Base station coverage anomaly detection method based on big data analysis

Abstract

The present invention discloses a base station coverage anomaly detection method based on big data analysis, which belongs to the field of base station coverage detection technology in the field of communication technology. In order to solve the problem that the existing base station coverage anomaly detection is not accurate enough and cannot adapt to dynamic environmental changes, the method collects the operating data such as the received signal strength indicator value of all base stations in the target area for 1-30 consecutive days and at intervals of 10-60 minutes, calculates the coverage feature value obtained by linear weighting of multi-dimensional indicators after dimensionality reduction processing, inputs it into the isolation forest model trained with historical data and dynamically updated, and outputs the coverage anomaly probability value. When the coverage anomaly probability value is greater than 0.7 and less than or equal to 0.9, the corresponding base station position is marked on the electronic map as a coverage anomaly area. The method can accurately locate the base station coverage anomaly area, is suitable for real-time monitoring and anomaly identification of base station coverage status in communication network maintenance, and improves network optimization efficiency and service quality.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a base station coverage anomaly detection method based on big data analysis. Background Art

[0002] In the operation and maintenance of wireless communication networks, accurate detection of base station coverage status is a core component in ensuring network service quality. With the evolution of communication technology, base station types are becoming increasingly diverse, including 4G / 5G dual-mode base stations and 5G NR base stations supporting millimeter wave frequency bands. The operational data they generate exhibits multi-source heterogeneous characteristics. This data is not only multi-dimensional but also continuously collected at minute-level intervals, forming massive amounts of time series data covering continuous time periods. This makes it difficult for traditional detection methods based on single indicator thresholds or simple statistical analysis to establish effective correlations at the heterogeneous data fusion level. These methods are unable to capture coverage anomaly patterns characterized by the collaborative representation of multi-dimensional features in real time, resulting in insufficient accuracy and timeliness in anomaly identification.

[0003] At the big data processing level, real-time analysis of massive amounts of time series data faces the dual challenges of computing resource consumption and processing efficiency. On the one hand, the raw data contains a large amount of redundant information, such as repeated signal features collected continuously over long periods of time. Direct processing significantly increases computational complexity and leads to increased detection latency. On the other hand, traditional dimensionality reduction techniques struggle to balance feature preservation and computational efficiency when processing high-dimensional time series data. These issues prevent existing technologies from rapidly processing data and identifying anomalies in scenarios such as sudden traffic spikes and weather changes, making it difficult to meet the minute-level response requirements of network operations and maintenance.

[0004] In dynamic network environments, base station coverage status exhibits nonlinear fluctuations due to multiple factors. Geographical differences, such as dense urban areas versus open suburban terrain, seasonal traffic load fluctuations, such as peak traffic flow during the morning and evening rush hours, and extreme weather conditions, such as strong winds and heavy rain during typhoons, can all alter the distribution patterns of base station operating characteristics. Traditional detection models typically use fixed weighting coefficients, such as a weighting coefficient range of 0.1-0.5 for each dimension, and static thresholds, such as a coverage anomaly probability threshold of 0.7. These models lack the ability to perceive and adaptively adjust to environmental variables. For example, during major regional events, base station traffic loads may exceed 70% of their designed capacity. Continuing to use conventional weighting coefficients in such situations may underestimate the contribution of traffic load characteristics to anomaly detection. Furthermore, during typhoons, environmental characteristics such as wind speed and rainfall intensity are strongly correlated with signal penetration loss and link stability. Traditional models may not accurately identify coverage anomalies caused by these environmental changes. Therefore, existing technologies have significant limitations in real-time processing of massive amounts of heterogeneous data and adaptive model updates in dynamic environments. Summary of the Invention

[0005] The present invention provides a base station coverage anomaly detection method based on big data analysis, which can accurately locate the base station coverage anomaly area, is suitable for real-time monitoring and anomaly identification of base station coverage status in communication network maintenance, improves network optimization efficiency and service quality, and can solve the problems of traditional base station coverage anomaly detection methods that cannot effectively integrate multi-dimensional operation data, lack real-time performance and have low detection accuracy.

[0006] In order to achieve these objectives and other advantages of the present invention, the present invention provides a base station coverage anomaly detection method based on big data analysis, comprising the following steps:

[0007] S1. Collect operating data of all base stations in the target area over a continuous time period, wherein the operating data includes: received signal strength indicator value, reference signal reception quality value, handover success rate value, and traffic load value. The time period is 1-30 days and the collection interval is 10-60 minutes.

[0008] S2. Calculate a coverage characteristic value for each base station based on the operating data, where the coverage characteristic value is linearly weighted by the 10th percentile of the received signal strength indicator value, the coefficient of variation of the reference signal reception quality value, the 24-hour sliding average of the handover success rate value, and the peak-to-valley difference of the traffic load value;

[0009] S3. Input the coverage feature value into an isolation forest model trained with historical data, and output a coverage anomaly probability value for each base station;

[0010] S4. When the coverage anomaly probability value is within a range of a probability threshold, mark the corresponding base station location as a coverage anomaly area on the electronic map, wherein the range of the probability threshold is greater than 0.7 and less than or equal to 0.9.

[0011] Preferably, in step S2, before calculating the coverage eigenvalue, the operating data needs to be subjected to dimensionality reduction processing, and the dimensionality reduction processing specifically includes:

[0012] S210, compressing the received signal strength indicator value, the reference signal reception quality value, the handover success rate value, and the traffic load value into a 2-4 dimensional feature space using a principal component analysis method;

[0013] S220, using the segmented aggregation approximation algorithm to compress time series data to 10%-20% of its original length;

[0014] In step S2, the weighting coefficient of the linear weighting is determined by:

[0015] Establish a dynamic adjustment model for weighted coefficients, with input parameters including base station type, geographical environment, and time period;

[0016] Adopting an adaptive weighting coefficient algorithm, taking the historical coverage anomaly detection accuracy as the objective function, the weighting coefficient of each feature is periodically optimized;

[0017] The weighting coefficients of each feature meet the following ranges and the total is 1.0: the weighting coefficient of the 10th percentile of the received signal strength indication value is 0.3-0.5, the weighting coefficient of the coefficient of variation of the reference signal reception quality value is 0.2-0.3, the weighting coefficient of the 24-hour sliding average of the switching success rate value is 0.1-0.2, and the weighting coefficient of the peak-to-valley difference of the traffic load value is 0.1-0.2.

[0018] Preferably, in step S3, the isolation forest model is further updated, and the updating method of the isolation forest model includes:

[0019] S310, periodically receiving incremental operation data;

[0020] S320: When the cumulative incremental running data volume reaches 5%-15% of the original training data volume, a subtree update operation is triggered;

[0021] S330: The subtree update operation includes: selecting a subtree to be updated with a depth of 3-5 in the current isolation forest model, generating a replacement subtree based on the incremental running data, and atomically replacing the subtree to be updated with the replacement subtree. Specifically, the operation includes: calculating the coverage difference of the subtree with respect to the incremental running data; selecting the top K subtrees with the highest coverage difference, where K is 10%-30% of the total number of current subtrees. The coverage difference is calculated as follows: , is the indicator function, when the condition is met, the output is 1, otherwise the output is 0, T is the abnormality judgment threshold, D i is the coverage difference of the i-th subtree, n is the number of incremental running data samples, Incrementally run data samples for the i-th subtree The anomaly score prediction value of Run data samples for increments The actual abnormal label is 1 for abnormal and 0 for normal;

[0022] The atomic replacement of the replacement subtree includes: creating a replacement subtree version identifier; and pointing the to-be-updated subtree pointer to the replacement subtree through an atomic transaction.

[0023] Preferably, in step S4, after marking the coverage abnormal area, a positioning compensation step is performed:

[0024] Retrieve user measurement report data within a 300-500m radius of the abnormal base station;

[0025] When the percentage of sampling points where the received signal strength indicator value in the user measurement report is continuously lower than -110dBm exceeds 20% and does not exceed 40%, the corresponding geographic coordinates will be added to the abnormal area boundary;

[0026] After step S4, the method further includes: S5, performing spatial cluster analysis and spatial cluster verification on the marked base stations with abnormal coverage, specifically:

[0027] S510, extracting the geographic coordinates of the abnormal base station and constructing a spatial feature matrix, including latitude and longitude, base station type and coverage feature value;

[0028] S520, using the DBSCAN algorithm to perform spatial clustering on abnormal base stations, setting parameters as follows: neighborhood radius ε, ranging from 500 to 1500 meters; minimum sample number MinPts, taking 3 to 5 base stations;

[0029] S530: When there is a cluster containing ≥MinPts base stations in the clustering result, it is determined to be continuous area coverage attenuation. The specific verification rules are as follows: a. Calculate the average coverage anomaly probability value of all base stations in the cluster ,like 0.7, and the spatial distribution of base stations within the cluster is continuous and sheet-like, a regional abnormality alarm is triggered; the spatial range of the cluster is output and marked as a continuous coverage attenuation area on the electronic map;

[0030] S540: For isolated abnormal base stations that do not form a valid cluster, retain the original single-point abnormal mark.

[0031] Preferably, in step S4, a method for adjusting a dynamic threshold value for the coverage anomaly probability value is further included, specifically comprising:

[0032] Calculate the standard deviation of the coverage anomaly probability value of base stations in the entire area within 24 hours before the current time;

[0033] When the standard deviation exceeds 0.15 and does not exceed 0.25, the probability threshold is increased by 0.05-0.1;

[0034] When the standard deviation was lower than 0.05, the probability threshold was adjusted downward by 0.03-0.08.

[0035] Preferably, when collecting operating data, millimeter wave frequency band feature collection is added for 5G NR base stations to obtain beamforming failure rate values ​​and millimeter wave signal penetration loss compensation values;

[0036] When calculating the coverage characteristic value, a millimeter wave characteristic weighting coefficient factor is added, and the weighting coefficient factor is 0.2-0.4.

[0037] Preferably, the user measurement report data processing includes:

[0038] When the abnormal base station is a 4G / 5G dual-mode base station, perform multi-mode data verification:

[0039] Compare the difference between the 4G reference signal received power value and the 5G synchronization signal reference signal received power value;

[0040] If the difference exceeds 8dB but does not exceed 12dB and the duration exceeds 30% but does not exceed 50%, neighboring base station collaborative detection is triggered. The neighboring base station collaborative detection includes: a. sending a collaboration request instruction to 1-3 base stations in the same frequency band adjacent to the target base station; b. obtaining beam scanning measurement data of the neighboring base station. The beam scanning measurement data includes: user equipment reference signal received power difference and signal delay spread value. The user equipment reference signal received power difference is the difference in reference signal received power measured by the same user equipment at the target base station and the neighboring base station, and is calculated as follows: , RSRP target is the reference signal received power of the target base station, RSRP neighbor The reference signal received power measured by the neighboring base station; c. The coverage anomaly is confirmed to be valid when the following conditions are met: the difference in user equipment reference signal received power between the target base station and the neighboring base station is greater than 6dB and does not exceed 10dB; and the proportion of sampling points with signal delay spread values ​​greater than 100 nanoseconds and less than 150 nanoseconds is greater than 25% and does not exceed 40%.

[0041] Preferably, the isolation forest model training includes:

[0042] The historical data includes a typhoon weather scene dataset, and meteorological feature injection is performed during training:

[0043] The wind speed value of 12-20m / s and the rainfall intensity value of 30-50mm / h are used as environmental characteristic dimensions;

[0044] Add a meteorological feature judgment branch at the model splitting node.

[0045] Preferably, the dynamic threshold adjustment method further includes:

[0046] To enable the load compensation algorithm during major regional events:

[0047] When the base station traffic load value exceeds 70% and does not exceed 90% of the design capacity, a temporary coverage attenuation coefficient of 0.6-0.8 is generated;

[0048] The coverage anomaly probability value is multiplied by the attenuation coefficient and then compared with the probability threshold.

[0049] Preferably, marking the coverage abnormality area includes:

[0050] Perform spatial topology validation on the labeling results:

[0051] Detect the wireless backhaul link quality values ​​between abnormal base stations and adjacent base stations;

[0052] If the bit error rate of the return link exceeds 10 -4 and no more than 10 -3 If the transmission delay increases by 30-50ms, the abnormal flag of the base station will be frozen.

[0053] The present invention has at least the following beneficial effects:

[0054] First, through multi-dimensional operational data fusion and isolation forest model detection, we break through the limitations of traditional single indicators and achieve multi-feature collaborative analysis of base station coverage anomalies. Continuous time period data collection combined with probability threshold judgment can capture long-term trends and short-term fluctuations, improving the comprehensiveness and accuracy of anomaly identification. This provides a quantitative basis for base station fault location, avoids the blindness of manual inspections, and reduces operation and maintenance costs.

[0055] Second, the combination of principal component analysis and segmented aggregation approximation algorithms effectively reduces the computational complexity of high-dimensional time series data, compressing data size while retaining key features, improving processing efficiency and meeting real-time detection requirements. A dynamic weighting coefficient model adaptively adjusts weighting coefficients based on base station type, environment, and time period, ensuring that feature importance in different scenarios matches actual anomaly contribution. This avoids detection bias caused by fixed weighting coefficients and improves model generalization.

[0056] Third, the subtree update mechanism based on incremental running data can optimize local model parameters without full retraining, shorten the update cycle, and ensure that the model responds to changes in network status in a timely manner. By screening the subtrees to be updated through coverage differences and replacing them atomically, computing resource consumption is reduced, while avoiding the interference of old data that may be introduced by full updates, maintaining the stability of model detection accuracy, and is particularly suitable for scenarios with dynamic traffic or environmental changes. The positioning compensation step combines user measured data to correct the anomaly boundary and enhance the accuracy of single-point detection; spatial cluster analysis uses the DBSCAN algorithm to identify continuous regional attenuation, distinguish single-point failures from regional problems, and provide a layered disposal strategy for operation and maintenance. It accurately repairs isolated anomalies and initiates joint optimization for regional attenuation, improving resource allocation efficiency and avoiding increased global maintenance costs due to local misjudgments.

[0057] Fourth, a dynamic threshold adjustment mechanism adaptively adjusts thresholds based on fluctuations in the probability distribution of anomalies across the entire region. When the network is stable, detection standards are tightened to reduce false positives, while when fluctuations intensify, thresholds are relaxed to avoid missed positives. The standard deviation metric quantifies the degree of global anomaly dispersion, aligning thresholds with real-time network characteristics, improving the flexibility and adaptability of detection strategies, and reducing the risk of misjudgment during different time periods. The collection and weighting of millimeter-wave frequency band features for 5G NR base stations fills a dimensional gap in high-frequency communication detection. The introduction of beamforming failure rates and millimeter-wave signal penetration loss compensation values ​​accurately identifies unique issues in millimeter-wave signal propagation, such as obstruction and penetration loss. Combined with dedicated weighting factors, this increases the proportion of high-frequency anomalies in the feature values, ensuring comprehensive coverage quality assessments for new 5G base stations and promoting unified operation and maintenance of heterogeneous networks.

[0058] Fifth, a multi-standard data verification and neighboring cell collaborative detection mechanism analyzes cross-standard signal differences and coordinates measurements with neighboring base stations to eliminate interference factors outside the base station and accurately locate anomaly sources. This prevents false alarms caused by single-standard data misjudgment or neighboring cell interference, improving the reliability of anomaly confirmation. A meteorological feature injection mechanism incorporates environmental variables such as wind speed and rainfall intensity into model training, enabling the Isolation Forest model to identify weather-sensitive anomalies. During extreme weather conditions such as typhoons, the model uses meteorological feature branches split from nodes to provide early warning of coverage risks caused by environmental factors, assisting operations and maintenance personnel in developing preventive maintenance strategies and mitigating the impact of natural disasters on network services. A load compensation algorithm applies probability corrections to high-traffic load scenarios during major events, preventing base station performance fluctuations caused by peak traffic from being misidentified as coverage anomalies. Spatial topology verification monitors backhaul link quality to eliminate false anomaly flags caused by transport layer failures, such as increased bit error rate and latency, and prevents coverage issues from being misidentified as base station coverage failures. Freezing the markers of base stations with abnormal links can guide operations and maintenance personnel to prioritize troubleshooting transport layer issues, improve the accuracy of fault classification, and ensure the efficiency of network maintenance.

[0059] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The figure is a flow chart of the base station coverage anomaly detection method based on big data analysis of the present invention. DETAILED DESCRIPTION

[0061] The present invention is described in further detail below so that those skilled in the art can implement the invention with reference to the description.

[0062] It should be understood that terms such as “having”, “including” and “comprising” used herein do not preclude the existence or addition of one or more other elements or combinations thereof.

[0063] like Figure 1 As shown, the present invention provides a base station coverage anomaly detection method based on big data analysis, comprising the following steps:

[0064] S1. Collect operating data of all base stations in the target area within a continuous time period, wherein the operating data includes: received signal strength indicator value, reference signal reception quality value, handover success rate value and traffic load value. The time period is 1-30 days and the collection time interval is 10-60 minutes.

[0065] S2. Calculate the coverage characteristic value of each base station based on the operating data, where the coverage characteristic value is linearly weighted by the 10th percentile of the received signal strength indicator value, the coefficient of variation of the reference signal reception quality value, the 24-hour sliding average of the switching success rate value, and the peak-to-valley difference of the traffic load value.

[0066] S3. Input the coverage feature value into an isolation forest model trained with historical data, and output a coverage anomaly probability value for each base station.

[0067] S4. When the coverage anomaly probability value is within a range of a probability threshold, mark the corresponding base station location as a coverage anomaly area on the electronic map, wherein the range of the probability threshold is greater than 0.7 and less than or equal to 0.9.

[0068] Among them, marking the coverage abnormal area includes:

[0069] Perform spatial topology validation on the labeling results:

[0070] Detect the wireless backhaul link quality values ​​between abnormal base stations and adjacent base stations;

[0071] If the bit error rate of the return link exceeds 10 -4 and no more than 10 -3 If the transmission delay increases by 30-50ms, the abnormal flag of the base station will be frozen.

[0072] In the above embodiment, in the data collection phase, a time period of 1 day, 7 days, or 30 days can be selected to collect the operating data of the base stations in the target area. The collection interval can be set to 10 minutes, 30 minutes, or 60 minutes. The collected operating data includes the received signal strength indicator value, the reference signal reception quality value, the handover success rate value, and the traffic load value. When calculating the coverage characteristic value, the received signal strength indicator value is selected from the 10th percentile, the reference signal reception quality value uses the coefficient of variation, the handover success rate value uses the 24-hour sliding average, and the traffic load value uses the peak-to-valley difference. Each indicator is linearly weighted and fused. The weighting coefficient can be dynamically adjusted according to the base station type, geographical environment, and time period. For example, the traffic load value weighting coefficient can be appropriately increased to 0.2 for urban base stations, and the weighting coefficient of the received signal strength indicator value can be increased to 0.5 for suburban base stations.

[0073] Data collection can be achieved through base station monitoring terminals deployed in existing networks. These terminals collect real-time operational data from various base station types and upload it to the core network data center via the transmission network. To calculate coverage eigenvalues, data analysis software in the server cluster, such as the Python Pandas library, can be used to perform statistical calculations on time series data and implement linear weighting through matrix operations. The raw data for each indicator is stored in a distributed database, such as HBase, to facilitate subsequent model training and retrospective analysis.

[0074] When training the isolation forest model, a historical dataset must first be prepared. This data consists of operational data collected from base stations in the target area over the past 1-6 months, including received signal strength indicator (RSSI), reference signal reception quality (RSRQ), handover success rate, and traffic load. The data collection interval is set to 10-60 minutes, consistent with real-time monitoring, to ensure that the model training data has the same temporal characteristics as subsequent actual application data. The historical data is cleaned and preprocessed to remove significant outliers and missing data. Features of different dimensions are also normalized to ensure that each feature falls within a similar numerical range. After data preprocessing, coverage characteristics are calculated for each base station using the method described in S2. These include the 10th percentile of the RSSI, the coefficient of variation of the RSRQ, the 24-hour moving average of the handover success rate, and the peak-to-valley difference of the traffic load. These characteristic values ​​are linearly weighted and combined to form the model input feature vector. During training, the number of trees in the isolation forest is set to 100-200, and each tree is constructed by randomly selecting features and randomly partitioning values. The depth of the tree is limited to an average of 8-15 layers to prevent overfitting. Unsupervised learning is used during training, and no manually labeled anomaly labels are required. The model identifies abnormal patterns by calculating the average path length of samples in the forest. After the model training is completed, its performance is evaluated using a validation set. The validation set contains known normal base station operation data and manually confirmed anomaly cases. By adjusting the anomaly score threshold of the isolation forest, the detection accuracy of the model on the validation set is optimized. The trained model parameters and structure are saved to the model library, including the feature selection rules and node partitioning values ​​for each tree. The model is regularly retrained using newly added historical data, and the update cycle is set to 1-3 months to maintain adaptability to network changes. During the model deployment phase, the trained model is integrated into the real-time detection system through the API interface, receiving the coverage feature value input from the S2 step and outputting the coverage anomaly probability value for each base station.

[0075] Model deployment can be achieved using existing cloud computing platforms such as the Alibaba Cloud Machine Learning API. The trained isolation forest model is encapsulated as an API interface, which receives preprocessed coverage feature values ​​as input and outputs coverage anomaly probabilities in real time. New base station operating data can be pushed to the model input in real time via message queues such as Kafka, enabling online calculation of anomaly probabilities. When a typhoon warning is detected, the model can be triggered to prioritize the training subset containing meteorological features, improving detection sensitivity in inclement weather.

[0076] When the coverage anomaly probability value is greater than 0.7 and less than or equal to 0.9, the system automatically marks the corresponding base station location as a coverage anomaly area on an electronic map such as Baidu Map API. After marking, spatial topology verification is performed, and the wireless backhaul link quality data of the abnormal base station and the adjacent base stations is obtained through the network management system NMS to detect the backhaul link bit error rate, with the threshold set to 10 -4 -10 -3 If link quality anomalies are detected, the abnormal flag for the base station is frozen and a transport layer troubleshooting work order is generated.

[0077] Electronic map marking can be implemented through front-end visualization components such as ECharts, supporting color-coded icons to distinguish anomaly levels. Wireless backhaul link quality monitoring utilizes the performance statistics interface of existing transmission equipment, such as Huawei's OptiX RTN series microwave equipment, to obtain real-time bit error rate and latency data. Verification is automatically triggered by the business process management (BPM) system to ensure the accuracy of anomaly marking and avoid misjudgments due to transmission layer failures.

[0078] This embodiment comprehensively reflects base station coverage status through multi-dimensional data collection and dynamic feature calculation. Its spatial topology verification mechanism effectively distinguishes coverage anomalies from transmission layer failures, reducing ineffective investment in operations and maintenance resources. This method can provide precise positioning for base station coverage optimization, assisting operations and maintenance personnel in rapidly responding to network anomalies, improving communication network stability and service quality, and resolving the difficulty of identifying coverage anomalies in real time in environments with massive amounts of heterogeneous data.

[0079] In one specific embodiment, in step S2, before calculating the coverage feature value, the operating data needs to be subjected to dimensionality reduction processing, and the dimensionality reduction processing specifically includes:

[0080] S210, compressing the received signal strength indicator value, the reference signal reception quality value, the handover success rate value, and the traffic load value into a 2-4 dimensional feature space using a principal component analysis method;

[0081] S220, using the segmented aggregation approximation algorithm to compress time series data to 10%-20% of its original length;

[0082] In step S2, the weighting coefficient of the linear weighting is determined by:

[0083] Establish a dynamic adjustment model for weighted coefficients, with input parameters including base station type, geographical environment, and time period;

[0084] Adopting an adaptive weighting coefficient algorithm, taking the historical coverage anomaly detection accuracy as the objective function, the weighting coefficient of each feature is periodically optimized;

[0085] The weighting coefficients of each feature meet the following ranges and the total is 1.0: the weighting coefficient of the 10th percentile of the received signal strength indication value is 0.3-0.5, the weighting coefficient of the coefficient of variation of the reference signal reception quality value is 0.2-0.3, the weighting coefficient of the 24-hour sliding average of the switching success rate value is 0.1-0.2, and the weighting coefficient of the peak-to-valley difference of the traffic load value is 0.1-0.2.

[0086] In the above implementation, a principal component analysis algorithm can be used to reduce the dimension of the received signal strength indicator value, reference signal reception quality value, handover success rate value, and traffic load value. The output dimension of the principal component analysis can be set to 2, 3, or 4 dimensions, depending on the deployment density of base stations. 3-4 dimensions are recommended for high-density areas, and 2-3 dimensions are recommended for low-density areas. The reduced feature vector retains at least 90% of the variance contribution of the original data.

[0087] You can use a segmented aggregation approximation algorithm to compress time series data. The compression ratio can be set to 10%, 15%, or 20% of the original length. For a 60-minute time window, 10% compression corresponds to a 6-minute aggregation, 15% to a 9-minute aggregation, and 20% to a 12-minute aggregation. The aggregation method uses the arithmetic mean to reduce high-frequency noise.

[0088] The calculations of principal component analysis and segmented aggregation approximation can be deployed on a distributed computing platform. The data input interface supports Kafka or MQTT protocols, and the dimensionality reduction results are output to the in-memory database for subsequent feature calculation calls, which can solve the problem of insufficient real-time performance caused by big data processing.

[0089] The input parameters for the dynamic weighting coefficient adjustment model include base station type, geographic environment, and time of day. Base station type is categorized as macro or micro, geographic environment is categorized as urban, suburban, and rural, and time of day is categorized as peak hours (8:00 AM - 8:00 PM) and off-peak hours (8:00 PM - 8:00 AM). The model updates the weighting coefficient every 24 hours, loading the last 30 days of historical detection accuracy data as the objective function.

[0090] A gradient descent algorithm can be used to optimize weighting coefficients. The weighting coefficient for the 10th percentile of the received signal strength indicator (RSSI) is in the range of 0.3, 0.4, or 0.5. The weighting coefficient for the coefficient of variation of the reference signal reception quality (RSRQ) is in the range of 0.2, 0.25, or 0.3. The weighting coefficient for the moving average of the handover success rate is in the range of 0.1, 0.15, or 0.2. The weighting coefficient for the peak-to-valley difference in traffic volume is in the range of 0.1, 0.15, or 0.2. During the optimization process, the sum of the weighting coefficients is constrained to 1 to avoid overfitting.

[0091] The weighting coefficient adjustment module can be integrated into the policy engine of the base station network management system. It reads base station operating status data in real time, dynamically adjusts it, and writes it to the configuration database. The adjustment results are pushed to the anomaly detection module via an API to ensure that the weighting coefficient matches the current network status.

[0092] This implementation utilizes principal component analysis and segmented aggregation for approximate dimensionality reduction, reducing data processing workload and improving computational efficiency while preserving key feature information. A dynamic weighting coefficient adjustment mechanism adapts to the needs of coverage anomaly detection in different scenarios, improving detection accuracy and adaptability. The overall solution can be integrated into existing network management systems, eliminating the need for additional hardware investment and reducing operational complexity.

[0093] In one specific embodiment, step S3 further includes updating the isolation forest model. The updating method of the isolation forest model includes:

[0094] S310, periodically receiving incremental operation data;

[0095] S320: When the cumulative incremental running data volume reaches 5%-15% of the original training data volume, a subtree update operation is triggered;

[0096] S330: The subtree update operation includes: selecting a subtree to be updated with a depth of 3-5 in the current isolation forest model, generating a replacement subtree based on the incremental running data, and atomically replacing the subtree to be updated with the replacement subtree. Specifically, the operation includes: calculating the coverage difference of the subtree with respect to the incremental running data; selecting the top K subtrees with the highest coverage difference, where K is 10%-30% of the total number of current subtrees. The coverage difference is calculated as follows: , is the indicator function, which outputs 1 when the condition is met, otherwise it outputs 0, and T is the abnormality judgment threshold, which is usually 0.5; i is the coverage difference of the i-th subtree, n is the number of incremental running data samples, is the predicted value of the abnormal score of the incremental running data sample of the i-th subtree, and is the actual abnormal label of the incremental running data sample, where abnormal is 1 and normal is 0;

[0097] The atomic replacement of the replacement subtree includes: creating a replacement subtree version identifier; and pointing the to-be-updated subtree pointer to the replacement subtree through an atomic transaction.

[0098] In the above implementation, the incremental operation data reception period can be configured to be 1 hour, 6 hours, or 24 hours, depending on the frequency of base station data generation. The data interface can use a Kafka message queue or a distributed file storage system to support real-time streaming data access.

[0099] When the cumulative incremental running data volume reaches 5%, 10%, or 15% of the original training data volume, a subtree update is triggered. The original training data volume can be set to 10,000 to 100,000 samples, and the update is triggered when the incremental running data reaches 500 to 1,500. The threshold judgment module can be deployed on edge computing nodes to count the incremental running data volume in real time and trigger the update signal.

[0100] Incremental operational data can be stored in a time series database. The data format includes timestamps, base station IDs, operational indicators, and anomaly labels. During the data preprocessing phase, linear interpolation is used to fill missing values ​​to ensure the integrity of the incremental operational data input into the model.

[0101] Subtrees with a depth of 3, 4, or 5 in the current isolation forest model can be selected as candidates for update. The coverage difference threshold can be set to 0.25, 0.3, or 0.35, with subtrees exceeding the coverage difference threshold being added to the list for update. The coverage difference reflects the subtree's prediction deviation for incremental data; larger values ​​indicate a subtree's poorer adaptability to new data. Subtrees with the highest coverage difference are prioritized for update.

[0102] K subtrees with the highest degree of difference can be selected for replacement, based on 10%, 20%, or 30% of the total number of subtrees. The replacement subtrees are retrained based on incremental runtime data, preserving the depth constraints of the original tree. The atomic replacement process creates version identifiers, such as UUIDs, and uses database transactions to ensure the atomicity of pointer switches, avoiding concurrent access conflicts.

[0103] The subtree update module can be deployed on a GPU server, leveraging CUDA (Compute Unified Device Architecture) to accelerate the training process. After verification, the replaced subtree is synchronized to the online inference service, ensuring seamless model update switching.

[0104] It should be noted that the anomaly judgment threshold T is used to predict the results of the isolation forest subtree for a single sample. It is used to calculate the subtree coverage difference and evaluate model performance. It is mainly used in the model update stage. The probability threshold is the overall output probability of the isolation forest model for the base station, which determines whether the base station is marked as an anomaly. For example, it is greater than 0.7 and less than or equal to 0.9 in S4.

[0105] This implementation reduces the computational overhead of full model retraining and improves update efficiency through an incremental data triggering mechanism. Subtree difference screening ensures that model iterations optimize only those areas with performance degradation. Atomic replacement ensures service continuity, avoiding interruptions during the update process. The overall solution is adaptable to the real-time anomaly detection needs of large-scale base station clusters.

[0106] In one embodiment, the isolation forest model training includes:

[0107] The historical data includes a typhoon weather scene dataset, and meteorological feature injection is performed during training:

[0108] The wind speed value of 12-20m / s and the rainfall intensity value of 30-50mm / h are used as environmental characteristic dimensions;

[0109] Add a meteorological feature judgment branch at the model splitting node.

[0110] In the above embodiment, when training the isolated forest model, the historical data also includes a typhoon weather scene data set. The typhoon weather scene feature data includes wind speed values ​​and rainfall intensity values. The wind speed value needs to cover the typical wind speed range in typhoon weather, such as 12m / s, 15m / s, 20m / s and other key thresholds, corresponding to the wind intensity of tropical storms, severe tropical storms, and typhoon levels. This type of data is used to characterize the interference intensity of typhoons on wireless signal propagation, such as antenna deviation caused by fallen trees and equipment vibration. Rainfall intensity value: including rainfall levels such as 30mm / h, 40mm / h, and 50mm / h, reflecting the impact of heavy rainfall associated with typhoons on signal penetration loss, such as the attenuation effect of rain on the millimeter wave frequency band. Wind speed values ​​and rainfall intensity values ​​are used as environmental feature dimensions, and meteorological feature judgment branches are added at the model splitting nodes to enable the model to identify the impact of wind speed and rainfall intensity on base station coverage. During the training process, the cross-validation method can be used to optimize the model parameters, such as setting the number of subtrees to 100 and the maximum depth of a single tree to 20 layers, to improve the model's ability to fit high-dimensional features. This implementation method can solve the problem of inaccuracy of traditional models under extreme weather conditions. It should be noted that the geographical environment refers to the long-term stable physical properties of the base station deployment location, such as terrain type, building density or vegetation coverage, and does not include meteorological characteristics. Meteorological characteristics refer to weather state parameters collected in real time, such as wind speed value, rainfall intensity, etc.

[0111] In one embodiment, in step S4, after marking the coverage abnormal area, a positioning compensation step is performed:

[0112] Retrieve user measurement report data within a 300-500m radius of the abnormal base station;

[0113] When the percentage of sampling points where the received signal strength indicator value in the user measurement report is continuously lower than -110dBm exceeds 20% and does not exceed 40%, the corresponding geographic coordinates will be added to the abnormal area boundary;

[0114] After step S4, the method further includes: S5, performing spatial cluster analysis and spatial cluster verification on the marked base stations with abnormal coverage, specifically:

[0115] S510, extracting the geographic coordinates of the abnormal base station and constructing a spatial feature matrix, including latitude and longitude, base station type and coverage feature value;

[0116] S520, using the DBSCAN algorithm to perform spatial clustering on abnormal base stations, setting parameters as follows: neighborhood radius ε, ranging from 500 to 1500 meters; minimum sample number MinPts, taking 3 to 5 base stations;

[0117] S530: When there is a cluster containing ≥MinPts base stations in the clustering result, it is determined to be continuous area coverage attenuation. The specific verification rules are as follows: a. Calculate the average coverage anomaly probability value of all base stations in the cluster ,like 0.7, and the spatial distribution of base stations within the cluster is continuous and sheet-like, a regional abnormality alarm is triggered; the spatial range of the cluster is output and marked as a continuous coverage attenuation area on the electronic map;

[0118] S540: For isolated abnormal base stations that do not form a valid cluster, retain the original single-point abnormal mark.

[0119] In the above implementation, positioning compensation is used to retrieve user measurement report data within 300-500 meters of the abnormal base station. Areas where the received signal strength is continuously below -110dBm and where the sampling points account for 20%-40% are selected to be added to the outlier boundary. This significantly improves the accuracy and efficiency of network optimization. For example, if the received signal strength is continuously below -110dBm, this can be understood as N consecutive sampling points with a 10-minute interval, where N = 12 corresponds to 2 hours. This method dynamically corrects the coverage outlier boundary based on actual user-side signal measurement data, avoiding the bias of traditional base station theoretical models. It is particularly suitable for identifying transitional areas where the signal is attenuated but not completely interrupted, providing early warning of weak signal blind spots. By setting a 20%-40% coverage threshold, it effectively distinguishes temporary interference from persistent coverage issues, reducing false positives while ensuring that edge anomalies are not missed, providing operators with a highly reliable optimization target area. Combined with automated user measurement report analysis, it can quickly locate small coverage holes in high-density urban environments, significantly reducing operation and maintenance costs compared to manual drive testing. In addition, the identification of areas where signal strength is at a critical deterioration point (e.g., close to 40%) can support preventive maintenance, proactively triggering antenna parameter adjustments or small base station deployments, thereby eliminating potential faults before user complaints arise and improving overall network service quality.

[0120] Spatial feature matrix construction can extract basic information about abnormal base stations from the base station management system. Longitude and latitude coordinates can be accurate to six decimal places, and base station types can be categorized as macro, micro, and indoor distributed systems. Coverage feature values ​​include indicators such as received signal strength and reference signal quality, with values ​​ranging from 0 to 1. Existing spatial databases can be used to store this feature data, enabling fast query and matrix construction.

[0121] The feature matrix construction process involves data standardization to ensure comparability across different dimensions. An automatic update mechanism can be configured to update the feature matrix in real time when new abnormal base stations are added. Matrix data can be stored in an in-memory database, improving access speed during cluster analysis. After construction is complete, a data verification report is generated to ensure matrix integrity and accuracy.

[0122] DBSCAN cluster analysis can be implemented using an existing spatial analysis algorithm library. The neighborhood radius ε can be set to 500 meters, 800 meters, or 1500 meters, depending on base station density. The minimum sample size, MinPts, can be set to 3, 4, or 5 base stations to control clustering sensitivity. The algorithm takes into account the geographic distribution and signal coverage of base stations.

[0123] The clustering process runs on a distributed computing platform, enabling rapid processing of large-scale base station data. Exception handling mechanisms can be configured to automatically optimize parameters or switch algorithms when clustering time exceeds a threshold. Clustering results are generated into a visual report, showcasing the spatial distribution of base stations and clustering. The entire process typically completes within minutes, meeting real-time requirements.

[0124] During the clustering verification phase, the mean probability of coverage anomalies for base stations within each cluster is calculated, with a fixed threshold of 0.7. Spatial continuity analysis analyzes the density and shape of base station distribution to eliminate false clustering of discrete distributions. Existing spatial analysis tools can be used for continuity analysis, with the results output as Boolean values. Detailed logs are generated during the verification process to support subsequent auditing and analysis.

[0125] When a cluster meets both the probability mean and continuity conditions, the system generates a regional anomaly alert. Alert information includes key metrics such as the cluster's center coordinates, impact radius, and number of base stations. A multi-level alert mechanism can be configured to trigger different levels of notification based on cluster size. Alert information is pushed to the operations management platform in real time and highlighted on an electronic map.

[0126] The anomaly marking process uniformly marks all verified clusters, displaying them as polygonal areas on an electronic map. Marker colors can be set to orange or red, distinguishing them from single-point anomaly markers. Regional marking can be implemented using existing GIS system interfaces, supporting multi-layer overlay display. Marker information is synchronized to the operation and maintenance knowledge base for historical records.

[0127] For isolated base stations that do not form a valid cluster, their original single-point anomaly flags are retained. The system periodically reassesses these isolated points and triggers re-clustering when new anomaly base stations appear in the surrounding area. An automatic cleanup mechanism can be configured to remove the corresponding flags when the anomaly is resolved. An operation log is generated for all flag changes to ensure full traceability.

[0128] Spatial cluster analysis effectively distinguishes single-point anomalies from regional coverage issues, improving fault location accuracy. The DBSCAN-based clustering method adapts to varying base station densities and identifies spatially continuous anomaly areas. Systematic verification rules ensure the reliability of alarm information and reduce false alarms. The overall solution can be integrated into existing network management systems, improving network operation and maintenance efficiency.

[0129] In one specific embodiment, in step S4, a method for adjusting a dynamic threshold value for the coverage anomaly probability value is further included, specifically comprising:

[0130] Calculate the standard deviation of the coverage anomaly probability value of base stations in the entire area within 24 hours before the current time;

[0131] When the standard deviation exceeds 0.15 and does not exceed 0.25, the probability threshold is increased by 0.05-0.1;

[0132] When the standard deviation was lower than 0.05, the probability threshold was adjusted downward by 0.03-0.08.

[0133] The dynamic threshold adjustment method further includes:

[0134] To enable the load compensation algorithm during major regional events:

[0135] When the base station traffic load value exceeds 70% and does not exceed 90% of the design capacity, a temporary coverage attenuation coefficient of 0.6-0.8 is generated;

[0136] The coverage anomaly probability value is multiplied by the attenuation coefficient and then compared with the probability threshold.

[0137] In the above implementation, the dynamic threshold adjustment mechanism can be implemented via a statistical analysis module deployed on the network management server. The standard deviation calculation window is fixed at 24 hours, and the calculation period can be set to 1 or 4 hours. The standard deviation thresholds are set to 0.15 and 0.25 as the upper adjustment range, and 0.05 as the lower adjustment threshold. An existing time series database can be used to store probability value data and support rolling window calculations.

[0138] When the standard deviation is between 0.15 and 0.25, the probability threshold can be adjusted upward by 0.05, 0.08, or 0.1. When the standard deviation is less than 0.05, the probability threshold can be adjusted downward by 0.03, 0.05, or 0.08. The adjusted probability threshold takes effect immediately, and version information is recorded. A probability threshold change alarm can be configured to notify operations and maintenance personnel for confirmation when the adjustment exceeds 0.1. It should be noted that the initial probability threshold is determined based on a historical data validation set. Because the probability threshold is a range value, when it is adjusted upward, the entire threshold range is adjusted uniformly. For example, if the probability threshold is increased by 0.05, the new probability threshold is greater than 0.75 and less than or equal to 0.95.

[0139] The threshold adjustment module can be deployed on the policy management server and interacts with the anomaly detection engine in real time. The adjustment process takes historical threshold trends into account to avoid frequent fluctuations. The system retains the last 30 days of threshold records to support retrospective analysis. Each adjustment generates an operation log, documenting the adjustment time, magnitude, and triggering conditions.

[0140] During major events, load compensation can be triggered through the event management platform. Traffic load monitoring cycles can be set to 5 or 15 minutes, with design capacity data retrieved from the base station configuration database. Load thresholds are set at 70% and 90%, with compensation automatically enabled when the load falls within these ranges. The load monitoring module of an off-the-shelf network performance management system can be used.

[0141] The temporary coverage attenuation factor can be set to 0.6, 0.7, or 0.8, depending on the activity level. When the factor is applied, the coverage anomaly probability value is corrected in real time, and the corrected value is then compared with the current probability threshold. The coefficient validity period can be configured, and the original calculation method is automatically restored after the activity ends. The compensation process marks the data source for subsequent analysis.

[0142] The load compensation module can be deployed at core network edge computing nodes to ensure low-latency processing. The system records all operations during the compensation period, including coefficient values, the number of affected base stations, and changes in alarms. A configurable compensation effectiveness evaluation mechanism generates a dedicated report after the activity. The entire process is automated, requiring no human intervention. This can address the surge in false coverage alarms in high-density user scenarios.

[0143] Dynamic threshold adjustment adapts to normal fluctuations in network status and reduces false alarms caused by environmental factors. A standard deviation-based adjustment strategy maintains the stability of the alarm system and avoids overly sensitive or insensitive threshold settings. A load compensation mechanism effectively distinguishes between true coverage issues and temporary capacity overloads during major events, improving the accuracy of anomaly detection. The overall solution flexibly addresses the needs of diverse scenarios and enhances the intelligence of network operations and maintenance.

[0144] In one specific embodiment, when collecting operating data, millimeter wave frequency band feature collection is added for 5G NR base stations to obtain beamforming failure rate values ​​and millimeter wave signal penetration loss compensation values;

[0145] When calculating the coverage characteristic value, a millimeter wave characteristic weighting coefficient factor is added, and the weighting coefficient factor is 0.2-0.4.

[0146] In the above implementation, millimeter-wave frequency band feature collection can obtain beam management data via the northbound interface of the 5G base station. Beamforming failure rates can be calculated every minute or every five minutes, with the statistical period set to 15 minutes or 30 minutes. Millimeter-wave signal penetration loss compensation values ​​can be read from the measurement module of the base station's radio frequency unit, typically ranging from 10 to 30 dB. A network management protocol that supports 3GPP standards can be used to collect this data.

[0147] The data collection system can be deployed on edge computing nodes and communicate with base stations via a standard NM interface. The data collection frequency can be synchronized with regular metrics to ensure time alignment. During the collection process, outliers, such as sudden spikes in transient failure rates, are filtered out. The collected data is stored in a time series database, tagged with the base station ID, timestamp, and frequency band information. The system regularly verifies data integrity to ensure reliable collection.

[0148] The weighting factor for millimeter wave characteristics can be set in the characteristic calculation engine. The weighting factor can be set to 0.2, 0.3, or 0.4, with the specific value adjusted based on the network deployment environment. The weighting factor adjustment period can be set to 24 hours or triggered on demand, and the adjustment is based on historical optimization results. Weighting factor constraints can be configured to ensure that millimeter wave characteristics do not excessively influence the overall evaluation.

[0149] The weighting coefficient application process normalizes millimeter-wave eigenvalues ​​to align their dimensions with those of other features. Calculations are first multiplied by the weighting coefficient factor and then summed with the other weighted features. A matrix operation library can be used to accelerate the calculation process and support real-time eigenvalue updates. The system records the parameters and effects of each weighting coefficient adjustment to form an optimization knowledge base. The calculated anomaly probability value is synchronized with the decision engine.

[0150] Adding millimeter-wave signature collection can more comprehensively reflect the coverage characteristics of 5G base stations, particularly the propagation characteristics of high-frequency bands. Dynamic weighting coefficient settings can balance the impact of different frequency band characteristics, improving the accuracy of anomaly detection. The overall solution is compatible with the management system of existing 4G / 5G hybrid networks, enabling functional upgrades without hardware modifications. Through fine-grained millimeter-wave signature analysis, high-frequency band coverage issues can be more accurately identified. This can also address coverage misjudgments caused by millimeter-wave base stations' environmental sensitivity.

[0151] In one specific implementation, the user measurement report data processing includes:

[0152] When the abnormal base station is a 4G / 5G dual-mode base station, perform multi-mode data verification:

[0153] Compare the difference between the 4G reference signal received power value and the 5G synchronization signal reference signal received power value;

[0154] If the difference exceeds 8dB but does not exceed 12dB and the duration exceeds 30% but does not exceed 50%, neighboring base station collaborative detection is triggered. The neighboring base station collaborative detection includes: a. sending a collaboration request instruction to 1-3 base stations in the same frequency band adjacent to the target base station; b. obtaining beam scanning measurement data of the neighboring base station. The beam scanning measurement data includes: user equipment reference signal received power difference and signal delay spread value. The user equipment reference signal received power difference is the difference in reference signal received power measured by the same user equipment at the target base station and the neighboring base station, and is calculated as follows: , RSRP target is the reference signal received power of the target base station, RSRP neighbor The reference signal received power measured by the neighboring base station; c. The coverage anomaly is confirmed to be valid when the following conditions are met: the difference in user equipment reference signal received power between the target base station and the neighboring base station is greater than 6dB and does not exceed 10dB; and the proportion of sampling points with signal delay spread values ​​greater than 100 nanoseconds and less than 150 nanoseconds is greater than 25% and does not exceed 40%.

[0155] In the above implementation, multi-mode data verification can be implemented by the measurement report processing module of the dual-mode base station. The 4G reference signal received power value can be extracted from the RRC measurement report, and the 5G synchronization signal reference signal received power value can be obtained from the NR measurement report. The difference comparison thresholds are set to 8dB and 12dB, and the duration ratio thresholds are set to 30% and 50%. The data comparison function of the existing multi-mode base station management system can be used.

[0156] The difference calculation period can be set to 1 or 5 seconds, and the duration statistics window can be set to 5 or 10 minutes. The system records the results of each comparison, forming time series data for trend analysis. When conditions are met, a verification report is automatically generated and subsequent processes are triggered. The verification process eliminates short-term discrepancies caused by normal scenarios such as device switching.

[0157] Neighboring base station collaborative detection can send collaborative request instructions over the X2 or Xn interface. The request instruction can be configured to include the target base station ID, the anomaly type, and the requested data type. The number of neighboring base stations can be set to one, two, or three, with the selection strategy based on the base station topology table. An inter-base station communication protocol stack that supports the 3GPP standard can be selected.

[0158] After receiving the coordination request, the neighboring base station will start the specified measurement task. The user equipment reference signal received power difference is calculated using the formula , where RSRP values ​​are accurate to 0.1dB. Signal delay spread values ​​can be obtained from the channel estimation module with a measurement accuracy of 1 nanosecond. The measurement data is transmitted back to the requester via an encrypted channel.

[0159] For the anomaly validity confirmation criteria, the reference signal received power difference thresholds are set at 6dB and 10dB. The signal delay spread thresholds are set at 100 nanoseconds and 150 nanoseconds, and the anomaly sampling point ratio thresholds are set at 25% and 40%. A library of readily available wireless signal analysis algorithms can be used for conditional judgment. The confirmation process integrates measurement results from multiple neighboring cells and employs a voting mechanism to improve accuracy.

[0160] After validation, a final anomaly diagnosis report is generated, including the anomaly type, impact scope, and confidence level assessment. This report is synchronized to the network optimization platform and operations management system. For confirmed anomalies, the system recommends appropriate optimization measures, such as antenna adjustment or parameter optimization. The entire process typically completes within minutes, meeting real-time requirements.

[0161] Multi-mode data verification effectively identifies coverage discrepancies between dual-mode base stations, avoiding the limitations of single-mode detection. Neighboring cell collaborative detection improves the accuracy of anomaly detection by cross-validating multi-dimensional measurement data. The overall solution leverages existing inter-base station collaboration mechanisms, enabling refined coverage analysis without the need for additional hardware. Standardized interfaces and processes ensure large-scale deployment and application. It can also address positioning distortion caused by differences in measurement standards for multi-mode base stations.

[0162] The number of devices and processing scales described herein are intended to simplify the description of the present invention. Applications, modifications, and variations of the present invention will be readily apparent to those skilled in the art.

[0163] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A base station coverage anomaly detection method based on big data analysis, characterized in that: The following steps are involved: S1. Collect operating data of all base stations in the target area over a continuous time period, wherein the operating data includes: received signal strength indicator value, reference signal reception quality value, handover success rate value, and traffic load value. The time period is 1-30 days and the collection interval is 10-60 minutes. S2. Calculate a coverage characteristic value for each base station based on the operating data, where the coverage characteristic value is linearly weighted by the 10th percentile of the received signal strength indicator value, the coefficient of variation of the reference signal reception quality value, the 24-hour sliding average of the handover success rate value, and the peak-to-valley difference of the traffic load value; S3. Input the coverage feature value into an isolation forest model trained with historical data, and output a coverage anomaly probability value for each base station; S4. When the coverage anomaly probability value is within a range of a probability threshold, marking the corresponding base station location as a coverage anomaly area on the electronic map, wherein the range of the probability threshold is greater than 0.7 and less than or equal to 0.9; The method further includes updating the isolation forest model, wherein the updating method of the isolation forest model includes: S310, periodically receiving incremental operation data; S320: When the cumulative incremental running data volume reaches 5%-15% of the original training data volume, a subtree update operation is triggered; S330: The subtree update operation includes: selecting a subtree to be updated with a depth of 3-5 in the current isolation forest model, generating a replacement subtree based on the incremental running data, and atomically replacing the subtree to be updated with the replacement subtree. Specifically, the operation includes: calculating the coverage difference of the subtree with respect to the incremental running data; selecting the top K subtrees with the highest coverage difference, where K is 10%-30% of the total number of current subtrees. The coverage difference is calculated as follows: , is the indicator function, when the condition is met, the output is 1, otherwise the output is 0, T is the abnormality judgment threshold, D i is the coverage difference of the i-th subtree, n is the number of incremental running data samples, Incrementally run data samples for the i-th subtree The anomaly score prediction value of Run data samples for increments The actual abnormal label is 1 for abnormal and 0 for normal; The atomic replacement of the replacement subtree includes: creating a replacement subtree version identifier; and pointing the to-be-updated subtree pointer to the replacement subtree through an atomic transaction.

2. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: In step S2, before calculating the coverage eigenvalue, the operating data needs to be subjected to dimensionality reduction processing, and the dimensionality reduction processing specifically includes: S210, compressing the received signal strength indicator value, the reference signal reception quality value, the handover success rate value, and the traffic load value into a 2-4 dimensional feature space using a principal component analysis method; S220, using the segmented aggregation approximation algorithm to compress time series data to 10%-20% of its original length; In step S2, the weighting coefficient of the linear weighting is determined by: Establish a dynamic adjustment model for weighted coefficients, with input parameters including base station type, geographical environment, and time period; Adopting an adaptive weighting coefficient algorithm, taking the historical coverage anomaly detection accuracy as the objective function, the weighting coefficient of each feature is periodically optimized; The weighting coefficients of each feature meet the following ranges and the total is 1.0: the weighting coefficient of the 10th percentile of the received signal strength indicator value is 0.3-0.5, the weighting coefficient of the coefficient of variation of the reference signal reception quality value is 0.2-0.3, the weighting coefficient of the 24-hour sliding average of the switching success rate value is 0.1-0.2, and the weighting coefficient of the peak-to-valley difference of the traffic load value is 0.1-0.

2.

3. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: In step S4, after marking the coverage abnormal area, the positioning compensation step is performed: Retrieve user measurement report data within a 300-500m radius of the abnormal base station; When the percentage of sampling points where the received signal strength indicator value in the user measurement report is continuously lower than -110dBm exceeds 20% and does not exceed 40%, the corresponding geographic coordinates will be added to the abnormal area boundary; After step S4, the method further includes: S5, performing spatial cluster analysis and spatial cluster verification on the marked base stations with abnormal coverage, specifically: S510, extracting the geographic coordinates of the abnormal base station and constructing a spatial feature matrix, including latitude and longitude, base station type and coverage feature value; S520, using the DBSCAN algorithm to perform spatial clustering on abnormal base stations, setting parameters as follows: neighborhood radius ε, ranging from 500 to 1500 meters; minimum sample number MinPts, taking 3 to 5 base stations; S530: When there is a cluster containing ≥MinPts base stations in the clustering result, it is determined to be continuous area coverage attenuation. The specific verification rules are as follows: a. Calculate the average coverage anomaly probability value of all base stations in the cluster ,like 0.7, and the spatial distribution of base stations within the cluster is continuous and sheet-like, a regional abnormality alarm is triggered; the spatial range of the cluster is output and marked as a continuous coverage attenuation area on the electronic map; S540: For isolated abnormal base stations that do not form a valid cluster, retain the original single-point abnormal mark.

4. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: In step S4, a method for adjusting a dynamic threshold value for the coverage anomaly probability value is further included, specifically including: Calculate the standard deviation of the coverage anomaly probability value of base stations in the entire area within 24 hours before the current time; When the standard deviation exceeds 0.15 and does not exceed 0.25, the probability threshold is increased by 0.05-0.1; When the standard deviation was lower than 0.05, the probability threshold was adjusted downward by 0.03-0.

08.

5. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: When collecting operational data, we added millimeter wave frequency band feature collection for 5G NR base stations to obtain beamforming failure rate values ​​and millimeter wave signal penetration loss compensation values. When calculating the coverage characteristic value, a millimeter wave characteristic weighting coefficient factor is added, and the weighting coefficient factor is 0.2-0.

4.

6. The base station coverage anomaly detection method based on big data analysis according to claim 3, characterized in that: The user measurement report data processing includes: When the abnormal base station is a 4G / 5G dual-mode base station, perform multi-mode data verification: Compare the difference between the 4G reference signal received power value and the 5G synchronization signal reference signal received power value; If the difference exceeds 8dB but does not exceed 12dB and the duration exceeds 30% but does not exceed 50%, neighboring base station collaborative detection is triggered. The neighboring base station collaborative detection includes: a. sending a collaborative request instruction to 1-3 base stations in the same frequency band adjacent to the target base station; b. obtaining the beam scanning measurement data of the neighboring base station, including: the user equipment reference signal received power difference and the signal delay spread value. The user equipment reference signal received power difference is the difference in reference signal received power measured by the same user equipment at the target base station and the neighboring base station, and the calculation formula is: , RSRP target is the reference signal received power of the target base station, RSRP neighbor The reference signal received power measured by the neighboring base station; c. The coverage anomaly is confirmed to be valid when the following conditions are met: the difference in user equipment reference signal received power between the target base station and the neighboring base station is greater than 6dB and does not exceed 10dB; and the proportion of sampling points with signal delay spread values ​​greater than 100 nanoseconds and less than 150 nanoseconds is greater than 25% and does not exceed 40%.

7. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: The isolation forest model training includes: The historical data also includes a typhoon weather scene dataset, and meteorological feature injection is performed during training: The wind speed value of 12-20m / s and the rainfall intensity value of 30-50mm / h are used as environmental characteristic dimensions; Add a meteorological feature judgment branch at the model splitting node.

8. The base station coverage anomaly detection method based on big data analysis according to claim 4, characterized in that: The dynamic threshold adjustment method further includes: To enable the load compensation algorithm during major regional events: When the base station traffic load value exceeds 70% and does not exceed 90% of the design capacity, a temporary coverage attenuation coefficient of 0.6-0.8 is generated; The coverage anomaly probability value is multiplied by the attenuation coefficient and then compared with the probability threshold.

9. The base station coverage anomaly detection method based on big data analysis according to claim 1, characterized in that: Marking of coverage anomaly areas includes: Perform spatial topology validation on the labeling results: Detect the wireless backhaul link quality values ​​between abnormal base stations and adjacent base stations; If the bit error rate of the return link exceeds 10 -4 and no more than 10 -3 If the transmission delay increases by 30-50ms, the abnormal flag of the base station will be frozen.

Images