance manual, and observe whether the network quality problem is restored.

[0004] However, poor wireless quality issues occur very frequently, and the large number of poor wireless quality work orders results in a heavy workload for maintenance personnel, low fault handling efficiency, and consequently affects customer perception. Summary of the Invention

[0005] This application provides a training method, apparatus, device, medium, and product for a network quality poor analysis model, which addresses the shortcomings of low fault handling efficiency in existing technologies for wireless quality poor work orders and improves the fault handling efficiency of wireless quality poor work orders.

[0006] Firstly, this application provides a training method for a network quality poor analysis model, comprising: Based on the operational metrics and alarm data of wireless cells, a time series network is trained to obtain a backbone model. Add anomaly detection branch, root cause localization branch, and indicator prediction branch to the back end of the backbone model; The output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model are trained in a supervised manner to obtain the network quality analysis model.

[0007] In one embodiment, a time-series network is trained based on the operational metrics and alarm data of the wireless cell to obtain a backbone model, including: Based on operational metrics at the same point in time, obtain metric vectors; Text encoding is performed on the descriptive text in the alarm data to obtain a fixed-dimensional alarm vector; Obtain the average value of alarm vectors within the same sampling period; The index vector and the average value are concatenated to obtain the encoding vector; The encoded vectors are sorted by time to obtain a time vector sequence; Input the time vector sequence of the previous time period into the time series network to obtain the predicted time vector sequence of the key operating indicators in the next time period from the operating indicators output by the time series network. The key operating indicators are the main operating indicators that affect the network quality of the wireless cell. Based on the mean square error of the predicted time vector sequence for the next time period and the actual key operating indicators for the next time period, a loss function for the time series network is constructed. The time series network is trained using the loss function of the time series network to obtain the backbone model.

[0008] In one embodiment, the anomaly detection branch is trained in a supervised manner as follows: Label the normal and abnormal data in the time vector sequence to obtain the first labeled time vector sequence; Based on the backbone model and the anomaly detection branch, the first preset network is obtained; The first labeled time vector sequence is input into the first preset network, and the first prediction probability output by the first preset network is obtained. The first prediction probability includes the probability that the predicted time vector sequence is normal data. Based on the label values ​​and the first predicted probability of the time vector sequence after the first annotation, a loss function of the first preset network is constructed. Based on the loss function of the first preset network, the output layer and anomaly detection branch of the backbone model are fine-tuned.

[0009] In one embodiment, the root cause localization branch is trained in a supervised manner as follows: When the time vector sequence after the first annotation is abnormal data, the reasons for poor network quality in the time vector sequence after the first annotation are annotated to obtain the time vector sequence after the second annotation. Based on the backbone model and root cause localization branch, a second preset network is obtained; The time vector sequence after the second annotation is input into the second preset network, and the second prediction probability output by the second preset network is obtained. The second prediction probability includes the probability of the network quality poorness of the predicted time vector sequence after the first annotation. Based on the label values ​​and second prediction probabilities of the time vector sequence after the second annotation, a loss function for the second preset network is constructed. Based on the loss function of the second preset network, the output layer and root cause localization branch of the backbone model are fine-tuned.

[0010] In one embodiment, the metric prediction branch is trained in a supervised manner as follows: Based on the time vector sequence after the second annotation, obtain the operation strategy of the wireless cell; After adjusting the operational indicators based on the operational strategy, a third-annotated time vector sequence is obtained based on the values ​​of the key operational indicators in the previous time period and the values ​​of the key operational indicators in the next time period. Based on the backbone model and the indicator prediction branch, a third preset network is obtained; The time vector sequence after the third annotation and the operation strategy are input into the third preset network to obtain the predicted value of the next time period output by the third preset network. The predicted value of the next time period is used to characterize the operation status of key operation indicators in the next time period. Based on the predicted values ​​for the next time period and the actual values ​​of the key operating indicators for the next time period, a loss function for the third preset network is constructed. Based on the loss function of the third preset network, the output layer and index prediction branch of the backbone model are fine-tuned.

[0011] Secondly, this application also provides a method for analyzing network quality issues, including: Input the indicators to be analyzed and alarm data of the wireless cell into the network quality poor analysis model, and obtain the analysis results of the indicators to be analyzed output by the network quality poor analysis model. The analysis results include abnormal results of the indicators to be analyzed, the causes of network quality poorness, and predicted key operating indicator values. The network quality analysis model is trained according to the training method of the network quality analysis model as described in any one of claims 1 to 5.

[0012] Thirdly, this application also provides a training device for a network quality analysis model, comprising: The first training module is used to train a time series network based on the operating indicators and alarm data of wireless cells to obtain a backbone model. The building blocks are used to add anomaly detection branches, root cause localization branches, and metric prediction branches to the back end of the backbone model. The second training module is used to perform supervised training on the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model to obtain the network quality analysis model.

[0013] Fourthly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a training method for any of the network quality analysis models described above, or a network quality analysis method.

[0014] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a training method for any of the network quality difference analysis models described above, or a network quality difference analysis method.

[0015] This application also provides a computer program product, including a computer program that, when executed by a processor, implements a training method for any of the network quality degradation analysis models described above, or a network quality degradation analysis method.

[0016] The training method, apparatus, equipment, medium, and product for the network quality defect analysis model provided in this application, through training a time network to obtain a backbone model, enables intelligent analysis and prediction of operational indicators and high-precision data, which is beneficial to improving the analysis efficiency of network quality defect work orders. This application, through supervised training, fine-tunes the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model, further realizing the prediction of abnormal quality defects, the causes of network quality defects, and adjusted key operational indicators. This achieves intelligent analysis, processing, and evaluation of processing results for network quality defect work orders, improving the efficiency of fault handling in wireless quality defect work orders. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the training method of the network quality analysis model provided in this application.

[0019] Figure 2 This is the second flowchart illustrating the training method of the network quality analysis model provided in this application.

[0020] Figure 3 This is a flowchart illustrating the network quality analysis process provided in this application.

[0021] Figure 4 This is a schematic diagram of the training device for the network quality analysis model provided in this application.

[0022] Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The application areas of this application include network quality analysis of wireless cells. For ease of understanding, the relevant terms and concepts that may be involved in the embodiments of this application will be introduced below.

[0025] A wireless cell is a logical coverage area defined by wireless receiving and transmitting equipment (such as base stations, antennas, etc.) at the end of a mobile communication network (e.g., the wireless access side). It is the basic unit by which the network provides wireless connectivity services to users.

[0026] Network quality degradation analysis is a systematic diagnostic and optimization process for network service quality degradation. Its core objective is to identify quality degradation phenomena, locate the root causes, and propose improvement measures through data-driven methods, ultimately restoring or improving network quality and user experience.

[0027] Time series networks are a combination of time series and network structure. At their core, they use time series data (sequences of observations arranged in chronological order) as input or structural elements of the network model. By mining the dependencies in the time dimension and the correlation characteristics of the network structure, they can model, predict, or analyze dynamic systems.

[0028] Supervised training uses labeled training data to teach a model the mapping relationship between input features and target labels, thereby enabling the model to predict or classify new input data. The essence of supervised training is "learning decision rules from labeled data."

[0029] Unsupervised training involves using unlabeled training data to uncover its inherent structure or patterns, allowing the model to autonomously learn the data's distribution. The goal of unsupervised training is to "explore the characteristics of the data itself," rather than "predict external labels." The essence of unsupervised training is "learning the implicit structure of data from unlabeled data."

[0030] A neural network (NN) is a computational model inspired by the neural structure of the biological brain. It learns and models complex data patterns by simulating the connection and information transmission mechanisms of neurons (nodes). Neural networks can automatically extract features from data and perform tasks such as classification, regression, and prediction.

[0031] The following is combined Figures 1-5This application describes the training method, apparatus, and electronic equipment for the network quality analysis model.

[0032] Figure 1 This is one of the flowcharts illustrating the training method of the network quality analysis model provided in this application, such as... Figure 1 As shown, the training method of the network quality difference analysis model includes steps S100 to S300, and the specific steps are as follows.

[0033] S100: Based on the operational metrics and alarm data of wireless cells, a time series network is trained to obtain the backbone model.

[0034] Based on received wireless quality issues work orders, a large amount of operational data from the wireless cells is collected. This operational data includes operational metrics and alarm data, both of which carry timestamps. A time vector sequence is constructed based on the collected timestamped operational metrics and alarm data; this time vector sequence is unsupervised data. A time series network is then trained unsupervised using this time vector sequence, enabling the trained network to accurately predict future time vector sequences based on known time vector sequences, thus obtaining the backbone model.

[0035] Optionally, the time vector sequence of all operating metrics for the next day (e.g., 96 operating metrics for 15 days, including 15×96=1440 timestamps) can be predicted based on the time vector sequence of all operating metrics for multiple days (e.g., 15 days) (e.g., 96 operating metrics for 1 day, including 1×96=96 timestamps).

[0036] Optionally, based on the time vector sequences of all operating metrics over multiple days (e.g., 96 operating metrics over 15 days, including 1440 timestamps), predict the time vector sequences of some key operating metrics for the next day (e.g., K key operating metrics over 1 day, including K timestamps).

[0037] S200: Add anomaly detection branch, root cause localization branch, and indicator prediction branch to the back end of the backbone model.

[0038] The trained backbone model has learned the general characteristics of the operating indicators and alarm data of wireless cells, and can accurately predict future time vector sequences based on known time vector sequences.

[0039] like Figure 2As shown, an anomaly detection branch, root cause localization branch, and metric prediction branch are added to the back end of the backbone model. The anomaly detection branch is used to determine whether the operating metrics are abnormal. That is, it determines whether the current problem of the operating metric corresponding to the poor wireless quality work order is an occasional disturbance or a real problem. If it is a real problem, the root cause localization branch is needed to further locate and analyze the cause of the poor network quality (e.g., interference, coverage, capacity, fault, or other reasons).

[0040] Furthermore, the root cause analysis branch is also used to set corresponding operational strategies based on the identified causes of network quality issues. Operations personnel implement the corresponding operational strategies based on the identified actual causes of the network quality problems. After implementing the operational strategies, the metric prediction branch predicts the changes in key operational metrics after the operations, thereby determining whether the quality issue has been resolved, ultimately completing the closed-loop processing of the wireless quality issue work order.

[0041] S300: Supervised training is performed on the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model to obtain the network quality analysis model.

[0042] Supervisory data is constructed based on operational metrics, alarm data, and tags. Tags include whether operational metrics are normal, the reasons for network quality deterioration caused by abnormal operational metrics, and changes in key operational metrics after implementing relevant operational strategies.

[0043] Supervised training is performed simultaneously on the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model using supervised data. This allows for fine-tuning of the parameters of the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch, ensuring that the fine-tuned parameters can adapt to the changes in the three corresponding tasks.

[0044] Furthermore, during supervised training, the parameters of the backbone model, except for the output layer, are fixed, and only the parameters of the output layer of the backbone model are fine-tuned.

[0045] Based on the output layer of the fine-tuned backbone model, other parameters of the backbone model, the fine-tuned anomaly detection branch, the fine-tuned root cause localization branch, and the fine-tuned index prediction branch, a network quality analysis model is obtained.

[0046] The network quality defect analysis model training method provided in this application, through training a time network to obtain a backbone model, enables intelligent analysis and prediction of operational indicators and high-level data, which is beneficial to improving the analysis efficiency of network quality defect work orders. This application, through supervised training, fine-tunes the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model, further realizing the prediction of abnormal quality defects, the causes of network quality defects, and adjusted key operational indicators. This achieves intelligent analysis, processing, and evaluation of processing results for network quality defect work orders, improving the efficiency of fault handling in wireless quality defect work orders.

[0047] Based on the above embodiments, a time series network is trained using unsupervised metrics and alarm data from wireless cells to obtain a backbone model, including the following steps: Based on operational metrics at the same point in time, obtain metric vectors; Text encoding is performed on the descriptive text in the alarm data to obtain a fixed-dimensional alarm vector; Obtain the average value of alarm vectors within the same sampling period; The index vector and the average value are concatenated to obtain the encoding vector; The encoded vectors are sorted by time to obtain a time vector sequence; Input the time vector sequence of the previous time period into the time series network to obtain the predicted time vector sequence of the key operating indicators in the next time period from the operating indicators output by the time series network. The key operating indicators are the main operating indicators that affect the network quality of the wireless cell. Based on the mean square error of the predicted time vector sequence for the next time period and the actual key operating indicators for the next time period, a loss function for the time series network is constructed. The time series network is trained using the loss function of the time series network to obtain the backbone model.

[0048] During data acquisition, the operational metrics of the wireless cell are collected periodically at a sampling frequency (e.g., 15 minutes), and the operational metrics at the same time point are concatenated into an metric vector according to the timestamp. Alarm data can be generated at any time, and the number of alarms N within a sampling period is not fixed. For each alarm data... The descriptive text of specific alarms in the alarm data is encoded to obtain a fixed-dimensional alarm vector. For example, a general vector model can be used to encode the descriptive text in the alarm data into a fixed-dimensional (e.g., 1024-dimensional) alarm vector. The general vector model includes the BAAI General Embedding (BGE) model.

[0049] The alarm vectors within the same sampling period are averaged, and this average value is concatenated with the indicator vector to obtain the encoded vector. The formula for the encoded vector is shown in formula (1). All encoded vectors are sorted by time to obtain a time vector sequence.

[0050] (1); in, for The encoding vector at time step, for The index vector at time, This represents the average value of the alarm vectors within the same sampling period. This represents the number of alarm vectors within the same sampling period. For the first One alarm vector.

[0051] Input the time vector sequence of the previous time period into the time series network to train the time series network.

[0052] The time vector sequence includes the time vector sequence of the previous time period (for all operating indicators) and the time vector sequence of the next time period (for all operating indicators). The time series network learns features from the time vector sequence of the previous time period and outputs the predicted time vector sequence of the key operating indicators (some operating indicators) for the next time period.

[0053] Furthermore, the next time period is taken as the previous time period, and the predicted time vector sequence of each time period of the key operating indicators is obtained in sequence according to the above steps.

[0054] The loss function of the time series network is constructed based on the mean squared error of the predicted time vector sequence for the next time period and the actual time vector sequence of the key operating indicators for the next time period. The expression of the loss function of the time series network is shown in Equation (2).

[0055] (2); in, Let be the loss function of the time series network. The number of key performance indicators, The key performance indicators are numbered. The time point number, The duration of the sampling period. For the position located at the The first time point The actual values ​​of key performance indicators (obtained through the time vector sequence of the next time period of the actual key performance indicators). For the position located at the The first time point Predicted values ​​of key performance indicators (obtained through the predicted time vector sequence for the next time period).

[0056] The time series network is trained using its loss function, and its parameters are updated to obtain the backbone model.

[0057] Optionally, the time series network includes an encoder-decoder based transformer structure. For example, a time series network with a 12-layer encoder and a 12-layer decoder can be designed. The specific number of model layers can be adjusted according to the amount of running metrics. When there is sufficient data from wireless cells, the time series network is usually a large time series model to ensure that the amount of data and the number of model parameters are matched. To train this time series network with unsupervised data (time vector sequences), a pre-training branch is temporarily designed at the end of the time series network. The entire time series network uses the time vector sequence of the previous time period as input to predict the predicted time vector sequence of the next time period. This data does not need to be labeled.

[0058] Optionally, the time vector sequence of the previous time period (for all operating metrics) can be input into the time series network to train the network. The predicted time vector sequence of all operating metrics for the next time period output by the time series network can then be obtained.

[0059] This application achieves unified encoding of operational indicators and alarm data by encoding alarm data and concatenating it with indicator vectors to obtain encoded vectors. This improves the efficiency of subsequent feature extraction and learning from time vector sequences. Based on the mean squared error of the predicted time vector sequence for the next time period and the actual time vector sequence of the key operational indicators for the next time period, this application constructs a loss function for the time series network, enabling effective training of the time series network and improving its learning efficiency for the features of key operational indicators.

[0060] Based on the above embodiments, the anomaly detection branch is trained in a supervised manner in the following way: Label the normal and abnormal data in the time vector sequence to obtain the first labeled time vector sequence; Based on the backbone model and the anomaly detection branch, the first preset network is obtained; The first labeled time vector sequence is input into the first preset network, and the first prediction probability output by the first preset network is obtained. The first prediction probability includes the probability that the predicted time vector sequence is normal data. Based on the label values ​​and the first predicted probability of the time vector sequence after the first annotation, a loss function of the first preset network is constructed. Based on the loss function of the first preset network, the output layer and anomaly detection branch of the backbone model are fine-tuned.

[0061] The anomaly detection branch consists of a neural network composed of several (e.g., three) fully connected layers. In the anomaly detection branch, the output dimension of the last fully connected layer is 2 (including the probability that the predicted time vector sequence is normal data and the probability that it is anomalous data).

[0062] The normal and abnormal data in the time vector sequence are labeled to obtain the first labeled time vector sequence. For example, when the encoded vector in the time vector sequence is normal data, the encoded vector is labeled as 1, and the label value is obtained. When the encoded vector in the time vector sequence is abnormal data, the encoded vector is labeled as 0, and the label value is obtained.

[0063] The backbone model and the anomaly detection branch are concatenated to obtain the first preset network. The first labeled time vector sequence is input into the first preset network to obtain the first prediction probability output by the first preset network. The first prediction probability includes the probability that the predicted time vector sequence is normal data (a decimal between 0 and 1). Based on the first prediction probability, the probability that the predicted time vector sequence is anomaly data (a decimal between 0 and 1) can be derived.

[0064] Based on the label values ​​and the first prediction probability of the time vector sequence after the first annotation, the loss function of the first preset network is constructed. The loss function of the first preset network is constructed using binary cross-entropy. The calculation formula of the loss function of the first preset network is shown in formula (3).

[0065] (3); in, The loss function of the first preset network, The label values ​​are the time vector sequences after the first annotation. This represents the first predicted probability.

[0066] Based on the loss function of the first preset network, the output layer and anomaly detection branch of the backbone model are fine-tuned.

[0067] This application constructs a loss function for a first preset network based on the label values ​​of the time vector sequence after the first annotation and the first prediction probability, thereby enabling fine-tuning of the output layer and anomaly detection branch of the backbone model, and thus realizing intelligent analysis of network quality poor work orders.

[0068] Based on the above embodiments, the root cause localization branch is trained in a supervised manner in the following way: When the time vector sequence after the first annotation is abnormal data, the reasons for poor network quality in the time vector sequence after the first annotation are annotated to obtain the time vector sequence after the second annotation. Based on the backbone model and root cause localization branch, a second preset network is obtained; The time vector sequence after the second annotation is input into the second preset network, and the second prediction probability output by the second preset network is obtained. The second prediction probability includes the probability of the network quality poorness of the predicted time vector sequence after the first annotation. Based on the label values ​​and second prediction probabilities of the time vector sequence after the second annotation, a loss function for the second preset network is constructed. Based on the loss function of the second preset network, the output layer and root cause localization branch of the backbone model are fine-tuned.

[0069] The root cause localization branch consists of a neural network composed of multiple (e.g., 3) fully connected layers. In the root cause localization branch, the output dimension of the last fully connected layer is equal to the total number of causes of network quality issues. For example, if the total number of causes of network quality issues is 5, namely interference, coverage, capacity, fault, or other causes, then the output dimension of the last fully connected layer in the root cause localization branch will be 5.

[0070] When the time vector sequence after the first annotation is abnormal data, the reasons for poor network quality in the time vector sequence after the first annotation are labeled to obtain the time vector sequence after the second annotation. For example, if the reason for poor network quality in the time vector sequence after the first annotation is interference, interference is labeled as 1, and network quality reasons such as coverage, capacity, fault, or other reasons are labeled as 0, thus obtaining the label value of the time vector sequence after the second annotation.

[0071] Based on the concatenation of the backbone model and root cause localization branches, a second preset network is obtained. The time vector sequence after the second annotation is input into the second preset network, and the second predicted probability output by the second preset network is obtained. The second predicted probability is used to characterize the probability of the network quality deterioration cause in the time vector sequence after the first annotation. For example, the second probability includes a probability of 0.8 for interference, 0.05 for coverage, 0.05 for capacity, 0.05 for fault, and 0.05 for other causes.

[0072] Based on the label values ​​and second prediction probabilities of the time vector sequence after the second annotation, a loss function for the second preset network is constructed. Cross-entropy is used to construct the loss function of the second preset network. The loss function of the second preset network is shown in Equation (4).

[0073] (4); in, The loss function of the second preset network, This represents the total number of reasons for poor network quality. The label value of the time vector sequence after the second annotation (belonging to the first) (Label values ​​for reasons for poor network quality) The second predicted probability (belonging to the first) (Probability of each cause of poor network quality). This is the number indicating the cause of poor network quality.

[0074] Based on the loss function of the second preset network, the output layer and root cause localization branch of the backbone model are fine-tuned.

[0075] Based on the label values ​​and second predicted probabilities of the time vector sequence after the second annotation, this application constructs the loss function of the second preset network, realizes the fine-tuning of the output layer and root cause localization branch of the backbone model, and thus realizes the intelligent localization of the network quality.

[0076] Based on the above embodiments, the indicator prediction branch is trained in a supervised manner in the following way: Based on the time vector sequence after the second annotation, obtain the operation strategy of the wireless cell; After adjusting the operational indicators based on the operational strategy, a third-annotated time vector sequence is obtained based on the values ​​of the key operational indicators in the previous time period and the values ​​of the key operational indicators in the next time period. Based on the backbone model and the indicator prediction branch, a third preset network is obtained; The time vector sequence after the third annotation and the operation strategy are input into the third preset network to obtain the predicted value of the next time period output by the third preset network. The predicted value of the next time period is used to characterize the operation status of key operation indicators in the next time period. Based on the predicted values ​​for the next time period and the actual values ​​of the key operating indicators for the next time period, a loss function for the third preset network is constructed. Based on the loss function of the third preset network, the output layer and index prediction branch of the backbone model are fine-tuned.

[0077] The metric prediction branch comprises a neural network consisting of multiple (e.g., three) fully connected layers. In this branch, the output dimension of the last fully connected layer equals the number of key operational metrics. Based on the second-annotated time vector sequence, the operational strategy corresponding to the network quality deterioration causes of the wireless cell is obtained. After adjusting the operational metrics in the time vector sequence according to the operational strategy, a third-annotated time vector sequence is obtained based on the values ​​of the key operational metrics from the previous time period and the values ​​of the key operational metrics from the next time period.

[0078] The backbone model and the indicator prediction branch are concatenated to obtain the third preset network. The time vector sequence after the third annotation and the operation strategy are input into the third preset network. The third preset network adjusts the key operating indicators in the time vector sequence after the third annotation according to the operation strategy and outputs the predicted value for the next time period.

[0079] Based on the predicted values ​​for the next time period and the actual values ​​of the key operating indicators for the next time period, the loss function of the third preset network is constructed. The calculation formula of the loss function of the third preset network is shown in formula (5).

[0080] (5); in, The loss function of the third preset network, For the next time period Predicted values ​​of key operational indicators, For the actual next time period The values ​​of key operating indicators, The key performance indicators are numbered. The number of key operational indicators.

[0081] Based on the loss function of the third preset network, the output layer and index prediction branch of the backbone model are fine-tuned.

[0082] This application constructs a loss function for a third preset network based on the predicted values ​​for the next time period and the actual values ​​of key operational indicators for the next time period. This enables fine-tuning of the output layer and indicator prediction branch of the backbone model, thereby achieving intelligent prediction of the values ​​of key operational indicators in the time vector sequence in which the operational strategy has been implemented.

[0083] Furthermore, the anomaly detection branch, root cause localization branch, and indicator prediction branch are fine-tuned simultaneously. Most parameters in the backbone model are fixed, and only the parameters of the output layer of the backbone model, the anomaly detection branch, the root cause localization branch, and the indicator prediction branch are trained, so that these parameters can adapt to the changes in the three tasks. The formula for calculating the overall loss function after fine-tuning is shown in formula (6).

[0084] (6); in, This characterizes the case where the time vector series is outlier; that is, the loss of the root cause localization branch is only calculated when the time vector series is outlier. The loss function of the third preset network, The loss function of the second preset network, The loss function of the first preset network, This is the overall loss function for fine-tuning.

[0085] This application also provides a method for analyzing poor network quality, including the following steps: Input the indicators to be analyzed and alarm data of the wireless cell into the network quality poor analysis model, and obtain the analysis results of the indicators to be analyzed output by the network quality poor analysis model. The analysis results include abnormal results of the indicators to be analyzed, the causes of network quality poorness, and predicted key operating indicator values. Among them, the network quality analysis model is trained according to the training method of any of the above network quality analysis models.

[0086] The network quality poor analysis model encodes the indicators to be analyzed and alarm data, generates a time vector sequence to be analyzed, predicts the key operating indicators in the time vector sequence, obtains the predicted time vector sequence, performs anomaly analysis on the predicted time vector sequence, analyzes the network quality poor when the indicator to be analyzed is abnormal, obtains adjustment strategies based on the reasons for the network quality poor, and predicts the key operating indicators based on the adjustment strategies to obtain the predicted key operating indicator values.

[0087] Following the steps outlined above, a network quality deterioration analysis model is trained. In practical applications, this model is used to analyze wireless quality deterioration work orders. The network quality deterioration analysis model includes anomaly detection branches, root cause localization branches, and indicator prediction branches.

[0088] like Figure 3 As shown, a poor wireless quality work order is obtained. The target metrics (operational data) and alarm data (poor wireless quality work orders) of the wireless cell are input into the network quality analysis model. The network quality analysis model encodes the target metrics and alarm data to generate a time vector sequence to be analyzed. The network quality analysis model predicts the key operational metrics in the time vector sequence to be analyzed, obtaining the predicted time vector sequence.

[0089] The anomaly detection branch of the network quality degradation analysis model performs anomaly analysis on the predicted time vector series to identify the authenticity of wireless cell performance degradation. The anomaly detection branch determines whether there is abnormal data in the analyzed indicators. If no abnormal data is found, the analysis ends.

[0090] If abnormal data is found, root cause analysis is used to determine the reasons for the network quality issues causing the anomalies in the analyzed metrics. Adjustment strategies are then derived based on these reasons.

[0091] After adjusting the indicators to be analyzed based on the adjustment strategy, the indicator prediction branch performs network performance indicator evaluation. The indicator prediction branch then re-predicts key operational indicators, obtaining the adjusted key operational indicator values ​​based on the adjustment strategy, and determines whether the predicted key operational indicator values ​​are abnormal. If the predicted key operational indicator values ​​are not abnormal, a closed-loop wireless quality poor performance work order is established. If the predicted key operational indicator values ​​are determined to be abnormal, the adjustment strategy is updated, and network performance indicator evaluation is performed again until the predicted key operational indicator values ​​are no longer abnormal.

[0092] This application uses a network quality poor analysis model to intelligently analyze anomalies in wireless cells, identify the causes of network quality poorness, and predict key operational indicators after adjustment based on adjustment strategies, thereby improving the fault handling efficiency of wireless quality poor work orders.

[0093] This application enables intelligent analysis and processing of wireless quality issues work orders. First, a time-series network is designed as the backbone model. A large amount of long-term performance and alarm operation data from wireless cells is collected, encoded, and integrated for pre-training the backbone model. For the three key stages of the wireless quality maintenance process—anomaly detection, root cause localization, and indicator prediction—corresponding processing branches are designed, and a relatively small amount of supervised data is used to fine-tune these branches, resulting in a network quality issue analysis model. This model can identify anomalies in wireless quality issues work orders, further provide specific reasons for network quality issues, push relevant information and solutions to maintenance personnel, predict network recovery status after processing, and achieve work order closure, thus realizing intelligent wireless quality issue analysis and work order closure.

[0094] The backbone model in this application is a key component of the entire network quality degradation analysis model. Unlike large language models that use text data for unsupervised pre-training, the backbone model in this application uses operational metrics and alarm data of wireless cells for unsupervised pre-training. First, data encoding technology is used to integrate this data, forming a time vector sequence as model input. Then, the time vector sequence of the previous time period is used to predict the time vector sequence of the next time period. This solves the problem of limited supervised data and the inability to fully utilize a large amount of unsupervised operational metric data, enabling the training of the initial parameters of the backbone model and obtaining a general representation for each wireless cell. In the downstream branches, this application designs multiple processing branches specifically for the wireless quality degradation operation and maintenance process, and fine-tunes the parameters of the output layer of the backbone model and the parameters of the three branches using supervised data. This allows the final network quality degradation analysis model to adapt to downstream scenarios, effectively improving the accuracy of anomaly analysis, root cause localization, and metric prediction in the network quality degradation analysis model.

[0095] The training apparatus for the network quality analysis model provided in this application is described below. The training apparatus for the network quality analysis model described below can be referred to in correspondence with the training method for the network quality analysis model described above.

[0096] like Figure 4 As shown, a training device for a network quality analysis model includes: The first training module 401 is used to train a time series network based on the operating indicators and alarm data of wireless cells to obtain a backbone model. Module 402 is used to add anomaly detection branches, root cause localization branches, and indicator prediction branches to the back end of the backbone model. The second training module 403 is used to perform supervised training on the output layer, anomaly detection branch, root cause localization branch and indicator prediction branch of the backbone model to obtain the network quality analysis model.

[0097] The network quality defect analysis model training device provided in this application, through training a time network, acquires a backbone model, enabling intelligent analysis and prediction of operational indicators and high-level data, which is beneficial to improving the analysis efficiency of network quality defect work orders. This application, through supervised training, fine-tunes the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model, further realizing the prediction of abnormal quality defects, the causes of network quality defects, and adjusted key operational indicators. This achieves intelligent analysis, processing, and evaluation of processing results for network quality defect work orders, improving the efficiency of fault handling in wireless quality defect work orders.

[0098] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0099] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540. The processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a training method for a network quality deterioration analysis model. This method includes: training a time-series network based on the operating indicators and alarm data of the wireless cell to obtain a backbone model; adding anomaly detection branches, root cause localization branches, and indicator prediction branches to the back end of the backbone model; and performing supervised training on the output layer, anomaly detection branches, root cause localization branches, and indicator prediction branches of the backbone model to obtain the network quality deterioration analysis model.

[0100] Alternatively, the processor 510 may call logic instructions in the memory 530 to execute the network quality deterioration analysis method provided by the above methods. The method includes: inputting the indicators to be analyzed and alarm data of the wireless cell into the network quality deterioration analysis model, obtaining the analysis results of the indicators to be analyzed output by the network quality deterioration analysis model, the analysis results including abnormal results of the indicators to be analyzed, the causes of network quality deterioration and predicted key operating indicator values; wherein, the network quality deterioration analysis model is trained according to the above-mentioned network quality deterioration analysis model training method.

[0101] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the training method of the network quality analysis model provided by the above methods. The method includes: training a time series network based on the operating indicators and alarm data of wireless cells to obtain a backbone model; adding an anomaly detection branch, a root cause localization branch, and an indicator prediction branch to the back end of the backbone model; and performing supervised training on the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model to obtain a network quality analysis model.

[0103] Alternatively, the computer can execute the network quality deterioration analysis method provided by the above methods, which includes: inputting the indicators to be analyzed and alarm data of the wireless cell into the network quality deterioration analysis model, obtaining the analysis results of the indicators to be analyzed output by the network quality deterioration analysis model, the analysis results including abnormal results of the indicators to be analyzed, the causes of network quality deterioration and predicted key operating indicator values; wherein, the network quality deterioration analysis model is trained according to the above-mentioned network quality deterioration analysis model training method.

[0104] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a training method for the network quality poor analysis model provided by the above methods. The method includes: training a time series network based on the operating indicators and alarm data of wireless cells to obtain a backbone model; adding an anomaly detection branch, a root cause localization branch, and an indicator prediction branch to the back end of the backbone model; and performing supervised training on the output layer, anomaly detection branch, root cause localization branch, and indicator prediction branch of the backbone model to obtain a network quality poor analysis model.

[0105] Alternatively, when the computer program is executed by the processor, it implements the network quality deterioration analysis method provided by the above methods. The method includes: inputting the indicators to be analyzed and alarm data of the wireless cell into the network quality deterioration analysis model, obtaining the analysis results of the indicators to be analyzed output by the network quality deterioration analysis model, and the analysis results include abnormal results of the indicators to be analyzed, the causes of network quality deterioration, and predicted key operating indicator values; wherein, the network quality deterioration analysis model is trained according to the above-mentioned network quality deterioration analysis model training method.

[0106] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0107] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A training method for a network quality analysis model, characterized in that, include: Based on the operational metrics and alarm data of wireless cells, a time series network is trained to obtain a backbone model. Add an anomaly detection branch, a root cause localization branch, and an indicator prediction branch to the back end of the backbone model; The output layer of the backbone model, the anomaly detection branch, the root cause localization branch, and the indicator prediction branch are trained in a supervised manner to obtain the network quality analysis model.

2. The training method for the network quality analysis model according to claim 1, characterized in that, The backbone model, obtained by training a time-series network based on the operational metrics and alarm data of the wireless cell, includes: Based on the operational metrics at the same point in time, obtain the metric vector; The descriptive text in the alarm data is text encoded to obtain a fixed-dimensional alarm vector; Obtain the average value of alarm vectors within the same sampling period; The index vector and the average value are concatenated to obtain the encoding vector; The encoded vectors are sorted by time to obtain a time vector sequence; The time vector sequence of the previous time period is input into the time series network to obtain the predicted time vector sequence of the key operating indicators in the next time period of the operating indicators output by the time series network. The key operating indicators are the main operating indicators that affect the network quality of the wireless cell. Based on the mean square error of the predicted time vector sequence of the next time period and the actual key operating indicators of the next time period, a loss function for the time series network is constructed. The time series network is trained based on its loss function to obtain the backbone model.

3. The training method for the network quality analysis model according to claim 1, characterized in that, The anomaly detection branch is trained in a supervised manner using the following method: Label the normal and abnormal data in the time vector sequence to obtain the first labeled time vector sequence; Based on the backbone model and the anomaly detection branch, a first preset network is obtained; The first labeled time vector sequence is input into the first preset network to obtain the first prediction probability output by the first preset network. The first prediction probability includes the probability that the predicted time vector sequence is the normal data. Based on the label values ​​of the first labeled time vector sequence and the first predicted probability, a loss function for the first preset network is constructed. Based on the loss function of the first preset network, the output layer of the backbone model and the anomaly detection branch are fine-tuned.

4. The training method for the network quality analysis model according to claim 3, characterized in that, The root cause localization branch is trained in a supervised manner as follows: When the first labeled time vector sequence is the abnormal data, the reasons for poor network quality in the first labeled time vector sequence are labeled to obtain the second labeled time vector sequence. Based on the backbone model and the root cause localization branch, a second preset network is obtained; The second labeled time vector sequence is input into the second preset network to obtain the second prediction probability output by the second preset network. The second prediction probability includes the probability of predicting the network quality poor cause of the first labeled time vector sequence. Based on the label values ​​of the second labeled time vector sequence and the second predicted probability, a loss function for the second preset network is constructed. Based on the loss function of the second preset network, the output layer of the backbone model and the root cause localization branch are fine-tuned.

5. The training method for the network quality analysis model according to claim 1, characterized in that, The indicator prediction branch is trained in a supervised manner using the following method: Based on the time vector sequence after the second annotation, the operation strategy of the wireless cell is obtained; After adjusting the operational indicators based on the operational strategy, a third-annotated time vector sequence is obtained based on the values ​​of the key operational indicators in the previous time period and the values ​​of the key operational indicators in the next time period. Based on the backbone model and the indicator prediction branch, a third preset network is obtained; The time vector sequence after the third annotation and the operation strategy are input into the third preset network to obtain the predicted value of the next time period output by the third preset network. The predicted value of the next time period is used to characterize the operating status of the key operation indicator in the next time period. Based on the predicted value of the next time period and the actual value of the key operating indicators of the next time period, a loss function of the third preset network is constructed. Based on the loss function of the third preset network, the output layer of the backbone model and the index prediction branch are fine-tuned.

6. A method for analyzing network quality issues, characterized in that, include: Input the indicators to be analyzed and alarm data of the wireless cell into the network quality poor analysis model, and obtain the analysis results of the indicators to be analyzed output by the network quality poor analysis model. The analysis results include the abnormal results of the indicators to be analyzed, the causes of network quality poorness, and the predicted key operating indicator values. The network quality analysis model is trained according to the training method of the network quality analysis model as described in any one of claims 1 to 5.

7. A training device for a network quality analysis model, characterized in that, include: The first training module is used to train a time series network based on the operating indicators and alarm data of wireless cells to obtain a backbone model. The module is used to add anomaly detection branches, root cause localization branches, and indicator prediction branches to the back end of the backbone model. The second training module is used to perform supervised training on the output layer of the backbone model, the anomaly detection branch, the root cause localization branch, and the indicator prediction branch to obtain the network quality analysis model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the training method of the network quality analysis model as described in any one of claims 1 to 5, or the network quality analysis method as described in claim 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the network quality analysis model as described in any one of claims 1 to 5, or the network quality analysis method as described in claim 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the network quality analysis model as described in any one of claims 1 to 5, or the network quality analysis method as described in claim 6.

Images