A method for automatic configuration recovery of a communication radio unit

Abstract

This invention relates to an automatic configuration and recovery method for communication radio frequency units, specifically in the field of communications. This method constructs a dynamic graph integrating configuration, performance, and topology relationships, and utilizes a spatiotemporal graph neural network to learn a network health benchmark. This enables the automatic discovery of implicit parameter conflicts introduced by manual configuration. Its community-based subgraph partitioning and consistency evaluation mechanism can accurately locate abnormal radio frequency unit groups causing performance degradation. By simulating and evaluating multiple repair schemes in parallel within a digital twin environment, the optimality and security of the recovery decision are ensured. Ultimately, a complete closed loop is formed from intelligent detection, precise location, simulation optimization to secure execution, significantly improving the automation level and problem-solving efficiency of network operation and maintenance, and effectively avoiding chronic performance degradation caused by configuration errors.

Description

Technical Field

[0001] This invention relates to the field of communications, and more specifically, to a method for automatic configuration recovery of a communication radio frequency unit. Background Technology

[0002] With the rapid development of mobile communication technology and the explosive growth of user traffic demand, especially in densely populated urban core areas, large business districts, or emerging large residential communities, communication networks face continuous expansion pressure. To cope with the surge in service capacity and coverage demand, operators need to complete the centralized deployment and network access activation of a large number of new radio frequency units, such as active antenna units or radio frequency remote units, in a short period of time. This scenario usually involves a mixed network of new equipment and existing old equipment, with complex and diverse configuration parameter systems. At the same time, the external environment presents extremely high user density and complex and ever-changing service models, which puts forward stringent requirements on network performance and stability. In this high-intensity, short-cycle network construction process, the massive and tedious equipment parameter configuration work relies heavily on manual operation by engineers. The tight time and heavy workload have become the main challenges.

[0003] However, existing network operation and maintenance technologies are significantly inadequate when facing the implicit configuration defects introduced by the above scenarios. Currently, ensuring the accuracy of network configuration mainly relies on traditional post-event verification methods such as manual drive testing and fixed-point testing, or on simple rule alarms preset in the network management system, such as physical cell identifier conflict alarms. These methods can usually only detect obvious configuration errors that violate hard rules. For more subtle parameter inconsistencies that may be introduced due to large-scale manual configuration, such as omissions in neighbor cell relationship configuration, failure to achieve optimal matching of power and antenna tilt angle settings in overlapping coverage areas, or improper setting of handover parameter thresholds based on multi-layer networks, existing technologies lack effective proactive discovery mechanisms. While advanced network management systems integrate root cause analysis based on key performance indicators, their analytical models primarily focus on explicit equipment hardware failures, transmission problems, or sudden external interference. Their diagnostic capabilities are weak for "sub-healthy" states—informative but persistent damage to network performance caused by the interaction of multiple network element configuration parameters under specific wireless propagation environments and service loads. Therefore, after expanding complex networks in densely populated urban areas, how to automatically and intelligently identify such deep parameter conflicts caused by manual configuration from a network-wide perspective, and further execute precise and automated configuration repairs to prevent chronic degradation of local network performance, has become a core challenge that urgently needs to be addressed in existing technologies. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an automatic configuration recovery method for communication radio frequency units. This method constructs a dynamic graph that integrates configuration, performance, and topology relationships, and utilizes a spatiotemporal graph neural network self-learning network health benchmark to automatically detect implicit parameter conflicts introduced by manual configuration, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: Collect RF unit configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data from the network management system, performance management system, and terminal measurement reports. Standardize and spatiotemporally align the collected configuration parameters, engineering parameters, KPIs, and MR data. Based on the processed configuration parameters, engineering parameters, KPIs, and MR data, construct a dynamic configuration-performance graph with RF units as nodes, switching and interference relationships between RF units as edges, and configuration parameters, KPIs, and MR data as attributes. Step S2: Obtain network state snapshots at multiple historical time points from the dynamic configuration-performance graph, use the spatiotemporal graph neural network model to perform unsupervised training on the multiple network state snapshots, obtain a graph encoder through training, and determine a health state representation vector that represents the benchmark of network health state based on the network state snapshots at multiple historical time points. Step S3: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2 to obtain the health status representation vector of the current network. Calculate the deviation between the current network health status representation vector and the health status benchmark vector determined in step S2. When the deviation exceeds a preset threshold, divide the dynamic configuration-performance graph into subgraphs based on the community detection algorithm. Calculate the structure-attribute consistency score of each subgraph. Determine the radio frequency units contained in the subgraphs with structure-attribute consistency scores lower than the consistency threshold as radio frequency units with implicit configuration conflicts, forming an abnormal unit group. Step S4: For the abnormal unit group identified in step S3, generate at least one set of candidate configuration repair schemes with different adjustment dimensions and magnitudes in the digital twin network environment and perform parallel simulation. Use the graph encoder obtained in step S2 to evaluate the health recovery effect of the network state after simulation, obtain the recovery effect evaluation value corresponding to each candidate configuration repair scheme, select the candidate configuration repair scheme with the best recovery effect evaluation value, and send it to the corresponding radio frequency unit in the real network for execution after security verification to complete the automatic recovery of configuration. In a preferred embodiment, the configuration parameters collected from the network management system in step S1 include at least the unique identifier of each radio frequency unit, physical cell identifier, frequency point, transmit power, antenna azimuth angle, downtilt angle, and neighbor cell relationship list. The engineering parameters collected from the network management system include at least the latitude, longitude, and altitude of each radio frequency unit; The key performance indicators (KPIs) data collected from the performance management system include at least the cell throughput, wireless connection rate, handover success rate, and drop rate of each radio frequency unit within a preset time window. The wireless measurement report (MR) data extracted from the terminal measurement report includes at least the reference signal received power and signal-to-interference-plus-noise ratio of the serving cell and all neighboring cells reported by the terminal.

[0006] In a preferred embodiment, the specific operations for standardizing and spatiotemporally aligning the collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are as follows: The collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are assigned a unified global timestamp and matched and integrated based on the unique identifier of the radio frequency unit, so that various types of data belonging to the same radio frequency unit and within the same time window form an associated data set.

[0007] In a preferred embodiment, the process of constructing a dynamic configuration-performance graph with radio frequency units as nodes, switching and interference relationships between radio frequency units as edges, and configuration parameters, key performance indicators (KPIs), and wireless measurement report (MR) data as attributes specifically includes: Radio frequency units are used as nodes in the dynamic configuration-performance graph. The attributes of each node consist of all the configuration parameters and engineering parameters of the corresponding radio frequency unit. According to the neighbor cell relationship list collected from the network management system, if the neighbor cell relationship list of the first radio frequency unit contains the unique identifier of the second radio frequency unit, then a directed edge representing the switching relationship from the first radio frequency unit to the second radio frequency unit is established between the two nodes representing the first radio frequency unit and the second radio frequency unit. For any two radio frequency units, based on the associated data set, within a preset time window, all measurement records of the reference signal received power of these two radio frequency units simultaneously reported by all terminals are selected from the wireless measurement report (MR) to form a measurement record set. For each measurement record in the measurement record set, perform the following calculation process: First, the value obtained by subtracting the reference signal received power value of the first radio frequency unit from the reference signal received power value of the second radio frequency unit and adding an intensity offset constant is used as the basis interference intensity component for the current measurement record. Secondly, calculate the absolute value of the difference between the signal-to-interference-plus-noise ratio of the terminal on the first radio frequency unit and a preset target signal quality value; then, divide the absolute value by a quality sensitivity adjustment factor and take the negative of the resulting quotient. Next, the value of the exponential function with the natural constant as the base and negative numbers as the exponent is calculated and used as the signal quality sensitivity weight of the current measurement record; Finally, the basic interference intensity component of the current measurement record is multiplied by the signal quality sensitivity weight to obtain the weighted interference contribution value of the current measurement record; After obtaining the weighted interference contribution values ​​of all measurement records in the measurement record set, these weighted interference contribution values ​​are summed and then divided by the total number of measurement records in the measurement record set. The result is the comprehensive interference intensity evaluation value between the first radio frequency unit and the second radio frequency unit. If the overall interference strength assessment value exceeds the preset interference threshold, an undirected edge representing the interference relationship is established between the two nodes representing the first and second radio frequency units, and the overall interference strength assessment value is used as the weight attribute of this undirected edge. For each node, the KPI quantile obtained within the preset time window is attached as the node's dynamic performance attribute. For each directed edge representing the handover relationship, the handover success rate obtained within the preset time window is attached as the directed edge's dynamic performance attribute. The dynamic performance attributes of the nodes and directed edges are updated as the preset time window slides, thereby constructing a dynamic configuration-performance graph.

[0008] In a preferred embodiment, in step S2, the specific operation of obtaining a graph encoder by unsupervised training on multiple network state snapshots using a spatiotemporal graph neural network model is as follows: From the dynamic configuration-performance graph constructed in step S1, a series of network state snapshots are extracted at fixed sampling intervals within a pre-selected historical time period to form a set of network state snapshots for training. Each network state snapshot contains nodes, directed edges, and undirected edges that make up the dynamic configuration-performance graph, as well as the attributes of the nodes and edges. The attributes of the nodes include configuration parameters, engineering parameters, and key performance indicators (KPIs) quantiles collected from the network management system and performance management system. The attributes of the edges include the comprehensive interference intensity assessment value calculated based on the wireless measurement report (MR) as the weight of the undirected edge and the handover success rate as the dynamic performance attribute of the directed edge. Nodes, directed edges, undirected edges, and their respective attributes together constitute a network state snapshot. Multiple network state snapshots are arranged in chronological order to form a training dataset. The spatiotemporal graph neural network model consists of multiple spatiotemporal convolutional layers stacked alternately, where each spatiotemporal convolutional layer contains a spatial graph convolutional sub-layer and a temporal convolutional layer; The spatiotemporal graph neural network model is trained in an unsupervised manner. The training objective function consists of two parts: the first part is the loss for maximizing mutual information of topological attributes; the second part is the temporal consistency comparison loss. For a network state snapshot used as an anchor point, network state snapshots that are temporally adjacent are considered positive samples, and network state snapshots that are not temporally adjacent and are randomly selected are considered negative samples. The training objective function is a weighted sum of the loss for maximizing mutual information of topological attributes and the loss for temporal consistency comparison. The two losses are jointly optimized by minimizing the training objective function through the backpropagation algorithm, thereby training a graph encoder that can encode any network state snapshot into a health state representation vector of a fixed dimension vector.

[0009] In a preferred embodiment, the process of determining a health status representation vector based on network state snapshots from multiple historical time points specifically involves: All network state snapshots used in training the graph encoder are input into the trained graph encoder to obtain the health state representation vector corresponding to each network state snapshot, forming a health state representation vector set. The robust estimation algorithm of minimum covariance determinant estimation is used to calculate the health state representation vector set to obtain a health state baseline vector that is insensitive to potential outliers in the set, which serves as the health state representation vector representing the baseline of the network health state. At the same time, the covariance matrix of the health state representation vector set is calculated.

[0010] In a preferred embodiment, step S3, specifically calculating the deviation between the current network's health status representation vector and the health status baseline vector determined in step S2, involves the following steps: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2, and output the health state representation vector of the current network; obtain the health state baseline vector and the corresponding covariance matrix calculated based on the network state snapshots at historical time points in step S2; Robust principal component analysis is performed on the set of all historical health status representation vectors used to determine the baseline health status vector. The first k principal component directions are extracted to form a projection matrix. Using the projection matrix, the difference vector between the current network's health status representation vector and the baseline health status vector is projected onto the first k principal component directions to obtain a dimension-reduced difference projection vector. For each principal component direction, a robust estimation method is used to calculate the dispersion of the values ​​obtained after projecting all historical health status representation vectors in that direction, which is used as the robust variance of that direction. Construct a diagonal matrix whose diagonal elements are composed of the robust variances corresponding to the directions of each principal component; calculate the transpose of the dimension-reduced difference projection vector, multiply it by the inverse of the diagonal matrix, and then multiply it by the dimension-reduced difference projection vector itself; finally, perform a square root operation on the multiplication result, and the resulting scalar value is the deviation.

[0011] In a preferred embodiment, the specific process of determining the radio frequency units contained in the subgraph with a structure-attribute consistency score lower than the consistency threshold as radio frequency units with implicit configuration conflicts, and forming an abnormal unit group, is as follows: When the deviation exceeds a preset threshold, a community detection algorithm based on modularity optimization is used to divide the real-time dynamic configuration-performance graph into multiple subgraphs. Each subgraph contains a subset of multiple radio frequency unit nodes and all edges between nodes within that subset. For each subgraph obtained by partitioning, its structure-attribute consistency score is calculated. The structure-attribute consistency score is determined by the weighted sum of configuration conflict component, performance anomaly component, and topology rationality component. A consistency score threshold is set, and the consistency score threshold is determined as follows: During the model training phase, the dynamic configuration-performance graph of the historical normal period is used to calculate the structure-attribute consistency scores of a large number of subgraphs based on the same method, and a specified low quantile of these structure-attribute consistency scores is taken as the consistency score threshold; subgraphs whose calculated structure-attribute consistency scores are lower than the consistency score threshold are judged as abnormal subgraphs; the sets of radio frequency unit nodes contained in all abnormal subgraphs are merged to form an abnormal unit group.

[0012] In a preferred embodiment, the specific process of generating at least one set of candidate configuration repair schemes containing different adjustment dimensions and magnitudes and performing parallel simulations for the abnormal unit group identified in step S3 in the digital twin network environment is as follows: For each radio frequency unit in the abnormal unit group, analyze at least one main reason that leads to a low structure-attribute consistency score in step S3. The main reasons include at least physical cell identifier conflict, improper transmit power setting, improper antenna tilt setting, or omission of neighbor cell relationship configuration. Based on the main reason analysis results, automatically generate a set of candidate configuration repair schemes, where each candidate configuration repair scheme explicitly specifies the adjustment value of one or more configuration parameters of one or more radio frequency units in the abnormal unit group. The adjustment value of the configuration parameters is generated according to preset adjustment rules and security boundaries. In a digital twin network environment, based on the dynamic configuration-performance graph at the current moment, a local network simulation model is constructed, which includes an anomalous unit group, first-order neighbor nodes directly connected to nodes in the anomalous unit group through edges, and second-order neighbor nodes directly connected to the first-order neighbor nodes through edges. Each of at least one set of candidate configuration repair schemes is independently injected into the corresponding local network simulation model in the digital twin network environment, and each local network simulation model is driven in parallel to run for a preset simulation duration. During the simulation, virtual wireless measurement reports and virtual key performance indicator data generated by the simulation are collected.

[0013] In a preferred embodiment, the process of evaluating the health recovery effect of the simulated network state using the graph encoder obtained in step S2, obtaining the recovery effect evaluation value corresponding to each candidate configuration repair scheme, selecting the candidate configuration repair scheme with the best recovery effect evaluation value, and sending it to the corresponding radio frequency unit in the real network after security verification is specifically as follows: For each candidate configuration repair scheme, based on the simulated wireless measurement report and key performance index data collected after the simulation run, a simulated dynamic configuration-performance graph is reconstructed according to the method in step S1. The simulated dynamic configuration-performance graph is input into the graph encoder obtained in step S2 to obtain a simulated health state representation vector. The Euclidean distance between the simulated health state representation vector and the health state reference vector determined in step S2 is calculated and denoted as the new distance. The Euclidean distance between the current network health state representation vector and the health state reference vector before the repair is performed is calculated and denoted as the original distance. The ratio of the new distance to the original distance is calculated and the result is used as the basic recovery score of the candidate configuration repair scheme. Simultaneously, calculate a penalty item for schemes that are penalized for excessively large adjustments in the scope and magnitude of the repair scheme; Subtract a penalty term weighted by a penalty coefficient from the basic recovery score, and the result is used as the recovery effect evaluation value of the candidate configuration repair scheme. Select the scheme with the highest recovery effect evaluation value from all candidate configuration repair schemes as the optimal candidate configuration repair scheme. Before sending the optimal candidate configuration repair scheme to the real network, perform a security verification. After the security verification is passed, the optimal candidate configuration repair scheme is converted into a specific configuration command and automatically sent in batches to the corresponding radio frequency units in the real network for configuration update through the standard network configuration management interface. After the configuration update is executed, return to step S1.

[0014] The beneficial effects of this invention are as follows: By constructing a dynamic graph that integrates configuration, performance, and topology relationships, and by leveraging a spatiotemporal graph neural network to learn a network health benchmark, this solution achieves automatic discovery of implicit parameter conflicts introduced by manual configuration. Its subgraph partitioning and consistency evaluation mechanism based on community discovery can accurately locate abnormal radio frequency unit groups that cause performance degradation. By simulating and evaluating multiple repair schemes in parallel in a digital twin environment, the optimality and security of recovery decisions are ensured. Ultimately, a complete closed loop is formed from intelligent detection, precise location, simulation optimization to secure execution, significantly improving the automation level and problem-solving efficiency of network operation and maintenance, and effectively avoiding chronic performance degradation caused by configuration errors. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1The method for automatic configuration recovery of a communication radio frequency unit, as shown, specifically includes the following steps: Step S1: Collect RF unit configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data from the network management system, performance management system, and terminal measurement reports. Standardize and spatiotemporally align the collected configuration parameters, engineering parameters, KPIs, and MR data. Based on the processed configuration parameters, engineering parameters, KPIs, and MR data, construct a dynamic configuration-performance graph with RF units as nodes, switching and interference relationships between RF units as edges, and configuration parameters, KPIs, and MR data as attributes. Step S2: Obtain network state snapshots at multiple historical time points from the dynamic configuration-performance graph, use the spatiotemporal graph neural network model to perform unsupervised training on the multiple network state snapshots, obtain a graph encoder through training, and determine a health state representation vector that represents the benchmark of network health state based on the network state snapshots at multiple historical time points. Step S3: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2 to obtain the health status representation vector of the current network. Calculate the deviation between the current network health status representation vector and the health status benchmark vector determined in step S2. When the deviation exceeds a preset threshold, divide the dynamic configuration-performance graph into subgraphs based on the community detection algorithm. Calculate the structure-attribute consistency score of each subgraph. Determine the radio frequency units contained in the subgraphs with structure-attribute consistency scores lower than the consistency threshold as radio frequency units with implicit configuration conflicts, forming an abnormal unit group. Step S4: For the abnormal unit group identified in Step S3, generate at least one set of candidate configuration repair schemes with different adjustment dimensions and magnitudes in the digital twin network environment and perform parallel simulation. Use the graph encoder obtained in Step S2 to evaluate the health recovery effect of the network state after simulation, obtain the recovery effect evaluation value corresponding to each candidate configuration repair scheme, select the candidate configuration repair scheme with the best recovery effect evaluation value, and send it to the corresponding radio frequency unit in the real network for execution after security verification to complete the automatic recovery of configuration.

[0020] In this embodiment, it is specifically noted that in step S1, the configuration parameters collected from the network management system include at least the unique identifier, physical cell identifier, frequency, transmit power, antenna azimuth, downtilt angle, and neighbor cell relationship list for each radio frequency unit. The unique identifier is used to unambiguously refer to a radio frequency unit within the system; the physical cell identifier and frequency are used to distinguish wireless air interface signals; the transmit power, antenna azimuth, and downtilt angle are key configurable parameters that directly affect wireless coverage and quality; the neighbor cell relationship list defines the set of target cells that the radio frequency unit is allowed to connect to during the handover process. The engineering parameters collected from the network management system include at least the latitude, longitude, and altitude of each radio frequency unit. The latitude and longitude are used to determine the geographical location of the radio frequency unit, and the altitude refers to the height of the antenna installation above the ground. Together, they are used to calculate the spatial relationship and propagation model between radio frequency units. The key performance indicators (KPIs) collected from the performance management system include at least the cell throughput, radio call success rate, handover success rate, and drop rate for each radio frequency unit within a preset time window. The preset time window is usually 15 minutes or 1 hour. Cell throughput reflects data capacity, radio call success rate reflects access performance, handover success rate reflects mobility performance, and drop rate reflects connection retention capability. These indicators are the direct basis for evaluating the quality of network performance. The wireless measurement report (MR) data extracted from the terminal measurement report includes at least the reference signal received power and signal-to-interference-plus-noise ratio (SNR) of the serving cell and all neighboring cells reported by the terminal. The reference signal received power is used to measure signal strength, and the SNR is used to measure signal quality. Both are the basic data for analyzing coverage and interference conditions. The specific steps for standardizing and spatiotemporally aligning the collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are as follows: The collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are assigned a unified global timestamp and matched and integrated based on the unique identifier of the radio frequency unit (RF unit). This allows various types of data belonging to the same RF unit and within the same time window to form an associated data set. The unified global timestamp originates from the network clock synchronization protocol. Matching and integrating based on the unique identifier of the RF unit means associating and binding configuration parameters, engineering parameters, KPIs within the same statistical period, and MR data within the same time period under the same unique identifier. This forms a complete state description of the RF unit at a specific moment, and the associated data set is the direct input for subsequent map construction. The process of constructing a dynamic configuration-performance graph, with radio frequency units as nodes, switching and interference relationships between radio frequency units as edges, and configuration parameters, key performance indicators (KPIs), and wireless measurement report (MR) data as attributes, is as follows: Radio frequency (RF) units are used as nodes in the dynamic configuration-performance graph. The attributes of each node consist of all configuration parameters and engineering parameters of the corresponding RF unit. Based on the neighbor cell relationship list collected from the network management system, if the neighbor cell relationship list of the first RF unit contains the unique identifier of the second RF unit, a directed edge representing the switching relationship from the first RF unit to the second RF unit is established between the two nodes representing the first RF unit and the second RF unit. The established directed edge records the preset switching path in the network design and is the basis for analyzing the correctness of the switching logic. For any two radio frequency units, based on the associated data set, within a preset time window, all measurement records of the reference signal received power of the two radio frequency units reported simultaneously by all terminals are selected from the wireless measurement report (MR) to form a measurement record set; "simultaneous reporting" means that the same measurement record contains the measurement results for the two radio frequency units, which ensures that the interference assessment is based on the actual perception of the terminals at the same time and place. For each measurement record in the measurement record set, perform the following calculation process: First, the value obtained by subtracting the reference signal received power value of the first radio frequency unit from the reference signal received power value of the second radio frequency unit and adding an intensity offset constant is used as the basis for the current measurement record. The intensity offset constant is used to adjust the reference for interference calculation. Its typical value can be set between 0 and 10 dBm, for example, 3 dBm, to compensate for measurement deviation or define the starting threshold for interference significance. Secondly, calculate the absolute value of the difference between the signal-to-interference-plus-noise ratio (SNR) of the terminal on the first radio frequency unit and a preset target signal quality value; then, divide the absolute value by a quality sensitivity adjustment factor and take a negative number for the quotient; the preset target signal quality value is the signal quality level that the network optimization expects to achieve, for example, set to -3dB; the quality sensitivity adjustment factor is used to control the sensitivity of the signal quality deviation to the final weight, and a typical value can be selected between 1 and 5, for example, 2. The core of this step is that when the signal quality of the serving cell is worse, the current measurement record is given a greater weight, because the impact of interference is more fatal in this scenario; Next, the value of an exponential function with a base of the natural constant and a negative exponent is calculated as the signal quality sensitivity weight of the current measurement record. This calculation transforms the relative degree of signal quality degradation into a weight value between 0 and 1, thus focusing on the "poor experience point". Finally, the basic interference intensity component of the current measurement record is multiplied by the signal quality sensitivity weight to obtain the weighted interference contribution value of the current measurement record. This multiplication operation makes the final interference assessment not only consider the relative strength of the signal, but also coupled with the actual signal quality perceived by the terminal, thereby more accurately reflecting the level of interference that has a substantial impact on the user experience. After obtaining the weighted interference contribution values ​​of all measurement records in the measurement record set, these weighted interference contribution values ​​are summed and then divided by the total number of measurement records in the measurement record set. The result is the comprehensive interference strength evaluation value between the first radio frequency unit and the second radio frequency unit. The summation and averaging operations, based on statistical laws, yield a stable and representative interference relationship strength between the two radio frequency units. If the overall interference intensity assessment value exceeds the preset interference threshold, an undirected edge representing the interference relationship is established between the two nodes representing the first and second radio frequency units, and the overall interference intensity assessment value is used as the weight attribute of this undirected edge. The preset interference threshold is a configurable parameter used to determine whether the interference is strong enough to require explicit characterization in the topology graph; for example, it can be set to 5dB. The established undirected edge and weight attribute quantitatively characterize the actual mutual interference relationship between radio frequency units that affects performance. This is fundamentally different from interference prediction based on engineering distance or simple co-frequency rules, and has higher accuracy. For each node, a key performance indicator (KPI) quantile obtained statistically within a preset time window is attached as the node's dynamic performance attribute. The KPI quantile is preferably the 50th percentile (median). This is used to reflect the typical performance level of the radio frequency unit within a time window, avoiding the influence of abnormal extreme values. For each directed edge representing a handover relationship, its handover success rate, statistically obtained within a preset time window, is attached as a dynamic performance attribute of the directed edge. The handover success rate attribute of the edge directly reflects the actual execution effect of the handover path. The dynamic performance attributes of nodes and directed edges are updated as the preset time window slides, thereby constructing a dynamic configuration-performance graph. By updating attributes through a sliding time window, the dynamic configuration-performance graph can continuously and in real time reflect the evolution of the network's operating status, providing a vivid data foundation for subsequent anomaly detection. In the graph, static configuration, dynamic performance, and topological relationships between units are organically unified, forming the core data model for realizing the identification of implicit configuration conflicts.

[0021] In this embodiment, it is specifically necessary to explain that in step S2, the unsupervised training of multiple network state snapshots using a spatiotemporal graph neural network model to obtain a graph encoder is performed as follows: From the dynamic configuration-performance graph constructed in step S1, a series of network state snapshots are extracted within a pre-selected historical time period at fixed sampling intervals to form a set of network state snapshots for training. The fixed sampling interval is hourly, half-hourly, or other time intervals set according to the actual network data collection frequency; hourly is preferred to balance the model's learning needs for the daily network variation patterns with the amount of data computation; the historical time period corresponds to a period of operation in which the network has no known major faults and the key performance indicators (KPIs) are stable; usually, a continuous time span of 30 days or longer is selected to ensure coverage of various business modes such as weekdays, weekends, peak hours, and off-peak hours. Each network state snapshot contains nodes, directed edges, and undirected edges that make up the dynamic configuration-performance graph, as well as the attributes of the nodes and edges. The attributes of the nodes include configuration parameters, engineering parameters, and key performance indicators (KPIs) quantiles collected from the network management system and performance management system. The attributes of the edges include the comprehensive interference intensity assessment value calculated based on the wireless measurement report (MR) as the weight of the undirected edge and the handover success rate as the dynamic performance attribute of the directed edge. Nodes, directed edges, undirected edges, and their respective attributes together constitute a network state snapshot. Multiple network state snapshots are arranged in chronological order to form a training dataset. The spatiotemporal graph neural network model consists of multiple spatiotemporal convolutional layers stacked alternately. Each spatiotemporal convolutional layer contains a spatial graph convolutional sub-layer and a temporal convolutional sub-layer. The spatial graph convolutional sub-layer employs a graph attention network mechanism. The specific calculation process is as follows: For any node in the network state snapshot, firstly, the feature vector of that node in the previous layer is multiplied by a trainable first weight matrix to obtain the current first feature projection of that node. Simultaneously, the feature vectors of each of the node's first-order neighbors in the previous layer are multiplied by the same first weight matrix to obtain the current second feature projections of each neighbor node. Then, the first feature projection and a second feature projection are concatenated. The concatenated vector is then multiplied by a trainable attention vector, and the result is input into a leaky linear rectified function for nonlinear activation to obtain an original attention score. This process is repeated for all the node's first-order neighbors to obtain multiple original attention scores. Attention scores are calculated. Finally, these raw attention scores are processed using a normalized exponential function to obtain the final attention weights between the node and each of its first-order neighbor nodes. Based on these final attention weights, the second feature projections of each neighbor node are weighted and summed, then combined with the first feature projection of the node and subjected to a nonlinear transformation. This outputs the node's feature vector updated by the current spatial graph convolutional sublayer, thereby learning the spatial dependencies between each node's attributes and its local network topology. First-order neighbor nodes consist of other nodes directly connected to the node via directed or undirected edges. The attention mechanism allows the model to dynamically focus on neighbors that have a greater impact on the current node's state; for example, neighbors with higher interference weights may be assigned higher weights during spatial information aggregation. The feature vector is a trainable, continuous numerical vector used within the model to represent the node's state, with its initial value mapped from the node's attribute information. The temporal convolutional sublayer employs gated causal convolution. The specific calculation process is as follows: For each node's feature vector in the network state snapshot sequence, a sequence formed by the feature vector in the temporal dimension is convolved using a one-dimensional causal convolution kernel to obtain the first convolution result; simultaneously, another one-dimensional causal convolution kernel is used to convolve this sequence to obtain the second convolution result; the first convolution result is input into the hyperbolic tangent activation function, and the second convolution result is input into the sigmoid growth curve activation function; finally, the first convolution result processed by the hyperbolic tangent activation function and the second convolution result processed by the sigmoid growth curve activation function are multiplied element-wise to output the feature vector of the node updated by the current temporal convolutional sublayer, thereby capturing the trend and periodic pattern of the performance indicators of each node's corresponding radio frequency unit over time. The gated causal convolutional structure ensures the effective use of historical information while avoiding the leakage of future information, enabling the model to effectively learn the regular fluctuations of performance indicators during busy / idle times and between day and night. The spatiotemporal graph neural network model is trained unsupervised. The training objective function consists of two parts. The first part is the topological attribute mutual information maximization loss, which aims to enable the node feature vectors learned by the spatiotemporal graph neural network model to simultaneously predict the node's own attribute information and the connectivity of its local network. This forces the node representation to encode its own configuration attributes and its neighboring (topological) information, making nodes with similar configurations and topological environments closer in the feature space. The second part is the temporal consistency contrast loss. For a network state snapshot used as an anchor point, temporally adjacent network state snapshots are considered positive samples, while temporally non-adjacent and randomly selected network state snapshots are considered negative samples. This part of the loss aims to make the graph-level representation vectors learned by the spatiotemporal graph neural network model from the anchor point snapshots closer to the graph-level representation vectors of positive sample snapshots in the feature space, while differing more from the graph-level representation vectors of negative sample snapshots. This allows the model to learn the smooth evolution of network states in a short period of time, enhancing the representation's ability to predict normal network states. The graph encoder is capable of characterizing continuous and stable network operation patterns. The graph-level representation vector is obtained by pooling the feature vectors of all nodes in the corresponding network state snapshot. Pooling operations are typically global average pooling or global max pooling. The training objective function is a weighted sum of the loss maximizing topological attribute mutual information and the loss comparing temporal consistency. The weight of the loss maximizing topological attribute mutual information is a first balance coefficient ranging from zero to one, and the weight of the loss comparing temporal consistency is one minus the first balance coefficient. The typical value range of the first balance coefficient is 0.3 to 0.7, for example, 0.5, to achieve a balanced learning of spatial dependency and temporal consistency. The backpropagation algorithm minimizes the training objective function to jointly optimize the two losses, thereby training a graph encoder capable of encoding any network state snapshot into a fixed-dimensional vector representing the health state. A fully trained graph encoder can compress complex network state snapshots into a low-dimensional vector containing configuration, topology, performance, and temporal health patterns, which is the basis for subsequent anomaly measurement. The process of determining a health status representation vector based on network state snapshots from multiple historical time points is as follows: All network state snapshots used in training the graph encoder are input into the trained graph encoder to obtain a health state representation vector corresponding to each network state snapshot, forming a health state representation vector set. The robust estimation algorithm of minimum covariance determinant estimation is used to calculate the health state representation vector set, resulting in a baseline health state vector that is insensitive to potential outliers in the set. This baseline health state representation vector serves as the benchmark for representing the network's health state. Minimum covariance determinant estimation estimates the mean and covariance by finding a subset of data (typically more than half the total number of samples) that minimizes the determinant of the covariance matrix estimated by this subset. The resulting mean vector is highly robust to unlabeled outliers (such as occasional minor performance fluctuations) that may exist in the data. This allows for a purer representation of the network's "health" center. Simultaneously, the covariance matrix of the health status representation vector set is calculated. This covariance matrix quantifies the dispersion of health status distribution within the space spanned by the health status representation vectors and the correlation between dimensions. In subsequent step S3, when calculating the deviation, this covariance matrix is ​​used to standardize the Mahalanobis distance, resulting in a statistically interpretable anomaly score. The preset threshold for deviation can be set based on the statistical distribution of the health status representation vector set. For example, the deviation threshold can be set to the 99th percentile of the corresponding Mahalanobis distance, or a fixed value such as 3.0 can be set empirically. When the Mahalanobis distance between the real-time network's health status representation vector and the health status benchmark vector exceeds this threshold, it is determined that the overall network state may deviate from the health benchmark, requiring further analysis.

[0022] In this embodiment, it is specifically necessary to explain that the process of calculating the deviation between the current network's health status representation vector and the health status baseline vector determined in step S3 is as follows: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2, and output the health state representation vector of the current network; obtain the health state baseline vector and the corresponding covariance matrix calculated based on the network state snapshots at historical time points in step S2; Robust principal component analysis (RPA) is performed on the set of all historical health status representation vectors used to determine the baseline health status vector. The top k principal component directions are extracted to form a projection matrix, where k is a positive integer smaller than the dimension of the health status representation vector. Robust principal component analysis effectively resists the influence of potential outliers in historical data and extracts the most important and robust patterns representing changes in the network's health status. The value of k can be determined by the cumulative variance contribution rate exceeding a preset proportion (e.g., 85%). Using the projection matrix, the difference vector obtained by subtracting the baseline health status vector from the current network's health status representation vector is projected onto the top k principal component directions to obtain a dimension-reduced difference projection vector. This projection operation removes redundant and unimportant dimensions from the original high-dimensional representation space, focusing on key change patterns. For each principal component direction, a robust estimation method is used to calculate the dispersion of the values ​​obtained after projecting all historical health status representation vectors in that direction, which is used as the robust variance for that direction. For example, the median absolute deviation can be used for robust estimation, which is insensitive to outliers and can more stably estimate the normal fluctuation range of health status in the main change directions. Construct a diagonal matrix whose diagonal elements are composed of the robust variances corresponding to the directions of each principal component; calculate the transpose of the dimension-reduced difference projection vector, multiply it by the inverse of the diagonal matrix, and then multiply it by the dimension-reduced difference projection vector itself; finally, perform a square root operation on the multiplication result, and the resulting scalar value is the deviation. The deviation is used to measure the overall degree of deviation of the current network state from the historical health state benchmark vector; the deviation is essentially the Mahalanobis distance after robust principal component dimensionality reduction and robust variance standardization. The larger the value, the further the current network state deviates from the historical normal benchmark in the main health change patterns. The preset threshold of the deviation can be set according to the deviation distribution calculated from the set of historical health state characterization vectors. For example, the 99th percentile of the historical deviation value can be taken as the threshold. When the deviation calculated in real time exceeds this threshold, it is determined that the network as a whole is in an abnormal state and subsequent localization analysis needs to be initiated. The specific process of identifying radio frequency (RF) units contained in subgraphs with structure-attribute consistency scores below the consistency threshold as having implicit configuration conflicts and forming an anomalous unit group is as follows: When the deviation exceeds a preset threshold, a community detection algorithm based on modularity optimization is used to divide the real-time dynamic configuration-performance graph into multiple subgraphs. Modularity optimization ensures that the connections between radio frequency unit nodes within each subgraph are tight, while the connections between different subgraphs are sparse. This division physically corresponds to dividing the network into multiple relatively independent radio frequency unit groups with frequent internal interactions, which helps to narrow the problem localization scope from the entire network to a specific local area. Each subgraph contains a subset of multiple radio frequency unit nodes and all edges between nodes within that subset. For each subgraph, its structure-attribute consistency score is calculated. This score is determined by a weighted sum of three components: configuration conflict, performance anomaly, and topology rationality. The sum of the first, second, and third weighting coefficients is equal to one. These three weighting coefficients can be determined based on network operation and maintenance experience or meshing parameter tuning; for example, they can be set to 0.4, 0.4, and 0.2 respectively to balance the impact of configuration, performance, and topology factors. The configuration conflict component quantifies whether there are conflicts or inconsistencies in the configuration parameters between radio frequency units within the subgraph. Its calculation process is as follows: the configuration conflict component is the sum of one minus a penalty term, where the sum of the penalty terms is calculated by... The first sub-weight is obtained by multiplying the ratio of the number of physical cell identifier collisions to the maximum possible number of collisions, the second sub-weight is multiplied by the overlap coverage violation ratio, and the third sub-weight is multiplied by the power setting dispersion. The number of physical cell identifier collisions refers to the number of paired radio frequency units with physical cell identifier collisions within the sub-map. The maximum possible number of collisions is calculated based on the total number of nodes in the radio frequency unit node set of the sub-map, specifically by multiplying the total number of nodes by (total number of nodes minus one) and then dividing by two. The overlap coverage violation ratio refers to the proportion of links that are judged to have excessive signal strength based on the reference signal received power, resulting in frequent ping-pong handovers, out of the total number of links. The power setting dispersion is used to measure the degree of dispersion of the transmit power settings of each radio frequency unit within the sub-map. The closer the configuration collision component value is to 1, the better the configuration coordination within the sub-map. The performance anomaly component is used to quantify the degree of degradation of key performance indicators (KPIs) of radio frequency units and connection edges within a subgraph relative to their historical normal levels. The calculation process is as follows: For each radio frequency unit node within the subgraph, calculate the standard score of its current KPI relative to the historical baseline value obtained from network state snapshot data at historical time points. Take a specified high percentile quantile of the absolute values ​​of the standard scores of all nodes within the subgraph as the performance anomaly aggregate value for that subgraph. For example, the 75th percentile can be chosen to reflect the degree of anomaly of the poorly performing nodes within the subgraph, avoiding interference from individual extreme values. The negative of the performance anomaly aggregate value is taken as the performance anomaly component. The closer the performance anomaly component value is to 0 (i.e., the smaller the absolute value), the closer the overall performance of the subgraph is to its historical normal level. The topology rationality component is used to quantify the degree of matching between the geographical distribution concentration of radio frequency units (RF units) within a subgraph and the strength of their connectivity relationships. Its calculation process is as follows: the reciprocal of the standard deviation of the latitude and longitude coordinates of all RF unit nodes within the subgraph is taken as the geographical concentration. A higher geographical concentration indicates a more compact geographical distribution of nodes. The average value of the comprehensive interference intensity assessment of all undirected edges within the subgraph is calculated as the average observed interference intensity. Based on the geographical distance between each pair of nodes within the subgraph, a propagation model is used to calculate the predicted interference intensity. The absolute value of the correlation coefficient between the average observed interference intensity and the predicted interference intensity sequence is calculated as the edge weight support. A higher edge weight support indicates a better match between the observed interference relationship and the geographically-based expectation, and a more rational topology. The topology rationality component is obtained by multiplying the geographical concentration and the edge weight support. A higher topology rationality component value indicates a better match between the geographical distribution of nodes within the subgraph and the observed interference relationship. A consistency score threshold is set, and the method for determining the consistency score threshold is as follows: During the model training phase, using the dynamic configuration-performance graph of the historical normal period, a large number of subgraphs are calculated based on the same method to obtain the structure-attribute consistency scores, and a specified low quantile of these structure-attribute consistency scores is taken as the consistency score threshold; the specified low quantile can be, for example, the 5th percentile, which means that the scores corresponding to the 5% of subgraphs with the lowest structure-attribute consistency scores are used as the threshold to filter out the abnormal subgraphs with the worst consistency during the detection phase; subgraphs whose calculated structure-attribute consistency scores are lower than the consistency score threshold are judged as abnormal subgraphs; the sets of radio frequency unit nodes contained in all abnormal subgraphs are merged to form an abnormal unit group; this abnormal unit group accurately indicates the set of suspected radio frequency units in the network that have degraded performance due to configuration inconsistency, providing a clear target for targeted repair in step S4.

[0023] In this embodiment, the specific process of generating at least one set of candidate configuration repair schemes with different adjustment dimensions and magnitudes and performing parallel simulations in step S4, for the abnormal unit group determined in step S3, is as follows: For each radio frequency unit in the abnormal unit group, analyze at least one main reason that leads to a low structure-attribute consistency score in step S3. The main reasons include at least physical cell identifier conflicts, improper transmit power settings, improper antenna tilt settings, or omissions in neighbor cell relationship configuration. By performing attribution analysis on each radio frequency unit in the abnormal unit group, the specific configuration dimension most likely to cause performance degradation can be identified from numerous possible causes, making the subsequently generated repair solutions more targeted and avoiding blind adjustments. Based on the main reason analysis results, a set of candidate configuration repair solutions is automatically generated, where each candidate configuration repair solution explicitly specifies the adjustment value of one or more configuration parameters for one or more radio frequency units in the abnormal unit group. The adjustment values ​​of the configuration parameters are generated according to preset adjustment rules and security boundaries, specifically including three generation strategies: for explicit configuration parameter conflicts such as physical cell identifier conflicts, the preset conflict resolution rule base is directly applied to generate correction values; the preset conflict resolution rule base contains standardized conflict resolution logic, such as reassigning a current geographic area to the conflicting radio frequency unit. Unused physical cell identifiers within the domain; for continuously adjustable configuration parameters such as transmit power and antenna tilt, starting from the current value of the configuration parameter, multiple different adjustment candidate values ​​are generated in both positive and negative directions within a preset safe adjustment range and according to a preset step size; the preset safe adjustment range is set based on the device hardware capabilities and network planning principles, for example, the transmit power adjustment range is set to ±3dB, and the downtilt adjustment range is set to ±2 degrees; the preset step size is set according to the parameter sensitivity, for example, the power step size can be 0.5dB, and the tilt step size can be 0.5 degrees; for adjustments involving multiple radio frequency units or multiple configuration parameters within an abnormal unit group, a conflict correlation graph is constructed based on the conflict correlation relationships between units within the abnormal unit group, and a joint adjustment scheme is generated using a heuristic search strategy based on the conflict correlation graph. The heuristic search strategy prioritizes adjusting radio frequency units with a large number of connected edges in the conflict correlation graph and located in areas with dense conflict relationships, or prioritizes adjusting configuration parameters that have the greatest impact on the configuration conflict component in the structure-attribute consistency score; this strategy aims to resolve the most critical conflicts with the fewest adjustment actions, thereby improving repair efficiency; In a digital twin network environment, based on the dynamic configuration-performance graph at the current moment, a local network simulation model is constructed, which includes an anomalous unit group, first-order neighbor nodes directly connected to nodes in the anomalous unit group via edges, and second-order neighbor nodes directly connected to the first-order neighbor nodes via edges. The inclusion of first-order and second-order neighbor nodes is to fully simulate the chain reaction of configuration adjustments on the surrounding network and ensure the accuracy of the simulation evaluation. This local network simulation model integrates wireless propagation model, network service traffic model, and communication protocol stack simulation functions to simulate the network operation state after the radio frequency unit configuration adjustment. Each of at least one set of candidate configuration repair schemes is independently injected into the corresponding local network simulation model in the digital twin network environment, and each local network simulation model is driven in parallel to run for a preset simulation duration. This simulation duration is sufficient to cover a complete short cycle of network services and observe the initial stable state of network performance after configuration adjustment. The preset simulation duration is to simulate the actual network operation time of 15 to 30 minutes. During the simulation, virtual wireless measurement reports and virtual key performance indicator data generated by the simulation are collected. Through parallel simulation, the potential effects of a large number of candidate schemes can be evaluated efficiently and synchronously. The graph encoder obtained in step S2 is used to evaluate the health recovery effect of the network state after simulation, and the recovery effect evaluation value corresponding to each candidate configuration repair scheme is obtained. The candidate configuration repair scheme with the best recovery effect evaluation value is selected and, after security verification, is sent to the corresponding radio frequency unit in the real network. The specific execution process is as follows: For each candidate configuration repair scheme, based on the simulated wireless measurement report and key performance index data collected after the simulation run, a simulated dynamic configuration-performance graph is reconstructed according to the method in step S1. The simulated dynamic configuration-performance graph is input into the graph encoder obtained in step S2 to obtain a simulated health state representation vector. The Euclidean distance between the simulated health state representation vector and the health state baseline vector determined in step S2 is calculated and denoted as the new distance. The Euclidean distance between the current network health state representation vector and the health state baseline vector before the repair is performed is calculated and denoted as the original distance. The ratio of the new distance to the original distance is calculated and the result is used as the basic recovery score of the candidate configuration repair scheme. The basic recovery score directly quantifies the proportion by which the repair scheme pulls the network health state back to the historical baseline. The closer the score is to 1, the more ideal the recovery effect. Simultaneously, a penalty term is calculated to penalize excessive adjustments in the scope and magnitude of the remediation plan. This penalty term is the sum of the first, second, and third penalty weights multiplied by three penalty factors. Typical values ​​for the first, second, and third penalty weights are between 0.1 and 0.5, and the specific values ​​can be configured based on tolerance for the adjustment scope, magnitude, and impact propagation. The three penalty factors are: the proportion of the number of adjusted radio frequency units to the total number of radio frequency units in the abnormal unit group, the sum of the absolute values ​​of the total adjustment magnitude of the adjusted configuration parameters, and an indicator value. This indicator value is set to one when the remediation plan involves adjusting radio frequency units outside the abnormal unit group, and zero otherwise. The introduction of the penalty term aims to encourage precise and minimal remediation, avoiding the potential risks to network stability caused by over-adjustment. The recovery performance evaluation value of the candidate configuration repair scheme is obtained by subtracting a scheme penalty term weighted by a penalty coefficient from the basic recovery score. The penalty coefficient is used to balance the recovery performance and the conservatism of the repair actions, and its typical value can be selected between 0.1 and 0.3. The scheme with the highest recovery performance evaluation value among all candidate configuration repair schemes is selected as the optimal candidate configuration repair scheme. Before the optimal candidate configuration repair scheme is distributed to the real network, it is subjected to security verification, which includes: boundary verification, confirming that all parameter adjustment values ​​in the optimal candidate configuration repair scheme are within the limits allowed by the device hardware and the boundaries allowed by network security specifications; the limits allowed by the device hardware include maximum transmit power, antenna mechanical adjustment range, etc.; consistency verification, verifying that the new configuration in the optimal candidate configuration repair scheme will not introduce new global parameter conflicts at the network management level; for example, verifying that the newly allocated physical cell identifier is consistent with the entire network. The system ensures that the power / tilt settings are unique within the management domain or conform to the regional coverage plan; and it prepares for rollback by generating and saving a snapshot of the current configuration of all relevant RF units before execution, and pre-setting a rollback plan. Rollback preparation is an important measure to ensure the security of the automatic recovery process, ensuring that the original state can be quickly restored in case of unforeseen problems after the repair. After the security verification is passed, the optimal candidate configuration repair plan is converted into specific configuration commands, which are automatically distributed in batches to the corresponding RF units in the real network through the standard network configuration management interface for configuration update. After the configuration update is executed, the system returns to step S1 to re-collect data to update the dynamic configuration-performance graph, and then proceeds to subsequent steps to verify the configuration recovery effect and network status, thus forming a closed-loop automatic recovery process of detection, location, decision-making, execution and verification. This closed-loop mechanism ensures the effectiveness and adaptability of automatic recovery and can continuously respond to network changes and new configuration problems.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automatic configuration recovery of a communication radio frequency unit, characterized in that, Specifically, the steps include the following: Step S1: Collect RF unit configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data from the network management system, performance management system, and terminal measurement reports. Standardize and spatiotemporally align the collected configuration parameters, engineering parameters, KPIs, and MR data. Based on the processed configuration parameters, engineering parameters, KPIs, and MR data, construct a dynamic configuration-performance graph with RF units as nodes, switching and interference relationships between RF units as edges, and configuration parameters, KPIs, and MR data as attributes. Step S2: Obtain network state snapshots at multiple historical time points from the dynamic configuration-performance graph, use the spatiotemporal graph neural network model to perform unsupervised training on the multiple network state snapshots, obtain a graph encoder through training, and determine a health state representation vector that represents the benchmark of network health state based on the network state snapshots at multiple historical time points. Step S3: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2 to obtain the health status representation vector of the current network. Calculate the deviation between the current network's health status representation vector and the health status benchmark vector determined in step S2. When the deviation exceeds a preset threshold, divide the dynamic configuration-performance graph into subgraphs based on the community detection algorithm. Calculate the structure-attribute consistency score of each subgraph. Identify the radio frequency units (RF units) contained in subgraphs with structure-attribute consistency scores lower than the consistency threshold as having implicit configuration conflicts, forming an abnormal unit group. The specific process is as follows: When the deviation exceeds a preset threshold, a community detection algorithm based on modularity optimization is used to divide the real-time dynamic configuration-performance graph into multiple subgraphs. Each subgraph contains a subset of multiple radio frequency unit nodes and all edges between nodes within that subset. For each subgraph obtained by partitioning, its structure-attribute consistency score is calculated. The structure-attribute consistency score is determined by the weighted sum of configuration conflict component, performance anomaly component, and topology rationality component. A consistency score threshold is set, and the consistency score threshold is determined as follows: During the model training phase, the dynamic configuration-performance graph of the historical normal period is used to calculate the structure-attribute consistency scores of a large number of subgraphs based on the same method, and a specified low quantile of these structure-attribute consistency scores is taken as the consistency score threshold; subgraphs whose calculated structure-attribute consistency scores are lower than the consistency score threshold are judged as abnormal subgraphs; the sets of radio frequency unit nodes contained in all abnormal subgraphs are merged to form an abnormal unit group; Step S4: For the abnormal unit group identified in Step S3, generate at least one set of candidate configuration repair schemes with different adjustment dimensions and magnitudes in the digital twin network environment and perform parallel simulation. Use the graph encoder obtained in Step S2 to evaluate the health recovery effect of the network state after simulation, obtain the recovery effect evaluation value corresponding to each candidate configuration repair scheme, select the candidate configuration repair scheme with the best recovery effect evaluation value, and send it to the corresponding radio frequency unit in the real network for execution after security verification to complete the automatic recovery of configuration.

2. The automatic configuration recovery method for a communication radio frequency unit according to claim 1, characterized in that: In step S1, the configuration parameters collected from the network management system include at least the unique identifier of each radio frequency unit, physical cell identifier, frequency point, transmit power, antenna azimuth angle, downtilt angle, and neighbor cell relationship list. The engineering parameters collected from the network management system include at least the latitude, longitude, and altitude of each radio frequency unit; The key performance indicators (KPIs) data collected from the performance management system include at least the cell throughput, wireless connection rate, handover success rate, and drop rate of each radio frequency unit within a preset time window. The wireless measurement report (MR) data extracted from the terminal measurement report includes at least the reference signal received power and signal-to-interference-plus-noise ratio of the serving cell and all neighboring cells reported by the terminal.

3. The automatic configuration recovery method for a communication radio frequency unit according to claim 2, characterized in that: The specific operations for standardizing and spatiotemporally aligning the collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are as follows: The collected configuration parameters, engineering parameters, key performance indicators (KPIs), and wireless measurement report (MR) data are assigned a unified global timestamp and matched and integrated based on the unique identifier of the radio frequency unit, so that various types of data belonging to the same radio frequency unit and within the same time window form an associated data set.

4. The automatic configuration recovery method for a communication radio frequency unit according to claim 3, characterized in that: The process of constructing a dynamic configuration-performance graph, with radio frequency units as nodes, switching and interference relationships between radio frequency units as edges, and configuration parameters, key performance indicators (KPIs), and wireless measurement report (MR) data as attributes, is as follows: Radio frequency units are used as nodes in the dynamic configuration-performance graph. The attributes of each node consist of all the configuration parameters and engineering parameters of the corresponding radio frequency unit. According to the neighbor cell relationship list collected from the network management system, if the neighbor cell relationship list of the first radio frequency unit contains the unique identifier of the second radio frequency unit, then a directed edge representing the switching relationship from the first radio frequency unit to the second radio frequency unit is established between the two nodes representing the first radio frequency unit and the second radio frequency unit. For any two radio frequency units, based on the associated data set, within a preset time window, all measurement records of the reference signal received power of these two radio frequency units simultaneously reported by all terminals are selected from the wireless measurement report (MR) to form a measurement record set. For each measurement record in the set of measurement records, perform the following calculation: First, the value obtained by subtracting the reference signal received power value of the first radio frequency unit from the reference signal received power value of the second radio frequency unit and adding an intensity offset constant is used as the basis interference intensity component for the current measurement record. Secondly, calculate the absolute value of the difference between the signal-to-interference-plus-noise ratio of the terminal on the first radio frequency unit and a preset target signal quality value; then, divide the absolute value by a quality sensitivity adjustment factor and take the negative of the resulting quotient. Next, the value of the exponential function with the natural constant as the base and negative numbers as the exponent is calculated and used as the signal quality sensitivity weight of the current measurement record; Finally, the basic interference intensity component of the current measurement record is multiplied by the signal quality sensitivity weight to obtain the weighted interference contribution value of the current measurement record; After obtaining the weighted interference contribution values ​​of all measurement records in the measurement record set, these weighted interference contribution values ​​are summed and then divided by the total number of measurement records in the measurement record set. The result is the comprehensive interference intensity assessment value between the first radio frequency unit and the second radio frequency unit. If the comprehensive interference strength assessment value exceeds the preset interference threshold, an undirected edge representing the interference relationship is established between the two nodes representing the first radio frequency unit and the second radio frequency unit, and the comprehensive interference strength assessment value is used as the weight attribute of this undirected edge; for each node, the key performance indicator (KPI) quantile obtained statistically within the preset time window is attached as the dynamic performance attribute of the node. For each directed edge representing a switching relationship, the switching success rate statistically obtained within a preset time window is attached as a dynamic performance attribute of the directed edge. The dynamic performance attributes of nodes and directed edges are updated as a preset time window slides, thereby constructing a dynamic configuration-performance graph.

5. The automatic configuration recovery method for a communication radio frequency unit according to claim 4, characterized in that: In step S2, the specific operation of obtaining a graph encoder by unsupervised training of multiple network state snapshots using a spatiotemporal graph neural network model is as follows: From the dynamic configuration-performance graph constructed in step S1, a series of network state snapshots are extracted at fixed sampling intervals within a pre-selected historical time period to form a set of network state snapshots for training. Each network state snapshot contains nodes, directed edges, and undirected edges that make up the dynamic configuration-performance graph, as well as the attributes of the nodes and edges. The attributes of the nodes include configuration parameters, engineering parameters, and key performance indicators (KPIs) quantiles collected from the network management system and performance management system. The attributes of the edges include the comprehensive interference intensity assessment value calculated based on the wireless measurement report (MR) as the weight of the undirected edge and the handover success rate as the dynamic performance attribute of the directed edge. Nodes, directed edges, undirected edges, and their respective attributes together constitute a network state snapshot. Multiple network state snapshots are arranged in chronological order to form a training dataset. The spatiotemporal graph neural network model is composed of multiple spatiotemporal convolutional layers stacked alternately, where each spatiotemporal convolutional layer contains a spatial graph convolutional sub-layer and a temporal convolutional layer; The spatiotemporal graph neural network model is trained in an unsupervised manner. The training objective function consists of two parts: the first part is the loss for maximizing mutual information of topological attributes; the second part is the temporal consistency comparison loss. For a network state snapshot that serves as an anchor point, network state snapshots that are adjacent in time are considered as positive samples, and network state snapshots that are not adjacent in time and are randomly selected are considered as negative samples. The training objective function is a weighted sum of the loss for maximizing mutual information of topological attributes and the loss for comparing temporal consistency. The two losses are jointly optimized by minimizing the training objective function through backpropagation, thereby training a graph encoder that can encode any network state snapshot into a health state representation vector of a fixed dimension.

6. The automatic configuration recovery method for a communication radio frequency unit according to claim 5, characterized in that: The process of determining a health status representation vector based on network state snapshots from multiple historical time points specifically involves: All network state snapshots used in training the graph encoder are input into the trained graph encoder to obtain the health state representation vector corresponding to each network state snapshot, forming a health state representation vector set. The robust estimation algorithm of minimum covariance determinant estimation is used to calculate the health state representation vector set to obtain a health state baseline vector that is insensitive to potential outliers in the set, which serves as the health state representation vector representing the baseline of the network health state. At the same time, the covariance matrix of the health state representation vector set is calculated.

7. The automatic configuration recovery method for a communication radio frequency unit according to claim 6, characterized in that: In step S3, the process of calculating the deviation between the current network's health status representation vector and the health status baseline vector determined in step S2 is as follows: Input the dynamic configuration-performance graph corresponding to the current network state obtained in real time into the graph encoder obtained in step S2, and output the health state representation vector of the current network; obtain the health state baseline vector and the corresponding covariance matrix calculated based on the network state snapshots at historical time points in step S2; Robust principal component analysis is performed on the set of all historical health state representation vectors used to determine the baseline vector of health state, and the directions of the first k principal components are extracted to form a projection matrix. Using a projection matrix, the difference vector obtained by subtracting the baseline health state vector from the current network's health state representation vector is projected onto the first k principal component directions to obtain a dimension-reduced difference projection vector. For each principal component direction, a robust estimation method is used to calculate the dispersion of the values ​​obtained after projecting all historical health state representation vectors in that direction, which is used as the robust variance of that direction. Construct a diagonal matrix whose diagonal elements are composed of the robust variances corresponding to the directions of each principal component; calculate the transpose of the dimension-reduced difference projection vector, multiply it by the inverse of the diagonal matrix, and then multiply it by the dimension-reduced difference projection vector itself; finally, perform a square root operation on the multiplication result, and the resulting scalar value is the deviation.

8. The automatic configuration recovery method for a communication radio frequency unit according to claim 7, characterized in that: In step S4, the specific process of generating at least one set of candidate configuration repair schemes containing different adjustment dimensions and magnitudes and performing parallel simulations for the abnormal unit group identified in step S3 is as follows: For each radio frequency unit in the abnormal unit group, analyze at least one main reason that leads to a low structure-attribute consistency score in step S3. The main reasons include at least physical cell identifier conflict, improper transmit power setting, improper antenna tilt setting, or omission of neighbor cell relationship configuration. Based on the main reason analysis results, automatically generate a set of candidate configuration repair schemes, where each candidate configuration repair scheme explicitly specifies the adjustment value of one or more configuration parameters of one or more radio frequency units in the abnormal unit group. The adjustment value of the configuration parameters is generated according to preset adjustment rules and security boundaries. In a digital twin network environment, based on the dynamic configuration-performance graph at the current moment, a local network simulation model is constructed, which includes an anomalous unit group, first-order neighbor nodes directly connected to nodes in the anomalous unit group through edges, and second-order neighbor nodes directly connected to the first-order neighbor nodes through edges. Each of at least one set of candidate configuration repair schemes is independently injected into the corresponding local network simulation model in the digital twin network environment, and each local network simulation model is driven in parallel to run for a preset simulation duration. During the simulation, virtual wireless measurement reports and virtual key performance indicator data generated by the simulation are collected.

9. The automatic configuration recovery method for a communication radio frequency unit according to claim 8, characterized in that: The process of using the graph encoder obtained in step S2 to evaluate the health recovery effect of the simulated network state, obtaining the recovery effect evaluation value corresponding to each candidate configuration repair scheme, selecting the candidate configuration repair scheme with the best recovery effect evaluation value, and sending it to the corresponding radio frequency unit in the real network after security verification is as follows: For each candidate configuration repair scheme, based on the simulated wireless measurement report and key performance index data collected after the simulation run, a simulated dynamic configuration-performance graph is reconstructed according to the method in step S1; the simulated dynamic configuration-performance graph is input into the graph encoder obtained in step S2 to obtain a simulated health state representation vector; the Euclidean distance between the simulated health state representation vector and the health state reference vector determined in step S2 is calculated and denoted as the new distance; Calculate the Euclidean distance between the current network health status representation vector and the health status baseline vector before the repair is performed, and denote it as the original distance; Calculate the ratio of the new distance to the original distance, and use the result as the basic recovery score for the candidate configuration repair scheme. Simultaneously, calculate a penalty item for schemes that are penalized for excessively large adjustments in the scope and magnitude of the repair scheme; Subtract a penalty term weighted by a penalty coefficient from the basic recovery score, and the result is used as the recovery effect evaluation value of the candidate configuration repair scheme. Select the scheme with the highest recovery effect evaluation value from all candidate configuration repair schemes as the optimal candidate configuration repair scheme. Before sending the optimal candidate configuration repair scheme to the real network, perform a security verification. After the security verification is passed, the optimal candidate configuration repair scheme is converted into a specific configuration command and automatically sent in batches to the corresponding radio frequency units in the real network for configuration update through the standard network configuration management interface. After the configuration update is executed, return to step S1.

Images