Training a model for detecting poor data quality in 5G networks and a method for detecting poor data quality.

Abstract

This invention provides a training model and method for detecting poor network data quality in 5G networks. Based on existing network stability performance data, it adds network user-related indicators such as the number of users and user traffic, and user experience-related indicators such as game login success rate and game login duration. It introduces the concept of semi-supervised learning, utilizing continuously increasing historical data to construct anomaly labels upstream using an unsupervised model and business rules, expanding the scope of poor-quality business scenarios. Downstream, it uses these labels to train a supervised machine learning tree model, resulting in a network quality defect identification model. Through multiple iterations and adjustments, the accuracy of the supervised model is improved for subsequent online real-time monitoring and identification.

Description

Technical Field

[0001] This invention relates to the field of monitoring and maintenance of operator communications, and in particular to a training model for detecting poor data quality in 5G networks and a method for detecting poor data quality. Background Technology

[0002] Existing operator centralized performance systems monitor and detect internet access services of communication equipment. The centralized performance platform collects unified data on operator services from communication equipment, focusing on network stability assessment metrics such as TCP latency, TCP request success rate, HTTP latency, TCP request success rate, and uplink / downlink RTT. By setting fixed thresholds for each metric, it identifies anomalies. When persistent anomalies are detected, maintenance personnel conduct further analysis to pinpoint whether there are any network abnormalities in the communication equipment. Manual analysis by maintenance personnel is time-consuming, labor-intensive, and requires a high level of expertise. Furthermore, manual analysis is challenging due to the difficulty of scenario analysis and high labor costs.

[0003] To address the issue of high manual labor costs and unsatisfactory results in network analysis, this paper proposes a method that, based on existing indicator threshold detection, summarizes expert business knowledge to formulate rules. When real-time indicator anomalies occur, a quality-defect rule is applied to determine the network quality, ultimately achieving the identification of poor network performance.

[0004] Due to the influence of external environmental factors, telecommunications equipment operators' networks are characterized by high frequency and large degree of change. Conventional threshold-based methods for detecting poor network quality are insufficient to cover all network scenarios, rule development is time-consuming, and there are numerous false positives and false negatives. When significant changes occur in network conditions, rules cannot be adjusted in a timely manner, failing to achieve prompt responses to network changes. Accumulating expert knowledge is necessary, and rule construction is time-consuming and labor-intensive. Furthermore, due to the complexity of network services, building rules covering all scenarios is difficult and the accuracy of judgments is not high. Summary of the Invention

[0005] In view of the complexity of network services in existing operator communication monitoring and maintenance technologies, the difficulty in constructing rules covering all scenarios and the low accuracy of judgment, this invention provides a method for training a 5G network data quality defect detection model and detecting quality defects.

[0006] In a first aspect, the present invention provides a method for training a 5G network data quality poor detection model, the method comprising:

[0007] The service data obtained from the 5G network data using the network management system includes stable performance data reflecting the performance of the 5G network, user experience index data, and user characterization data used to indirectly determine the poor quality of the 5G network.

[0008] The service data is preprocessed to obtain key features, year-on-year and month-on-month comparison features, statistical features, time features, and data smoothing features for the 5G network's poor quality detection; a first training set and a first test set are determined based on the key features, year-on-year and month-on-month comparison features, statistical features, and time features; a second training set and a second test set are determined based on the key features and statistical features for poor quality detection.

[0009] The first training set is input into the random forest model for training to obtain the first supervised poor quality detection model; the second training set is input into the DNN neural network model for training to obtain the second supervised poor quality detection model.

[0010] The stability performance data and user sensation index data are input into an unsupervised model for training to obtain unlabeled data with anomaly labels; the unlabeled data with anomaly labels are used to iteratively train the supervised first and second quality defect detection models to obtain the first and second quality defect detection models.

[0011] The first test set and the second test set are respectively input into the first poor quality detection model and the second poor quality detection model to obtain the trained first poor quality detection model and the second poor quality detection model.

[0012] The initial parameters of the third poor quality detection model are obtained based on the data smoothing features, and the third poor quality detection model is determined by grid search on the initial parameters.

[0013] The 5G network data quality detection results are determined by integrating prediction using the first, second, and third quality defect detection models.

[0014] On the other hand, embodiments of the present invention provide a method for detecting poor quality of 5G network data, applied to the aforementioned poor quality detection model, the method comprising:

[0015] The network management system is used to obtain stable performance data, user experience index data, and user characterization data of the 5G network data.

[0016] Preprocessing the stability performance data, user experience index data, and user characterization data yields key features for detecting poor quality in the 5G network, year-on-year and month-on-month characteristics, statistical characteristics, temporal characteristics, and data smoothing characteristics.

[0017] The first prediction result is obtained by inputting the key features of poor quality detection, year-on-year and month-on-month comparison features, statistical features, and time features into the first poor quality detection model; the second prediction result is obtained by inputting the key features of poor quality detection and statistical features into the second poor quality detection model; and the third prediction result is determined by inputting the data smoothing features into the third poor quality detection model.

[0018] The quality defect detection result of the 5G network data is determined by integrating the first prediction result, the second prediction result, and the third prediction result.

[0019] Compared to existing technologies, this invention provides a training model and method for detecting poor network data quality in 5G networks. Building upon existing network stability performance data, it adds network user-related indicators such as the number of users and user traffic, and user experience-related indicators such as game login success rate and game login duration. It introduces the concept of semi-supervised learning, utilizing continuously increasing historical data to construct anomaly labels upstream using unsupervised models and business rules, expanding the scope of poor-quality business scenarios. Downstream, it uses these labels to train a supervised machine learning tree model, resulting in a network quality identification model. Through multiple iterations and adjustments, the accuracy of the supervised model is improved for subsequent online real-time monitoring and identification. Attached Figure Description

[0020] Figure 1 This is a flowchart of the training method for a 5G network data quality poor detection model;

[0021] Figure 2 This is a schematic diagram of data indicators for a 5G network data quality poor detection model;

[0022] Figure 3 This is a diagram illustrating the correlation of data indicators in the poor quality detection model;

[0023] Figure 4 This is a schematic diagram illustrating the iteration of a supervised model using unlabeled data with anomaly labels;

[0024] Figure 5 This is a diagram illustrating the labels on unlabeled data with anomaly labels.

[0025] Figure 6 This is a flowchart illustrating the method for detecting poor data quality in 5G networks.

[0026] Figure 7 This is a schematic diagram of an integrated prediction method for detecting poor data quality in 5G networks. Detailed Implementation

[0027] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0028] This invention provides a training model and method for detecting quality issues in 5G network data. First, it obtains the optimal model through feature engineering and supervised model tuning, thereby improving the accuracy of result judgment. Then, the model can be iteratively updated using only historical data and minimal manual adjustment of anomaly labels, allowing for rapid iteration as subsequent 5G network performance data is generated. Finally, by combining classification models and time-series prediction algorithms, it judges whether anomalies exist in user experience metrics data (KQI, Key Quality Indicator, and Key Performance Indicator), improving the reliability of the results.

[0029] Figure 1 The flowchart of the training method for the 5G network data quality defect detection model is shown in the figure. The method specifically trains the quality defect detection model through steps S110-S140.

[0030] As shown in box S110, firstly, the service data of the 5G communication equipment network is obtained through the network management system.

[0031] Figure 2 This is a schematic diagram of the data indicators of the 5G network data quality poor detection model. As shown in the figure, the business data indicators are designed as stable performance data, user experience data, and user characterization data.

[0032] Stability performance data reflects the specific performance of 5G communication equipment networks, including metrics such as TCP latency, TCP request success rate, HTTP latency, TCP request success rate, and uplink / downlink RTT.

[0033] User experience metrics data are derived from real-world communication equipment network user scenarios and refined to the actual user services. In addition to the original communication equipment network stability assessment metrics, user-perceived KQI metrics are added, distinguishing between services such as video, games, web pages, and video VR.

[0034] Video services include metrics such as video playback success rate, average number of stutters per playback session, percentage of stutter duration, video playback waiting time, video download speed, average video bitrate, and frequency and bitrate.

[0035] The gaming business includes metrics such as game login success rate, game login duration, average latency above the interface, latency below the interface, game battle duration, and battle latency.

[0036] Web page services include metrics such as page response success rate, average page response time, page display success rate, average page display time, and page download speed.

[0037] VR services include metrics such as average latency of M2P upstream, average latency of M2P downstream, VR / AR video playback waiting time, VR / AR video download speed, and VR / AR bitrate ratio.

[0038] User characterization data is used to prevent issues such as missed data collection in network monitoring metrics, including user volume and real-time traffic. By using user behavior characterization data, a basic assessment can be made of whether there are quality issues affecting user experience.

[0039] As shown in box S120, in order to ensure the accuracy and reliability of the data, the service data of the 5G network is preprocessed to obtain the key features of the 5G network quality defect detection, the year-on-year and month-on-month characteristics, the statistical characteristics, the time characteristics, and the data smoothing characteristics.

[0040] First, the collected stability performance data, user experience index data, and user characterization data are cleaned to remove invalid or abnormal data points. Data cleaning uses methods such as interpolation, deletion, or replacement to remove missing data, outliers, and duplicate data to ensure data integrity and consistency.

[0041] Secondly, the cleaned user representation data is normalized to eliminate differences in dimensions or nonlinear relationships between data, thereby improving the effectiveness of subsequent data analysis and modeling.

[0042] Next, correlation analysis was performed on the normalized user representation data, the cleaned stable performance data, and the user experience index data, and feature selection was conducted to identify key features for poor quality detection in 5G networks. Figure 3This is a diagram illustrating the correlation of data metrics in the poor quality detection model. As shown, key features for poor quality detection include user count, traffic, TCP latency (both first and second hops), HTTP success rate, and uplink RTT. The colored blocks in the diagram represent the degree of correlation between two features, with lighter to darker blocks indicating a change from no correlation to a positive correlation. For example, when comparing the HTTP request count to its own latency, the correlation is "1" (black); when compared to the TCP latency (both first and second hops), the correlation is "0" (white). Correlation analysis and feature selection of these features yield the key features for poor quality detection in 5G networks.

[0043] Then, the embodiments disclosed in this invention, combined with professional experience and some feasible solutions, performed a series of operations and supplements on historical data to extract more representative and information-rich features. Specifically, based on the key features for poor quality detection obtained through screening, feature engineering was constructed to determine the year-on-year and month-on-month features, statistical features, and time features of the 5G network.

[0044] The year-on-year and month-on-month comparison features include feature data determined by calculating the ratios and differences between the stability performance data and user experience index data for different time periods. Considering that 5G network stability performance data and user experience index data may exhibit significant trends across different time periods, this embodiment introduces year-on-year and month-on-month comparison features to capture these changes. Specifically, network stability performance data is reflected by the KQI indicator; user experience index data is reflected by the KPI indicator.

[0045] By calculating the ratios or differences between characteristic values ​​over different time periods, the implementation can reflect the degree of change in 5G network performance. In some possible implementations, calculating the ratios of KPIs such as uplink RTT, TCP latency, and TCP success rate between the current time point and the previous time point can provide dynamic information on changes in network quality.

[0046] Statistical features are used to describe the distribution and variation of data, including statistics such as the mean and standard deviation of the data. In some possible implementations, statistical features include the mean and standard deviation of the number of TCP requests and the number of HTTP requests.

[0047] Temporal features are included to reflect the behavioral patterns of the 5G network at different time periods. Time information such as hour, day of the week, and month is extracted from each piece of feature data to reveal the variation patterns of network performance at different time granularities.

[0048] Finally, the exponential smoothing method was used to eliminate noise, reduce data fluctuations, and determine the data smoothing characteristics of the 5G network in the key features of poor quality detection, including the number of users, traffic characteristics, and statistical characteristics.

[0049] The specific implementation includes determining the smoothing coefficient of the features based on the number of users, traffic characteristics, and statistical characteristics in the preset key features for time quality difference detection. In some possible implementations, the initial values ​​of metrics such as the number of users, traffic, number of HTTP requests, and number of TCP requests from the historical data of the past three months are selected, and a smoothing coefficient α is chosen.

[0050] Based on the smoothing coefficient, the smoothing values ​​of the number of users, traffic characteristics, and statistical characteristics in the preset time quality difference detection key features are determined at each time t. This process continues until the smoothed values ​​for all time points are obtained. Smoothed values The mathematical expression is,

[0051]

[0052] in, The original data values, The smoothed value at the previous time step. In some possible implementations, the corresponding ARIMA model is selected based on the smoothed value at each of the above time steps.

[0053] As shown in box S130, since the proportion of abnormal data that actually trigger alarms is scarce in the collected 5G network quality poor data, this invention adopts the concept of semi-supervised learning to handle the network quality poor detection problem in order to ensure the accuracy of the quality poor detection model. Semi-supervised learning is a method that uses labeled and unlabeled data for model training, which can make full use of limited labeled data and a large amount of unlabeled data to improve model performance.

[0054] First, by preprocessing the business data, a limited set of key features for detecting poor quality with labels, including year-on-year and month-on-month characteristics, statistical features, and time features, have been identified. These features are then used to model network quality issues.

[0055] In some possible implementations, two supervised learning models are trained by using random forests and DNN neural networks respectively to learn the features and patterns of poor network quality.

[0056] The input data for Random Forest (a supervised first-class network quality defect identification model) includes stable performance data of 5G network, user experience index data, key features for quality defect detection such as TCP latency, HTTP success rate, uplink RTT, KQI index data, and year-on-year and month-on-month features, time features, statistical features, etc. generated by feature engineering.

[0057] The input data for the DNN neural network (supervised second network poor quality identification model) uses stable performance data of 5G network, KQI index data, and statistical feature data generated by feature engineering.

[0058] Random forests consist of multiple decision trees, which are easier to interpret and understand than DNN neural networks. Each decision tree independently classifies or regresses stable performance data, user experience index data, key features for detecting poor quality, year-on-year and month-on-month comparison features, statistical features, and time features. The final prediction result is derived by combining the results of all trees.

[0059] DNN neural networks consist of multiple neural network layers, each containing multiple neurons. Based on the statistical feature data in the key features of quality defect detection, the network processes and learns through multiple non-linear transformation layers to classify network quality defects.

[0060] The supervised model described above can be used as an initial model to classify and identify poor network quality. By using random forests and DNN neural networks for modeling separately, the strong coupling of a single algorithm can be avoided.

[0061] In some possible implementations, a first dataset is obtained based on the key features of poor quality detection, year-on-year and month-on-month comparison features, statistical features, and time features, and the first dataset is divided into a first training set and a first test set in an 8:2 ratio; a second dataset is obtained based on the key features of poor quality detection and statistical features, and the second dataset is also divided into a second training set and a second test set in an 8:2 ratio.

[0062] In some possible implementations, a training set is input into a random forest model to train a supervised first quality defect detection model; a second training set is input into a DNN neural network model to train a supervised second quality defect detection model.

[0063] Secondly, based on supervised model training, unsupervised model learning methods, such as Isolation Forest and Mean-shift, are used with a large amount of unlabeled data to train on stable performance data and user experience index data to identify outlier data. According to preset rules, the outlier data are labeled to determine the unlabeled data with anomaly labels.

[0064] Figure 4 The figure illustrates the iterative training of a supervised model using unlabeled data with anomalous labels. In some possible implementations, an unsupervised model is used to label discrete unlabeled data, and the supervised model is iteratively trained using the aforementioned unlabeled data with anomalous labels. Figure 5This is a diagram illustrating the labeling of unlabeled data with anomaly labels. As shown in the figure, the preset rules determine a portion of unlabeled data with anomaly labels by observing historical data with poor network quality, defining some anomaly patterns and rules based on business rules and domain knowledge, and manually adjusting some of the anomaly data found by the labeling algorithm according to the business rules.

[0065] Then, the quality defect detection model is iteratively adjusted using unlabeled data with anomaly labels and labeled data. By using unlabeled data with anomaly labels and labeled data together as training samples, the performance of the supervised quality defect detection model is continuously optimized through self-training.

[0066] Finally, through multiple iterations and adjustments, the model's performance and accuracy are further improved. Each iteration updates the model using unlabeled data with anomaly labels, and the new defect detection model is used for classifying and identifying network defects. As iterations proceed, the model's performance gradually improves, and its ability to detect and identify network defects is enhanced.

[0067] As shown in box S140, data smoothing features are determined in step S120. In some possible implementations, the Auto-ARIMA algorithm is used based on these data smoothing feature indicators to automatically select the best ARIMA model (third-order quality detection model) for time series prediction. These data smoothing features include indicators such as the number of users, traffic, number of HTTP requests, and number of TCP requests.

[0068] First, determine the range of characteristic sequences such as the number of users, traffic, number of HTTP requests, and number of TCP requests, and the corresponding initial parameters of the ARIMA model. These model parameters include information such as the difference order and MA order.

[0069] Then, a grid search is performed on the initial parameters. Specifically, within the preset parameter range, the mean square error, sum of squared residuals, and other indicators are used to evaluate the model's fitting effect for each set of parameters, and the best ARIMA model (third quality defect detection model) is selected based on the above fitting effect.

[0070] In summary, the embodiments disclosed in this invention have trained a first quality defect detection model, a second quality defect detection model, and a third quality defect detection model using a 5G network data quality defect detection model training method.

[0071] Figure 6 This is a flowchart illustrating a method for detecting poor quality of 5G network data. As shown in the figure, this method specifically implements the detection of poor quality of 5G network data through steps S510 to S540.

[0072] As shown in box S510, the network management system is used to obtain the stability performance data, user experience index data, and user characterization data of the 5G network data.

[0073] As shown in box S510, the key features for detecting poor quality of the 5G network, the year-on-year and month-on-month characteristics, statistical characteristics, time characteristics, and data smoothing characteristics are obtained by preprocessing the stable performance data, user experience index data, and user characterization data.

[0074] As shown in box S530, firstly, the key features for poor quality detection, year-on-year and month-on-month comparison features, statistical features, and time features are input into the first poor quality detection model to obtain the first prediction result. Then, the key features for poor quality detection and statistical features are input into the second poor quality detection model to obtain the second prediction result.

[0075] In some possible implementations, the scheme selects historical data from three months of key features for quality defect detection, including the number of users, traffic characteristics, and statistical characteristics, and performs smoothing processing to determine the data smoothing characteristics of the 5G network.

[0076] The data smoothing features are input into the third poor quality detection model to obtain the predicted value for a preset time; in this embodiment, the preset time is set to one week.

[0077] Set upper and lower thresholds, and determine the third prediction result based on the predicted value according to the upper and lower thresholds.

[0078] As shown in box S540, in order to further improve the accuracy of the scheme, this embodiment of the invention uses an integrated prediction method to determine the quality defect detection result of 5G network data by combining the first prediction result, the second prediction result and the third prediction result. Figure 7 This is a schematic diagram of an integrated prediction method for detecting poor data quality in 5G network data. As shown in the figure, the embodiment uses a weighted voting classification approach for integration. For a given sample, each supervised learning algorithm performs a classification prediction, and then votes are taken based on the prediction results of each algorithm. Finally, the integrated prediction of poor data quality detection result is determined by majority vote based on the voting results.

[0079] To improve the accuracy and reliability of operator communication monitoring and maintenance, this invention provides a training model and method for detecting poor network quality in 5G network data. This method departs from traditional manual review or fixed-rule judgment for work order quality inspection, instead employing a semi-supervised learning approach. It utilizes historical data and industry-accumulated business rules to construct anomaly labels, expanding the scope of poor quality service scenarios. Simultaneously, it uses supervised tree models such as random forests and neural networks to classify and detect anomalies in 5G network quality data from communication devices. The Auto-ARIMA time-series prediction algorithm is used to determine anomaly indicators based on data trends. By integrating multiple algorithms and models through ensemble learning, the prediction results of each model are voted on, and their classification decisions are comprehensively utilized to improve the accuracy and robustness of 5G network quality detection. Through real-time detection and monitoring of network quality issues, it promptly identifies and recognizes network quality problems, providing operators with accurate and reliable network optimization solutions, thereby improving user experience and satisfaction.

[0080] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a 5G network data quality poor detection model, characterized in that, include: The service data obtained from the 5G network data using the network management system includes stable performance data reflecting the performance of 5G network communication equipment, user experience index data reflecting the quality of 5G network services, and user characterization data used to make a basic judgment on the poor quality of the 5G network. Preprocessing the business data yields key features, year-on-year and month-on-month characteristics, statistical characteristics, and time characteristics for detecting poor quality in the 5G network. Based on the user representation data, key features for poor quality detection, user experience index data, year-on-year and month-on-month comparison features, statistical features, and time features, a first training set and a first test set are determined; based on the stable performance data, user experience index data, and statistical features, a second training set and a second test set are determined. The first training set is input into the random forest model for training to obtain the first supervised poor quality detection model; the second training set is input into the DNN neural network model for training to obtain the second supervised poor quality detection model. The stability performance data and user sensation index data are input into an unsupervised model for training to obtain unlabeled data with anomaly labels; The first and second poor quality detection models are obtained by iteratively training supervised first and second poor quality detection models using unlabeled data with anomaly labels. The first test set and the second test set are respectively input into the first poor quality detection model and the second poor quality detection model to obtain the trained first poor quality detection model and the second poor quality detection model. The quality defect detection results of the 5G network data are determined by ensemble prediction using a first quality defect detection model and a second quality defect detection model, including: Data smoothing processing is performed on key features of poor quality detection to determine the data smoothing features of 5G networks; The initial parameters of the third poor quality detection model are obtained based on the data smoothing features, and the third poor quality detection model is determined by grid search on the initial parameters. The 5G network data quality detection results are determined by integrating prediction using the first, second, and third quality defect detection models.

2. The training method for the 5G network data quality deterioration detection model according to claim 1, characterized in that: The stability performance data includes metrics such as TCP latency, TCP request success rate, HTTP latency, TCP request success rate, and uplink / downlink RTT. The user experience metrics data include: for video services, metrics such as video playback success rate, average number of stutters per playback session, percentage of stutter duration, video playback waiting time, video download speed, average video bitrate, and bitrate / rate ratio; for gaming services, metrics such as game login success rate, game login duration, average latency above the interface, latency below the interface, game battle duration, and battle latency; for web services, metrics such as page response success rate, average page response time, page display success rate, average page display time, and page download speed; and for VR services, metrics such as average upstream latency of M2P, average downstream latency of M2P, VR / AR video playback waiting time, VR / AR video download speed, and VR / AR bitrate / rate ratio. The user representation data includes metrics such as user volume and traffic.

3. The training method for the 5G network data quality deterioration detection model according to claim 1, characterized in that, The preprocessing method for the business data includes: The stable performance data, user sensation index data, and user characterization data are cleaned to obtain cleaned stable performance data, user sensation index data, and user characterization data. The cleaned user representation data is processed to determine the normalized user representation data; Correlation analysis was performed on the normalized user characterization data, the cleaned stable performance data, and the user experience index data to determine the key features for poor quality detection of the 5G network. Based on the key features of the poor quality detection, feature engineering is constructed to determine the year-on-year and month-on-month characteristics, statistical characteristics, and time characteristics of the 5G network.

4. The training method for the 5G network data quality deterioration detection model according to claim 3, characterized in that, The key features for poor quality detection include the number of users, traffic, TCP latency, HTTP success rate, and related features of uplink RTT.

5. The training method for the 5G network data quality deterioration detection model according to claim 3, characterized in that: The year-on-year and month-on-month comparison features include feature data obtained by calculating the ratio and difference of the stability performance data and user experience index data in different time periods. The statistical features include feature data determined by statistically analyzing the number of TCP requests and the number of HTTP requests; The time features include feature data extracted and determined based on time information for behavioral patterns in different time periods.

6. The training method for the 5G network data quality deterioration detection model according to claim 1, characterized in that, The data smoothing processing method includes: The smoothing coefficient of the feature is determined based on the number of users, traffic characteristics, and statistical characteristics in the preset key features for time quality difference detection. ; Based on the smoothing coefficient, the smoothing values ​​of the number of users, traffic characteristics, and statistical characteristics in the preset time quality difference detection key features are determined at each time t. Continue until smoothed values ​​for all time points are obtained.

7. The method for training a 5G network data quality deterioration detection model according to claim 6, characterized in that, The smoothing value The mathematical expression is, in, The original data values, The smoothed value from the previous time step.

8. The training method for the 5G network data quality deterioration detection model according to claim 1, characterized in that, The method for determining unlabeled data with anomalous labels includes: The isolated forest and Mean-shift algorithms were used to process the cleaned user experience index data and user characterization data to identify outlier data points. According to preset rules, outlier data points in the data are marked to identify unlabeled data with outlier labels.

9. A method for detecting poor data quality in 5G networks, applied to the poor data quality detection model according to any one of claims 1-8, characterized in that, include: The network management system is used to obtain stable performance data, user experience index data, and user characterization data of the 5G network data. Preprocessing the stability performance data, user experience index data, and user characterization data yields key features, year-on-year and month-on-month comparison features, statistical features, and time features for detecting poor performance in the 5G network. User characterization data, key features for poor quality detection, user experience index data, year-on-year and month-on-month comparison features, statistical features, and time features are input into the first poor quality detection model to obtain the first prediction result; stable performance data, user experience index data, and statistical features are input into the second poor quality detection model to obtain the second prediction result. Historical data of user count, traffic characteristics, and statistical characteristics from the key features of poor quality detection are selected and smoothed to determine the data smoothing characteristics of the 5G network; the data smoothing characteristics are input into the third poor quality detection model to obtain the predicted value at a preset time. The predicted value is judged based on the upper and lower threshold values ​​to determine the third prediction result; Based on the first prediction result, the second prediction result and the third prediction result, the integrated prediction is performed by majority voting to determine the quality defect detection result of the 5G network data, including: performing data smoothing processing on the key features of quality defect detection to determine the data smoothing features of the 5G network. The initial parameters of the third poor quality detection model are obtained based on the data smoothing features, and the third poor quality detection model is determined by grid search on the initial parameters. The 5G network data quality detection results are determined by integrating prediction using the first, second, and third quality defect detection models.

Images