Communication base station operation state monitoring method based on big data

Abstract

The invention belongs to the technical field of communication base station operation monitoring, and discloses a communication base station operation state monitoring method based on big data, and the method comprises the steps: employing a dynamic threshold isolation forest and time sequence noise adaptive filtering combination scheme in a data processing stage, and carrying out the statistics of different business abnormal value distribution rules through combining historical business data; in the feature extraction stage, a time-varying mutual information entropy is calculated through a sliding window, a parameter dynamic association map is constructed, then a GAT network is used for highlighting the influence of core parameters, and Bi-LSTM is used for extracting parameter gradual change and sudden change features; the dynamic coupling relation of parameters under different service time slots can be accurately captured, hidden fault premonition can be effectively recognized in advance, and fault prediction lag is avoided; through a four-dimensional scene label and a scene embedding layer, discrete scene features are converted into continuous vectors, and weights are dynamically adjusted; a base station scene element knowledge base is established to store fault and parameter association rules, similar scenes are retrieved in combination with element learning, and initial parameters are loaded to quickly adapt to new scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication base station operation monitoring, and particularly relates to a communication base station operation state monitoring method based on big data. BACKGROUND

[0002] The communication base station is the core of the digital infrastructure, and its operation stability directly affects the network service quality and is related to the actual effect of the industry digitization; with the large-scale application of network slicing, edge computing and other technologies, the parameter dimension is greatly increased, so that the parameters monitored by a single base station are greatly increased; the state evolution is not linear; the scene difference is also obvious, and the needs of dense urban areas, industrial Internet of Things and remote pastoral areas are different; the prior art is difficult to adapt to these complex needs, and the following technical problems exist: The prior art only analyzes the time-varying coupling relationship between parameters by using static threshold or single-dimensional time sequence characteristics, and cannot capture the dynamic change of the coupling strength in different business time slots such as voice, video and Internet of Things for the Massive MIMO antenna load and transmission link delay of the 5G base station; the traditional static correlation analysis will cause the fault prediction to lag, and cannot discover the precursor of the hidden fault in advance.

[0003] Most of the existing models are in the mode of offline training and fixed deployment, and cannot adapt to the dynamic changes of heterogeneous scenes; for example, in the industrial Internet of Things scene, the base station needs to withstand electromagnetic interference; the base station in the remote pastoral area needs to adapt to low-power operation; the traditional model does not have a scene self-adaptation mechanism, and once it enters a new scene or iterates the business, such as adding VR / AR business, the monitoring accuracy will be greatly reduced, and the model needs to be retrained manually, which has a high operation and maintenance cost. SUMMARY

[0004] The purpose of the present application is to provide a communication base station operation state monitoring method based on big data to solve the problems in the background art.

[0005] In order to achieve the above purpose, the present application provides the following technical scheme: a communication base station operation state monitoring method based on big data, and the specific steps of the method are as follows: Data acquisition stage: collect multiple types of base station parameters, set a network interruption fault tolerance mechanism, and make each data have a scene label; Data processing stage: based on the data and scene labels collected in the data acquisition stage, the abnormal value distribution rule is counted and the noise is filtered, the normalization weight is set according to the business, and the preprocessed data is output; Feature extraction stage: based on the preprocessed data and scene labels, the time-varying mutual information entropy is calculated by using a sliding window, a GAT network is established and the weight is allocated according to the parameter priority, and then the time sequence features are extracted by Bi-LSTM, and finally a double-dimensional feature set is formed; Model training stage: Combine the two-dimensional feature set and scene label to establish the training data set, build a hybrid model containing scene embedding layer, GAT and Bi-LSTM double branch and reinforcement learning, adjust the feature weight according to the scene, dynamically update the reinforcement learning coefficient, and finally train through GPU cluster; Evaluation stage: Incremental calculation according to parameter changes, generate evaluation report and push to operation and maintenance platform; Hierarchical early warning stage: Establish cross-layer parameter association graph, trace back abnormal parameter root cause, and execute hierarchical early warning according to double indicators; Model self-evolution stage: Based on evaluation data and trace results, implement edge and cloud layer distillation on hybrid model; Establish base station scene meta-knowledge base, adapt to new scene with meta-learning; Adjust distillation strategy according to early warning accuracy and operation and maintenance cost, and optimize the feedback model to the evaluation stage.

[0006] Preferably, the data collection stage is as follows: Establish four-level synchronous collection architecture of base station ontology, network link, business scene and operation and maintenance indicators; Base station ontology parameters are collected by built-in sensors in base station; Network link parameters are synchronously obtained through core network gateway; Business feature parameters are obtained from business management platform; Operation and maintenance indicator parameters are synchronously called from operation and maintenance management system; Environment parameters are collected through external sensors deployed around the base station; Accurately align through timestamp during collection, so that each parameter data packet carries four-dimensional scene label; At the same time, set network fault tolerance mechanism, automatically supplement data after network recovery, and finally output collected data and scene label.

[0007] Preferably, the data processing stage is as follows: Based on the data and bound scene label of the data collection stage, according to the data noise characteristics and parameter association requirements in different business scenes, adopt dynamic threshold isolated forest and time series noise adaptive filtering combination scheme; Dynamic threshold isolated forest first calculates the abnormal value distribution law of business including voice and Internet of Things according to historical business data; Time series noise filtering stage, adjust the mode number according to different scenes through variational mode decomposition, and design business time slot weighted time-varying normalization algorithm combined with business time slot characteristics to update normalization benchmark value regularly; Set normalization weight for voice business and Internet of Things business respectively, adjust high-dimensional parameters to unified dimension and retain dynamic association relationship between parameters, and output preprocessed high-dimensional data.

[0008] Preferably, the feature extraction stage is as follows: Based on the high-dimensional data processed in the data processing stage, combined with the business type information in the scene label in the data collection stage, a parameter dynamic association graph combining the time-varying mutual information entropy of the business time slot weighting and the graph attention network is established; the time-varying mutual information entropy calculation adopts a sliding window mechanism, and after the end of each window, the association strength is updated in real time through the formula; After the base station parameter priority matrix is introduced into the graph attention network, when calculating the attention weight of the first-level parameter, the parameter importance coefficient is set to highlight the influence of the core parameter. In the time sequence feature extraction link, Bi-LSTM is used to extract features including the gradual trend of beamforming accuracy and the sudden change of parameters caused by the burst of Internet of Things business, and finally a double-dimensional feature set is obtained.

[0009] Preferably, the model training stage is specifically as follows: Using the double-dimensional feature set in the feature extraction stage, combined with the scene label in the data collection stage, a scene and state double-dimensional training data set is established, which contains multiple scene types and fault cases; A hybrid model of four-dimensional scene embedding layer, graph attention network and Bi-LSTM double branch and reinforcement learning is established. 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 double branch features through the full connection layer, and adjusts the weight of the corresponding feature vector according to the scene. The graph attention network branch loads the dynamic association features in the feature extraction stage for training, and the Bi-LSTM branch learns the time sequence evolution features. The reward function of the reinforcement learning module aims to balance the early warning effect and the operation and maintenance cost, and realizes optimization by dynamically adjusting the weights of the two types of key influence factors; the model training is based on GPU cluster, and after the training is completed, it has the ability to evaluate the health degree of the running state and the probability of implicit failure precursor.

[0010] Preferably, the evaluation stage is specifically as follows: 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, and the Flink streaming computing framework is used to establish an incremental calculation and dynamic caching mechanism; the judgment logic of incremental calculation is: comparing the difference between the current parameter value and the last evaluation value in real time, when the difference of the core parameter exceeds the preset threshold or the difference of the non-core parameter exceeds the preset threshold, triggering re-feature extraction and model evaluation; stable parameters are stored in the dynamic cache; The evaluation report generated in real time includes the running state health score and fluctuation curve, the probability of implicit failure precursor and the top three risk parameters, and the time-varying association relationship diagram of key abnormal parameters, and the report is pushed to the operation and maintenance platform and the visual large screen.

[0011] Preferably, the hierarchical early warning stage is specifically as follows: Based on the real-time evaluation report of the evaluation stage, combined with the parameter dynamic correlation graph of the feature extraction stage, the cross-layer parameter correlation graph of the physical layer, network layer and service layer is expanded; when the graph is established, the interlayer parameter influence weight is calculated by adopting the Bayesian network according to the causal relationship of the historical fault cases, and the weight value is updated in real time; The abnormality tracing stage is based on the abnormal parameters marked in the report, traces the propagation path along the graph in reverse, finally locates the root cause, and generates a visual tracing graph; The hierarchical early warning takes the running state health degree and the implicit failure precursor probability as the determination index, executes different response strategies for different risk levels, and outputs the abnormality tracing result and the early warning data.

[0012] Preferably, the model self-evolution stage is specifically as follows: Based on the real-time evaluation data of the evaluation stage and the abnormality tracing result of the hierarchical early warning stage, the hybrid model of the model training stage is subjected to edge and cloud hierarchical online distillation optimization; the edge node layer is deployed in the local server of the base station, and light distillation is adopted for the stable data of a single base station: only the output features of the Bi-LSTM layer are transmitted to the student model, and the loss function adopts the mean square error; the cloud aggregation layer is deployed in the regional cloud center, and full distillation is implemented for the common abnormal data in the multi-base station scene: the parameter knowledge and attention weight of the teacher model are transmitted to the student model through the KL divergence loss function, and the graph attention network correlation branch and the scene embedding layer parameters are updated synchronously; Meanwhile, a base station scene meta-knowledge base is established, stores the fault and parameter correlation rules, automatically mines new correlation rules and updates through the Apriori algorithm; when meta-learning adapts to a new scene, similar scenes are first searched from the knowledge base, the corresponding model parameters are loaded as initial values, and then the new scene samples are fine-tuned, so that the adaptation time is shortened; A new self-evolution effect evaluation module is added, which regularly calculates the historical early warning accuracy, operation and maintenance cost and fault repair time, and adjusts the distillation strategy according to the operation and maintenance cost and the early warning accuracy; a collection, analysis and optimization process is formed, and the optimized model is fed back to the real-time evaluation of the evaluation stage.

[0013] The beneficial effects of the present application are as follows: 1、The application adopts a dynamic threshold isolated forest and time series noise adaptive filtering combined scheme in the data processing stage, combines historical business data statistics different business abnormal value distribution law;In the feature extraction stage, the time-varying mutual information entropy is calculated through the sliding window, the parameter dynamic correlation graph is constructed, and then the GAT network is used to highlight the influence of core parameters such as antenna power and transmission delay;Bi-LSTM extracts parameter gradual change and sudden change characteristics;It can accurately capture the dynamic coupling relationship of parameters under different business time slots, such as the correlation change of 5G base station Massive MIMO antenna load and transmission link delay, effectively identify the implicit fault precursor in advance, avoid fault prediction lag, and improve the accuracy of base station operation state monitoring.

[0014] 2、The application converts discrete scene characteristics into continuous vectors and dynamically adjusts the weight through four-dimensional scene labels and scene embedding layers;A base station scene meta-knowledge base is established to store fault and parameter correlation rules, similar scenes are retrieved and initial parameters are loaded to quickly adapt to new scenes through meta-learning;At the same time, the model is optimized through edge and cloud layer distillation, which can adapt to new services without manual retraining;Greatly improve the scene adaptability, avoid manual intervention, reduce operation and maintenance cost, ensure the stability of monitoring accuracy in different scenes, and adapt to network technology iteration and business expansion demand.

[0015] 3、The application constructs a cross-layer parameter correlation graph based on Bayesian network physical layer, network layer and service layer, updates the interlayer parameter influence weight in real time, can track the abnormal parameter propagation path in reverse along the graph, locates the fault root cause and generates a visual graph;In the hierarchical early warning stage, the operation state health degree and implicit fault precursor probability are taken as double indexes, and different risk levels correspond to different response strategies, such as health degree≥90 points only record log, health degree<60 points automatically trigger work order dispatching and standby link switching;Realize accurate positioning and efficient disposal of abnormalities, shorten the fault repair time, optimize operation and maintenance resource scheduling, significantly improve the pertinence and efficiency of operation and maintenance response. BRIEF DESCRIPTION OF DRAWINGS

[0016] Fig. 1 The flow chart of the communication base station operation state monitoring method based on big data of the application; Fig. 2 The flow chart of feature extraction and model training of the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the application.

[0018] AsFigs. 1-2 As shown, the embodiment of the application provides a communication base station running state monitoring method based on big data. The specific steps of the method are as follows: The data acquisition stage is specifically as follows: A four-level synchronous acquisition architecture of base station ontology, network link, service scenario and operation and maintenance index is established. Base station ontology parameters (such as Massive MIMO antenna port power) are acquired in real time by a built-in sensor of the base station, and the sampling frequency is set to 10 Hz. Network link parameters (such as slice bandwidth allocation) are synchronously acquired through a core network gateway, and the delay is controlled within 50 ms. Service characteristic parameters (such as service type and time slot occupancy rate) are acquired in real time from a service management platform. Operation and maintenance index parameters (such as fault repair time length, operation and maintenance work order processing progress and base station energy consumption cost) are synchronously called from an operation and maintenance management system. Environmental parameters (such as electromagnetic interference intensity and pasture low-temperature endurance index) are acquired by an external sensor deployed around the base station, and the protection level reaches IP65. The architecture supports concurrent synchronous acquisition of 100,000-level base station high-dimensional data. The acquisition is accurately aligned through a time stamp, so that each parameter data packet carries four-dimensional scenario labels (environment-service-network-operation and maintenance), avoiding misplacement of data and labels. At the same time, a network outage fault tolerance mechanism is set. After network recovery, the data is automatically transmitted to ensure integrity, and finally the acquired data and scenario labels are output.

[0019] The data processing stage is specifically as follows: Based on the data and bound scenario labels in the data acquisition stage, a dynamic threshold isolated forest and time series noise adaptive filtering combined scheme is adopted according to the data noise characteristics and parameter association requirements in different service scenarios. The dynamic threshold isolated forest first calculates the abnormal value distribution law of voice, Internet of Things and other services according to historical service data of three months. The abnormal value threshold of voice service is set to ±3σ, which is 20% tighter than the conventional threshold. The threshold of Internet of Things service is set to ±5σ, which is 15% looser than the conventional threshold. In the time series noise filtering stage, the number of modes is adjusted through variational mode decomposition (VMD) for different scenarios. The electromagnetic interference noise frequency of industrial scenarios is concentrated in 200-500 MHz, and the number of VMD modes is set to 6. The multi-base station mutual interference noise frequency in urban areas is 800-1200 MHz, and 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 according to the service time slot characteristics. The normalization reference value is updated every 24 hours to avoid long-term data drift. The normalization weights of voice service and Internet of Things service are set respectively, the high-dimensional parameters are adjusted to a unified dimension, and the dynamic association relationship between parameters is retained, and the preprocessed high-dimensional data is output.

[0020] Dynamic threshold abnormal value judgment formula: ; In the formula: represents the parameter abnormal value judgment threshold, and the parameter value exceeding this range is determined as an abnormal value and enters the noise filtering process; represents the parameter mean value calculated based on historical three-month business data, which can reflect the benchmark level of long-term stable operation of the parameter; represents the standard deviation of the parameter, reflecting the data dispersion degree and embodying the normal fluctuation range of the parameter; represents the threshold coefficient, which is adjusted according to the business type, and the voice business = 3, the threshold is tightened by 20% compared with the conventional one; the Internet of Things business = 5, the threshold is relaxed by 15% compared with the conventional one; and the data characteristics of different businesses are adapted.

[0021] The feature extraction stage is specifically as follows: Based on the high-dimensional data processed in the data processing stage, combined with the business type information in the scene label in the data collection stage, a parameter dynamic correlation graph of time-varying mutual information entropy combined with a graph attention network (GAT) is established for the business time slot; the time-varying mutual information entropy is calculated using a sliding window mechanism, and a 50ms window of the voice business contains 10 sampling points, and a 500ms window of the Internet of Things contains 100 sampling points; the correlation strength is updated in real time after the end of each window through the formula, to ensure the capture of parameter correlation changes under different businesses; The business feature parameters (such as time slot occupancy rate) are statistical parameters, and the sampling frequency is 1Hz; the sampling points of the voice business are base station real-time running parameters (such as antenna power instantaneous value), and the sampling frequency is 200Hz, and the two types of parameters belong to different monitoring dimensions, and the sampling frequency is set according to the parameter function requirement.

[0022] The GAT network is designed as a 3-layer structure, containing 3 attention heads, and the hidden layer dimension of each attention head is 64; after introducing the base station parameter priority matrix, the parameter importance coefficient (antenna power 0.8, transmission delay 0.7) is additionally superimposed when calculating the attention weight of the first-level parameter (antenna power, transmission delay), and the finally allocated weight is 3 times that of the third-level parameter (such as environmental temperature), highlighting the influence of core parameters; In the time sequence feature extraction link, the bidirectional LSTM (Bi-LSTM) adopts a 2-layer structure, the hidden layer dimension is 128, the optimizer is selected as Adam, the learning rate is set as 0.001, the early stopping mechanism (stopping when the validation set loss does not decrease for 3 rounds) is used to avoid overfitting, the gradual trend of beam forming accuracy within 10 minutes, parameter sudden change caused by Internet of Things business burst and other features are extracted, and finally a double-dimensional feature set is obtained.

[0023] Time-varying mutual information entropy formula: ; In the formula: indicates the service time slot The parameter The time-varying mutual information entropy of the parameter , that is, the real-time correlation strength, directly reflects the dynamic change of the coupling relationship between parameters; indicates the parameter The initial mutual information entropy (basic correlation strength) of the parameter is calculated based on historical stable data; indicates the service time slot weight, which is dynamically adjusted according to the service type. The voice service = 1.2, the real-time requirement is high, and the weight is higher; the Internet of Things service = 0.8, the data fluctuation is large, and the weight is slightly lower; , indicates two base station parameters that need to be analyzed for correlation, which are often core parameters such as Massive MIMO antenna load and transmission link delay; indicates the service time slot type, such as voice service time slot and Internet of Things service time slot, which accurately corresponds to the service period of the actual operation of the base station.

[0024] Among them, the model training stage is specifically as follows: Use the double-dimensional feature set of the feature extraction stage to establish a scene and state double-dimensional training data set in combination with the scene labels of the data collection stage, which contains typical scenes such as dense urban areas, suburbs, industrial Internet of Things, remote pastoral areas, and transportation hubs. 300 normal samples and 50 fault samples are collected for each scene; more than 3000 fault cases are collected to cover common faults such as antenna failure, transmission link failure, and service overload failure, ensuring data diversity; Establish a four-dimensional scene embedding layer, a GAT and Bi-LSTM double branch, and a reinforcement learning hybrid model. The scene embedding layer first converts the four-dimensional discrete labels of the environment (electromagnetic interference level), service (type, time slot), network (slice configuration), and operation and maintenance (cost, repair time) into 64-dimensional continuous feature vectors through the Embedding layer, and then adjusts them to the same dimension (256-dimensional) as the double-branch features through the full connection layer. When the electromagnetic interference of the industrial scene is > 80dB and the endurance index of the pastoral area is < 60, the weight of the corresponding feature vector is automatically increased by 1.5 times; the GAT branch loads the dynamic correlation features of the feature extraction stage for training, and the Bi-LSTM branch learns the time evolution features; The reward function of the reinforcement learning module balances the early warning effect and the operation and maintenance cost, and optimizes by dynamically adjusting the weights of two key influence factors; the model training is based on GPU cluster, and after training, it has the ability to output the evaluation of the health degree of running state and the probability of implicit failure precursor.

[0025] The reinforcement learning reward function is: ; In the formula: represents the real-time reward value of the reinforcement learning module, guiding the model parameter update direction, the higher the value, the better the current model strategy; represents the early warning accuracy weight coefficient, which is dynamically adjusted according to the scene; in the high-fault period, such as holiday business peak =0.7, priority is given to the accuracy of early warning; for example, during routine operation and maintenance =0.5, updated every second according to real-time data; represents the real-time early warning accuracy of the model, with a value range of 0-100%, calculated by comparing the prediction results with the actual failure situation; represents the operation and maintenance cost weight coefficient, which is dynamically adjusted according to the scene; in the cost-sensitive period, such as routine operation and maintenance =0.6, to control the cost of manual intervention; in the high-fault period =0.4, also updated every second; represents the real-time operation and maintenance cost of the model, including the actual cost of manual troubleshooting, fault repair device scheduling, etc.; and are dynamically adjusted weights, which do not need to satisfy and be 1, the core goal is to balance the early warning effect and the operation and maintenance cost according to the scene priority (priority is given to the accuracy of early warning in the high-fault period, and priority is given to the cost control in the routine operation and maintenance period), and the weight value is determined by iterative optimization of historical operation and maintenance data.

[0026] The evaluation stage is as follows: Based on the pre-processed 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, and the Flink streaming computing framework is used to establish an incremental calculation and dynamic caching mechanism; the judgment logic of incremental calculation is: real-time comparison of the difference between the current parameter value and the last evaluation value, when the core parameter difference (antenna power, transmission delay) exceeds ±5% or the non-core parameter (environmental temperature, fan speed) difference 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 cache data validity period is set to 5 minutes, and the cache data is automatically invalidated when the time is up, to avoid the influence of outdated data on the evaluation result; The Flink cluster adopts a 3 JobManager (master node) and 10 TaskManager (worker node) architecture, each TaskManager is allocated 4 cores and 8G resources, and the evaluation delay is controlled to be ≤30ms through parallel computing, meeting the high-frequency monitoring requirements of 5G base stations; the real-time evaluation report generated contains the running state health score and fluctuation curve, the probability of implicit failure precursor and the top three risk parameters, and the time-varying correlation graph of key abnormal parameters, such as the change of the correlation strength between antenna load and transmission delay, the report is pushed to the operation and maintenance platform and visual large screen through the API interface, and supports SMS alarm triggering.

[0027] The hierarchical early warning stage is specifically as follows: Based on the real-time evaluation report in the evaluation stage, combined with the parameter dynamic correlation graph in the feature extraction stage, the cross-layer parameter correlation graph of the physical layer, network layer and service layer is expanded; when the graph is established, the causal relationship of more than 1000 historical fault cases is analyzed, and the influence weight between layers is calculated by using the Bayesian network, and the weight value is updated in real time: for example, the influence weight of the beam forming precision of the physical layer on the network layer slice bandwidth is 0.6, and the influence weight of the network layer bandwidth allocation on the service layer user QoS is 0.8, and the weight value is updated in real time; In the abnormality tracing stage, based on the abnormal parameters marked in the report, such as low beam forming precision, the propagation path is traced back along the graph: low beam precision→ imbalance of antenna load and transmission delay (network layer)→ insufficient transmission link bandwidth allocation (network layer)→ video service time slot occupancy exceeds threshold (service layer), and finally the root cause of unreasonable service time slot allocation is located, and a force-directed graph form visual tracing graph is generated, the node size represents the importance of the parameter, and the edge thickness represents the influence weight; The hierarchical early warning takes the running state health degree and the implicit failure precursor probability as the judgment index, and executes different response strategies for different risk levels: green (health degree ≥90 points, precursor probability <5%) only records the log; yellow (80≤health degree <90 points, 5%≤precursor probability <20%) pushes specific adjustment suggestions, such as temporarily expanding the video service slice bandwidth by 50Mbps; orange (60≤health degree <80 points, 20%≤precursor probability <50%) notifies the operation and maintenance team to prepare for troubleshooting; red (health degree <60 points, precursor probability ≥50%) automatically triggers the work order system to assign orders to the operation and maintenance personnel within 3 kilometers, synchronously starts the standby transmission link switching, and outputs the abnormality tracing result and early warning data; the operation and maintenance personnel within 3 kilometers are the operation and maintenance personnel within the preset operation and maintenance radius (reference optimal scheduling range of historical fault average repair time ≤1 hour) associated with the base station geographical position.

[0028] The model self-evolution stage is specifically as follows: Based on the real-time evaluation data of the evaluation stage and the abnormality traceability results of the hierarchical early warning stage, the mixed model of the model training stage is subjected to edge and cloud hierarchical online distillation optimization; the edge node layer is deployed on the local server of the base station, and light distillation is adopted for single-base-station stable data (parameter change frequency < 1 Hz): only the output features of the Bi-LSTM layer are transmitted to the student model, the distillation temperature is set to 5, the loss function is mean square error, and the time consumption of each distillation is ≤ 30 seconds, so as to avoid occupying too much edge computing power; The edge node is only deployed on the single-base-station local server, and there is no multi-base-station parameter correlation data, so only the time sequence features of the Bi-LSTM layer (which can be independently calculated by the single base station) are transmitted; the GAT layer parameter correlation features need to rely on multi-base-station data, which are synchronized to the edge after being uniformly optimized by the cloud aggregation layer, taking into account the edge computing power limitation and feature effectiveness.

[0029] The cloud aggregation layer is deployed in the regional cloud center, and full distillation is implemented for multi-base-station common abnormal data (such as mutual interference of multiple base stations in dense urban areas and concentrated electromagnetic interference in industrial areas): the parameter knowledge and attention weight of the teacher model (complete GAT and Bi-LSTM model) are transmitted to the student model through the KL divergence loss function, and the parameters of the graph attention network correlation branch and the scene embedding layer are updated synchronously, and the distillation termination condition is set to the base station warning accuracy of a certain area being > 95% for 72 consecutive hours, so as to reduce invalid calculation; At the same time, a base station scene meta-knowledge base is established to store the fault and parameter correlation rules of typical scenes, for example, electromagnetic interference > 90 dB in an industrial scene → antenna power fluctuation ± 10%, and the knowledge base automatically mines new correlation rules and updates every time 100 new fault cases are added; when adapting to a new scene through meta-learning, for example, adding an industrial Internet of Things slice business, first search for similar scenes (such as existing industrial scenes) from the knowledge base, load the corresponding model parameters as initial values, and then fine-tune with 500 new scene samples, which shortens the adaptation time from the original 10 minutes to 5 minutes; A new self-evolution effect evaluation module is added to regularly calculate the warning accuracy, operation and maintenance cost, and fault repair time of the past 24 hours, and if the operation and maintenance cost decreases by > 10% and the warning accuracy is ≥ 95%, the current distillation strategy is solidified; if the warning accuracy decreases by > 5% or the fault repair time increases by > 20%, it is automatically rolled back to the previous version; a collection, analysis, and optimization process is formed, and the optimized model is fed back to the real-time evaluation of the evaluation stage to continuously improve the monitoring adaptability and accuracy.

[0030] Cloud full distillation KL divergence loss function: ; In the formula: represents the KL divergence, and represents the probability distribution of the teacher model and the probability distribution of the student model The smaller the value is, the closer the two model distributions are, and the better the distillation effect is. represents the output probability distribution of the teacher model on the th feature dimension, containing the weight distribution of GAT correlation features and Bi-LSTM time sequence features. represents the output probability distribution of the student model on the th feature dimension, which is the target distribution to be aligned with the teacher model in the distillation process. represents the feature dimension index output by the model, corresponding to the 256-dimensional two-dimensional feature set in the model training stage, containing all output dimensions of the GAT correlation branch and the Bi-LSTM time sequence branch, ensuring full knowledge transfer.

[0031] It should be noted that in this paper, 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 that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0032] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application 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 the method are as follows: Data acquisition stage: collect multiple types of base station parameters, set a network fault tolerance mechanism, and make each data carry a scene tag; Data processing stage: based on the data collected in the data acquisition stage and the scene tag, the distribution of abnormal values is counted and noise is filtered, the normalization weight is set according to the business, and the preprocessed data is output; Feature extraction stage: based on the preprocessed data and the scene tag, the time-varying mutual information entropy is calculated by using a sliding window, the GAT network is established and the weight is allocated according to the parameter priority, and the time sequence features are extracted by Bi-LSTM to form a double-dimensional feature set; Model training stage: combined with the double-dimensional feature set and the scene tag, a training data set is established, a hybrid model containing a scene embedding layer, a GAT and Bi-LSTM double branch and reinforcement learning is built, the feature weight is adjusted according to the scene and the reinforcement learning coefficient is dynamically updated, and the training is completed by a GPU cluster; Evaluation stage: incremental calculation is performed according to parameter changes, an evaluation report is generated and pushed to the operation and maintenance platform; Hierarchical early warning stage: a cross-layer parameter correlation graph is established, the root cause of abnormal parameters is traced back, and hierarchical early warning is performed based on double indicators; Model self-evolution stage: based on the evaluation data and the traceability result, the hybrid model is implemented by layered distillation of the edge and the cloud; a base station scene meta-knowledge base is established, and a new scene is adapted by meta-learning; the distillation strategy is adjusted according to the early warning accuracy and the operation and maintenance cost, and the optimized model is fed back to the evaluation stage. 2.The big data based communication base station operation state monitoring method of claim 1, wherein: The data acquisition stage is as follows: A four-level synchronous acquisition architecture of base station ontology, network link, business scene and operation and maintenance indicators is established; base station ontology parameters are collected by built-in sensors; network link parameters are obtained through core network gateways; Business feature parameters are obtained from the business management platform; operation and maintenance indicator parameters are called from the operation and maintenance management system; environmental parameters are collected by external sensors; During acquisition, time stamps are aligned to make each parameter data packet carry four-dimensional scene tags; at the same time, a network fault tolerance mechanism is set, and data is automatically transmitted after network recovery, finally outputting the collected data and scene tags. 3.The big data based communication base station operation state monitoring method of claim 2, wherein: The data processing stage is as follows: Based on the data and bound scene tags in the data acquisition stage, a dynamic threshold isolation forest and time series noise adaptive filtering combination scheme is used according to the data noise characteristics and parameter correlation requirements in different business scenes; the dynamic threshold isolation forest first counts the distribution of abnormal values according to historical business data; Through variational modal decomposition, the number of modes is adjusted for different scenes, and a time-varying normalization algorithm with business time slot weighting is designed combining with the business time slot characteristics; the normalization weight is set for voice business and Internet of Things business respectively, the high-dimensional parameters are adjusted to a unified dimension while retaining the dynamic correlation between parameters, and the preprocessed high-dimensional data is output. 4.The big data based communication base station operation state monitoring method of claim 3, wherein: The feature extraction stage is as follows: Based on the high-dimensional data processed in the data processing stage, the business type in the scene tag is combined to establish a parameter dynamic correlation graph of time-varying mutual information entropy combined with a graph attention network; the correlation strength is updated by a sliding window mechanism. The graph attention network sets a parameter importance coefficient when calculating the attention weight of the first-level parameter after introducing the base station parameter priority matrix, and highlights the influence of the core parameter. The Bi-LSTM extracts the gradual change trend and sudden change characteristics of the parameter to form a double-dimensional feature set. 5.The big data based communication base station operation state monitoring method of claim 4, wherein: The model training stage is as follows: Using the double-dimensional feature set, a scene and state double-dimensional training data set is established by combining scene labels, which contains various scene types and fault cases; A hybrid model is established, in which the scene embedding layer converts four-dimensional discrete labels into continuous feature vectors and adjusts the dimensions to adjust the weights of the corresponding feature vectors according to the scene; the graph attention network branch loads dynamic correlation features for training; the Bi-LSTM branch learns the time evolution characteristics; The reinforcement learning module balances the warning effect and operation and maintenance cost by dynamically adjusting the weights of two key influence factors to achieve optimization; after the model is trained by the GPU cluster, it has the ability to output the evaluation of the running state health degree and the probability of implicit fault precursor. 6.The big data based communication base station operation state monitoring method of claim 5, wherein: The evaluation stage is as follows: Based on the preprocessed high-dimensional data stream, the feature extraction logic of the feature extraction stage and the trained hybrid model are called, and the Flink streaming computing framework is used to establish an incremental calculation and dynamic caching mechanism; incremental calculation is triggered by comparing parameter differences, and stable parameters are stored in the cache; The real-time evaluation report includes the running state health score and fluctuation curve, the implicit fault precursor probability and risk parameters, and the time-varying correlation graph of abnormal parameters, which are pushed to the operation and maintenance platform and visual large screen. 7.The big data based communication base station operational status monitoring method of claim 6, wherein: The hierarchical early warning stage is as follows: Based on the evaluation report and parameter dynamic correlation graph, a cross-layer parameter correlation graph of the physical layer, network layer and business layer is established; the influence weight of interlayer parameters is calculated using Bayesian network, and the weight value is updated in real time; Based on the abnormal parameters marked in the report, the propagation path is traced back along the graph, and the root cause is finally located, and a visual traceability graph is generated; Taking the running state health degree and the implicit fault precursor probability as the judgment index, different response strategies are executed for different risk levels, and the abnormal traceability result and early warning data are output. 8.The big data based communication base station operation state monitoring method of claim 7, wherein: The model self-evolution stage is as follows: Based on the real-time evaluation data of the evaluation stage and the abnormal traceability result of the hierarchical early warning stage, the hybrid model of the model training stage is implemented by edge and cloud layered online distillation optimization; A base station scene meta-knowledge base is established to store fault and parameter correlation rules, and new correlation rules are automatically mined and updated by algorithms; meta-learning retrieves similar scenes, loads corresponding model parameters as initial values, and fine-tunes them to adapt to new scenes; The newly added self-evolution effect evaluation module regularly calculates the warning accuracy, operation and maintenance cost and fault repair time, and adjusts the distillation strategy accordingly; a collection, analysis and optimization process is formed, and the optimized model is fed back to the evaluation stage.

Images