Intelligent identification and optimization system for VORR high-packet-loss cell based on machine learning

Abstract

The invention relates to the technical field of 5G communication network optimization, in particular to a VORR high packet loss cell intelligent identification and optimization system based on machine learning. The system comprises a data acquisition and preprocessing module, an index feature mining and modeling analysis module, a high packet loss cell intelligent identification module, an optimization strategy recommendation module and a system integration and tool output module. Wherein the data acquisition and preprocessing module gathers VORR cell multi-source heterogeneous data and completes preprocessing, the index feature mining and modeling analysis module screens core indexes through feature engineering and trains and optimizes a machine learning model, and the high packet loss cell intelligent identification module realizes whole network cell packet loss risk scoring and problem sorting based on the model. The optimization strategy recommendation module outputs targeted optimization suggestions based on a knowledge graph, and the system integration module realizes instrumentalized output and is in butt joint with an operator network optimization platform. The 5G voice service quality and the network operation and maintenance efficiency are remarkably improved, the manual intervention cost is reduced, and cross-manufacturer and cross-region generalization deployment is supported.

Description

Technical Field

[0001] This invention relates to the technical field of 5G communication network optimization, and more particularly to a machine learning-based intelligent identification and optimization system for VoNR high packet loss cells. Background Technology

[0002] With the accelerated global commercialization of 5G, VoNR (Voice over New Radio), as the native voice technology of 5G networks, is gradually replacing VoLTE as the mainstream voice transmission method. Internationally, major telecommunications equipment vendors such as Ericsson and Nokia have conducted research on VoNR voice quality assurance, while operators such as Verizon and NTT Docomo have already implemented commercial VoNR deployments and are continuously optimizing their networks.

[0003] In China, although the three major operators have completed the deployment of VoNR in some cities, the VoNR network still faces key performance problems in actual operation, such as low call connection rate, high call drop rate, and excessive packet loss rate, due to the limitations of the current 5G network NSA networking architecture, SA scheduling strategy adaptation issues, and insufficient NR coverage. These problems are particularly prominent in high packet loss scenarios, weak coverage areas, and handover boundary areas, which seriously affect the voice service experience of 5G users.

[0004] Traditional VoNR network optimization methods mainly rely on the experience and judgment of network optimization experts and fixed threshold screening. This approach has obvious limitations: on the one hand, it is difficult to adapt to the complex and ever-changing topology and dynamic operating scenarios of 5G networks, and cannot cope with the surge in data volume brought about by the expansion of network scale; on the other hand, manual analysis is inefficient and highly subjective, making it difficult to accurately identify and optimize high packet loss cells in real time, resulting in unstable optimization effects and high operation and maintenance costs.

[0005] To address this, while domestic research institutions and operators have gradually attempted to introduce technologies such as machine learning and data mining for intelligent network optimization, existing solutions have not yet formed a complete closed-loop system from multi-source data fusion and model training optimization to strategy recommendation. They still have shortcomings in terms of cross-vendor compatibility, scenario generalization ability, and strategy interpretability.

[0006] Therefore, there is an urgent need for a system that can intelligently identify and accurately optimize VoNR high packet loss cells in order to solve the shortcomings of existing technologies and support the continuous improvement of 5G voice service quality. Summary of the Invention

[0007] This invention proposes a distribution network planning optimization method and system oriented towards multi-energy coordination and low-carbon goals, realizing intelligent identification and accurate optimization of VoNR high packet loss cells.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On one hand, a machine learning-based VoNR high packet loss cell intelligent identification and optimization system is provided, including a data acquisition and preprocessing module, an indicator feature mining and modeling analysis module, a high packet loss cell intelligent identification module, an optimization strategy recommendation module, and a system integration and tool output module that interacts with the high packet loss cell intelligent identification module and the optimization strategy recommendation module respectively. The data acquisition and preprocessing module is used to aggregate multi-source heterogeneous data from VoNR cells, preprocess the multi-source heterogeneous data, and construct a high packet loss sample library and training set. The indicator feature mining and modeling analysis module is used to identify core indicators that are highly correlated with VoNR packet loss using feature engineering methods, train and optimize various machine learning models, and improve the adaptability of the models in different network scenarios through generalization testing. The high packet loss cell intelligent identification module is used to score the packet loss risk of VoNR cells across the entire network based on an optimized machine learning model, generate a sorted list of problematic cells, and output the screening criteria to support visual analysis and manual verification. The optimization strategy recommendation module is used to establish a strategy library based on optimization knowledge graphs or expert experience rules, match the packet loss cause type of the problem cell with the strategy library, output targeted optimization suggestions and support interpretable output; The system integration and tool output module is used to form a graphical, semi-automated tool system that supports historical problem tracking, strategy execution recording, effect comparison and evaluation, and connects to the operator's network optimization platform.

[0009] Preferably, the multi-source heterogeneous data collected by the data acquisition and preprocessing module includes: KQI / KPI data, MR measurement data, alarm data, parameter data, and user perception data. The KQI / KPI data includes VoNR voice uplink PDCP layer user plane packet loss rate, VoNR voice downlink PDCP layer user plane packet loss rate, wireless call success rate, call drop rate, and handover success rate; the MR measurement data includes RSRP, SINR, BLER, and MR coverage; the alarm data includes cell outage alarms, BBU optical module anomaly alarms, and cell VSWR alarms; the parameter data includes cell frequency band, frequency point configuration, PCI planning, handover threshold, and power control parameters; and the user perception data includes user complaint records, user rating feedback, and VoNR voice quality assessment data.

[0010] Preferably, the preprocessing operations of the data acquisition and preprocessing module include: data cleaning, removing invalid data, duplicate data, and missing values ​​to ensure data integrity and consistency; data normalization, standardizing data from different sources and in different formats, including converting KPI data to a unified unit and unifying timestamps to the ISO standard format; label construction, using whether the VoNR voice packet loss rate exceeds a preset threshold as the criterion to mark cells as having high packet loss or normal, forming supervised learning labels, where the preset threshold is 1.5%; outlier handling, using the Z-score method to identify and process outliers; and time alignment and spatial merging, performing time dimension alignment and spatial dimension merging on KQI / KPI data, MR measurement data, alarm data, and parameter data to ensure data spatiotemporal consistency.

[0011] Preferably, the feature engineering method of the indicator feature mining and modeling analysis module includes: feature extraction, extracting core features from the preprocessed data. The core features include coverage quality features, network performance features, alarm features, parameter configuration features, and user perception features. The coverage quality features include RSRP, SINR, BLER, and MR coverage. The network performance features include at least wireless call success rate, call drop rate, handover success rate, and PRB utilization. The alarm features include the number of alarms, alarm type, and alarm duration. The parameter configuration features include frequency band, frequency point, PCI, and handover threshold. The user perception features include user ratings and number of complaints.

[0012] Preferably, the model training and optimization process of the module includes: The high packet loss sample database was divided into a training set, a validation set, and a test set in a ratio of 7:2:1. A machine learning model was trained using the training set, and its hyperparameters were optimized through cross-validation. Model performance was evaluated on the validation set to select the optimal model. The model's generalization ability was validated on the test set. Model evaluation metrics included accuracy, precision, recall, F1 score, AUC-ROC curve, and feature importance, which was obtained through SHAP value or feature contribution analysis.

[0013] Preferably, the workflow of the high packet loss cell intelligent identification module includes: data input, receiving real-time collected and preprocessed KQI / KPI data, MR measurement data, alarm data, and parameter data; calling the optimized machine learning model to infer the input data and outputting a prediction result of high packet loss or normal; classifying high packet loss cells into three levels—high risk, medium risk, and low risk—based on the prediction result and packet loss rate, and sorting them according to risk level; visually displaying the geographical distribution of high packet loss cells through a risk map, generating a problem list including cell ID, region, network standard, equipment manufacturer, frequency band, and packet loss rate, and outputting a historical packet loss trend map of high packet loss cells.

[0014] Preferably, the knowledge graph construction of the optimization strategy recommendation module includes: constructing a knowledge graph using a graph database, where nodes include problem type, optimization strategy, equipment manufacturer, parameter configuration, and historical cases; the data sources for the knowledge graph include network optimization expert experience, historical optimization cases, and equipment manufacturer recommendation strategies, where historical optimization cases include site replacement cases, parameter adjustment cases, and interference troubleshooting cases; and the knowledge graph update mechanism involves connecting with the operator's network optimization platform via an API interface to periodically synchronize optimization strategy execution results and new cases, thereby updating the knowledge graph content.

[0015] Preferably, the strategy matching and output of the optimization strategy recommendation module includes: The strategy matching algorithm matches and associates optimization strategies in a knowledge graph based on the problem types of high packet loss cells. The problem types include weak coverage, severe interference, handover failure, insufficient capacity, and abnormal parameter configuration. Strategy scoring evaluates the matched optimization strategies based on three dimensions: historical execution success rate, strategy complexity, and execution cost. The strategy output lists optimization strategies in descending order of scores. These optimization strategies include parameter adjustment, base station replenishment, antenna and feeder power adjustment, PCI replanning, neighbor cell optimization, bandwidth adjustment, multi-frequency splitting, and interference investigation. Interpretable output, displaying the rationale for strategy recommendations, historical execution results, and success rate data through a visual interface.

[0016] Preferably, the system integration and tool output module includes: The API interface submodule connects with the operator's network optimization platform through the API interface to realize automatic issuance of problem work orders, synchronization of policy execution records, and feedback of optimization effect data. The edge computing submodule deploys the model inference submodule and the strategy recommendation submodule on the edge side to improve the data processing response speed; The permission hierarchical management submodule assigns different system access permissions and operation permissions to wireless network engineers and administrators. The report export submodule provides standardized templates and supports exporting analysis results and optimization suggestions.

[0017] As a preferred approach, the model input features and policy matching logic are adjusted according to the characteristics of different network environments to improve system adaptability.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. Achieve full automation of VoNR high packet loss cell optimization process, significantly improving operation and maintenance efficiency: This invention achieves automatic access and standardized processing of multi-source heterogeneous data through data acquisition and preprocessing modules, combined with machine learning model training of indicator feature mining and modeling analysis modules, and real-time risk scoring of high packet loss cell intelligent identification modules. Finally, targeted suggestions are output through optimization strategy recommendation modules, forming a full-process automated system of "data acquisition - model identification - strategy recommendation - effect evaluation". It eliminates the need to rely on human experience to complete problem location and optimization solution formulation, significantly reducing the cost of manual intervention and improving network operation and maintenance efficiency by more than 30%.

[0019] 2. Improve the accuracy and generalization ability of high packet loss cell identification: This invention adopts multi-source data fusion technology (integrating KQI / KPI, MR, alarm, parameter and user perception data), selects core indicators that are highly correlated with VoNR packet loss through feature engineering, and compares and trains various machine learning models such as XGBoost, LightGBM, and random forest. Combined with cross-validation and generalization testing to optimize model parameters, the accuracy and recall rate of high packet loss cell identification are improved to over 90%. At the same time, it can adapt to different network standards (NSA / SA), different equipment manufacturers (Huawei, ZTE, Ericsson) and different regional network scenarios, solving the problem of poor adaptability of traditional threshold screening.

[0020] 3. The optimization strategy is highly scientific and feasible, ensuring the quality of 5G voice services: The optimization strategy recommendation module of this invention is based on knowledge graph construction, integrating the experience of network optimization experts, historical optimization cases (site replacement, parameter adjustment, interference investigation, etc.) and equipment manufacturer recommended strategies. It can match the optimal strategy according to the specific problem type of high packet loss cells (weak coverage, interference, insufficient capacity, etc.) and output the basis for strategy recommendation, historical execution success rate and expected effect, which is convenient for wireless network engineers to understand and implement. Actual verification shows that after adopting the optimization strategy recommended by this invention, the average packet loss rate of VoNR voice PDCP layer is reduced by 40%-60%, and the user voice perception score is improved by 15%-25%.

[0021] 4. Supports cross-vendor and cross-regional deployment and closed-loop continuous optimization, with wide applicability: This invention, through standardized data interface and API integration design, is compatible with network data formats of different equipment manufacturers and operator network optimization platforms, enabling generalized deployment across manufacturers and regions. At the same time, the system has a built-in closed-loop management and continuous learning mechanism, which can feed back the optimization strategy execution effect to the model and knowledge graph, dynamically adjust model parameters and update the strategy library, so that the system can continuously improve recognition accuracy and strategy adaptability in long-term operation, avoiding the decay of optimization effect due to changes in the network environment.

[0022] 5. Promoting the transformation of 5G network operation and maintenance from "passive optimization" to "proactive discovery": Traditional optimization methods require passive investigation after user complaints or abnormal network indicators. This invention can identify high-risk cells for packet loss in advance and issue tiered warnings through real-time data collection and model inference. Combined with trend analysis functions, it can determine whether the problem is sudden or long-term, realizing proactive monitoring and prediction of VoNR network quality. This shortens the problem handling cycle from "days" to "hours", effectively reducing user complaints caused by packet loss and improving the overall service level of 5G networks. Attached Figure Description

[0023] Figure 1 This is a flowchart of a machine learning-based intelligent identification and optimization system for high packet loss cells in VoNR, as described in a specific embodiment of the present invention. Figure 2 This is a network performance data diagram in a specific embodiment of the present invention.

[0024] Figure 3 This is a graph of measurement data from a specific embodiment of the present invention.

[0025] Figure 4 This is an alarm data diagram from a specific embodiment of the present invention.

[0026] Figure 5 This is a diagram of neighboring cell parameters in a specific embodiment of the present invention.

[0027] Figure 6 This is a summary data chart of weekly or monthly granularity of the cell in a specific embodiment of the present invention.

[0028] Figure 7 This is a problem attribution diagram in a specific embodiment of the present invention.

[0029] Figure 8 This is a recommended result diagram in a specific embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please refer to Figure 1 As shown, this application proposes a machine learning-based intelligent identification and optimization system for VoNR high packet loss cells, including a data acquisition and preprocessing module, an indicator feature mining and modeling analysis module, a high packet loss cell intelligent identification module, an optimization strategy recommendation module, and a system integration and tool output module that interacts with the high packet loss cell intelligent identification module and the optimization strategy recommendation module respectively. The data acquisition and preprocessing module is used to aggregate multi-source heterogeneous data from VoNR cells, preprocess the multi-source heterogeneous data, and construct a high packet loss sample library and training set. This module is responsible for collecting multi-source heterogeneous data from the 5G network and performing preprocessing operations such as data cleaning, normalization, and label construction to provide a high-quality data foundation for subsequent model training and optimization strategy recommendations.

[0032] (1) Data collection KQI / KPI data: Collect key quality indicators (KQI) and key performance indicators (KPI) for VoNR voice services, such as VoNR voice uplink / downlink PDCP layer user plane packet loss rate, wireless call completion rate, call drop rate, handover success rate, etc.

[0033] MR measurement data: collect coverage quality data from the Mobility Measurement Report (MR), such as RSRP (Reference Signal Received Power) (dBm), SINR (Signal Interference-Noise Ratio) (dB), BLER (Block Error Rate), etc.

[0034] Alarm data: Collects alarm information from network devices, including cell outages, BBU optical module malfunctions, cell VSWR alarms, etc., to identify abnormal events in the network.

[0035] Parameter data: Collect cell configuration parameters, such as frequency bands (N1, N78, N5, etc.), frequency point configuration, PCI planning, handover threshold, power control parameters, etc., to analyze the impact of parameter configuration on packet loss.

[0036] User perception data: Collects data such as user complaint records, rating feedback, and VoNR voice quality assessment to evaluate the actual effect of optimization strategies.

[0037] (2) Data preprocessing Data cleaning: Remove invalid data, duplicate data, and missing values ​​to ensure data integrity and consistency.

[0038] Data normalization: Standardize data from different sources and in different formats, such as converting KPI data to a unified unit and unifying timestamps to the ISO standard format.

[0039] Label construction: Based on whether the VoNR voice packet loss rate exceeds a set threshold (e.g., 1.5%), the cell is labeled as either "high packet loss" or "normal" as labels for supervised learning.

[0040] Outlier handling: Outliers are identified and handled using methods such as Z-score and IQR to avoid model training bias.

[0041] Time alignment and spatial merging: Time alignment and spatial merging are performed on the collected KPI, MR, alarm, parameter and other data to ensure the consistency of data in time and space dimensions.

[0042] The indicator feature mining and modeling analysis module is used to identify core indicators that are highly correlated with VoNR packet loss using feature engineering methods, train and optimize various machine learning models, and improve the adaptability of the models in different network scenarios through generalization testing. This module is responsible for extracting key features from the preprocessed data and building machine learning models based on these features to identify VoNR high packet loss cells.

[0043] (1) Feature Engineering Feature extraction: Extract features related to VoNR packet loss rate from the raw data, including: Coverage quality characteristics: such as RSRP, SINR, BLER, MR coverage, etc.

[0044] Network performance characteristics: such as wireless call success rate, call drop rate, handover success rate, PRB utilization, etc.

[0045] Alarm characteristics: such as the number of alarms, alarm type, alarm duration, etc.

[0046] Parameter configuration features: such as frequency band, frequency point, PCI, switching threshold, power control parameters, etc.

[0047] User-perceived characteristics: such as user ratings, number of complaints, voice quality assessment, etc.

[0048] Feature selection: Feature importance analysis methods (such as Shapley value-based interpretation and recursive feature elimination method RFE) are used to screen features that are highly correlated with VoNR packet loss rate, thereby improving the model's generalization ability and recognition accuracy.

[0049] Feature encoding: One-hot encoding or label encoding is performed on categorical features (such as frequency band, manufacturer, region, etc.) to adapt to the input requirements of machine learning models.

[0050] (2) Model training and evaluation Model selection: Supervised learning algorithms (such as XGBoost, LightGBM, Random Forest, Logistic Regression, etc.) are used to model and train the preprocessed data.

[0051] Model training process: The dataset is divided into training, validation, and test sets.

[0052] The model is trained using the training set, and the model parameters are optimized through cross-validation.

[0053] Evaluate model performance on the validation set and select the optimal model.

[0054] The model's generalization ability is evaluated on the test set to ensure its adaptability in different network scenarios.

[0055] Model evaluation metrics: accuracy, precision, recall, F1 score, AUC-ROC curve, feature importance analysis (such as SHAP value, feature contribution).

[0056] The high packet loss cell intelligent identification module is used to score the packet loss risk of VoNR cells across the entire network based on an optimized machine learning model, generate a sorted list of problematic cells, and output the screening criteria to support visual analysis and manual verification. This module uses a pre-trained machine learning model to identify VoNR cells in real time, determine whether they are high packet loss cells, and output the identification results.

[0057] (1) Identification process Data input: Input real-time collected data such as KQI / KPI, MR, alarms, and parameters into the model.

[0058] Model inference: The model infers from the input data and outputs a prediction result (high packet loss or normal).

[0059] Results output: The recognition results are output to a visualization platform for wireless network engineers to view and process.

[0060] Risk level classification: Based on the identification results, high packet loss cells are classified into different risk levels (e.g., high risk, medium risk, low risk) and prioritized.

[0061] (2) Display of recognition results Risk Map: Displays the distribution of cells with high packet loss on a map, making it easy to quickly locate problem areas.

[0062] Problem List: List detailed information about cells with high packet loss, including time, region, network type, manufacturer, cell ID, frequency band, packet loss rate, etc.

[0063] Trend Analysis: Displays the historical trends of cells with high packet loss, facilitating analysis of whether the problem is sudden or long-term.

[0064] The optimization strategy recommendation module is used to establish a strategy library based on optimization knowledge graphs or expert experience rules, match the packet loss cause type of the problem cell with the strategy library, output targeted optimization suggestions and support interpretable output; Based on knowledge graph technology, the module combines recognition results with historical optimization cases to recommend targeted optimization strategies to improve the quality of VoNR voice services.

[0065] (1) Knowledge Graph Construction Knowledge graph structure: The knowledge graph is constructed using a graph database (such as Neo4j or TigerGraph), and the nodes include "problem type", "optimization strategy", "equipment manufacturer", "parameter configuration", "historical cases", etc.

[0066] The knowledge graph data sources include: network optimization expert experience, historical optimization cases (such as site replacement, parameter adjustment, interference troubleshooting, etc.), and equipment manufacturer recommendation strategies (such as Huawei, ZTE, Ericsson, etc.).

[0067] Knowledge graph update mechanism: The knowledge graph content is updated regularly through API interface with the network optimization platform to ensure the timeliness and accuracy of the strategy base.

[0068] (2) Recommended optimization strategies Policy matching algorithm: Based on the problem type in the identification results (e.g., weak coverage, severe interference, handover failure, etc.), the most relevant optimization strategy is matched in the knowledge graph.

[0069] Strategy recommendation logic: Based on the question type in the recognition results, query the matching strategy in the knowledge graph.

[0070] The matched strategies are scored, with scoring dimensions including historical execution success rate, strategy complexity, and execution cost.

[0071] Output a list of recommended strategies, sorted by rating, for wireless network engineers to choose from and implement.

[0072] Interpretable output: When recommending strategies, the system displays information such as the basis for the recommendation, historical execution results, and success rate through a visual interface, making it easier for wireless network engineers to understand and implement the strategy.

[0073] The system integration and tool output module is used to form a graphical, semi-automated tool system that supports historical problem tracking, strategy execution recording, effect comparison and evaluation, and connects to the operator's network optimization platform.

[0074] This module is responsible for integrating the system with a network optimization platform of a certain operator, and outputting optimization strategies and analysis results to achieve closed-loop management of the entire process.

[0075] (1) System integration API Interface: By connecting with a network optimization platform of a certain operator through the API interface, it can realize the automatic issuance of problem work orders, the automatic synchronization of execution records, and the comparative analysis of optimization effects.

[0076] Edge computing module: Deploy model inference and policy recommendation modules at the edge to improve response speed and processing power.

[0077] Data synchronization mechanism: Ensures data synchronization between the system and the network optimization platform, including optimization strategy execution results and optimization effect evaluation data.

[0078] (2) Tool output Automatic issue ticket distribution: Identified high packet loss cells and recommended optimization strategies are automatically distributed as work tickets for wireless network engineers to execute.

[0079] Optimization effect comparison and analysis: By comparing data such as VoNR packet loss rate, user rating, and network performance indicators before and after optimization, the actual effect of the optimization strategy is evaluated.

[0080] Report Export: Provides standardized templates and supports exporting analysis results and optimization suggestions to PDF, Excel and other formats for reporting or archiving.

[0081] Hierarchical access control: Hierarchical access control ensures the access and operation permissions of users with different roles (such as wireless network engineers, administrators, etc.) in the system, thus guaranteeing data security.

[0082] The system of this invention also includes a closed-loop management and continuous learning mechanism, which can adjust the model input features and policy matching logic according to the characteristics of different network environments, thereby improving the system's adaptability, as detailed below: The results of the optimization strategy are fed back to the model for continuous model optimization and knowledge graph updates, thus achieving closed-loop system management.

[0083] (1) Closed-loop management process Problem identification: High packet loss cells are identified using machine learning models.

[0084] Strategy Recommendation: Optimization strategy based on knowledge graph.

[0085] Policy Execution: Wireless network engineers execute recommended optimization policies.

[0086] Performance evaluation: The system collects data such as VoNR packet loss rate, user ratings, and network performance indicators after execution to evaluate the optimization effect.

[0087] Model update: Feedback the optimization results to the model for continuous model optimization and knowledge graph updates.

[0088] (2) Continuous learning mechanism Model optimization: Based on the optimization results and user feedback, dynamically adjust the model parameters to improve the model's recognition accuracy.

[0089] Knowledge graph update: Update the knowledge graph with the results of optimization strategy execution and user feedback data to enrich the content of the strategy library.

[0090] Adaptive learning: The system has adaptive learning capabilities, which can automatically adjust the model and strategy recommendation logic according to different network environments and data characteristics, thereby improving the system's generalization ability and adaptability.

[0091] This embodiment provides a specific implementation process for a machine learning-based intelligent identification and optimization system for VoNR cells with high packet loss. Based on large-scale VoNR network data, the system automatically identifies cells with high packet loss risk using machine learning technology and combines this with knowledge graphs for intelligent optimization recommendations. The specific steps are as follows: 1. Data Acquisition and Preprocessing Stage By unifying the access to operator data, including but not limited to KQI / KPI performance data, MR measurement data, alarm data, neighbor cell parameter tables, and user complaint data, standardized interfaces are used for automated collection. The data preprocessing module cleans, deduplicates, fills in missing values, removes outliers, aligns time, and spatially merges the raw data to ensure data quality and analyzability, forming a unified standard data table as the basis for subsequent analysis.

[0092] First, by unifying the access of multi-source heterogeneous data from the operator's existing network, the data types include, but are not limited to: Network performance data (KQI / KPI): Wireless call success rate, call drop rate, handover success rate, number of RRC users, PRB utilization, VoNR voice downlink PDCP layer user plane packet loss rate, VoNR voice uplink PDCP layer user plane packet loss rate, etc.; network performance data graphs are shown below. Figure 2 As shown, the measurement data is as follows Figure 3As shown, the alarm data is as follows Figure 4 As shown, the neighboring cell parameters are as follows: Figure 5 As shown.

[0093] Neighbor cell parameter table: neighbor cell relationships, frequency configuration, PCI planning, etc.; User perception data: complaint records, rating feedback, etc.

[0094] Data is automatically collected through standardized interfaces, supporting batch processing, API, streaming data and other access methods.

[0095] The above data undergoes preprocessing operations such as cleaning, deduplication, missing value imputation, outlier removal, time alignment, and spatial merging to form a unified and standardized data table, ensuring data quality and the accuracy of subsequent analysis.

[0096] 2. Feature extraction and sample construction stage Based on multi-source data fusion, through statistical analysis and feature engineering methods, core feature indicators closely related to the packet loss rate of the VoNR voice PDCP layer user plane are screened, such as RSRP, SINR, BLER, PRB utilization, call connection rate, call drop rate, number of RRC users, alarm frequency, user rating, etc., to construct a training sample set. The label is set according to whether the cell belongs to the high packet loss risk, thus forming a supervised learning dataset.

[0097] Based on the preprocessed data, combined with industry experience and statistical methods, core features closely related to VoNR packet loss were identified, including: Wireless coverage parameters: RSRP, SINR, neighboring cell coverage quality; Access capabilities include: number of RRC users, PRB utilization, throughput, etc. Stability-related metrics: alarm count, abnormal events; User perception metrics: Complaint frequency, ratings; Derivative categories: volatility index, load index, alarm weight, etc.

[0098] Based on the high packet loss criteria set by the operator, all cells in the network are labeled to form a supervised learning dataset, which is divided into a training set, a validation set, and a test set (e.g., 7:2:1) as the basis for model input.

[0099] A certain operator's 5G VoNR high packet loss cell judgment criteria: A cell with either the downlink PDCP layer user plane packet loss rate or the uplink PDCP layer user plane packet loss rate > 1.5% is a VoNR high packet loss cell (precondition: "total number of downlink PDCP layer packets" > 100,000 or "total number of uplink PDCP layer packets" > 100,000; only cells that meet this condition participate in the poor quality screening).

[0100] Data is taken from the weekly or monthly granular summary table of the community, such as... Figure 6 As shown: V.9.2 requires the following rules when retrieving data: If VoNR_PDCP_LossPktUl is not empty and VoNR_PDCP_RxLossPktUl is empty, then take the value of VoNR_PDCP_LossPktUl and put it into VoNR_PDCP_RxLossPktUl; otherwise, directly take the value of VoNR_PDCP_RxLossPktUl.

[0101] 3. Model Training and Validation Phase An ensemble learning algorithm, including XGBoost and RandomForest, was employed to construct a high-packet-loss cell identification model based on the aforementioned feature data. Cross-validation and hyperparameter tuning were used to improve the model's generalization ability. Finally, the optimal model was selected and deployed on the platform, outputting a network-wide cell packet loss risk score and classification results. This model can adaptively adjust to regional differences to ensure high identification accuracy.

[0102] We selected mainstream algorithms such as XGBoost, LightGBM, and RandomForest for modeling, and compared and analyzed them to select the optimal model. The training process includes: Grid search hyperparameter tuning; Cross-validation prevents overfitting; Feature importance assessment to assist in feature optimization.

[0103] Evaluation metrics include accuracy, recall, precision, F1 score, AUC, KS score, etc., to ensure the robustness and generalization ability of the model.

[0104] Use interpretability methods such as SHAP to analyze feature contributions, aid in understanding model results, and improve transparency and traceability.

[0105] Finally, the best-performing model (e.g., XGBoost) is selected and saved in pkl / joblib format for later use.

[0106] 4. High packet loss cell screening and risk assessment stage The platform sorts all cells in the network according to their packet loss risk scores, filters out a list of cells with high packet loss risk, and supports multi-dimensional visualization (geographic heatmap, correlation matrix, trend analysis chart, etc.). Simultaneously, it automatically generates a problem attribution report by combining model output with expert experience rules, revealing the main causes of high packet loss (such as interference from neighboring cells, weak coverage, insufficient capacity, frequent handover, etc.).

[0107] Perform risk scoring (Score, 0-1) on all cells in the network, categorizing them into high, medium, and low risk based on the score. Output a list of problematic cells, including: Basic information such as cell ID, region, and frequency band; Risk scoring and key issue indicators; Problem attribution (e.g., weak coverage, insufficient capacity, severe interference, etc.).

[0108] It provides multi-dimensional visualizations (heatmaps, trend charts, distribution maps, etc.) to help wireless network engineers quickly locate problem areas and causes.

[0109] By combining model and rule analysis, the system automatically attributes problem types, generating clear causal labels for each problem, thus providing a basis for optimization strategies. Figure 7 As shown.

[0110] 5. Optimization Strategy Intelligent Recommendation Stage The platform has a built-in knowledge graph that integrates historical optimization cases, expert rules, industry standards, etc. Based on characteristics such as cell type, frequency band, coverage scenario, and load conditions, it automatically matches and recommends optimization strategies (such as adjusting parameters, adding base stations, optimizing PCI planning, optimizing neighbor cell relationships, adjusting scheduling strategies, etc.) and provides predictions of the effects before and after optimization, as well as interpretability evidence, for wireless network engineers to make decisions.

[0111] Based on knowledge graphs and expert rule bases, and combined with factors such as cell type, problem cause, and risk level, optimization strategies are automatically recommended. These strategies cover: Coverage improvement: base station replenishment, antenna and feeder adjustments, and power optimization; Capacity expansion: bandwidth adjustment, multi-frequency traffic splitting; Interference cancellation: PCI replanning, neighbor cell optimization, interference investigation; Alarm clearing: troubleshooting and configuration correction; Perception Enhancement: Adjustment of business parameters and optimization of scheduling.

[0112] The recommendation results include: community information, problem attribution, suggested measures, expected effects, and interpretability evidence. Output can be in Excel, PDF, or via API for use. The recommendation results are as follows: Figure 8 As shown.

[0113] 6. Results Output and Platform Integration Phase The system outputs analysis results and optimization suggestions through a user-friendly platform, supports report export (Excel, PDF), and connects to a network optimization platform of a telecom operator via API. This enables closed-loop management functions such as automatic issuance of problematic work orders, tracking of historical issues, and comparative analysis of results, thereby improving operational efficiency and reducing labor costs.

[0114] As can be seen from the above embodiments, the system provided by the present invention can realize an automated closed loop from data collection, modeling and recognition to optimization and recommendation, significantly improve the efficiency of 5G VoNR network quality optimization, reduce manual intervention, and has good generalization and practical application value.

[0115] The system outputs analysis results through a visualization platform, supporting: Interactive interface: risk map, issue list, trend analysis, etc.; Report export: Standardized templates for reporting or archiving; API Interface: Connects to a network optimization platform of a certain operator to realize work order flow.

[0116] It features closed-loop management functions including historical issue tracking, effectiveness evaluation, and recurrence warning. It supports hierarchical access control to ensure data security.

[0117] The system has continuous learning capabilities, continuously optimizing the model and knowledge base by combining optimization effect data, achieving self-learning and self-evolution, and improving long-term recognition accuracy and recommendation effect.

[0118] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A machine learning-based intelligent identification and optimization system for VoNR high packet loss cells, characterized in that, It includes a data acquisition and preprocessing module that interacts sequentially and is logically related, an indicator feature mining and modeling analysis module, a high packet loss cell intelligent identification module, an optimization strategy recommendation module, and a system integration and tool output module that interacts with the high packet loss cell intelligent identification module and the optimization strategy recommendation module respectively; The data acquisition and preprocessing module is used to aggregate multi-source heterogeneous data from VoNR cells, preprocess the multi-source heterogeneous data, and construct a high packet loss sample library and training set. The indicator feature mining and modeling analysis module is used to identify core indicators that are highly correlated with VoNR packet loss using feature engineering methods, train and optimize various machine learning models, and improve the adaptability of the models in different network scenarios through generalization testing. The high packet loss cell intelligent identification module is used to score the packet loss risk of VoNR cells across the entire network based on an optimized machine learning model, generate a sorted list of problematic cells, and output the screening criteria to support visual analysis and manual verification. The optimization strategy recommendation module is used to establish a strategy library based on optimization knowledge graphs or expert experience rules, match the packet loss cause type of the problem cell with the strategy library, output targeted optimization suggestions and support interpretable output; The system integration and tool output module is used to form a graphical, semi-automated tool system that supports historical problem tracking, strategy execution recording, effect comparison and evaluation, and connects to the operator's network optimization platform.

2. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The data acquisition and preprocessing module aggregates multi-source heterogeneous data, including: KQI / KPI data, MR measurement data, alarm data, parameter data, and user-perceived data. The KQI / KPI data includes VoNR voice uplink PDCP layer user plane packet loss rate, VoNR voice downlink PDCP layer user plane packet loss rate, wireless call success rate, call drop rate, and handover success rate; the MR measurement data includes RSRP, SINR, BLER, and MR coverage; the alarm data includes cell outage alarms, BBU optical module anomaly alarms, and cell VSWR alarms; the parameter data includes cell frequency band, frequency point configuration, PCI planning, handover threshold, and power control parameters; and the user perception data includes user complaint records, user rating feedback, and VoNR voice quality assessment data.

3. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The preprocessing operations of the data acquisition and preprocessing module include: data cleaning, removing invalid data, duplicate data, and missing values ​​to ensure data integrity and consistency; data normalization, standardizing data from different sources and in different formats, including converting KPI data to a unified unit and unifying timestamps to the ISO standard format; label construction, using whether the VoNR voice packet loss rate exceeds a preset threshold as the criterion to mark cells as having high packet loss or normal, forming supervised learning labels, where the preset threshold is 1.5%; outlier handling, using the Z-score method to identify and process outliers; and time alignment and spatial merging, performing time dimension alignment and spatial dimension merging on KQI / KPI data, MR measurement data, alarm data, and parameter data to ensure data spatiotemporal consistency.

4. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The feature engineering method of the indicator feature mining and modeling analysis module includes: feature extraction, extracting core features from preprocessed data. The core features include coverage quality features, network performance features, alarm features, parameter configuration features, and user perception features. The coverage quality features include RSRP, SINR, BLER, and MR coverage. The network performance features include at least wireless call success rate, call drop rate, handover success rate, and PRB utilization. The alarm features include alarm quantity, alarm type, and alarm duration. The parameter configuration features include frequency band, frequency point, PCI, and handover threshold. The user perception features include user ratings and number of complaints.

5. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The model training and optimization process of the module includes: The high packet loss sample database was divided into a training set, a validation set, and a test set in a ratio of 7:2:

1. A machine learning model was trained using the training set, and its hyperparameters were optimized through cross-validation. Model performance was evaluated on the validation set to select the optimal model. The model's generalization ability was validated on the test set. Model evaluation metrics included accuracy, precision, recall, F1 score, AUC-ROC curve, and feature importance, which was obtained through SHAP value or feature contribution analysis.

6. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The workflow of the high packet loss cell intelligent identification module includes: data input, receiving real-time collected and preprocessed KQI / KPI data, MR measurement data, alarm data, and parameter data; calling the optimized machine learning model to infer the input data and outputting a prediction result of high packet loss or normal; classifying high packet loss cells into three levels—high risk, medium risk, and low risk—based on the prediction result and packet loss rate, and sorting them according to risk level; visually displaying the geographical distribution of high packet loss cells through a risk map, generating a problem list including cell ID, region, network standard, equipment manufacturer, frequency band, and packet loss rate, and outputting a historical packet loss trend map of high packet loss cells.

7. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The knowledge graph construction of the optimization strategy recommendation module includes: constructing a knowledge graph using a graph database, with nodes including problem type, optimization strategy, equipment manufacturer, parameter configuration, and historical cases; data sources for the knowledge graph include network optimization expert experience, historical optimization cases, and equipment manufacturer recommendation strategies, with historical optimization cases including site replacement cases, parameter adjustment cases, and interference troubleshooting cases; the knowledge graph update mechanism involves connecting with the operator's network optimization platform via an API interface to periodically synchronize optimization strategy execution results and new cases, thereby updating the knowledge graph content.

8. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The strategy matching and output of the optimization strategy recommendation module include: The strategy matching algorithm matches and associates optimization strategies in a knowledge graph based on the problem types of high packet loss cells. The problem types include weak coverage, severe interference, handover failure, insufficient capacity, and abnormal parameter configuration. Strategy scoring evaluates the matched optimization strategies based on three dimensions: historical execution success rate, strategy complexity, and execution cost. The strategy output lists optimization strategies in descending order of scores. These optimization strategies include parameter adjustment, base station replenishment, antenna and feeder power adjustment, PCI replanning, neighbor cell optimization, bandwidth adjustment, multi-frequency splitting, and interference investigation. Interpretable output, displaying the rationale for strategy recommendations, historical execution results, and success rate data through a visual interface.

9. The intelligent identification and optimization system for high packet loss cells in VoNR based on machine learning according to claim 1, characterized in that, The system integration and tool output module includes: The API interface submodule connects with the operator's network optimization platform through the API interface to realize automatic issuance of problem work orders, synchronization of policy execution records, and feedback of optimization effect data. The edge computing submodule deploys the model inference submodule and the strategy recommendation submodule on the edge side to improve the data processing response speed; The permission hierarchical management submodule assigns different system access permissions and operation permissions to wireless network engineers and administrators. The report export submodule provides standardized templates and supports exporting analysis results and optimization suggestions.

10. A machine learning-based intelligent identification and optimization system for VoNR high packet loss cells according to claim 1, characterized in that, Adjust the model input features and policy matching logic according to the characteristics of different network environments to improve system adaptability.

Images