Network fault positioning method and device, electronic equipment and computer program product

Abstract

The invention provides a network fault positioning method and device, electronic equipment and a computer program product, and belongs to the field of wireless networks. The method comprises the following steps: acquiring operation data of a plurality of network devices in a target network area; generating a data set according to the operation data, and calculating correlation coefficients between various types of operation state information and key performance indexes; according to a preset correlation coefficient threshold value, screening operation state information which is strongly correlated with the key performance index to form a strongly correlated feature set; based on the strong correlation feature set, multiple integrated learning strategies are adopted to carry out parallel model training, and multiple candidate fault prediction models are generated; evaluating the accuracy rate and the recall rate of each candidate fault prediction model, and selecting an optimal fault prediction model from the candidate fault prediction models according to a preset performance criterion; and checking the optimal fault prediction model by using newly added data, and updating and iterating the model according to a checking result.

Description

Methods, devices, electronic equipment, and computer program products for locating network faults Technical Field

[0001] This application relates to the field of wireless networks, and more particularly to a method, apparatus, electronic device, and computer program product for locating network faults. Background Technology

[0002] Currently, the field of wireless network optimization faces the challenge of accurately identifying and locating latent equipment faults. Latent equipment faults refer to situations where potential performance issues exist but have not yet reached alarm thresholds. These faults are difficult to detect and accurately locate, severely impacting the efficiency of root cause analysis of network performance degradation. Existing methods, such as fault time difference judgment and multi-dimensional clustering prediction based on human experience, rely on labels of already occurred faults or specific dimensional data. These methods are difficult to apply to latent fault prediction scenarios lacking clear fault identifiers, and suffer from limited applicability and reliance on human experience for prediction accuracy. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and computer program product for locating network faults, which can solve the problems of limited applicability and reliance on human experience for prediction accuracy in existing methods.

[0004] In a first aspect, embodiments of this application provide a method for locating network faults. The method includes the following steps: acquiring operational data from multiple network devices in a target network area, the operational data including operational status information and key performance indicators; generating a dataset based on the operational data, and calculating the correlation coefficients between various operational status information and key performance indicators; filtering operational status information strongly correlated with key performance indicators according to a preset correlation coefficient threshold to form a strongly correlated feature set; training multiple candidate fault prediction models in parallel using various ensemble learning strategies based on the strongly correlated feature set; evaluating the accuracy and recall of each candidate fault prediction model, and selecting the optimal fault prediction model based on preset performance criteria; testing the optimal fault prediction model using new data, and updating and iterating the model based on the test results.

[0005] Secondly, embodiments of this application provide a network fault location device, which includes: a data acquisition module for acquiring operational data of multiple network devices in a target network area, the operational data including operational status information and key performance indicators; a data calculation module for generating a dataset based on the operational data and calculating the correlation coefficients between various operational status information and key performance indicators; a feature construction module for filtering operational status information strongly correlated with key performance indicators according to a preset correlation coefficient threshold to form a strongly correlated feature set; a model training module for performing parallel model training based on the strongly correlated feature set using multiple ensemble learning strategies to generate multiple candidate fault prediction models; a model evaluation module for evaluating the accuracy and recall of each candidate fault prediction model and selecting the optimal fault prediction model based on a preset performance criterion; and a model verification module for verifying the optimal fault prediction model using new data and updating and iterating the model based on the verification results.

[0006] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the network fault location method as described in the first aspect.

[0007] Fourthly, embodiments of this application provide a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, the program instructions being executed by a computer to implement the steps of the network fault location method as described in the first aspect.

[0008] This application's embodiments use correlation coefficients to screen operational status features strongly correlated with key performance indicators, effectively eliminating irrelevant or weakly correlated noise data, allowing the model to focus on the core signals that truly affect network status. Employing a multi-strategy ensemble learning framework to train multiple models in parallel and select the best one integrates the advantages of different algorithms, overcoming potential biases or overfitting issues with single models, thereby significantly improving the ability to detect and locate latent faults that have not reached alarm thresholds. By introducing a model verification and iterative update mechanism based on new data, it can dynamically respond to changes in the network environment. When model performance deteriorates, retraining or adjustment can be triggered, ensuring that predictive capabilities are maintained or even improved over time, helping to reduce long-term maintenance costs and guaranteeing the continuous effectiveness of network quality analysis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of 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 only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 is a flowchart illustrating a network fault location method provided in an embodiment of this application; Figure 2 is a schematic diagram illustrating the monotonicity of observations and KPIs provided in an embodiment of this application; Figure 3 is a schematic diagram illustrating a strongly correlated feature set provided in an embodiment of this application; Figure 4 is a flowchart illustrating the training process of the Bagging strategy provided in an embodiment of this application; Figure 5 is a flowchart illustrating the AdaBoost algorithm of the Boosting strategy provided in an embodiment of this application; Figure 6 is a schematic diagram illustrating the AdaBoost algorithm of the Boosting strategy provided in an embodiment of this application; Figure 7 is a schematic diagram illustrating the combined basic classifier of the Boosting strategy provided in an embodiment of this application; Figure 8 is a flowchart illustrating the training process of the Stacking strategy provided in an embodiment of this application; Figure 9 is a schematic diagram illustrating the output training model of the Stacking strategy provided in an embodiment of this application; Figure 10 is a structural schematic diagram illustrating a network fault location device provided in an embodiment of this application; Figure 11 is a flowchart illustrating the application of latent fault location provided in an embodiment of this application; Figure 12 is a structural schematic diagram illustrating an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments 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, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] Improving customer experience is a key objective of network optimization. Currently, in the process of wireless network optimization, it is difficult to accurately identify and locate hidden equipment faults that affect customer experience. Hidden equipment faults refer to potential problems that exist in equipment but have not yet reached the alarm threshold to trigger an alarm. The existence of hidden equipment faults significantly hinders the analysis and localization of the causes of equipment performance, network indicators, and customer experience degradation. Accurate identification and localization of hidden problems is a crucial and proactive task for ensuring stable network operation and improving customer experience. The difficulty of hidden problems lies in their difficulty in discovery and localization; the investigation and resolution of hidden problems is a pressing challenge in practical work. Currently, there are two common solutions in the industry to address this problem: the first is the fault time difference judgment method, which infers the time and type of impending fault based on the time difference of various related fault data; the second is the clustering prediction method, which predicts the time and type of impending fault based on various real-time key indicators and fault types.

[0014] The existing base station fault prediction technology has the following problems.

[0015] Question 1: The fault time difference judgment method first requires determining the contextual information such as the fault that has already occurred, the time, and other faults that occurred before and after the fault to predict the subsequent faults and their timing. This method is not applicable to the prediction of hidden faults.

[0016] Question 2: When using clustering prediction methods, machine learning is often performed on data from multiple dimensions or based on human experience. During training, each type of data must correspond to a fault label. The accuracy of this type of machine learning mainly depends on human experience in grasping the data from each dimension. If the dimensions are not precise enough, it is difficult to achieve a high prediction accuracy. At the same time, because it requires corresponding fault labels, it is not suitable for judging hidden faults.

[0017] This application proposes a method and system for locating latent faults. Addressing the issue of poor KPI indicators caused by various operational data problems during equipment operation, this method employs KPI correlation evaluation, strong feature selection, and multi-model machine learning training to ultimately select a model with high accuracy for predicting latent faults caused by these performance issues. This application can predict both latent and explicit faults, providing accurate and integrated technical analysis support for network performance fluctuations caused by various faults in actual network production.

[0018] The following description, in conjunction with Figures 1 to 12, details a network fault location method, apparatus, electronic device, and computer program product provided in this application through specific embodiments and application scenarios.

[0019] Figure 1 illustrates a network fault location method provided by an embodiment of this application. This method can be executed by an electronic device, which may include a server and / or terminal devices. In other words, the method can be executed by software or hardware installed on the server and / or terminal devices. The method includes the following steps: Step 110, acquiring operational data of multiple network devices in a target network area, the operational data including operational status information and key performance indicators.

[0020] The target network area includes the logical or physical scope to which the method is applied. It can be a complete mobile communication network (e.g., 4G / 5G), a city-level network, a specific base station cluster, or any other subset of the network divided for centralized performance monitoring and fault management. The target network area contains multiple network devices that require coordinated monitoring.

[0021] The network devices can be the infrastructure units that make up the wireless access network, the most common entities being base stations and their subordinate cells. Each device is an independent data acquisition point and fault prediction unit.

[0022] The operational data can be a set of raw indicators collected from the aforementioned network devices that reflect their working status and performance. Operational data can have multi-dimensional characteristics.

[0023] The operational status information includes raw measurement values ​​or event records that directly reflect the working status of the device's internal hardware, software, and radio frequency units. Examples include transmit power, receive sensitivity, carrier frequency offset, hardware temperature, main control board status, VSWR, and bit error rate. This information typically comes from the device's operation and maintenance unit or performance management counters.

[0024] Key Performance Indicators (KPIs) include aggregated and calculated service-level metrics used to measure the quality of service provided by network devices. Examples include wireless call success rate, call drop rate, handover success rate, user plane latency, and uplink / downlink throughput. Deterioration in KPIs directly impacts the end-user's network experience and can directly indicate whether network problems exist.

[0025] Step 120: Generate a dataset based on the running data, and calculate the correlation coefficients between various running status information and key performance indicators.

[0026] The dataset can be integrated, cleaned, and reconstructed from the raw operational data obtained in step 110 to form a unified, standardized set of analytical data. Specific operations typically include the following: Data alignment and fusion involves associating and merging operational status information tables and key performance indicator tables from different sources based on timestamps and unique device identifiers (such as base station IDs and cell CGIs) to ensure that the device status at each point in time corresponds one-to-one with its performance.

[0027] Data cleaning involves removing missing values, obvious errors, or unreliable outliers from the data to ensure data quality.

[0028] Structured reconstruction involves organizing data into a time-device indexed format, where each row represents all observations of a specific device at a given point in time, and columns contain various operational status information, various KPI values, and possible label information.

[0029] The correlation coefficient is used to measure the strength and direction of the statistical association between a single piece of operational status information (independent variable) and a single key performance indicator (dependent variable). Preferably, the Spearman rank correlation coefficient is used for calculation. Unlike the common Pearson correlation coefficient, which requires a linear relationship and normal distribution, the Spearman coefficient is a non-parametric method. It is calculated based on the rank (or ordering position) of the variable, which can more effectively capture monotonic relationships (i.e., the trend of one variable increasing or decreasing as another variable increases, whether linear or non-linear), and is insensitive to outliers in the data, resulting in more robust results. The signs of latent faults in network equipment often manifest as a co-occurrence of a trend of deteriorating operational status (e.g., a sustained, slow decline in a power value) and a trend of declining KPIs. The Spearman coefficient is suitable for detecting such monotonic trend associations.

[0030] Specifically, in the calculation process, for each device (cell) in the dataset, the Spearman rank correlation coefficient of each of its operational status information (e.g., average transmit power) and each of its key performance indicators (e.g., wireless connection rate) is calculated over the entire time series to obtain a correlation coefficient matrix. This matrix can quantify the correlation strength between all potential features and performance results.

[0031] Step 130: Based on the preset correlation coefficient threshold, filter the operating status information that is strongly correlated with the key performance indicators to form a strongly correlated feature set.

[0032] The preset correlation coefficient threshold can be a pre-defined threshold value used to determine whether the correlation is strong enough. This threshold can be set based on the statistical significance of the Spearman rank correlation coefficient and practical engineering experience. For example, the threshold can be set to 0.4. When the absolute value (|ρ|) of the Spearman correlation coefficient between the running status information and the KPI is greater than this threshold, it is considered that there is a moderate or higher degree of monotonic correlation between the two; otherwise, the correlation is considered weak or negligible. Setting the threshold can achieve preliminary filtering of massive features, focusing on key signals.

[0033] The strongly correlated operational status information is obtained by comparing the absolute value of each Spearman correlation coefficient ρ calculated in step 120 with a preset threshold. If |ρ| > the threshold, the corresponding operational status information (feature) is determined to have strong (or significant) predictive or explanatory significance for the key performance indicator. This feature will be retained and marked. If |ρ| ≤ the threshold, the feature is determined to have insufficient correlation with the current KPI and will be removed in subsequent modeling of this KPI to avoid noise interference.

[0034] The strongly correlated feature set is the output of step 130. For each network device (or cell) and each key performance indicator to be predicted, all the operating status information that has been screened and determined to be strongly correlated, along with their corresponding KPI data (including historical values ​​and labels, such as "normal" / "abnormal"), are recombined into a new, refined dataset. This new set is the strongly correlated feature set. Compared to the original complete dataset, this feature set has the advantages of reduced dimensionality, improved signal-to-noise ratio, and enhanced interpretability. It can eliminate a large number of irrelevant or weakly correlated features to improve model training efficiency. Targeting the signal features most relevant to the target KPI, the machine learning algorithm can more clearly learn the fault modes. The selected features themselves are device parameters statistically correlated with network performance degradation, providing direct clues for maintenance personnel to analyze the root causes of faults.

[0035] Step 140: Based on the strongly correlated feature set, multiple ensemble learning strategies are used to train parallel models and generate multiple candidate fault prediction models.

[0036] Various ensemble learning strategies can be machine learning paradigms that accomplish learning tasks by building and combining multiple base learners. For example, the following three mainstream strategies can be applied in parallel to cover different model optimization objectives: Bagging reduces model variance and improves generalization ability by training multiple base learners in parallel; Boosting reduces model bias and improves prediction accuracy by training multiple weak learners sequentially; and Stacking achieves better generalization performance by learning how to combine the prediction results of different base learners.

[0037] Parallel model training involves executing training processes under the different ensemble learning strategies described above simultaneously or sequentially but independently. For example, for the same strongly correlated feature set, model training pipelines based on Bagging, Boosting, and Stacking strategies will be initiated separately. Each strategy can be further configured with or nested with multiple specific machine learning algorithms (e.g., Bagging can use random forests as base learners; Boosting can use AdaBoost or GBDT; the first layer of Stacking can contain multiple heterogeneous algorithms). After training is completed for each strategy, a candidate fault prediction model will be produced (e.g., Bagging_Model, Boosting_Model, Stacking_Model, etc.).

[0038] These candidate models differ fundamentally in their training data, algorithmic principles, or combinations, resulting in varying interpretations and predictive tendencies for the same failure modes. This diversity provides multiple comparable and selectable high-quality solutions for subsequent steps, avoiding the local optima that might result from a single model.

[0039] Step 150: Evaluate the accuracy and recall of each candidate fault prediction model, and select the optimal fault prediction model based on preset performance criteria.

[0040] Precision and recall are core metrics used to evaluate the performance of binary classification models (e.g., predicting "fault" / "normal"). Precision is the proportion of correct predictions among all model predictions, reflecting the overall reliability of the model. Recall, or completeness, refers to the proportion of samples that are actually faults that are correctly predicted by the model, reflecting the completeness of the model's fault detection. In scenarios involving the prediction of hidden network faults, recall is particularly important because missing a potential fault (low recall) can lead to serious network performance degradation or user complaints.

[0041] The preset performance criterion refers to selecting the model with the highest recall rate from candidate models whose accuracy reaches a certain high threshold (e.g., 95%) as the optimal model. Setting an accuracy threshold first ensures that the selected model has basic overall reliability, excluding unstable models that, while having high recall, also have many false positives. Among models that meet reliability requirements, the most sensitive and least likely to miss detections are prioritized, ensuring that latent faults can be detected as early as possible.

[0042] Based on the above performance criteria, all candidate models were evaluated and ranked. Models that met the accuracy threshold entered the final candidate pool, and the model with the highest recall was officially selected as the optimal fault prediction model for subsequent online prediction or root cause localization.

[0043] Step 160: Use the newly added data to test the optimal fault prediction model, and update and iterate the model based on the test results.

[0044] The new data includes recently generated network device operational data collected according to the same specifications, following the historical training data used in generating the optimal model. A fixed validation period is typically set (e.g., using data from the most recent two weeks). This new data is used as a separate validation set and fed into the deployed optimal fault prediction model for prediction. The model's predictions are then compared with actual observed network performance or confirmed fault records (as ground truth) during this period. Evaluating the model's real-world performance in the current or recent network environment can diagnose whether model performance has degraded due to network updates, seasonal changes, the introduction of new services, or other factors.

[0045] The update and iteration process is a decision-making and retraining process based on performance evaluation.

[0046] If the test results show that the model's accuracy and recall both remain above the preset high-level thresholds (e.g., accuracy > 95% and recall > 75%), then the model is considered still effective and fit. In this case, the newly labeled data increments can be incorporated into the existing training dataset for fine-tuning or retraining the model, thus updating and consolidating its knowledge.

[0047] If the test results show that any key metric of the model (especially recall) drops significantly below the threshold, a model performance degradation warning is issued. This indicates that the current model may no longer be able to effectively capture new failure modes. At this point, a complete model reconstruction process will be triggered, which may include the following: data inspection, firstly analyzing whether there are quality problems or collection anomalies in the newly added data; feature re-evaluation, re-performing the calculation and analysis of feature correlations (which can be traced back to steps 120 and / or 130) to check whether new strongly correlated features have appeared or whether the original features have become invalid; model retraining, based on a larger dataset containing both new and old data, restarting the entire process from feature selection to multi-strategy model training and evaluation (i.e., starting from step 130 or 140), generating a new generation of optimal failure prediction models to replace the old model.

[0048] This embodiment uses correlation coefficients to filter operational status features strongly correlated with key performance indicators, effectively eliminating irrelevant or weakly correlated noise data, allowing the model to focus on the core signals that truly affect network status. Employing a multi-strategy ensemble learning framework, multiple models are trained in parallel and the best one is selected. This integrates the advantages of different algorithms, overcoming potential biases or overfitting issues with single models, thus significantly improving the ability to detect and locate latent faults that have not reached alarm thresholds. By introducing a model verification and iterative update mechanism based on new data, it can dynamically respond to changes in the network environment. When model performance deteriorates, retraining or adjustment can be triggered, ensuring that predictive capabilities are maintained or even improved over time. This helps reduce long-term maintenance costs and ensures the continuous effectiveness of network quality analysis.

[0049] In yet another exemplary embodiment, based on step 110 of the above embodiment, operational data of multiple network devices in the target network area is obtained. The operational data includes operational status information and key performance indicators. The method of this embodiment may further include the following specific steps: collecting device operational data according to a required format and storing it in a database; formatting the stored data; extracting and rearranging the data according to time nodes, base station and cell information; removing abnormal data; and normalizing the data. Device operational data is collected (hourly level) and stored in a database according to the requirements of the device operational status information table, as shown in the table below:

[0050] Performance data is collected and entered into the database according to the requirements of the performance data information table, as shown in the table below:

[0051] The cells with poor KPI quality are entered into the database. This table shows the cells whose root causes have been clearly identified, and these cells do not need to participate in subsequent training and learning, as shown in the table below:

[0052] Based on the dataTime, stationId, and cgi columns in the equipment operation status information table and the performance data information table, the data in the two tables are transformed into rows and columns as follows to generate the HidFaultDF dataset. Here, value_1 to value_n represent different values ​​(observations) corresponding to different names (information names) in the equipment operation status information table; kpi_1 to kpi_n represent different values ​​(indicator values) corresponding to different kpiNames (KPI names) in the performance data information table; and Label_1 to Label_n represent the classification labels for different values ​​(indicator values) corresponding to different kpiNames (KPI names) in the performance data information table.

[0053] The following table shows the HidFaultDF dataset:

[0054] Since latent faults are not the only factor causing performance degradation, relevant records of the poor KPI cells in the KPI poor quality cell table that have been identified as having clear causes should be removed from the HidFaultDF dataset (e.g., power outages, base station failures, service outages, interference, high loads, insufficient deep coverage in residential areas, etc.). These poor cells are those that must be rectified, and since the root causes of the problems have been identified and have low correlation with latent faults, they should not be used as sample data for machine learning.

[0055] In yet another exemplary embodiment, based on step 120 of the above embodiment, a dataset is generated according to the running data, and the correlation coefficients between various types of running status information and key performance indicators are calculated. The method of this embodiment may further include the following specific steps: aligning and fusing the running status information and the key performance indicator data according to time sequence and device identification to generate an initial dataset; removing data records from the initial dataset whose causes of key performance indicator degradation have been clearly identified and marked; and calculating the Spearman rank correlation coefficient between each piece of running status information and each key performance indicator for the processed data.

[0056] Specifically, this embodiment uses the Spearman correlation coefficient to assess the correlation between latent fault observations and KPIs. The Spearman correlation coefficient is a non-parametric statistical method used to measure the monotonic relationship between two variables.

[0057] Referring to Figure 2, the Spearman correlation coefficient is applicable to any type of data, including non-linear and non-normally distributed data. Compared to the Pearson correlation coefficient, it is more accurate. When outliers in the data have a significant impact on the results, the Spearman correlation coefficient is more stable and accurate. The Spearman correlation coefficient is used to calculate the correlation between different types of observations and a specific KPI, allowing for the subsequent selection of observation types and KPIs with high correlation for use as data in subsequent machine learning.

[0058] HidFaultDF_Cell_Xi is constructed by iterating through a specific cell's value_i (a type of hazard observation value) in the HidFaultDF dataset, along with the index value of a specific KPI and its corresponding label (0 indicates normal, 1 indicates abnormal). The HidFaultDF_Cell_Xi dataset is shown in the table below:

[0059] Calculate the correlation ρ between feature observations value_i and kpi_i in the HidFaultDF_Cell_Xi data. The correlation ρ ranges from -1 to 1. The closer the value is to 1 or -1, the stronger the monotonic relationship (correlation) between the two variables. 0 indicates that there is no monotonic relationship (correlation) between the two variables. The formula for calculating the correlation ρ is as follows:

[0060] Where R(x) and R(y) are the rank of x and y, respectively (rank in statistics and data analysis refers to the position of each data point after the dataset is arranged in ascending or descending order). and Let x represent the average rank, y represent the observed value, and y represent the kpi index value.

[0061] In this embodiment, by aligning and fusing data according to time sequence and device identifier, the consistency of the status and performance of the analysis unit (each device / cell at each time) can be ensured, thus building a high-quality analysis foundation. Furthermore, by removing degraded data records with known root causes, model training focuses on true latent fault modes, effectively avoiding interference from known explicit problems in the learning process and enhancing the model's specificity in detecting unknown and potential fault features. Using the Spearman rank correlation coefficient as the core tool for correlation evaluation, it can robustly handle the large number of nonlinear correlations and outliers present in real-world networks. Compared to the traditional Pearson correlation coefficient, it provides more stable and accurate results in complex and noisy operational data environments, and can more realistically reveal the potential monotonic trend correlation between operational status information and KPIs, providing a reliable quantitative basis for subsequent feature selection. Through iterative traversal and refined calculation, a quantified correlation degree is generated for each pair of "feature-KPI" combinations. The resulting strongly correlated feature set helps subsequent multi-strategy machine learning models achieve high-precision and highly interpretable predictions.

[0062] In another exemplary embodiment, based on step 130 of the above embodiment, according to the preset correlation coefficient threshold, the operating status information that is strongly correlated with the key performance indicators is screened to form a strongly correlated feature set. The method of this embodiment may also include the following specific steps: Referring to Figure 3, feature correlation calculation is performed on each cell of the HidFaultDF_Cell_Xi dataset to obtain the correlation degree ρ. The features and data corresponding to ρ that satisfy the condition (|ρ|>crl_thd) are combined into a set FeaturesSet, which serves as sample data for subsequent machine learning of the cell.

[0063] `crl_thd` represents the feature selection threshold, with a default value of 0.4 (where [0.8~1.0] is extremely strong correlation; [0.6~0.8] is strongly correlated; [0.4~0.6] is moderately correlated; [0.2~0.4] is weakly correlated; and [0.0~0.2] is extremely weakly correlated or uncorrelated). In the FeaturesSet dataset, HidFaultDF_Cell_Xi data with strong correlations are retained, while HidFaultDF_Cell_Xi data with poor correlations are deleted.

[0064] In yet another exemplary embodiment, based on step 140 of the above embodiment, based on the strongly correlated feature set, multiple ensemble learning strategies are used to train the model in parallel to generate multiple candidate fault prediction models. The method of this embodiment may also include the following specific steps: the multiple ensemble learning strategies include at least two or three of the three strategies of Bagging, Boosting and Stacking.

[0065] Specifically, this embodiment integrates three major mainstream machine learning frameworks—Bagging, Boosting, and Stacking—for in-depth learning to enhance the model's generalization ability, reduce errors, minimize bias, and improve prediction accuracy. Each framework can be used to integrate algorithms such as decision trees, random forests, AdaBoost, GBDT, support vector machines, and neural networks on the FeaturesSet dataset. The best model is selected as the latent fault location model for the cell based on accuracy and recall.

[0066] In yet another exemplary embodiment, when the Bagging strategy is adopted, the method of generating multiple candidate fault prediction models by using multiple ensemble learning strategies based on the strongly correlated feature set may further include the following specific steps: training multiple different base learners using the strongly correlated feature set; combining the prediction results of the multiple base learners and forming the final Bagging model through a voting or averaging mechanism.

[0067] Specifically, referring to Figure 4, the Bagging strategy trains several different models separately, then allows all models to vote on the output of test samples. By training multiple types of models on the training set and making predictions on the test set, the accuracy of the predicted values ​​is judged by comparing them with the true values. The Top N models with the highest accuracy (N by default is all models) are selected to form a new model. The predictions of each sub-model within this new model are then voted on, and the model with the most votes becomes the final prediction. This strategy effectively solves the problems of data overfitting, large algorithm variance, and low model generalization.

[0068] Split the test set: For the generated sample dataset HidFaultDF_Cell_Xi, split the test set and training set according to a certain ratio (by default, 80% is used as the training set and 20% as the test set).

[0069] The training and testing data are formatted as follows: all preferred strong feature observations associated with a certain KPI, and the label corresponding to that KPI.

[0070]

[0071] Multi-model training: Decision tree, random forest, DBSCAN, and KNN algorithms are used to train models on the HidFaultDF_Cell_Xi training set. Each algorithm corresponds to sub-models 1 to n. The trained sub-models 1 to n are tested using the HidFaultDF_Cell_Xi test set. By default, all models are combined to form a new model (or, based on accuracy and recall, the TOPN models with high recall and high accuracy can be selected to form a new model).

[0072] Output training model: The predicted values ​​of each sub-model in the new model are voted on and counted. The one with the most votes is the final prediction result of the new model. If the votes are the same, they are sorted in descending order of recall and precision. The sub-model with high recall and high precision is selected as the output of the new model Bagging_ResultModel.

[0073] In this embodiment, the Bagging strategy constructs multiple slightly different training subsets through bootstrapping and trains multiple base learners in parallel, effectively reducing the model's sensitivity and dependence on specific training datasets. Even if individual base learners overfit or have high variance due to data sampling noise, these random errors can be smoothed out through final voting or averaging mechanisms, enabling the integrated Bagging model to exhibit stronger generalization ability and predictive stability when facing unknown data (new time periods, new neighborhoods). By employing various heterogeneous algorithms such as decision trees, random forests, DBSCAN, and KNN as base learners, different algorithms learn from the same strongly correlated feature set from different mathematical perspectives, enabling a more comprehensive capture of the various complex and nonlinear patterns that latent faults may exhibit. This ensures that when one algorithm is not good at handling a specific fault mode, other algorithms can supplement or correct it, thereby improving the overall ability to represent various potential faults and the fault tolerance. Building upon the basic Bagging model, a sub-model selection mechanism based on accuracy and recall is introduced. This mechanism selects the best-performing model based on its performance (Accuracy and Recall) on the test set for final integration. This allows the final Bagging model to maintain high stability while further improving the overall accuracy and reliability of fault prediction.

[0074] In yet another exemplary embodiment, when the Boosting strategy is adopted, the method of generating multiple candidate fault prediction models by using multiple ensemble learning strategies based on the strongly correlated feature set may further include the following specific steps: sequentially training a series of weak learners, wherein each subsequent weak learner focuses on learning samples that were predicted incorrectly by the preceding weak learner; and weighting and combining the prediction results of each weak learner to form the final Boosting model.

[0075] Specifically, the Boosting strategy combines several weak classifier models into a strong classifier model through weighted averaging to improve classification accuracy. Boosting implements isomorphic ML (machine learning) algorithms in a sequential manner and transforms weak learners into strong learners by emphasizing the errors of previous models (see Figures 5 and 6). The default algorithm used is AdaBoost, which conducts multiple rounds of training using its weak classifiers (decision stumps). In each round, the classification error of the previous round is calculated, and the weights of the training set data are readjusted for the current round of training. By iterating through multiple rounds, the error of the previous model is continuously reduced, thereby improving the stability of the entire process and solving problems such as underfitting, large algorithm bias, and low model training accuracy.

[0076] Split the test set: For the generated sample dataset HidFaultDF_Cell_Xi, split the test set and training set according to a certain ratio (by default, 80% is used as the training set and 20% as the test set).

[0077] The training and testing data are formatted as follows: all preferred strong feature observations associated with a certain KPI, and the label corresponding to that KPI.

[0078]

[0079] Multi-model training: As shown in Figure 7, decision stumps are used as weak classifiers for the AdaBoost algorithm to train the model on the HidFaultDF_Cell_Xi training set data.

[0080] Initialize sample weights: Assign each sample an equal weight, i.e., wi = 1 / N, where N is the number of samples.

[0081] Iterative training of the base classifier: Train a base classifier using the current sample weights and calculate the error rate of the classifier, which is defined as the number of misclassified samples divided by the total number of samples.

[0082] Calculate the weights of the base classifiers: Calculate the corresponding weights based on the classifier's error rate, with classifiers having higher weights for lower error rates. The specific calculation formula is as follows: ,in, It is the error rate of the j-th classifier.

[0083] Update sample weights: For each sample, decrease its weight if it is correctly classified, and increase its weight if it is misclassified. The formula is as follows: ,in, Let yi be the classifier weight in the t-th iteration, yi be the true label of sample i, ht(xi) be the prediction result of the classifier for sample i in the t-th iteration, and Zt be the normalization factor that makes the sum of the new sample weights equal to 1.

[0084] Combine basic classifiers: Combine all basic classifiers into a strong classifier, where the weight of each classifier is equal to its corresponding weight.

[0085] Output the trained model: Output the generated strong learner model, Boosting_ResultModel.

[0086] In this embodiment, the Boosting strategy trains subsequent weak learners sequentially, allowing them to focus on learning from samples mispredicted by the preceding model. This enables the model to continuously focus on misclassified or difficult-to-determine fault cases. Through multiple iterations, the model gradually corrects its errors, integrating a series of simple weak learners (e.g., shallow decision stumps) into a powerful classifier. This effectively addresses the underfitting problem inherent in single, simple models, helping to reduce the overall bias of the model. Consequently, when facing nonlinear, ambiguous, and latent fault patterns, it achieves deeper and more accurate fitting, significantly improving prediction accuracy. In fault prediction scenarios, true fault samples (positive samples) are often far fewer than normal samples, and fault patterns can be diverse and hidden. Boosting strategies (e.g., AdaBoost) automatically and intelligently focus on data by dynamically updating sample weights. Misclassified samples (especially those potential faults that are missed) will receive higher weights in the next training round, forcing the new model to pay more attention to learning difficult and blind spots. This ensures that the model does not ignore the capture of a few but crucial fault patterns because normal samples make up the majority, thereby improving the recall rate for rare fault events and complex hidden faults. By combining a series of weak learners trained under different data weight distributions and weighting them according to their respective classification error rates, the resulting Boosting ensemble model has strong robustness.

[0087] In yet another exemplary embodiment, when the Stacking strategy is adopted, the method of generating multiple candidate fault prediction models by training multiple ensemble learning strategies in parallel based on the strongly correlated feature set may further include the following specific steps: training multiple different base learners using the strongly correlated feature set as a first-layer model; using the output prediction results of the multiple base learners as new feature inputs; and training a meta-learner based on the new feature inputs as a second-layer model, wherein the final output of the Stacking model is determined by the meta-learner.

[0088] Specifically, referring to Figure 8, the analysis of latent problems in equipment involves many factors, and the performance of a single model has certain bottlenecks. The Stacking strategy can effectively improve the accuracy of the model. The Stacking structure has two layers of algorithms connected in series. The first layer is called level 0, which may contain one or more base learners. The second layer is called level 1, which can only contain one meta learner.

[0089] During training, the HidFaultDF_Cell_Xi dataset of a certain cell is first input into level 0 for training. After training, each base learner in level 0 (which can be a decision tree, random forest, AdaBoost, GBDT, support vector machine, etc.) outputs its corresponding prediction result. These prediction results are concatenated into a new feature matrix, which is then input into the meta-learner in level 1 (which can be a neural network, decision tree, random forest, AdaBoost, GBDT, support vector machine, etc.) for training. The final prediction result output by the fusion model is the result of the level 1 learner output.

[0090] Divide the test set: For the generated cell sample dataset HidFaultDF_Cell_Xi, divide the test set and training set according to a certain ratio (default 80% as training set and 20% as test set). Divide the training set into 5 parts: FStrain1, FStrain2, FStrain3, FStrain4, and FStrain5.

[0091] The training and testing data are formatted as follows: all preferred strong feature observations associated with a certain KPI, and the label corresponding to that KPI.

[0092]

[0093] Multi-model training: XGBoost, LightGBM, and RandomForest were selected as the base models.

[0094] The XGBoost model is trained using FStrain1, FStrain2, FStrain3, FStrain4, and FStrain5 as validation sets and the remaining four sets as training sets, undergoing five rounds of cross-validation. Predictions are then made on the test set. This yields five predictions trained on the training set and one prediction (B1) on the test set. These five predictions are then vertically overlapped and merged to obtain A1. The same logic applies to the LightGBM and RandomForest models.

[0095] The predicted values ​​of the three models (xgboost, lightgbm, randomforest) on the training set are used as three "features" A1, A2, and A3, respectively. The logistic regression (LR) model is then used to train the model and establish the logistic regression (LR) model.

[0096] Output training model: See Figure 9. Using the trained LR model, make predictions on the values ​​of three "features" B1, B2, and B3 constructed from the predictions on the test set before the three base models, and obtain the final predicted category or probability. This LR model is called Stacking_ResultModel.

[0097] In this embodiment, the Stacking strategy provides a two-layer model structure. The first layer utilizes multiple base learners of different types and principles (e.g., XGBoost, LightGBM, RandomForest) to perform initial learning and prediction on the original strongly correlated feature set. These base learners mine data patterns from different perspectives, and their outputs (predicted probabilities or categories) constitute a new, higher-order meta-feature matrix. The second layer's meta-learner (e.g., Logistic Regression (LR)) specifically learns how to optimally combine these initial predictions. This structure enables the model to automatically discover and utilize the complementarity between different base models, helping to overcome the inherent limitations of any single model in terms of accuracy or generalization ability, and enhancing the ensemble effect. When training the first base model, a K-fold cross-validation method was used instead of a simple one-time partition. Each base model was trained and validated multiple times on different subsets of data. The final prediction for the training set was the concatenation of the validation predictions of each fold, and the prediction for the test set was the average of the predictions of each fold model on the complete test set. This method can make the most of the limited training data, reduce the model variance caused by the randomness of a single data partition, and ensure that the generated meta-features (A1-A3) and the test set predictions (B1-B3) used for the final evaluation are more robust and unbiased, thereby improving the generalization ability and stable performance of the entire Stacking ensemble model on new data.

[0098] In yet another exemplary embodiment, based on step 150 of the above embodiment, the accuracy and recall of each candidate fault prediction model are evaluated, and the optimal fault prediction model is selected from them according to a preset performance criterion. The method of this embodiment may also include the following specific steps: Specifically, the latent fault model evaluation is carried out by accuracy and recall. Accuracy is the probability of correctly predicting in all samples; recall refers to the probability of how many positive samples (potential hazards) are predicted as positive (potential hazards). The rationality of the model is mainly evaluated by combining accuracy and recall.

[0099] Accuracy reflects the classifier's or model's ability to correctly classify all samples, that is, its ability to correctly classify positive samples as positive and negative samples as negative. A higher value indicates better performance, as shown in the table below.

[0100]

[0101]

[0102] Recall: Reflects the ability of a classifier or model to correctly predict the full extent of positive samples (potentially problematic). It increases the proportion of positive samples (potentially problematic) that are correctly predicted as positive samples out of the total number of positive samples (potentially problematic). A higher recall value indicates better performance.

[0103]

[0104] When the accuracy is high (default 95%), it means that the model performs well on the entire dataset. However, this does not mean that the model has a high success rate in predicting potential problems when there are more normal samples and fewer samples with potential problems. To avoid this problem, recall (default 70%) is used to avoid inaccurate predictions due to a small number of potential problem samples. In terms of recall threshold setting, it is more inclined to predict more negative samples rather than miss any.

[0105] Based on the models Bagging_ResultModel, Boosting_ResultModel, and Stacking_ResultModel output by the three machine learning strategies Bagging, Boosting, and Stacking, the best model is selected according to accuracy and recall requirements: Best_Model = {model|model∈{x|x∈S and x.accuracy>95%} and model.recall=max(model.recall)}. In another exemplary embodiment, based on step 160 of the above embodiment, the best model is selected according to accuracy and recall requirements: Best_Model = {model|model∈{x|x∈S and x.accuracy>95%} and model.recall=max(model.recall)}. The optimal fault prediction model is tested using new data, and the model is updated and iterated based on the test results. This embodiment may further include the following specific steps: acquiring new network device operating status information and key performance indicator data within a preset time period as new data; using the optimal fault prediction model to predict the new data and calculating its accuracy and recall on the new data; if the calculated accuracy and recall both meet preset iteration maintenance conditions, then the new data is added to the training set, and the optimal fault prediction model is incrementally updated; if the calculated accuracy or recall does not meet the iteration maintenance conditions, then the data quality is checked or the model structure is adjusted, and the model is retrained and selected.

[0106] Specifically, in mobile communication networks, there are many factors that affect the indicators, such as weather, terminal problems, user habits, and unexpected problems. The generated Best_Model model is continuously tested with the latest two weeks of data and the model is iterated and upgraded.

[0107] High accuracy (accuracy > 95% and recall > 75%): New data is incorporated into the training model for model training.

[0108] Low accuracy (accuracy < 95% or recall < 75%): New data quality issues or other problems, in which case the data should be corrected or removed.

[0109] Model issues include models trained on historical data that are not suitable for new data, requiring the extraction of historical data; or the emergence of new dimensions in the model, requiring the addition of new feature factors, etc., necessitating model readjustment and retraining for prediction.

[0110] In this embodiment, the optimal fault prediction model is tested by periodically acquiring new data, which is equivalent to setting up a continuous health check for the model. By calculating the model's accuracy and recall on new data, it is possible to evaluate in real time whether the model is still suitable for the current network state. This enables the model to perceive environmental changes and adjust itself, thereby ensuring that its predictive ability will not naturally decay over time, and providing a guarantee for long-term and stable fault early warning.

[0111] This embodiment predefines clear iteration maintenance conditions (e.g., accuracy > 95% and recall > 75%) and formulates differentiated processing strategies based on the test results. If the test criteria are met, the newly added data is used as new knowledge to incrementally update the model; if the test criteria are not met, a diagnostic process is triggered. The closed-loop iteration mechanism transforms manual operation and maintenance experience (e.g., judging when the model fails and how to adjust it) into automated operations based on clear rules, reducing continuous reliance on expert experience and lowering long-term maintenance costs. Simultaneously, through proactive, periodic testing and timely model updates / reconstruction, self-repair can be completed before model performance significantly impacts business operations, avoiding a surge in false negatives or false positives due to model failure.

[0112] Corresponding to the network fault location method provided in the above embodiments, based on the same technical concept, this application also provides a network fault location device. Referring to FIG10, the device 400 includes a data acquisition module 410, a data calculation module 420, a feature construction module 430, a model training module 440, a model evaluation module 450, and a model verification module 460.

[0113] The data acquisition module 410 is used to acquire operational data of multiple network devices in a target network area, including operational status information and key performance indicators; the data calculation module 420 is used to generate a dataset based on the operational data and calculate the correlation coefficients between various operational status information and key performance indicators; the feature construction module 430 is used to filter operational status information strongly correlated with key performance indicators according to a preset correlation coefficient threshold to form a strongly correlated feature set; the model training module 440 is used to perform parallel model training based on the strongly correlated feature set using multiple ensemble learning strategies to generate multiple candidate fault prediction models; the model evaluation module 450 is used to evaluate the accuracy and recall of each candidate fault prediction model and select the optimal fault prediction model according to a preset performance criterion; and the model validation module 460 is used to validate the optimal fault prediction model using new data and update and iterate the model according to the validation results.

[0114] It should be noted that the network fault location device and the network fault location method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned network fault location method, and the repeated parts will not be described again.

[0115] The network fault location method and apparatus provided in this application are applied to the location method and system of latent faults. They use equipment operating status information, cell KPI performance data, and cells with poor quality issues for data preprocessing to generate the HidFaultDF dataset. By calculating the Spearman correlation coefficient between latent fault observations and KPIs, the strong correlation between different types of latent fault observations and KPIs is evaluated. Strongly correlated datasets with correlation coefficients not lower than the crl_thd threshold are combined into a FeaturesSet dataset as sample data for subsequent cell training. For each cell's FeaturesSet data, multiple models are trained using three machine learning strategies: Bagging, Boosting, and Stacking. The generated training models are then selected based on accuracy and recall requirements, with the best model chosen as the optimal model for predicting latent faults in that cell. The optimal model is evaluated based on the latest data, continuously iterating and upgrading the best model.

[0116] Referring to Figure 11, the main process includes the following steps: 1. Data preprocessing: Collect the equipment operation status information table, performance data information table, and KPI poor quality community table according to the template and put them into the database. Then, perform row and column processing on the data to generate the HidFaultDF dataset. The dataset does not include the relevant records involved in the kpi poor quality community table.

[0117] 2. Correlation assessment between faults and KPIs: The correlation between different types of observations and a certain KPI is calculated using the Spearman correlation coefficient, so as to select observation types and KPIs with high correlation in the future.

[0118] 3. Select strong features: Filter the correlation ρ between different types of observations and KPIs, and combine the observations and KPIs corresponding to ρ that meet the condition (|ρ|>crl_thd) into a set called FeaturesSet, which will be used as sample data for subsequent machine learning in this community.

[0119] 4. Multi-model training: For the FeaturesSet sample data of each cell, multi-model training is carried out under the three major machine learning framework strategies of Bagging, Boosting, and Stacking. By dividing the test set, training multiple models, and outputting the training model, the generalization ability of the model is improved, the error is reduced, the bias is reduced, and the prediction accuracy is improved.

[0120] 5. Model Evaluation: Based on the accuracy and recall requirements, select the best model from the final generated models Bagging_ResultModel, Boosting_ResultModel, and Stacking_ResultModel under the Bagging, Boosting, and Stacking strategies as the latent fault location model for the cell.

[0121] 6. Model Validation: The generated Best_Model model is continuously validated using the latest two weeks of data, and the model is iterated and upgraded.

[0122] The latent fault localization method and system provided in this application calculate the correlation between different types of latent fault observations and KPIs, select latent fault observations with strong correlations and corresponding KPI label data to form learning samples, train multiple models according to Bagging, Boosting, and Stacking machine learning strategies, select the best model for predicting latent faults in the cell based on accuracy and recall standards, and perform model validation based on new data to achieve iterative upgrades of the model. Multi-model training uses different combinations of Bagging, Boosting, and Stacking strategies on the FeaturesSet dataset to find the optimal solution for latent fault localization in the cell.

[0123] The method and system for locating latent faults provided in this application are more accurate, offer more options, and provide predictability. By evaluating the correlation between latent faults and KPIs, selecting strong features, and training multi-model machine learning, it identifies potential latent faults that may exist in equipment when a certain performance indicator in a region deteriorates. This strongly supports network optimization practices, assists in indicator degradation analysis, and ensures stable operation and maintenance of network indicators.

[0124] Accuracy includes analyzing the correlation between multi-dimensional equipment operation information and KPIs to uncover the relationship between hidden faults and KPIs, and to obtain various data characteristics when hidden faults occur for accurate judgment.

[0125] Options include multi-model training, with the option to choose one or more of the following frameworks for machine learning: Bagging, Boosting, and Stacking. The sub-models under each framework can be freely selected and allocated according to the actual situation, and the optimal solution under all frameworks is finally selected as the optimal model.

[0126] Predictability includes using the optimal model and taking real-time performance KPIs and equipment operating status information of the cell. When the cell performance KPI data fluctuates, it can accurately predict whether it is caused by a hidden fault, which greatly improves the efficiency of analyzing and locating the causes of cell performance index degradation due to hidden faults.

[0127] The operational status of network equipment directly impacts network performance, user experience, and customer satisfaction. Existing network problem analyses rarely address hidden network issues, primarily due to their high concealment and lack of alarm triggers, making effective analysis impossible. This application utilizes data mining—including equipment operational status information, cell KPI performance data, and data from cells with poor network quality—to determine the correlation between hidden fault observations and KPIs. It employs three machine learning strategies—Bagging, Boosting, and Stacking—for multi-model training. The optimal model for predicting hidden faults in cells is selected based on accuracy and recall, and continuously iterated and upgraded using the latest data. The proposed solution offers high accuracy, diverse options, and predictable results, reducing significant manpower investment and planning errors. It strongly supports network optimization practices and possesses broad application prospects.

[0128] Corresponding to the network fault location method provided in the above embodiments, based on the same technical concept, this application also provides an electronic device for executing the above method. Figure 12 is a schematic diagram of the structure of an electronic device implementing various embodiments of this application. As shown in Figure 12, the electronic device 500 can vary considerably due to different configurations or performance, and may include one or more processors 510 and a memory 520. The memory 520 may store one or more application programs or data. The memory 520 may be temporary or persistent storage. The application programs stored in the memory 520 may include one or more modules (not shown in the figure), and each module may include a series of computer-executable instructions in the electronic device. Furthermore, the processor 510 may be configured to communicate with the memory 520 and execute the series of computer-executable instructions in the memory 520 on the electronic device.

[0129] This application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, implement the steps of the network fault location method described above.

[0130] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems, devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0134] In a typical configuration, an electronic device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0135] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0136] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0137] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0138] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0140] It should be understood that the training and prediction processes of the artificial intelligence (AI) models involved in the various embodiments of this specification all adhere to multiple legal and compliant principles, including legal data sources, compliant data content, compliant data governance, compliant training objectives and schemes, compliant training processes, compliant training environments and tools, and compliant ethical verification of training results, and comply with the requirements of Article 5 of the Patent Law. Specifically: Legality of data sources: The datasets used for AI model training are all obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, and the copyright ownership and scope of authorization are fully marked, with no unauthorized open-source code or data reuse. Data authorized by partners has a formal data usage agreement, clearly defining the scope, duration, and confidentiality obligations of data use, and possessing a complete authorization chain. For self-collected data involving personal information, the informed consent procedure has been strictly followed, and personally identifiable information has been removed through de-identification processing (including but not limited to field masking, feature anonymization, and differential privacy technology application), fully complying with the requirements of the "Interim Measures for the Administration of Generative Artificial Intelligence Services," the "Personal Information Protection Law," and other relevant laws and regulations.

[0141] Data Content Compliance: The AI ​​model's dataset undergoes multiple screening and cleaning processes to remove all content that may violate social morality or harm public interests. It contains no obscene, pornographic, violent, discriminatory, or information that endangers national or public safety, nor does it involve the illegal acquisition or use of genetic resources. For data in sensitive areas (such as healthcare and finance), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technologies) ensures that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0142] Data governance compliance: A complete data traceability system is established during the AI ​​model training process to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure that the data is verifiable throughout its entire lifecycle. The dataset annotation process for AI models is completed by a professional human R&D team, clearly defining the proportion of human creative contributions, avoiding reliance on AI-generated data that has not undergone substantial human modification, and complying with the examination requirements for "human main contributions" in AI patent applications.

[0143] Training objectives and scheme compliance: The training objectives of the AI ​​model focus on [specific technical scenarios that can be supplemented, such as intelligent driving decision optimization, multimodal information interaction, etc., which can be replaced based on specific content], and the training scheme and final output results do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risks of being used for illegal activities, privacy infringement, or public safety disruption. The ethical principle of "intelligent for good" is strictly practiced.

[0144] Compliance of the training process: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on the feedback from the expert system to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop control of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0145] Training Environment and Tools Compliance: AI model training is implemented based on nationally licensed chips and a compliant training platform. All open-source frameworks and components used in the training process have obtained their corresponding licenses, and copyright statements and patent citation information are fully retained, with no instances of infringement or reuse. The training environment is constructed using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. Furthermore, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0146] Training results ethical verification and compliance: After the model is trained, it undergoes additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism is established to ensure that the model always complies with Article 5 of the Patent Law and relevant laws and regulations in practical applications.

[0147] In summary, the data and training process used in the AI ​​model of this specification strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there are no violations of laws, social ethics, public interests, or illegal use of genetic resources. It fully meets the compliance requirements for patent authorization.

Claims

1. A method for locating network faults, characterized in that, The method includes the following steps: acquiring operational data of multiple network devices in a target network area, the operational data including operational status information and key performance indicators; generating a dataset based on the operational data; and calculating the correlation coefficients between various types of operational status information and key performance indicators. Based on a preset correlation coefficient threshold, operational status information strongly correlated with key performance indicators is selected to form a strongly correlated feature set; based on the strongly correlated feature set, multiple ensemble learning strategies are used to train parallel models to generate multiple candidate fault prediction models; the accuracy and recall of each candidate fault prediction model are evaluated, and the optimal fault prediction model is selected from them according to a preset performance criterion. The optimal fault prediction model is tested using the new data, and the model is updated and iterated based on the test results.

2. The method according to claim 1, characterized in that, The step of generating a dataset based on the operational data and calculating the correlation coefficients between various operational status information and key performance indicators includes the following steps: aligning and fusing the operational status information and key performance indicator data according to time sequence and device identification to generate an initial dataset; removing data records from the initial dataset whose causes of key performance indicator degradation have been clearly identified and labeled; and calculating the Spearman rank correlation coefficient between each operational status information and each key performance indicator for the processed data.

3. The method according to claim 1, characterized in that, The step of training multiple candidate fault prediction models in parallel using multiple ensemble learning strategies based on the strongly correlated feature set includes the following steps: the multiple ensemble learning strategies include at least two or three of the three strategies of Bagging, Boosting and Stacking.

4. The method according to claim 3, characterized in that, When using the Bagging strategy, the process of training multiple candidate fault prediction models in parallel using multiple ensemble learning strategies based on the strongly correlated feature set includes the following steps: training multiple different base learners using the strongly correlated feature set; combining the prediction results of the multiple base learners and forming the final Bagging model through a voting or averaging mechanism.

5. The method according to claim 3, characterized in that, When using the Boosting strategy, the process of training multiple candidate fault prediction models in parallel using multiple ensemble learning strategies based on the strongly correlated feature set includes the following steps: sequentially training a series of weak learners, where each subsequent weak learner focuses on learning samples that were predicted incorrectly by the preceding weak learner; and weighting and combining the prediction results of each weak learner to form the final Boosting model.

6. The method according to claim 3, characterized in that, When using the Stacking strategy, the process of training multiple candidate fault prediction models in parallel using multiple ensemble learning strategies based on the strongly correlated feature set includes the following steps: training multiple different base learners using the strongly correlated feature set as the first layer model; using the output prediction results of the multiple base learners as new feature inputs; and training a meta-learner based on the new feature inputs as the second layer model, wherein the final output of the Stacking model is determined by the meta-learner.

7. The method according to claim 1, characterized in that, The process of testing the optimal fault prediction model using new data and updating and iterating the model based on the test results includes the following steps: acquiring new network device operating status information and key performance indicator data within a preset time period as new data; using the optimal fault prediction model to predict the new data and calculating its accuracy and recall on the new data; if the calculated accuracy and recall both meet the preset iteration maintenance conditions, then adding the new data to the training set and incrementally updating the optimal fault prediction model; if the calculated accuracy or recall does not meet the iteration maintenance conditions, then checking the data quality or adjusting the model structure, and retraining and selecting the model.

8. A network fault location device, characterized in that, The device includes: a data acquisition module for acquiring operational data of multiple network devices in a target network area, the operational data including operational status information and key performance indicators; and a data calculation module for generating a dataset based on the operational data and calculating the correlation coefficients between various types of operational status information and key performance indicators. The feature construction module is used to filter operating status information that is strongly correlated with key performance indicators based on preset correlation coefficient thresholds, and to form a strongly correlated feature set. The model training module is used to train the model in parallel using multiple ensemble learning strategies based on the strongly correlated feature set, and generate multiple candidate fault prediction models. The model evaluation module is used to evaluate the accuracy and recall of each candidate fault prediction model and select the optimal fault prediction model based on preset performance criteria. In addition, there is a model validation module, which is used to validate the optimal fault prediction model using the new data and update and iterate the model based on the validation results.

9. An electronic device, characterized in that, It includes a processor, a memory, a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions that, when executed by a computer, implement the steps of the method as described in any one of claims 1 to 7.

Images