A communication base station operation state monitoring method based on big data

Abstract

This invention belongs to the field of communication base station operation monitoring technology and discloses a communication base station operation status monitoring method based on big data. In the data processing stage, a combination of dynamic threshold isolated forest and time-series noise adaptive filtering is used, combined with historical service data to statistically analyze the distribution patterns of abnormal values ​​for different services. In the feature extraction stage, time-varying mutual information entropy is calculated using a sliding window to construct a dynamic parameter correlation graph. Then, a GAT network is used to highlight the influence of core parameters, and Bi-LSTM is used to extract gradual and abrupt parameter changes. This method can accurately capture the dynamic coupling relationship of parameters under different service time slots, effectively identify latent fault precursors in advance, and avoid fault prediction lag. Through four-dimensional scene labels and scene embedding layers, discrete scene features are transformed into continuous vectors and their weights are dynamically adjusted. A base station scene meta-knowledge base is established to store fault and parameter correlation rules, and meta-learning is used to retrieve similar scenes and load initial parameters to quickly adapt to new scenes.

Description

Technical Field

[0001] This invention belongs to the field of communication base station operation monitoring technology, specifically a method for monitoring the operation status of communication base stations based on big data. Background Technology

[0002] Communication base stations are the core of digital infrastructure. Their operational stability directly affects network service quality and the actual effectiveness of industry digitalization. With the large-scale application of technologies such as network slicing and edge computing, the number of parameters has increased significantly, resulting in a substantial increase in the parameters that a single base station needs to monitor. State evolution is not linear, and the differences in scenarios are also significant. For example, the needs of densely populated urban areas, industrial IoT, and remote pastoral areas are all different. Existing technologies are difficult to adapt to these complex needs, mainly due to the following technical problems:

[0003] Existing methods rely solely on static thresholds or single-dimensional time-series features for analysis, failing to capture the time-varying coupling relationships between parameters. For example, the coupling strength of the Massive MIMO antenna load and transmission link latency in 5G base stations changes dynamically in different service time slots such as voice, video, and IoT. Traditional static correlation analysis leads to delayed fault prediction and fails to detect early signs of hidden faults.

[0004] Most existing models are trained offline and deployed in a fixed manner, which cannot adapt to the dynamic changes of heterogeneous scenarios. For example, in the industrial Internet of Things scenario, base stations need to be resistant to electromagnetic interference; base stations in remote pastoral areas need to be adapted to low power consumption operation. Traditional models do not have a scenario adaptation mechanism. Once they enter a new scenario or business iteration, such as the addition of VR / AR business, the monitoring accuracy will drop significantly, requiring manual retraining of the model, which results in high operation and maintenance costs. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring the operational status of communication base stations based on big data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the operational status of communication base stations based on big data, the specific steps of which are as follows:

[0007] Data acquisition phase: Collect parameters of various types of base stations, set up a network outage fault tolerance mechanism, and assign a scene label to each data point;

[0008] Data processing stage: Based on the data and scene tags collected in the data acquisition stage, the distribution pattern of outliers is statistically analyzed and noise is filtered out. Normalized weights are set according to business requirements, and preprocessed data is output.

[0009] Feature extraction stage: Based on the preprocessed data and scene labels, the time-varying mutual information entropy is calculated using a sliding window, a GAT network is established and weights are assigned according to parameter priority, and then the temporal features are extracted through Bi-LSTM, finally forming a two-dimensional feature set;

[0010] Model training phase: Combine the two-dimensional feature set and scene labels to build a training dataset, construct a hybrid model including scene embedding layer, GAT and Bi-LSTM dual branches and reinforcement learning, adjust feature weights according to scene, dynamically update reinforcement learning coefficients, and finally train through GPU cluster;

[0011] Evaluation phase: Incremental calculations are performed based on parameter changes, and an evaluation report is generated and pushed to the operation and maintenance platform;

[0012] In the tiered early warning stage: establish cross-level parameter correlation maps, trace the root causes of abnormal parameters in reverse, and implement tiered early warning based on dual indicators;

[0013] Model self-evolution stage: Based on evaluation data and source tracing results, the hybrid model is distilled in a layered manner at the edge and cloud; a meta-knowledge base for base station scenarios is established, and meta-learning is used to adapt to new scenarios; the distillation strategy is adjusted according to the early warning accuracy and operation and maintenance costs, and the optimized model is fed back to the evaluation stage.

[0014] Preferably, the data acquisition stage is as follows:

[0015] A four-level synchronous data collection architecture is established, encompassing base station parameters, network link parameters, service scenarios, and operational metrics. Base station parameters are collected by built-in sensors within the base station. Network link parameters are synchronously obtained through the core network gateway. Service characteristic parameters are obtained from the service management platform. Operational metric parameters are synchronously retrieved from the operation and maintenance management system. Environmental parameters are collected through external sensors deployed around the base station.

[0016] During data collection, precise alignment is achieved using timestamps, ensuring that each parameter data packet carries a four-dimensional scene label. Simultaneously, a network outage fault tolerance mechanism is set up, automatically retransmitting data after network recovery, ultimately outputting the collected data and scene labels.

[0017] Preferably, the data processing stage is as follows:

[0018] Based on the data collected and the associated scene tags, and considering the data noise characteristics and parameter correlation requirements under different business scenarios, a combination of dynamic threshold isolated forest and time-series noise adaptive filtering is adopted. The dynamic threshold isolated forest first uses historical business data to statistically analyze the distribution patterns of outliers in businesses including voice and IoT.

[0019] In the temporal noise filtering stage, the number of modes is adjusted for different scenarios through variational mode decomposition. At the same time, a time-varying normalization algorithm with business time slot weighting is designed in combination with business time slot characteristics, and the normalization benchmark value is updated regularly. Normalization weights are set for voice services and IoT services respectively, and high-dimensional parameters are adjusted to a unified dimension while retaining the dynamic correlation between parameters, and the preprocessed high-dimensional data is output.

[0020] Preferably, the feature extraction stage is as follows:

[0021] Based on the high-dimensional data processed in the data processing stage, and combined with the business type information in the scene labels in the data collection stage, a dynamic association graph of time-varying mutual information entropy weighted by business time slots and combined with the parameters of the graph attention network is established; the time-varying mutual information entropy calculation adopts a sliding window mechanism, and the association strength is updated in real time through a formula after each window ends.

[0022] After introducing the base station parameter priority matrix into the graph attention network, when calculating the attention weights for the first-level parameters, parameter importance coefficients are set to highlight the influence of core parameters.

[0023] In the temporal feature extraction stage, Bi-LSTM is used to extract features including the gradual trend of beamforming accuracy and the sudden changes in parameters caused by the sudden outbreak of IoT services, ultimately obtaining a two-dimensional feature set.

[0024] Preferably, the model training phase is as follows:

[0025] By utilizing the two-dimensional feature set from the feature extraction stage and combining it with the scene labels from the data acquisition stage, a two-dimensional training dataset of scene and state is established, which includes various scene types and fault cases.

[0026] A hybrid model is established, consisting of a four-dimensional scene embedding layer, a graph attention network and a Bi-LSTM dual-branch, and reinforcement learning. The scene embedding layer first converts the four-dimensional discrete labels of environment, business, network, and operation and maintenance into continuous feature vectors through the embedding layer, and then adjusts them to the same dimension as the features of the dual branches through a fully connected layer. The weights of the corresponding feature vectors are adjusted according to the scene. The graph attention network branch loads the dynamically associated features from the feature extraction stage for training, and the Bi-LSTM branch learns the temporal evolution features.

[0027] The reward function of the reinforcement learning module aims to balance the early warning effect with the operation and maintenance cost, and is optimized by dynamically adjusting the weights of two key influencing factors. The model training is based on a GPU cluster, and after training, it has the ability to output the evaluation of the health status of the operation and the probability of latent fault precursors.

[0028] Preferably, the evaluation stage is as follows:

[0029] Based on the preprocessed high-dimensional data stream in the data processing stage, the feature extraction logic in the feature extraction stage and the hybrid model trained in the model training stage are called. The Flink streaming computing framework is used to establish an incremental computing and dynamic caching mechanism. The judgment logic of incremental computing is as follows: compare the difference between the current parameter value and the previous evaluation value in real time. When the difference of core parameters exceeds the preset threshold or the difference of non-core parameters exceeds the preset threshold, re-feature extraction and model evaluation are triggered. Stable parameters are stored in the dynamic cache.

[0030] The real-time generated assessment report includes the operational health score and fluctuation curve, the probability of latent fault precursors and the time-varying correlation diagram of the top three risk parameters and key abnormal parameters, and pushes the report to the operation and maintenance platform and visualization dashboard.

[0031] Preferably, the tiered early warning stage is as follows:

[0032] Based on the real-time evaluation report during the evaluation phase, and combined with the parameter dynamic correlation graph during the feature extraction phase, a cross-layer parameter correlation graph is formed for the physical layer, network layer, and business layer. When the graph is established, the causal relationship of historical failure cases is analyzed, and the influence weight of inter-layer parameters is calculated using a Bayesian network, with the weight values ​​updated in real time.

[0033] During the anomaly tracing phase, based on the anomaly parameters marked in the report, the propagation path is traced backward along the graph to ultimately locate the root cause and generate a visual tracing graph.

[0034] The tiered early warning system uses operational health status and the probability of latent fault precursors as indicators, implements different response strategies for different risk levels, and outputs anomaly tracing results and early warning data.

[0035] Preferably, the self-evolutionary stage of the model is as follows:

[0036] Based on real-time evaluation data from the evaluation phase and anomaly tracing results from the graded early warning phase, the hybrid model in the model training phase is optimized through edge and cloud-based hierarchical online distillation. The edge node layer is deployed on the local server of the base station, and lightweight distillation is used for stable data from a single base station: only the output features of the Bi-LSTM layer are passed to the student model, and the mean squared error is used as the loss function. The cloud aggregation layer is deployed in the regional cloud center, and full distillation is performed for common anomaly data in multi-base station scenarios: the parameter knowledge and attention weights of the teacher model are passed to the student model through the KL divergence loss function, and the parameters of the graph attention network association branch and the scene embedding layer are updated synchronously.

[0037] Simultaneously, a meta-knowledge base for base station scenarios is established to store the rules for associating faults and parameters. New association rules are automatically mined and updated using the Apriori algorithm. When adapting to new scenarios through meta-learning, similar scenarios are first retrieved from the knowledge base, and the corresponding model parameters are loaded as initial values. Then, the new scenario samples are used for fine-tuning to shorten the adaptation time.

[0038] The newly added self-evolution effect evaluation module regularly calculates historical early warning accuracy, operation and maintenance costs, and fault repair time, and adjusts the distillation strategy based on operation and maintenance costs and early warning accuracy; forming a collection, analysis, and optimization process, and the optimized model is fed back to the real-time evaluation in the evaluation stage.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. In the data processing stage, this invention employs a combination of dynamic threshold isolated forest and time-series noise adaptive filtering, combining historical service data to statistically analyze the distribution patterns of outliers in different services. In the feature extraction stage, time-varying mutual information entropy is calculated using a sliding window to construct a dynamic parameter correlation graph. Then, the GAT network is used to highlight the impact of core parameters such as antenna power and transmission delay, and Bi-LSTM is used to extract the gradual and abrupt changes in parameters. This invention can accurately capture the dynamic coupling relationship of parameters under different service time slots, such as the correlation changes between the 5G base station Massive MIMO antenna load and transmission link delay, effectively identifying latent fault precursors in advance, avoiding fault prediction lag, and improving the accuracy of base station operation status monitoring.

[0041] 2. This invention transforms discrete scene features into continuous vectors and dynamically adjusts weights through four-dimensional scene labels and scene embedding layers; it establishes a base station scene meta-knowledge base to store fault and parameter association rules, and combines meta-learning to retrieve similar scenes and load initial parameters to quickly adapt to new scenes; at the same time, it optimizes the model through edge and cloud-based layered distillation, which can adapt to new services without manual retraining; it greatly improves scene adaptability, avoids manual intervention, reduces operation and maintenance costs, ensures stable monitoring accuracy under different scenarios, and adapts to the needs of network technology iteration and business expansion.

[0042] 3. This invention constructs a cross-layer parameter correlation graph based on Bayesian networks, connecting the physical, network, and service layers. It updates the influence weights of parameters between layers in real time, allowing for the reverse tracing of abnormal parameter propagation paths along the graph to pinpoint the root cause of faults and generate a visual graph. In the tiered early warning stage, operational health and the probability of latent fault precursors are used as dual indicators. Different risk levels correspond to differentiated response strategies; for example, if the health score is ≥90, only logs are recorded, while if the health score is <60, work orders are automatically dispatched and backup links are switched. This achieves accurate anomaly location and efficient handling, shortens fault repair time, optimizes maintenance resource scheduling, and significantly improves the targeting and efficiency of maintenance responses. Attached Figure Description

[0043] Figure 1 This is a flowchart of the communication base station operation status monitoring method based on big data according to the present invention;

[0044] Figure 2 This is a flowchart of the feature extraction and model training process of the present invention. Detailed Implementation

[0045] 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.

[0046] like Figures 1 to 2 As shown in the figure, this embodiment of the invention provides a method for monitoring the operational status of communication base stations based on big data. The specific steps of the method are as follows:

[0047] The data acquisition phase is specifically as follows:

[0048] A four-level synchronous data collection architecture is established, encompassing base station parameters, network links, service scenarios, and operational metrics. Base station parameters (such as Massive MIMO antenna port power) are collected in real-time by built-in sensors at a sampling frequency of 10Hz. Network link parameters (such as slice bandwidth allocation) are synchronously acquired through the core network gateway, with latency controlled within 50ms. Service characteristic parameters (such as service type and time slot occupancy rate) are obtained in real-time from the service management platform. Operational metric parameters (such as fault repair time, operation and maintenance work order processing progress, and base station energy consumption cost) are synchronously retrieved from the operation and maintenance management system. Environmental parameters (such as electromagnetic interference intensity and low-temperature battery life index in pastoral areas) are collected through external sensors deployed around the base station, with a protection level of IP65.

[0049] This architecture supports concurrent synchronous acquisition of high-dimensional data from 100,000 base stations. During acquisition, it uses timestamps for precise alignment, ensuring that each parameter data packet carries a four-dimensional scene label (environment-service-network-operation and maintenance), thus avoiding data and label misalignment. At the same time, it sets up a network outage fault tolerance mechanism, automatically retransmitting data after the network is restored to ensure integrity, and finally outputs the acquired data and scene labels.

[0050] The data processing stage is specifically as follows:

[0051] Based on the data collected and the associated scene tags, and considering the data noise characteristics and parameter correlation requirements under different business scenarios, a combined scheme of dynamic threshold isolated forest and time-series noise adaptive filtering is adopted. The dynamic threshold isolated forest first analyzes the distribution patterns of outliers in voice, IoT, and other services based on three months of historical business data. For voice services, due to high real-time requirements, the outlier threshold is set at ±3σ, 20% tighter than the conventional threshold. For IoT services, due to large data fluctuations, the threshold is set at ±5σ, 15% more lenient than the conventional threshold.

[0052] In the temporal noise filtering stage, variational mode decomposition (VMD) is used to adjust the number of modes for different scenarios. In industrial scenarios, the frequency of electromagnetic interference noise is concentrated in 200-500MHz, so the number of VMD modes is set to 6. In urban areas, the frequency of mutual interference noise between multiple base stations is 800-1200MHz, so the number of modes is set to 8 to accurately filter specific noise. At the same time, a time-varying normalization algorithm with service time slot weighting is designed in combination with service time slot characteristics. The normalization benchmark value is updated with a 24-hour cycle to avoid long-term data drift. Normalization weights are set for voice services and IoT services respectively to adjust high-dimensional parameters to a unified dimension and retain the dynamic correlation between parameters, and output preprocessed high-dimensional data.

[0053] Formula for judging outliers in dynamic threshold:

[0054] ;

[0055] In the formula: This indicates the threshold for judging abnormal values ​​of parameters. If the parameter value exceeds this range, it is judged as an abnormal value and enters the noise filtering process.

[0056] This represents the average value of parameters calculated based on three months of historical business data, reflecting the baseline level of long-term stable operation of the parameters;

[0057] The standard deviation of a parameter reflects the degree of data dispersion and represents the normal fluctuation range of the parameter.

[0058] This represents the threshold coefficient, which adjusts according to the service type; for voice services... =3, high real-time requirements, threshold tightened by 20% compared to the conventional threshold; IoT business =5, with large data fluctuations, the threshold is relaxed by 15% compared to the usual threshold; adaptable to the data characteristics of different businesses.

[0059] The feature extraction stage is specifically as follows:

[0060] Based on the high-dimensional data processed in the data processing stage, and combined with the business type information in the scene labels from the data acquisition stage, a dynamic parameter association graph of business time slot weighted time-varying mutual information entropy combined with graph attention network (GAT) is established. The time-varying mutual information entropy calculation adopts a sliding window mechanism, with 10 sampling points in a 50ms window for voice services and 100 sampling points in a 500ms window for IoT. After each window ends, the association strength is updated in real time through a formula to ensure that the parameter association changes under different services are captured.

[0061] Service characteristic parameters (such as time slot occupancy rate) are statistical parameters with a sampling frequency of 1Hz; voice service sampling points are real-time operating parameters of the base station (such as instantaneous antenna power value) with a sampling frequency of 200Hz. The two types of parameters belong to different monitoring dimensions, and the sampling frequency is set according to the parameter function requirements.

[0062] The GAT network is designed with a 3-layer structure, containing 3 attention heads, each with a hidden layer dimension of 64. After introducing the base station parameter priority matrix, when calculating the attention weight for the first-level parameters (antenna power, transmission delay), an additional parameter importance coefficient (0.8 for antenna power, 0.7 for transmission delay) is added. The final assigned weight is 3 times that of the third-level parameters (such as ambient temperature), highlighting the influence of core parameters.

[0063] In the temporal feature extraction stage, the bidirectional LSTM (Bi-LSTM) adopts a 2-layer structure with a hidden layer dimension of 128. The optimizer is Adam, and the learning rate is set to 0.001. Overfitting is avoided by using an early stopping mechanism (if the validation set loss does not decrease after 3 rounds). Features such as the gradual trend of beamforming accuracy within 10 minutes and the sudden changes in parameters caused by the sudden outbreak of IoT business are extracted, and finally a two-dimensional feature set is obtained.

[0064] Time-varying mutual information entropy formula:

[0065] ;

[0066] In the formula: Indicates business time slot Below, parameters With parameters The time-varying mutual information entropy, i.e. the real-time correlation strength, directly reflects the dynamic changes in the coupling relationship between parameters;

[0067] Indicates parameters With parameters The initial mutual information entropy (basic association strength) is calculated based on historical stable data;

[0068] This indicates the weight of service time slots, which is dynamically adjusted according to the service type; for voice services... =1.2, high real-time requirements, higher weight; IoT business =0.8, the data fluctuates greatly, and the weight is slightly low;

[0069] , This indicates that the parameters of the two base stations whose relationship needs to be analyzed are often core parameters, such as Massive MIMO antenna load and transmission link delay.

[0070] It indicates the type of service time slot, such as voice service time slot or IoT service time slot, which accurately corresponds to the actual service period of the base station.

[0071] The model training phase is specifically as follows:

[0072] By utilizing the two-dimensional feature set from the feature extraction stage and combining it with the scene labels from the data collection stage, a two-dimensional training dataset of scene and state is established, including typical scenes such as dense urban areas, suburbs, industrial IoT, remote pastoral areas, and transportation hubs. For each scene, 300 normal samples and 50 fault samples are collected; a total of more than 3,000 fault cases are collected, covering common faults such as antenna faults, transmission link faults, and service overload faults, to ensure data diversity.

[0073] A hybrid model is established, consisting of a four-dimensional scene embedding layer, a dual-branch GAT and Bi-LSTM architecture, and reinforcement learning. The scene embedding layer first transforms four discrete labels—environment (electromagnetic interference level), business (type, time slot), network (slice configuration), and operation and maintenance (cost, repair time)—into 64-dimensional continuous feature vectors through an embedding layer. These vectors are then adjusted to the same dimension (256-dimensional) as the dual-branch features through a fully connected layer. In industrial scenarios with electromagnetic interference >80dB and pastoral scenarios with a battery life index <60, the weights of the corresponding feature vectors are automatically increased by 1.5 times. The GAT branch loads the dynamically associated features from the feature extraction stage for training, while the Bi-LSTM branch learns temporal evolution features.

[0074] The reward function of the reinforcement learning module aims to balance the early warning effect with the operation and maintenance cost, and is optimized by dynamically adjusting the weights of two key influencing factors. The model training is based on a GPU cluster, and after training, it has the ability to output the evaluation of the health status of the operation and the probability of latent fault precursors.

[0075] Reinforcement learning reward function:

[0076] ;

[0077] In the formula: This represents the real-time reward value of the reinforcement learning module, which guides the direction of model parameter updates. A higher value indicates a better current model policy.

[0078] This indicates the weighting coefficient for the accuracy of early warnings, which is dynamically adjusted according to the scenario; periods with high failure rates, such as peak business periods during holidays. =0.7, prioritizing the accuracy of early warnings; for example, during routine maintenance. =0.5, updated every second based on real-time data;

[0079] This represents the real-time early warning accuracy of the model, with a value ranging from 0 to 100%, calculated by comparing the prediction results with the actual fault conditions.

[0080] This represents the weighting coefficient for operation and maintenance costs, which is dynamically adjusted according to the scenario; cost-sensitive periods, such as routine operation and maintenance. =0.6, controlling the cost of manual intervention; high failure rate period =0.4, also updated every second;

[0081] This represents the real-time operation and maintenance cost corresponding to the model, including actual expenses such as manual troubleshooting costs and equipment scheduling costs for fault repair.

[0082] and To dynamically adjust the weights, it is not necessary to satisfy the sum of 1. The core objective is to balance the early warning effect and operation and maintenance cost based on the scenario priority (prioritizing the accuracy of early warning during periods of high failure incidence and prioritizing cost control during routine operation and maintenance). The weight values ​​are determined through iterative optimization using historical operation and maintenance data.

[0083] The evaluation phase is specifically as follows:

[0084] Based on the preprocessed high-dimensional data stream in the data processing stage, the feature extraction logic of the feature extraction stage and the hybrid model trained in the model training stage are called. The Flink streaming computing framework is used to establish an incremental computing and dynamic caching mechanism. The judgment logic of incremental computing is as follows: compare the difference between the current parameter value and the previous evaluation value in real time. When the difference of core parameters (antenna power, transmission delay) exceeds ±5% or the difference of non-core parameters (ambient temperature, fan speed) exceeds ±10%, re-feature extraction and model evaluation are triggered. Stable parameters with a change frequency of <1Hz are stored in the dynamic cache. The validity period of the cached data is set to 5 minutes, and it will automatically expire after the timeout to avoid outdated data affecting the evaluation results.

[0085] The Flink cluster adopts an architecture of 3 JobManagers (master nodes) and 10 TaskManagers (worker nodes). Each TaskManager is allocated 4 cores and 8GB of resources. Through parallel computing, the evaluation latency is controlled to ≤30ms, which meets the high-frequency monitoring requirements of 5G base stations. The real-time generated evaluation report includes the operational status health score and fluctuation curve, the probability of latent fault precursors and the top three risk parameters, and the time-varying correlation diagram of key abnormal parameters, such as the correlation strength change between antenna load and transmission latency. The report is pushed to the operation and maintenance platform and visualization dashboard through the API interface, and supports SMS alarm triggering.

[0086] The specific details of the tiered early warning stage are as follows:

[0087] Based on the real-time evaluation report during the evaluation phase, and combined with the parameter dynamic correlation map during the feature extraction phase, a cross-layer parameter correlation map is formed for the physical layer, network layer, and service layer. When the map is established, the causal relationship of more than 1,000 historical fault cases is analyzed, and the influence weight of inter-layer parameters is calculated using Bayesian network. The weight values ​​are updated in real time: for example, the influence weight of physical layer beamforming accuracy on network layer slice bandwidth is 0.6, and the influence weight of network layer bandwidth allocation on service layer user QoS is 0.8. The weight values ​​are updated in real time.

[0088] During the anomaly tracing phase, based on the abnormal parameters marked in the report, such as low beamforming accuracy, the propagation path is traced backward along the graph: low beam accuracy → imbalance between antenna load and transmission delay (network layer) → insufficient bandwidth allocation of transmission link (network layer) → video service time slot occupancy exceeds the threshold (service layer). The root cause is finally located as unreasonable service time slot allocation, and a visual tracing graph in the form of a force-directed graph is generated. The node size represents the importance of the parameter, and the edge thickness represents the influence weight.

[0089] The tiered early warning system uses operational health and the probability of latent fault precursors as indicators. Different response strategies are implemented for different risk levels: Green (health score ≥ 90, precursor probability < 5%) only logs are recorded; Yellow (health score 80 ≤ health score < 90, 5% ≤ precursor probability < 20%) pushes specific adjustment suggestions, such as temporarily expanding the video service slice bandwidth by 50Mbps; Orange (health score 60 ≤ health score < 80, 20% ≤ precursor probability < 50%) notifies the operations and maintenance team to prepare for investigation; Red (health score < 60, precursor probability ≥ 50%) automatically triggers the work order system to dispatch work orders to operations and maintenance personnel within 3 kilometers, simultaneously initiates backup transmission link switching, and outputs anomaly tracing results and early warning data; operations and maintenance personnel within 3 kilometers are those within a preset operations and maintenance radius (referring to the optimal scheduling range with a historical average fault repair time ≤ 1 hour) based on the geographical location of the base station.

[0090] The self-evolutionary stage of the model is as follows:

[0091] Based on real-time evaluation data during the evaluation phase and anomaly tracing results during the graded early warning phase, edge and cloud-based layered online distillation optimization is implemented for the hybrid model during the model training phase. The edge node layer is deployed on the local server of the base station. Lightweight distillation is adopted for stable data of a single base station (parameter change frequency < 1Hz): only the output features of the Bi-LSTM layer are passed to the student model, the distillation temperature is set to 5, the loss function is the mean square error, and the time for each distillation is ≤ 30 seconds to avoid consuming too much edge computing power.

[0092] Edge nodes only deploy local servers for a single base station and do not have multi-base station parameter association data. Therefore, they only transmit the time series features of the Bi-LSTM layer (which can be calculated independently by a single base station). The parameter association features of the GAT layer need to rely on multi-base station data. They are uniformly optimized by the cloud aggregation layer and then synchronized to the edge, taking into account both edge computing power limitations and feature effectiveness.

[0093] The cloud aggregation layer is deployed in the regional cloud center. It performs full distillation on common abnormal data in multi-base station scenarios (such as mutual interference between multiple base stations in dense urban areas and concentrated electromagnetic interference in industrial areas). The parameter knowledge and attention weights of the teacher model (complete GAT and Bi-LSTM model) are passed to the student model through the KL divergence loss function. The parameters of the graph attention network association branch and the scene embedding layer are updated synchronously. The distillation termination condition is set to the base station early warning accuracy of a certain area being >95% for 72 consecutive hours to reduce invalid calculations.

[0094] Simultaneously, a base station scenario meta-knowledge base is established to store fault and parameter association rules for typical scenarios. For example, in an industrial scenario, electromagnetic interference > 90dB → antenna power fluctuation ±10%. For every 100 new fault cases added to the knowledge base, new association rules are automatically mined and updated using the Apriori algorithm. When meta-learning adapts to new scenarios, such as adding industrial IoT slicing services, similar scenarios (such as existing industrial scenarios) are first retrieved from the knowledge base, and the corresponding model parameters are loaded as initial values. Then, 500 new scenario samples are used for fine-tuning, reducing the adaptation time from the original 10 minutes to 5 minutes.

[0095] The newly added self-evolution effect evaluation module periodically calculates the early warning accuracy, operation and maintenance costs, and fault repair time over the past 24 hours. If the operation and maintenance costs decrease by more than 10% and the early warning accuracy is ≥95%, the current distillation strategy is solidified. If the early warning accuracy decreases by more than 5% or the fault repair time increases by more than 20%, it automatically rolls back to the previous version. This forms a collection, analysis, and optimization process, and the optimized model is fed back to the real-time evaluation in the evaluation phase to continuously improve the monitoring adaptability and accuracy.

[0096] Cloud-based full distillation KL divergence loss function:

[0097] ;

[0098] In the formula: Let KL divergence be denoted as and let represent the probability distribution of the teacher model. With student model probability distribution The smaller the value, the closer the distributions of the two models are, and the better the distillation effect.

[0099] The teacher model is represented in the first... Output probability distribution across each feature dimension, including weight distributions of GAT association features and Bi-LSTM temporal features;

[0100] The student model is represented in the first... The output probability distribution on each feature dimension is the target distribution that needs to be aligned with the teacher model during the distillation process;

[0101] This represents the feature dimension index of the model output, corresponding to the 256-dimensional two-dimensional feature set during the model training phase. It includes all output dimensions of the GAT association branch and the Bi-LSTM temporal branch, ensuring full knowledge transfer.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the operational status of communication base stations based on big data, characterized in that: The specific steps of this method are as follows: Data collection phase: Establish a four-level synchronous collection architecture for base station body, network link, service scenario and operation and maintenance indicators. Collect base station body parameters, network link parameters, service characteristic parameters, operation and maintenance indicator parameters and environmental parameters. Set up a network outage fault tolerance mechanism to automatically retransmit data after the network is restored. Make each data carry a four-dimensional scenario label of environment-service-network-operation and maintenance. Data processing stage: Based on the data collected in the data acquisition stage and the four-dimensional scene labels, the distribution pattern of outliers is statistically analyzed and noise is filtered out. Normalized weights are set according to business requirements, and preprocessed data is output. Feature extraction stage: Based on preprocessed data and four-dimensional scene labels, time-varying mutual information entropy is calculated using a sliding window, a GAT network is established and weights are assigned according to parameter priority, and then Bi-LSTM is used to extract parameter gradual change trend and sudden change features, forming a two-dimensional feature set composed of parameter gradual change trend features and parameter sudden change features. Model training phase: Combine the two-dimensional feature set with the four-dimensional scene label to build a training dataset, build a hybrid model including scene embedding layer, GAT and Bi-LSTM dual branches and reinforcement learning, adjust feature weights according to scene and dynamically update reinforcement learning coefficients, and complete training through GPU cluster. Evaluation phase: Incremental calculations are performed based on parameter changes, and an evaluation report is generated and pushed to the operation and maintenance platform; Tiered early warning stage: Based on the assessment report and parameter dynamic correlation map, establish cross-layer parameter correlation map of physical layer, network layer and business layer, trace the root cause of abnormal parameters in reverse, and implement tiered early warning with the health status of operation and the probability of hidden fault precursors as dual indicators. Model self-evolution stage: Based on evaluation data and source tracing results, the hybrid model is subjected to edge and cloud-layered distillation; a base station scenario meta-knowledge base is established, and meta-learning is used to adapt to new scenarios; the distillation strategy is adjusted according to the early warning accuracy and operation and maintenance costs, and the optimized model is fed back to the evaluation stage.

2. The method for monitoring the operational status of communication base stations based on big data according to claim 1, characterized in that: The data acquisition phase is described in detail below: A four-level synchronous data collection architecture is established, encompassing base station parameters, network link parameters, service scenarios, and operational metrics. Base station parameters are collected by built-in sensors within the base station, while network link parameters are obtained through the core network gateway. Business characteristic parameters are obtained from the business management platform; operation and maintenance indicator parameters are retrieved from the operation and maintenance management system; environmental parameters are collected through external sensors. During data acquisition, timestamp alignment is used to ensure that each parameter data packet carries a four-dimensional scene label. At the same time, a network outage fault tolerance mechanism is set up so that data is automatically retransmitted after the network is restored. Finally, the acquired data and scene labels are output.

3. The method for monitoring the operational status of communication base stations based on big data according to claim 2, characterized in that: The data processing stage is as follows: Based on the data collected and the associated scene tags, and considering the data noise characteristics and parameter correlation requirements under different business scenarios, a combination of dynamic threshold isolated forest and time-series noise adaptive filtering is adopted. The dynamic threshold isolated forest first statistically analyzes the distribution pattern of outliers based on historical business data. Variational mode decomposition is used to adjust the number of modes for different scenarios. At the same time, a time-varying normalization algorithm with business time slot weighting is designed in combination with business time slot characteristics. Normalization weights are set for voice business and IoT business respectively, and high-dimensional parameters are adjusted to a unified dimension while retaining the dynamic correlation between parameters, and preprocessed high-dimensional data is output.

4. The method for monitoring the operational status of communication base stations based on big data according to claim 3, characterized in that: The feature extraction stage is as follows: Based on the high-dimensional data processed in the data processing stage, and combined with the business type in the scene label, a dynamic association graph of time-varying mutual information entropy weighted by business time slots and combined with the parameters of the graph attention network is established; the time-varying mutual information entropy calculation updates the association strength through a sliding window mechanism. After introducing the base station parameter priority matrix into the graph attention network, when calculating the attention weights for the first-level parameters, parameter importance coefficients are set to highlight the influence of core parameters. By extracting the gradual trend and abrupt change features of parameters using Bi-LSTM, a two-dimensional feature set is formed.

5. The method for monitoring the operational status of communication base stations based on big data according to claim 4, characterized in that: The model training phase is detailed as follows: A dual-dimensional training dataset of scene and state is established by using a dual-dimensional feature set and combining scene labels, which includes a variety of scene types and fault cases. A hybrid model is established, in which the scene embedding layer transforms four-dimensional discrete labels into continuous feature vectors and adjusts the dimensions, and adjusts the weights of the corresponding feature vectors according to the scene; the graph attention network branch loads dynamically associated features for training; and the Bi-LSTM branch learns temporal evolution features. The reward function of the reinforcement learning module aims to balance the early warning effect and the operation and maintenance cost, and is optimized by dynamically adjusting the weights of two key influencing factors. After the model is trained by the GPU cluster, it has the ability to output the evaluation of the health status of the operation and the probability of latent fault precursors.

6. The method for monitoring the operational status of communication base stations based on big data according to claim 5, characterized in that: The specific evaluation phase is as follows: Based on the preprocessed high-dimensional data stream, the feature extraction logic of the feature extraction stage is called and the hybrid model after training is completed. The Flink streaming computing framework is used to establish an incremental computing and dynamic caching mechanism. Incremental computing is triggered by comparing parameter differences, while stable parameters are stored in the cache. The real-time assessment report includes the operational health score and fluctuation curve, the probability of latent fault precursors and risk parameters, and the time-varying correlation diagram of abnormal parameters. The report is pushed to the operation and maintenance platform and the visualization dashboard.

7. The method for monitoring the operational status of communication base stations based on big data according to claim 6, characterized in that: The specific details of the tiered early warning stages are as follows: Based on the evaluation report and the dynamic parameter correlation graph, a cross-layer parameter correlation graph is established for the physical layer, network layer and business layer; Bayesian network is used to calculate the influence weight of inter-layer parameters, and the weight values ​​are updated in real time. Based on the abnormal parameters marked in the report, the propagation path was traced backward along the graph to finally locate the root cause and generate a visual source tracing graph. Using operational health status and the probability of latent fault precursors as indicators, different response strategies are implemented for different risk levels, and anomaly tracing results and early warning data are output.

8. The method for monitoring the operational status of communication base stations based on big data according to claim 7, characterized in that: The specific self-evolutionary stages of the model are as follows: Based on real-time evaluation data from the evaluation phase and anomaly tracing results from the graded early warning phase, edge and cloud-layered online distillation optimization is implemented for the hybrid model during the model training phase. A meta-knowledge base for base station scenarios is established to store the rules for associating faults and parameters. New association rules are automatically mined and updated through algorithms. Meta-learning retrieves similar scenarios, loads the corresponding model parameters as initial values, and fine-tunes them to adapt to new scenarios. The newly added self-evolution effect evaluation module regularly calculates the early warning accuracy, operation and maintenance costs and fault repair time, and adjusts the distillation strategy accordingly; forming a collection, analysis and optimization process, and feeding the optimized model back to the evaluation stage.

Images