Charging flow acquisition quality evaluation method based on machine learning

Abstract

The invention relates to a billing flow acquisition quality evaluation method based on machine learning. The method comprises the following steps: acquiring a structured data packet of 5G traffic data through a preset traffic acquisition point; performing data cleaning on the original billing traffic data to obtain a traffic data feature set; constructing a feature mapping learning model and an outlier mode isolation model based on the traffic data feature set, carrying out feature set linkage, and obtaining a first identifier and a second identifier of traffic anomaly feature detection by training the feature mapping learning model and the outlier mode isolation model; performing dynamic association according to the first identifier and the second identifier to obtain a multi-dimensional flow collection quality evaluation result; and obtaining a hierarchical visual monitoring interface based on the stream multi-dimensional quality assessment result, and performing closed-loop processing on the billing flow acquisition quality assessment, thereby realizing assessment and processing on the billing flow acquisition quality by using a machine learning method.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a method for evaluating the quality of billing traffic collection based on machine learning. Background Technology

[0002] In the field of data acquisition quality inspection in current traffic billing systems, machine learning technology is gradually becoming a key means to improve real-time performance and accuracy. It achieves near real-time identification and early warning of acquisition deviations by automatically analyzing the temporal characteristics of traffic data (such as periodic fluctuations and sudden increase / decrease patterns) and the correlation of device behavior (such as the matching degree between device type and traffic range), combined with supervised learning (such as XGBoost, LSTM) or unsupervised learning (such as Isolation Forest, Autoencoder) models. In practical terms, the industry has explored optimization schemes such as lightweight model deployment (such as quantized LSTM), dynamic threshold adjustment (adaptive optimization based on historical distribution), and multi-source data fusion (integrating features such as time, device, and network) to balance detection efficiency and false alarm rate. In terms of application scenarios, this technology has penetrated into fields such as telecommunications billing and IoT traffic monitoring, effectively shortening the deviation detection time (from several hours to minutes), reducing the risk of billing disputes caused by data distortion, and becoming an important technical means to ensure the reliability of traffic billing system data.

[0003] The current use of machine learning for data collection quality inspection in traffic metering systems has the following main drawbacks and limitations: 1. Data Quality Risks: Machine learning models are highly sensitive to data noise, missing values, or errors. For example, if traffic data contains outliers caused by equipment malfunctions or noise generated by network fluctuations, the model may learn incorrect patterns, leading to misjudgments (such as misidentifying normal traffic fluctuations as data collection bias). Furthermore, the accuracy of data labels directly affects the effectiveness of supervised learning; if historical bias events are incorrectly labeled, the model will inherit the bias.

[0004] 2. Overfitting risk: The model may overfit noise or specific scene features in the training data. For example, if traffic in the training data surges during a certain period due to a promotional activity, the model may treat such an anomaly as normal, leading to missed reports; conversely, if the model is too sensitive to small fluctuations, it may falsely report normal traffic. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a machine learning-based method for evaluating the quality of billing traffic collection; The objective of this invention can be achieved through the following technical solutions: S1: Obtain structured data packets of 5G traffic data through preset traffic collection points; S2: Calculate the data distribution characteristics of the original billing traffic data in the structured data packet, extract basic statistical features, and extract time-series smoothing features by combining time series trend analysis logic; periodically mine the traffic data trend, associate and map the trends of different traffic data with a preset multi-dimensional scene coordinate system, and obtain a multi-dimensional feature set of association identifiers; based on the multi-dimensional feature set, use an adaptive threshold generation mechanism to generate a scene-based adaptive threshold library, correct outliers, and obtain a corrected traffic data feature set; S3: Construct a feature mapping learning model and an outlier pattern isolation model based on the traffic data feature set and link the feature set together. By training the feature mapping learning model and the outlier pattern isolation model, obtain the first and second identifiers for traffic anomaly feature detection. Dynamically associate the first and second identifiers to obtain the evaluation results of multi-dimensional traffic acquisition quality. S4: Based on the evaluation results of the multi-dimensional traffic acquisition quality, a coordinate system is constructed to obtain a hierarchical visual monitoring interface, and the evaluation of the billing traffic acquisition quality is processed in a closed loop.

[0006] Specifically, the method for extracting the basic statistical features is as follows: based on the original billing traffic data in the structured data packet, calculate the range, interquartile range, standard deviation, and variance to obtain the data dispersion, and calculate the skewness coefficient and kurtosis coefficient to characterize the symmetry and steepness of the data distribution shape, thereby obtaining the basic statistical features. Specifically, the process of obtaining the time-series smoothing features includes filling missing values ​​and filtering outliers in the time-series billing traffic data in the structured data packets, processing missing data using linear interpolation, and identifying and isolating outliers using appropriate criteria. Construct sliding windows with different time scales, and smooth the data sequences within the corresponding windows using an exponentially weighted moving average algorithm, then calculate the residuals between the original data sequences and the smoothed sequences. The first and second differences are calculated on the smoothed sequence to obtain the rate of change and acceleration characteristics of the data trend and identify the trend inflection point. The smoothed sequence features, residual features and trend change features at different time scales are fused and dimensionality is reduced by principal component analysis algorithm to obtain the time series smoothing feature set.

[0007] Specifically, the dimensionality reduction process includes optimizing the dimensionality of the original time-series smooth feature set based on the principal component analysis algorithm, while retaining principal components whose cumulative variance contribution rate reaches a preset threshold, and obtaining low-dimensional features with corresponding trend and fluctuation information.

[0008] Specifically, the process of periodically mining the trend of traffic data is as follows: based on the time-series smoothing feature set, the time-domain data is converted to the frequency domain using fast Fourier transform, significant periodic components in the traffic data are identified, and the amplitude and phase characteristics of each periodic component are calculated. Based on the identification period as the basic unit, the time series data is divided into continuous analysis segments using the sliding window technique. For each analysis segment, the overall slope of the data is calculated by linear regression, and the trend type is obtained by combining the second-order difference feature of the smooth sequence. The cosine similarity of the trend feature vectors between the corresponding analysis segments is calculated, and the density clustering algorithm is used to cluster segments with similar trend features into one class, forming a trend pattern cluster. The boundaries of the initially defined trend segments are corrected by detecting abrupt changes in the first-order difference of the smooth sequence. Feature parameters are calculated for each defined trend segment to obtain a standardized trend feature descriptor and acquire the periodic trend features of the traffic data.

[0009] Specifically, the process of obtaining the multi-dimensional feature set of associated identifiers includes: pre-setting a multi-dimensional scene coordinate system, dividing into multi-dimensional categories, and obtaining a standardized scene coordinate library; forming a mapping rule library based on the matching conditions between the corresponding trends and scene coordinates, and generating a unique identifier for the associated feature trends and scene coordinates, including corresponding trend type codes, scene coordinate codes, and timestamps; and aggregating the trend feature data, basic statistical features, and time-series smoothing features based on the unique identifiers to obtain a multi-dimensional feature set of data self-features and scene-related features.

[0010] Specifically, the process of correcting outliers includes recording the corresponding candidate outliers according to a preset threshold, applying differentiated corrections for different outlier types, and then substituting the corrected data into the multi-dimensional feature set to obtain corrected data for secondary outliers.

[0011] Specifically, obtaining the first identifier includes: labeling samples and dividing the dataset based on the traffic data feature set; performing feature mapping on the labeled traffic data feature set to capture the nonlinear correlation between features; and obtaining the first identifier for traffic anomaly feature detection through probability normalization processing.

[0012] Specifically, obtaining the second identifier includes: identifying outliers based on unannotated data through a data outlier isolation mechanism; adjusting the model sensitivity by setting an outlier ratio adaptation parameters; and obtaining a standardized second identifier for outlier traffic data anomaly detection through training.

[0013] Specifically, the feature is that obtaining the multidimensional traffic quality assessment result includes: constructing a scenario-based verification set for the corresponding network scenario based on the first identifier and the second identifier; dynamically fusion the first identifier and the second identifier with weights; introducing a preset weight calculation model; and calculating the multidimensional traffic quality assessment result by combining the dynamically corrected weights.

[0014] Specifically, the hierarchical visualization monitoring interface includes: a near real-time identification and early warning module for acquisition deviation added to the scene layer, which captures real-time data at a preset frequency and compares it with the corresponding threshold in the scene-based adaptive threshold library to trigger the corresponding early warning indicator; in the indicator layer, a periodic detection data display area is used to periodically report the periodic traffic of the corresponding devices, and the deviation between the actual traffic and the benchmark curve is obtained by superimposing the periodic benchmark curve on the time axis; the hierarchical visualization monitoring interface is obtained by superimposing the scene layer and the indicator layer.

[0015] Specifically, the closed-loop processing of the billing traffic collection quality assessment involves establishing a multi-dimensional weight allocation mechanism based on the multi-dimensional traffic collection quality assessment results presented by the hierarchical visual monitoring interface, quantifying the priority of the impact of assessment indicators on collection quality, obtaining the iteration of traffic collection quality, and thus obtaining the closed-loop processing of the billing traffic collection quality assessment.

[0016] The beneficial effects of this invention are as follows: Based on the raw billing traffic data in the structured data packets, the range, interquartile range, standard deviation, and variance are calculated to characterize the degree of data dispersion. Skewness and kurtosis coefficients are calculated to reflect the symmetry and steepness of the data distribution, forming a basic set of statistical features.

[0017] Missing values ​​were filled using linear interpolation, and outliers were identified and isolated using the 3σ criterion to complete data preprocessing. A multi-timescale sliding window was constructed, and the data was smoothed using an exponentially weighted moving average algorithm. Residuals, first-order differences, and second-order differences were calculated. After fusing multi-scale features, dimensionality reduction was performed using principal component analysis, retaining principal components whose cumulative variance contribution rate reached a preset threshold, resulting in a time-series smoothing feature set.

[0018] The time-domain data is transformed to the frequency domain using Fast Fourier Transform (FFT) to identify significant periodic components and calculate amplitude and phase characteristics. Analysis segments are divided using a sliding window approach, with periods as the unit. Trend types are determined through linear regression and second-order differencing, and density clustering is used to form trend pattern clusters. First-order differencing abrupt changes are detected to correct segment boundaries, standardized feature parameters are calculated, and trend feature descriptors are generated.

[0019] First label (supervised learning): Label the feature set of traffic data and divide the dataset, capture non-linear correlations through feature mapping, and generate the label through probability normalization.

[0020] Second label (unsupervised learning): Based on unlabeled data, outliers are identified through an outlier isolation mechanism, the outlier ratio is adjusted to adapt parameters to optimize model sensitivity, and generated after training.

[0021] Multi-model fusion detection: Random forests excel at capturing non-linear relationships between features, while isolated forests can detect unknown abnormal patterns; the two complement each other to improve robustness. A weighted voting mechanism increases the model's AUC on the validation set from 0.85 (single model) to 0.92 (fusion model). Attached Figure Description

[0022] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating a machine learning-based method for evaluating the quality of billing traffic collection according to the present invention. Figure 2 This is a structural block diagram of the fusion model in this invention. Detailed Implementation

[0024] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0025] Please see Figure 1-2 A machine learning-based method for assessing the quality of billing traffic collection includes: S1: Obtain structured data packets of 5G traffic data through preset traffic collection points; S2: Calculate the data distribution characteristics of the original billing traffic data in the structured data packet, extract basic statistical features, and extract time-series smoothing features by combining time series trend analysis logic; periodically mine the traffic data trend, associate and map the trends of different traffic data with preset multi-dimensional scenarios, and obtain a multi-dimensional feature set of association identifiers; based on the multi-dimensional feature set, use an adaptive threshold generation mechanism to generate a scenario-based adaptive threshold library, correct outliers, and obtain a corrected traffic data feature set; S3: Construct a feature mapping learning model and an outlier pattern isolation model based on the traffic data feature set and link the feature set together. By training the feature mapping learning model and the outlier pattern isolation model, obtain the first and second identifiers for traffic anomaly feature detection. Dynamically associate the first and second identifiers to obtain the evaluation results of multi-dimensional traffic acquisition quality. S4: Construct a coordinate system based on the evaluation result of the multi-dimensional traffic collection quality, obtain a hierarchical visualization monitoring interface, and perform a closed-loop process on the evaluation of the charging traffic collection quality.

[0026] As a preferred technical solution of the present invention, the method for extracting the basic statistical features is to calculate the range, interquartile range, standard deviation, and variance based on the original charging traffic data in the structured data packet to obtain the data dispersion degree, and calculate the skewness coefficient and kurtosis coefficient to characterize the symmetry and steepness of the data distribution form, so as to obtain the basic statistical features.

[0027] In this embodiment, for the original charging traffic data in the structured data packet, the following method is used to extract basic statistical features to comprehensively characterize the data features. First, calculate the data dispersion degree, which is reflected by the range, interquartile range, standard deviation, and variance. The formula for the range is: R = max(x) - min(x), where x is the traffic data sequence, max(x) represents the maximum value in the traffic data sequence, and min(x) represents the minimum value. This index can quickly reflect the overall fluctuation range of the data; the formula for the interquartile range is: IQR = Q3 - Q1, where Q1 (lower quartile) is the value at the 25% position after the data is sorted, and Q3 (upper quartile) is the value at the 75% position after the data is sorted. It can effectively exclude the influence of extreme values and reflect the dispersion of the middle 50% of the data. Specifically, the process of obtaining the time-series smoothing features includes filling missing values and filtering outlier values for the time-series charging traffic data in the structured data packet, using the linear interpolation method to process the missing data, and identifying and isolating outlier points through corresponding criteria. Construct sliding windows of different time scales, and use the exponentially weighted moving average algorithm to smooth the data sequence within the corresponding window, and calculate the residual between the original data sequence and the smoothed sequence. Calculate the first-order difference and second-order difference of the smoothed sequence to obtain the change rate and acceleration characteristics of the data trend, and identify the trend turning points; fuse the smoothed sequence features, residual features, and trend change features at different time scales, and perform dimensionality reduction processing through the principal component analysis algorithm to obtain the time-series smoothing feature set.

[0028] In this embodiment, missing values are filled and outlier values are filtered for the time-series charging traffic data in the structured data packet. For missing data, the linear interpolation method is used. Assume that in the time series, the traffic values x1 corresponding to the known adjacent two time instants t1 and x2 corresponding to the traffic value x2 are known. For the intermediate missing time instant t (t1 < t < t2), the formula for its traffic value x is: (x = x1 + (x2-x1) can be used to reasonably fill in missing data and ensure data continuity. Outliers are identified and isolated using the 3σ criterion. First, the mean and standard deviation s of the data sequence are calculated. If a data point (xi) satisfies... If the value is not found, then the data is considered an outlier.

[0029] Calculate the first and second differences for the smoothed sequence. First difference The calculation formula is: It can reflect the rate of change of data trends; second-order difference The calculation formula is: This can further reflect the acceleration characteristics of data trend changes, and the changes in first-order and second-order differences can effectively identify trend turning points. Finally, the smoothed sequence features, residual features, and trend change features at different time scales are fused, and principal component analysis is used for dimensionality reduction. Principal component analysis constructs a covariance matrix, solves for eigenvalues ​​and eigenvectors, and selects the top k principal components with larger eigenvalues ​​as new features. While retaining the main information of the data, it reduces the feature dimensionality, ultimately obtaining a set of time-series smoothed features.

[0030] Specifically, the process of periodically mining the trend of traffic data is as follows: based on the time-series smoothing feature set, the time-domain data is converted to the frequency domain using fast Fourier transform, significant periodic components in the traffic data are identified, and the amplitude and phase characteristics of each periodic component are calculated. Based on the identification period as the basic unit, the time series data is divided into continuous analysis segments using the sliding window technique. For each analysis segment, the overall slope of the data is calculated by linear regression, and the trend type is obtained by combining the second-order difference feature of the smooth sequence. The cosine similarity of the trend feature vectors between the corresponding analysis segments is calculated, and the density clustering algorithm is used to cluster segments with similar trend features into one class, forming a trend pattern cluster. The boundaries of the initially defined trend segments are corrected by detecting abrupt changes in the first-order difference of the smooth sequence. Feature parameters are calculated for each defined trend segment to obtain a standardized trend feature descriptor and acquire the periodic trend features of the traffic data.

[0031] Based on the acquired time-series smoothing feature set, the fast Fourier transform is used to convert the time-domain data to the frequency domain. The core formula of the fast Fourier transform is: Where (x(n)) is the time-domain data sequence, N is the data length, (X(k)) is the frequency-domain data, and j is the imaginary unit. This transformation can clearly present the periodic components that are difficult to detect in the time domain in the frequency domain, thereby identifying significant periodic components in the flow data, such as daily, weekly, and monthly cycles. For each identified periodic component, its amplitude and phase characteristics are calculated. The amplitude reflects the intensity of the periodic fluctuation, while the phase reflects the starting position of the periodic fluctuation.

[0032] Using the identified cycle as the basic unit, time-series data is divided into continuous analytical segments using a sliding window technique. For example, if a daily cycle is identified, the data is divided into multiple continuous 24-hour analytical segments using a 1-hour sliding window, with 24 hours as the basic unit. For each analytical segment, the overall slope of the data is calculated using linear regression. The linear regression equation is: (y = ax + b), where y is the flow data, x is the time variable, a is the slope, and b is the intercept. The slope can be used to determine the overall trend of the data segment (rising, falling, or stable). Combining the second-order difference characteristics of the smoothed series, the trend type is further refined, such as rapid rise, slow rise, rapid fall, slow fall, and stable fluctuations.

[0033] The boundaries of the initially defined trend segments are corrected by detecting abrupt changes in the first-order difference of the smoothed sequence. The rate of change of the first-order difference sequence is calculated. If the rate of change of the first-order difference exceeds a set threshold at a certain moment, that moment is identified as an abrupt change point, and the trend segment boundary is adjusted to the position of the abrupt change point to ensure the consistency of the data trend within each trend segment. Feature parameters, such as the mean, standard deviation, slope, and amplitude of the data within the segment, are calculated for each defined trend segment. These feature parameters are then standardized to obtain standardized trend feature descriptors, ultimately obtaining the periodic trend features of the traffic data, providing a basis for subsequent construction of multi-dimensional feature sets and anomaly detection.

[0034] Specifically, the process of obtaining the multi-dimensional feature set of associated identifiers includes: pre-setting a multi-dimensional scene coordinate system, dividing into multi-dimensional categories, and obtaining a standardized scene coordinate library; forming a mapping rule library based on the matching conditions between the corresponding trends and scene coordinates, and generating a unique identifier for the associated feature trends and scene coordinates, including corresponding trend type codes, scene coordinate codes, and timestamps; and aggregating the trend feature data, basic statistical features, and time-series smoothing features based on the unique identifiers to obtain a multi-dimensional feature set of data self-features and scene-related features.

[0035] Specifically, the process of correcting outliers includes recording the corresponding candidate outliers according to a preset threshold, applying differentiated corrections for different outlier types, and then substituting the corrected data into the multi-dimensional feature set to obtain corrected data for secondary outliers.

[0036] In this embodiment, differentiated correction methods are used for different types of outliers. For abnormal acquisition latency, if the investigation reveals it is due to temporary network congestion, a timestamp calibration method is used for correction. The network logs of the data transmission process are queried to obtain the actual data generation time and transmission latency. The acquisition time is corrected to the data generation time plus the normal transmission latency. The normal transmission latency is determined by statistically analyzing the historical average transmission latency of the same time period and network link. For abnormal traffic data deviation, if it is determined to be pulse interference caused by a momentary device malfunction, a scenario-based mean correction method is used. Normal traffic data from the same device during the same scenario time period is selected, and its weighted average is calculated.

[0037] The corrected data is then fed back into the multi-dimensional feature set to recalculate its correlation with other features and its consistency within the overall feature set. If no new anomalies are detected during the second verification, the correction is considered effective. If anomalies still occur, a deep correction process is initiated. This process uses historical data from the device's recent complete cycle to train a local regression model. Using time, ambient temperature, and number of users as independent variables and traffic data as the dependent variable, a regression equation is constructed. This model predicts the reasonable values ​​that the corrected data should have, and the correction is repeated until the data conforms to the characteristics of the scenario. This ensures that the final traffic data feature set is accurate and reliable, providing high-quality data for subsequent model training and quality assessment.

[0038] Specifically, obtaining the first identifier includes: labeling samples and dividing the dataset based on the traffic data feature set; performing feature mapping on the labeled traffic data feature set to capture the nonlinear correlation between features; and obtaining the first identifier for traffic anomaly feature detection through probability normalization processing.

[0039] Specifically, obtaining the second identifier includes: identifying outliers based on unannotated data through a data outlier isolation mechanism; adjusting the model sensitivity by setting an outlier ratio adaptation parameters; and obtaining a standardized second identifier for outlier traffic data anomaly detection through training.

[0040] Specifically, obtaining the multidimensional traffic quality assessment result includes: constructing a scenario-based verification set for the corresponding network scenario based on the first identifier and the second identifier; dynamically fusion the weights of the first identifier and the second identifier; introducing a preset weight calculation model; and calculating the multidimensional traffic quality assessment result by combining the dynamically corrected weights.

[0041] In this embodiment, the processing procedure is as follows: Random Forest Input Features: Basic Features: Cleaned traffic values, device type (encoded as categorical variables), timestamps (extracting periodic features such as hour / day of the week); Temporal Features: Sliding window statistics (5-minute mean, slope, standard deviation), historical mean for the same period, sudden increase / decrease threshold deviation; Business Features: Network region to which the device belongs, number of historical abnormal records (e.g., number of abnormalities in the past 7 days). Output: Deviation Probability: The probability of predicting that the current traffic deviates from the normal range (0~1), normalized by the Sigmoid activation function; Training Strategy: Using historical labeled data (manually confirmed abnormal / normal samples), the training set / validation set is divided in an 8:2 ratio; Adjusting hyperparameters (e.g., tree depth, sample sampling ratio) to balance recall and precision; Unsupervised Learning Model: Isolation Forest; Input Features: Raw Traffic Data: Uncleaned raw values ​​(preserving potential unknown abnormal patterns); Temporal Smoothing Features: Moving average, exponentially weighted average (suppressing short-term noise).

[0042] Output: Anomaly score: range (0,1), the lower the score, the more abnormal (e.g., a score <0.5 is considered abnormal); Training strategy: use all historical data (no annotation required), construct multiple isolation trees by random partitioning; set the anomaly sample ratio parameter (e.g., contamination=0.01) to adapt to the anomaly density in different scenarios.

[0043] Weight allocation: Weights are dynamically adjusted based on the model's performance on the validation set. The formula for random forest weights is as follows: The formula for the weights of an isolated forest is as follows: The AUC value of the first identifier on the validation set is (AUC1), and the AUC value of the second identifier is (AUC2).

[0044] After model training, predictions are made on the validation set data, outputting the raw anomaly probability value (range 0-100) for each sample. The probability is then normalized using the Sigmoid function, as shown below: Where x is the original anomaly probability, and P is the normalized probability. RF The range is compressed to 0-1, serving as the primary identifier for detecting traffic anomalies. For example, if the original anomaly probability of a sample is 75, after normalization, P... RF =0.92, indicating a high confidence level that the sample belongs to the anomalous category.

[0045] For the isolated forest model, the model outputs the anomaly score S for each sample. IF (Range 0-1), the lower the score, the higher the degree of outlier. For example, a sample S of an industrial electricity meter. IF =0.23, which is lower than the critical value of 0.3 corresponding to contamination=0.05, and is therefore judged as an outlier. The outlier score S... IF The data is standardized by using Min-Max normalization to map it to the 0-1 interval (which already satisfies the original output range of the model; this is only for consistency verification), resulting in the second identifier. For example, the original S of a residential electricity meter sample... IF =0.65, which remains 0.65 after standardization, indicating that the sample has a low degree of outlier and belongs to normal data.

[0046] In the scenario of evaluating the quality of billing traffic collection, a scenario-based validation set is constructed for the morning peak (9:00-12:00) and nighttime closing (22:00-6:00) traffic scenarios of commercial complexes. The AUC value of the random forest model in the corresponding scenarios is then used. Taking an air conditioning billing traffic collection device in a shopping mall as an example, the first identifier PRF output by the random forest is 0.75 (representing the probability that the current traffic data of this device deviates from the normal range, after normalization, is 0.75), and the second identifier SIF output by the isolated forest is 0.3 (after transformation, 1−SIF=0.7, reflecting the probability of anomalies caused by outliers). Assuming that the weight of the random forest is 0.55 and the weight of the isolated forest is 0.45 calculated from the AUC value, the final deviation probability of the traffic quality evaluation = 0.55×0.75 + 0.45×0.7 = 0.7275. If the preset traffic quality anomaly threshold for this scenario is 0.7, then this result is higher than the threshold, and it is determined to be an abnormal traffic collection quality, triggering an alarm to prompt maintenance personnel to check the equipment collection link.

[0047] Specifically, the hierarchical visualization monitoring interface includes: a near real-time identification and early warning module for acquisition deviation added to the scene layer, which captures real-time data at a preset frequency and compares it with the corresponding threshold in the scene-based adaptive threshold library to trigger the corresponding early warning indicator; in the indicator layer, a periodic detection data display area is used to periodically report the periodic traffic of the corresponding devices, and the deviation between the actual traffic and the benchmark curve is obtained by overlaying the periodic benchmark curve on the time axis; the hierarchical visualization monitoring interface is obtained by overlaying the scene layer and the indicator layer.

[0048] Specifically, the closed-loop processing of the billing traffic collection quality assessment involves establishing a multi-dimensional weight allocation mechanism based on the multi-dimensional traffic collection quality assessment results presented by the hierarchical visual monitoring interface, quantifying the priority of the impact of assessment indicators on collection quality, obtaining the iteration of traffic collection quality, and thus obtaining the closed-loop processing of the billing traffic collection quality assessment.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for evaluating the quality of billing traffic collection based on machine learning, characterized in that, include: S1: Obtain structured data packets of 5G traffic data through preset traffic collection points; S2: Calculate the data distribution characteristics of the original billing traffic data in the structured data packet, extract basic statistical features, and extract time-series smoothing features by combining time series trend analysis logic; Periodically mine traffic data trends, associate and map the trends of different traffic data with preset multi-dimensional scenarios, and obtain a multi-dimensional feature set of association identifiers; Based on the multi-dimensional feature set, an adaptive threshold generation mechanism is used to generate a scenario-based adaptive threshold library, and outliers are corrected to obtain a corrected traffic data feature set. S3: Construct a feature mapping learning model and an outlier pattern isolation model based on the traffic data feature set and link the feature set together. By training the feature mapping learning model and the outlier pattern isolation model, obtain the first and second identifiers for traffic anomaly feature detection. Dynamically associate the first and second identifiers to obtain the evaluation results of multi-dimensional traffic acquisition quality. S4: Based on the evaluation results of the multi-dimensional traffic acquisition quality, a coordinate system is constructed to obtain a hierarchical visual monitoring interface, and the evaluation of the billing traffic acquisition quality is processed in a closed loop.

2. The method according to claim 1, characterized in that, The method for extracting the basic statistical features is as follows: based on the original billing traffic data in the structured data packet, the range, interquartile range, standard deviation, and variance are calculated to obtain the data dispersion, and the skewness coefficient and kurtosis coefficient are calculated to characterize the symmetry and steepness of the data distribution pattern, thereby obtaining the basic statistical features.

3. The method according to claim 1, characterized in that, The process of obtaining the time-series smoothing features includes filling missing values ​​and filtering outliers in the time-series billing traffic data in the structured data packets, processing missing data using linear interpolation, and identifying and isolating outliers using corresponding criteria. Construct sliding windows with different time scales, and smooth the data sequences within the corresponding windows using an exponentially weighted moving average algorithm, then calculate the residuals between the original data sequences and the smoothed sequences. The first and second differences are calculated on the smoothed sequence to obtain the rate of change and acceleration characteristics of the data trend and identify the trend inflection point. The smoothed sequence features, residual features and trend change features at different time scales are fused and dimensionality is reduced by principal component analysis algorithm to obtain the time series smoothing feature set.

4. The method according to claim 3, characterized in that, The dimensionality reduction process specifically includes optimizing the dimensionality of the original time-series smooth feature set based on the principal component analysis algorithm, while retaining principal components whose cumulative variance contribution rate reaches a preset threshold, and obtaining low-dimensional features with corresponding trend and fluctuation information.

5. The method according to claim 1, characterized in that, The specific process of periodically mining the trend of traffic data is as follows: based on the time-series smoothing feature set, the time-domain data is converted to the frequency domain using fast Fourier transform, significant periodic components in the traffic data are identified, and the amplitude and phase characteristics of each periodic component are calculated. Based on the identification period as the basic unit, the time series data is divided into continuous analysis segments using the sliding window technique. For each analysis segment, the overall slope of the data is calculated by linear regression, and the trend type is obtained by combining the second-order difference feature of the smooth sequence. The cosine similarity of the trend feature vectors between the corresponding analysis segments is calculated, and the density clustering algorithm is used to cluster segments with similar trend features into one class, forming a trend pattern cluster. The boundaries of the initially defined trend segments are corrected by detecting abrupt changes in the first-order difference of the smooth sequence. Feature parameters are calculated for each defined trend segment to obtain a standardized trend feature descriptor and acquire the periodic trend features of the traffic data.

6. The method according to claim 1, characterized in that, The acquisition of the multi-dimensional feature set of the associated identifier specifically includes: pre-setting a multi-dimensional scene coordinate system, dividing multi-dimensional categories, and acquiring a standardized scene coordinate library; forming a mapping rule library based on the matching conditions of the corresponding trends and scene coordinates, and generating a unique identifier of the associated feature trends and scene coordinates, including the corresponding trend type code, scene coordinate code, and timestamp. Based on the unique identifier, the trend feature data, basic statistical features, and time-series smoothing features are aggregated to obtain a multi-dimensional feature set of data self-features and scene-related features.

7. The method according to claim 1, characterized in that, The specific process of correcting outliers includes recording the corresponding candidate outliers according to a preset threshold, applying differentiated corrections for different outlier types, and then substituting the corrected data into the multi-dimensional feature set to obtain corrected data for secondary outliers.

8. The method according to claim 1, characterized in that, The process of obtaining the first identifier specifically includes: labeling samples and dividing the dataset based on the traffic data feature set; performing feature mapping on the labeled traffic data feature set to capture the nonlinear correlation between features; and obtaining the first identifier for traffic anomaly feature detection through probability normalization processing.

9. The method according to claim 1, characterized in that, The acquisition of the second identifier specifically includes: identifying outliers based on unannotated data through a data outlier isolation mechanism; adjusting the model sensitivity by setting an outlier ratio adaptation parameters; and acquiring a standardized second identifier for outlier traffic data anomaly detection through training.

10. The method according to claim 1, characterized in that, The process of obtaining the multidimensional traffic quality assessment result includes: constructing a scenario-based verification set for the corresponding network scenario based on the first identifier and the second identifier; dynamically fusion the first identifier and the second identifier with weights; introducing a preset weight calculation model; and calculating the multidimensional traffic quality assessment result by combining the dynamically corrected weights.

11. The method according to claim 1, characterized in that, The hierarchical visualization monitoring interface includes a near real-time identification and early warning module for acquisition deviation added to the scene layer. It captures real-time data at a preset frequency and compares it with the corresponding threshold in the scene-based adaptive threshold library to trigger the corresponding early warning indicator. In the indicator layer, the periodic detection data display area is used to report the periodic flow of the corresponding devices at regular intervals. By superimposing the periodic benchmark curve on the time axis, the deviation between the actual flow and the benchmark curve is obtained. A hierarchical visual monitoring interface is obtained by superimposing the scene layer and the indicator layer.

12. The method according to claim 1, characterized in that, The closed-loop processing of the billing traffic collection quality assessment is as follows: based on the multi-dimensional traffic collection quality assessment results presented by the hierarchical visual monitoring interface, a multi-dimensional weight allocation mechanism is established to quantify the priority of the impact of assessment indicators on collection quality, obtain the iteration of traffic collection quality, and obtain the closed-loop processing of the billing traffic collection quality assessment.

Images