Network performance hierarchical evaluation method based on multi-agent system

Abstract

The invention discloses a network performance hierarchical evaluation method based on a multi-agent system, and the method comprises the following steps: carrying out the hierarchical division of a to-be-evaluated wireless communication network, and constructing a hierarchical evaluation system; collecting performance data of the corresponding wireless network object; constructing a hierarchical evaluation graph; performing input embedding processing on each evaluation node in the hierarchical evaluation graph; generating a cross-level semantic distance coding result and a conflict perception bias result; inputting the node representation set, the edge attribute coding result, the cross-level semantic distance coding result and the conflict perception bias result into an improved Grapporter model, and executing multi-head attention calculation; carrying out step-by-step convergence on the evaluation node update representation set by utilizing a hierarchical virtual agent node chain; and correcting the overall performance evaluation result of the network, and generating a resource regulation and control strategy or an operation parameter optimization strategy. According to the method, multi-agent layered evaluation is adopted, and accurate optimization of the wireless network performance is realized.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network performance evaluation and optimization technology, and in particular to a hierarchical evaluation method for network performance based on a multi-agent system. Background Technology

[0002] As wireless communication networks continue to expand, their structures are becoming increasingly multi-layered, heterogeneous, and dynamically changing. Link status, access node load, regional network operation, and overall network performance are all correlated. Current technologies for wireless communication network performance analysis typically employ single-layer indicator statistics, local node monitoring, or comprehensive evaluation methods based on fixed rules. These methods monitor indicators such as latency, throughput, packet loss rate, jitter, resource utilization, channel quality, and handover success rate, and output network operation status results accordingly.

[0003] Existing technologies often lack a unified hierarchical evaluation mechanism when dealing with the relationships, convergence, and feedback between cross-level wireless network objects, resulting in insufficient expression of the coupling between local and global evaluation results. Existing solutions struggle to jointly model performance conflicts, time skewness, and anomaly propagation risks between nodes at different levels, limiting the ability of global performance evaluation results to characterize complex network states and consequently affecting the targeted nature of subsequent resource allocation and operational parameter optimization.

[0004] Therefore, how to provide a hierarchical evaluation method for network performance based on multi-agent systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a hierarchical evaluation method for network performance based on a multi-agent system. This invention comprehensively utilizes multi-agent collaborative perception, hierarchical evaluation graph construction, cross-level semantic distance coding, conflict-aware bias modeling, and improved Graphormer model fusion analysis techniques to uniformly model, hierarchically evaluate, and globally fuse the multi-level performance states of wireless communication networks. It also performs verification and correction in the event of evaluation conflicts, possessing advantages such as strong hierarchical evaluation capability, high global representation accuracy, strong conflict identification capability, and highly targeted resource regulation.

[0006] A hierarchical evaluation method for network performance based on a multi-agent system according to an embodiment of the present invention includes the following steps:

[0007] The wireless communication network to be evaluated is divided into layers, a hierarchical evaluation system is constructed, and multiple evaluation agents are deployed at the corresponding layers to obtain a set of evaluation agents;

[0008] Each evaluation agent in the evaluation agent set collects performance data of the corresponding wireless network object, generating a performance data set and a local evaluation result set.

[0009] A hierarchical evaluation graph is constructed based on the hierarchical evaluation system, the set of evaluation agents, the set of performance data, and the set of local evaluation results.

[0010] Input embedding processing is performed on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results;

[0011] Based on any two evaluation nodes in the hierarchical evaluation graph, generate cross-level semantic distance encoding results, and obtain conflict-aware bias results according to evaluation nodes at different levels.

[0012] The node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are input into the improved Graphormer model, and multi-head attention computation is performed to obtain the evaluation node update representation set.

[0013] By using a hierarchical virtual agent node chain, the updated representation set of the evaluation nodes is aggregated level by level to obtain the global network performance evaluation result;

[0014] The system detects conflicts between the network global performance evaluation results and the evaluation results, corrects the network global performance evaluation results, and generates resource control strategies or operating parameter optimization strategies.

[0015] Optionally, the hierarchical evaluation system includes a wireless link layer, a wireless access node layer, a regional wireless network layer, and a global wireless network layer.

[0016] Optionally, the step of collecting performance data of the corresponding wireless network object by each evaluation agent in the evaluation agent set and generating a performance data set and a local evaluation result set specifically includes:

[0017] The evaluation agents in the overall set of evaluation agents collect performance data on the network objects at the corresponding level according to the corresponding relationship matrix.

[0018] Construct performance data vectors based on the hierarchy and business type of the hierarchical network objects;

[0019] Time alignment and validity filtering are performed on each indicator component in the performance data vector to obtain a standardized collection sequence;

[0020] For each indicator component in the standardized collection sequence, an indicator normalization result is constructed, and for negative performance indicators, a corresponding reverse normalization result is constructed.

[0021] Based on the index weights corresponding to each performance index component, a local evaluation result is generated.

[0022] The performance data vectors, standardized acquisition sequences, and local evaluation results output by each evaluation agent in the overall evaluation agent set are aggregated according to hierarchical sets and hierarchical affiliations to generate performance data sets and local evaluation result sets.

[0023] Optionally, constructing the hierarchical evaluation graph based on the hierarchical evaluation system, the set of evaluation agents, the performance data set, and the set of local evaluation results specifically includes:

[0024] For each evaluation agent in the overall set of evaluation agents, a corresponding evaluation node is established, resulting in a set of evaluation nodes;

[0025] For each level in the hierarchy set, a corresponding virtual proxy node is established to obtain a set of virtual proxy nodes;

[0026] Based on the set of evaluation nodes and the set of virtual agent nodes, a set of nodes for a hierarchical evaluation graph is generated, and a hierarchical index, node type identifier, and object ownership identifier are assigned to each node.

[0027] Based on the set of intra-layer association relationships and the set of inter-layer convergence relationships, we obtain the set of intra-layer association edges and the set of inter-layer convergence edges;

[0028] Based on the result transmission requirements and correction transmission requirements between adjacent levels, establish inter-layer feedback relationships, and establish an inter-layer feedback edge set based on the inter-layer feedback relationships;

[0029] Based on the performance data set and the local evaluation result set, calculate the conflict determination value between each pair of evaluation nodes to obtain the conflict determination result.

[0030] Based on the conflict determination results of each evaluation node pair, determine the set of conflict node pairs, establish conflict verification relationships between each conflict node pair, and establish a set of conflict verification edges.

[0031] Based on the set of intra-layer associated edges, inter-layer convergence edges, proxy convergence edges, inter-layer feedback edges, proxy feedback edges, and conflict verification edges, generate the edge set of the hierarchical evaluation graph.

[0032] Construct a hierarchical evaluation graph based on the set of nodes and the set of edges of the hierarchical evaluation graph.

[0033] Optionally, the step of performing input embedding processing on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results specifically includes:

[0034] Obtain the hierarchical evaluation graph, and perform input embedding processing on each evaluation node in the hierarchical evaluation graph to generate performance evaluation input vector, agent role vector, hierarchical identifier vector, time state vector, wireless network topology relationship vector, and centrality vector;

[0035] Construct a spliced ​​input vector based on the performance evaluation input vector, agent role vector, hierarchy identifier vector, time state vector, wireless network topology vector, and centrality vector;

[0036] Linear mapping is performed on the concatenated input vectors of each evaluation node to obtain the node representation of the corresponding evaluation node, and a set of node representations is generated.

[0037] For each edge in the hierarchical evaluation graph, construct an edge attribute vector based on the edge relationship type, starting node level, target node level, edge direction, edge connection strength, and conflict determination value.

[0038] Linear mapping is performed on the edge attribute vectors of each edge to obtain the edge attribute codes of the corresponding edges, and the edge attribute code results are generated.

[0039] Optionally, the step of generating cross-level semantic distance encoding results based on any two evaluation nodes in the hierarchical evaluation graph, and obtaining conflict-aware bias results according to different levels of evaluation nodes, specifically includes:

[0040] For any two evaluation nodes in the hierarchical evaluation graph, construct a set of node pairs based on the node representation set, edge attribute encoding results, performance data set, and local evaluation result set;

[0041] For any node pair in the set of node pairs, based on the hierarchical evaluation graph, generate the topological association quantity, hierarchical span quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity;

[0042] Based on the topological association quantity, hierarchical cross-level quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity, a cross-level semantic distance value is generated.

[0043] Based on the cross-level semantic distance values, a cross-level semantic distance code is generated, and the cross-level semantic distance codes are aggregated to generate a cross-level semantic distance coding result.

[0044] For any node pair in the node pair set, generate a local performance score difference based on the local performance scores of the two corresponding evaluation nodes in the local evaluation result set at the current time.

[0045] Based on the local performance score sequence of the two corresponding evaluation nodes in the local evaluation result set, a trend correlation quantity is generated, and based on the time state vector of the two corresponding evaluation nodes, a time deviation quantity is generated.

[0046] Based on the difference in local performance scores, the correlation of trends, and the time deviation, conflict degree information is generated. Based on the conflict degree information and the conflict judgment value, a conflict perception bias is generated.

[0047] The conflict degree information and conflict perception bias of each node pair in the node pair set are aggregated to generate conflict perception bias results.

[0048] Optionally, the step of inputting the node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results into the improved Graphormer model, performing multi-head attention computation, and obtaining the evaluation node update representation set specifically includes:

[0049] Input the node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results into the improved Graphormer model to construct the input node feature matrix;

[0050] Construct a relation-enhancing attention unit, and for any node pair, generate a relation-enhancing bias term based on the edge attribute encoding result, the cross-level semantic distance encoding result, and the conflict-aware bias result;

[0051] In the improved Graphormer model, we construct intra-layer local consistency feature extraction paths and inter-layer convergence feature extraction paths to generate intra-layer local consistency masks and inter-layer convergence masks.

[0052] In each coding layer, query mapping, key mapping, and value mapping are performed on the feature matrix of the input node of the current coding layer to obtain the query matrix, key matrix, and value matrix, respectively.

[0053] In the intra-layer local consistency feature extraction path, the intra-layer attention weights are calculated and weighted summation is performed to obtain the intra-layer local consistency features;

[0054] In the inter-layer convergence feature extraction path, the inter-layer attention weights are calculated and weighted summation is performed to obtain the inter-layer convergence features;

[0055] The intra-layer local consistency features output by each attention head are concatenated to obtain the intra-layer local consistency representation;

[0056] The inter-layer convergence features output by each attention head are concatenated to obtain the inter-layer convergence representation;

[0057] In the dual-path fusion update unit, the intra-layer local consistency representation and the inter-layer convergence representation are concatenated and then input into the feature fusion mapping matrix to obtain the fused representation;

[0058] Based on the fusion representation, residual updates and feedforward updates are performed to obtain the output representation of each evaluation node in the current coding layer, and a set of updated representations for the evaluation nodes is generated.

[0059] Optionally, the step of using a hierarchical virtual proxy node chain to aggregate the updated representation sets of evaluation nodes level by level to obtain the global network performance evaluation result specifically includes:

[0060] Obtain the set of evaluation node update representations and extract the hierarchical index corresponding to each evaluation node to obtain a subset of node update representations;

[0061] In the hierarchical virtual agent node chain, for the virtual agent node corresponding to the wireless link layer, the updated representation subset of the node corresponding to the wireless link layer is received and aggregated step by step to obtain the agent aggregation result of each layer.

[0062] For any virtual proxy node at any level, calculate the proxy aggregation weight and obtain the proxy aggregation result for the current level;

[0063] Based on the agent aggregation results corresponding to each level in the hierarchical virtual agent node chain, a global performance representation of the wireless communication network is obtained;

[0064] The global performance representation of the wireless communication network is input into the global evaluation output layer to obtain the global performance evaluation result of the network.

[0065] Optionally, the step of detecting conflicts between the network global performance evaluation results and the evaluation results, correcting the network global performance evaluation results, and generating resource regulation strategies or operating parameter optimization strategies specifically includes:

[0066] Obtain the global network performance evaluation results and the proxy aggregation results corresponding to virtual proxy nodes at each level, and construct the evaluation conflict detection input set;

[0067] Based on the network global performance evaluation results, calculate the evaluation conflict detection value corresponding to each level;

[0068] The existence of an assessment conflict is determined based on the assessment conflict detection value corresponding to each level. When an assessment conflict is determined to exist, the conflict verification process is triggered.

[0069] Based on the hierarchical virtual agent node that generates the evaluation conflict, combined with the conflict perception bias result and the conflict verification edge, the corresponding conflict node pair is located to obtain the set of conflict node pairs.

[0070] For each conflict node pair in the set of conflict node pairs, update the attention weight, edge relationship, and local evaluation result of the corresponding conflict node pair;

[0071] Based on the updated attention weights, updated edge relationships, and updated local evaluation results, the corresponding edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are regenerated, and the hierarchical evaluation graph is reconstructed.

[0072] The updated hierarchical evaluation graph is re-input into the improved Graphormer model to regenerate the corrected global network performance evaluation results.

[0073] Based on the revised global network performance evaluation results, resource regulation strategies or operating parameter optimization strategies for wireless communication networks are generated.

[0074] The beneficial effects of this invention are:

[0075] This invention constructs a hierarchical evaluation system consisting of a wireless link layer, a wireless access node layer, a regional wireless network layer, and a global wireless network layer. Evaluation agents are deployed at each corresponding layer to collaboratively collect and locally evaluate performance data of wireless network objects at different layers, enabling network performance information to be organized and expressed according to a hierarchical structure. This enhances the ability to represent the correlation between local and global network states, resulting in a more complete hierarchy and consistency in the wireless communication network performance evaluation process.

[0076] This invention further constructs a hierarchical evaluation graph comprising evaluation nodes, virtual agent nodes, intra-layer association edges, inter-layer convergence edges, inter-layer feedback edges, and conflict verification edges. It combines performance data, local evaluation results, agent role information, hierarchical identification information, time state information, wireless network topology relationship information, and centrality information to form node representations. Simultaneously, it introduces cross-level semantic distance coding and conflict-aware bias to jointly model the topological associations, hierarchical spans, indicator correlations, time synchronization deviations, local performance differences, and anomaly propagation risks between nodes at different levels. Therefore, it can improve the ability to characterize cross-level coupling relationships, performance conflict relationships, and dynamic changes, enabling the global performance evaluation results to better reflect the actual operating state of the wireless communication network.

[0077] This invention employs an improved Graphormer model for fusion analysis of hierarchical evaluation graphs. Compared to existing models, its improvements lie in incorporating edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results into attention calculation, and dividing feature extraction into intra-layer local consistency feature extraction paths and inter-layer convergence feature extraction paths. Through these improvements, the model can simultaneously represent local relationships between nodes at the same level and convergence relationships between nodes at different levels, enhancing its ability to identify hierarchical span, temporal deviations, and performance conflicts. This, in turn, improves the accuracy of global performance evaluation results for wireless communication networks and the targeted nature of subsequent resource regulation strategy generation. Attached Figure Description

[0078] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0079] Figure 1 This is a flowchart of a hierarchical evaluation method for network performance based on a multi-agent system proposed in this invention;

[0080] Figure 2 This is a schematic diagram illustrating the construction of the hierarchical evaluation graph in a hierarchical evaluation method for network performance based on a multi-agent system proposed in this invention.

[0081] Figure 3 This is a schematic diagram of the improved Graphormer model in the hierarchical evaluation method for network performance based on multi-agent systems proposed in this invention. Detailed Implementation

[0082] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0083] refer to Figures 1-3 A hierarchical evaluation method for network performance based on multi-agent systems includes the following steps:

[0084] The wireless communication network to be evaluated is divided into layers, a hierarchical evaluation system is constructed, and multiple evaluation agents are deployed at the corresponding layers to obtain a set of evaluation agents;

[0085] Each evaluation agent in the evaluation agent set collects performance data of the corresponding wireless network object, generating a performance data set and a local evaluation result set.

[0086] A hierarchical evaluation graph is constructed based on the hierarchical evaluation system, the set of evaluation agents, the set of performance data, and the set of local evaluation results.

[0087] Input embedding processing is performed on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results;

[0088] Based on any two evaluation nodes in the hierarchical evaluation graph, generate cross-level semantic distance encoding results, and obtain conflict-aware bias results according to evaluation nodes at different levels.

[0089] The node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are input into the improved Graphormer model, and multi-head attention computation is performed to obtain the evaluation node update representation set.

[0090] By using a hierarchical virtual agent node chain, the updated representation set of the evaluation nodes is aggregated level by level to obtain the global network performance evaluation result;

[0091] The system detects conflicts between the network global performance evaluation results and the evaluation results, corrects the network global performance evaluation results, and generates resource control strategies or operating parameter optimization strategies.

[0092] In this embodiment, the wireless communication network to be evaluated is hierarchically divided to construct a hierarchical evaluation system, and multiple evaluation agents are deployed at the corresponding levels. The resulting set of evaluation agents specifically includes:

[0093] Obtain network topology data, device connection data, regional coverage data, and service bearer data of the wireless communication network to be evaluated. Based on the link connection relationship, node affiliation relationship, and regional aggregation relationship in the network topology data, perform hierarchical structure analysis on the wireless communication network to be evaluated and construct a basic hierarchical structure.

[0094] Based on the basic hierarchical structure, hierarchical mapping is performed on the network objects in the wireless communication network to be evaluated to generate a hierarchical set;

[0095] The hierarchical mapping includes: mapping the network object corresponding to the wireless transmission link to the wireless link layer, mapping the network object corresponding to the wireless access device to the wireless access node layer, mapping the area network object consisting of multiple wireless access devices and their coverage area to the area wireless network layer, and mapping the overall network object consisting of multiple area network objects to the global wireless network layer.

[0096] The hierarchical set includes the wireless link layer node set, the wireless access node layer node set, the regional wireless network layer node set, and the global wireless network layer node set;

[0097] Based on the hierarchical set, establish the hierarchical affiliation relationship of each network object in the wireless communication network to be evaluated, and establish intra-layer association relationship based on the connection relationship, coordination relationship and service association relationship between network objects in the same layer, and establish inter-layer aggregation relationship based on the merging relationship, convergence relationship and ownership relationship between lower-level network objects and upper-level network objects in adjacent layers.

[0098] The number of evaluation agents is configured according to the network objects at each level in the hierarchical set. Specifically, the number of evaluation agents deployed at each level is determined based on the number of wireless transmission links in the wireless link layer node set, the number of wireless access devices in the wireless access node layer node set, the number of regional network units in the regional wireless network layer node set, and the number of global network management objects in the global wireless network layer node set.

[0099] Based on the monitoring scope, performance index type, and management granularity of network objects at each level, the evaluation agents are deployed at the wireless link layer, wireless access node layer, regional wireless network layer, and global wireless network layer to form an evaluation agent set.

[0100] Establish a correspondence matrix between evaluation agents and hierarchical network objects, and determine the hierarchical identifier, monitoring scope, and associated objects of each evaluation agent based on the correspondence matrix;

[0101] Based on the hierarchical set, hierarchical affiliation, intra-layer association, inter-layer convergence, evaluation agent set, and corresponding relationship matrix, a hierarchical evaluation system is generated, including the wireless link layer, wireless access node layer, regional wireless network layer, and global wireless network layer.

[0102] In this embodiment, the process of each evaluation agent in the evaluation agent set collecting performance data of the corresponding wireless network object and generating a performance data set and a local evaluation result set specifically includes:

[0103] Each evaluation agent in the overall set of evaluation agents performs performance data collection on the corresponding network objects according to the correspondence matrix; wherein, when the correspondence matrix indicates that there is a correspondence between a certain evaluation agent and a certain network object, the evaluation agent collects the performance data vector of the network object at the current collection time.

[0104] Based on the hierarchical network object's hierarchy and business type, construct a corresponding set of performance indicator components to obtain a performance data vector.

[0105] The performance data vector should include at least latency, throughput, packet loss rate, jitter, bandwidth utilization, wireless retransmission rate, resource block utilization, channel quality, interference intensity, access success rate, handover success rate, and resource utilization.

[0106] Among them, each component in the performance data vector corresponds to the collected values ​​of latency, throughput, packet loss rate, jitter, bandwidth utilization, wireless retransmission rate, resource block utilization, channel quality, interference intensity, access success rate, handover success rate, and resource utilization at the current collection time.

[0107] Time alignment and validity filtering are performed on each indicator component in the performance data vector to obtain a standardized collection sequence;

[0108] The standardized acquisition sequence is formed by arranging the performance data vectors corresponding to each sampling moment within the acquisition time window in chronological order. The number of samplings within the acquisition time window is a preset number of samplings, and each sampling moment is a sampling moment arranged in chronological order.

[0109] For each indicator component in the standardized collection sequence, construct the indicator normalization result;

[0110] The normalized result of any performance index component at any sampling time is the result obtained by subtracting the minimum value of the performance index component within the preset statistical range from the sampled value of the performance index component at the sampling time, and then dividing by the difference between the maximum and minimum values ​​of the performance index component within the preset statistical range.

[0111] For negative performance metrics, construct the corresponding reverse normalization result, which is one minus the metric normalization result;

[0112] Among them, negative performance indicators include at least one or more of the following: latency, packet loss rate, jitter, bandwidth utilization, wireless retransmission rate, and interference intensity.

[0113] Based on the weights of each performance index component, the local performance score of the evaluation agent for the hierarchical network object at each sampling time is generated.

[0114] The local performance score is the sum of the product of the normalized or inverse normalized result of each performance index component and the corresponding index weight; where the total number of performance indexes participating in the local evaluation is the total number of preset performance indexes, and the sum of the index weights corresponding to each performance index component is one.

[0115] Generate local evaluation results for the evaluation agent based on local performance scores;

[0116] The local evaluation results include at least a local performance score, a local performance level, and an abnormal state marker; wherein, when the local performance score is less than the first preset performance level classification threshold, the local performance level is determined to be the first level.

[0117] When the local performance score is greater than or equal to the first preset performance level threshold and less than the second preset performance level threshold, the local performance level is determined to be the second level.

[0118] When the local performance score is greater than or equal to the second preset performance level threshold and less than the third preset performance level threshold, the local performance level is determined to be the third level.

[0119] When the local performance score is greater than or equal to the third preset performance level threshold, the local performance level is determined to be the fourth level.

[0120] Among them, the first preset performance level classification threshold, the second preset performance level classification threshold, and the third preset performance level classification threshold satisfy a sequentially increasing relationship; when the normalized result or reverse normalized result corresponding to any performance index component is less than the anomaly judgment threshold corresponding to that performance index component, the anomaly state is determined to be marked as one; otherwise, the anomaly state is determined to be marked as zero.

[0121] The performance data vectors, standardized acquisition sequences, and local evaluation results output by each evaluation agent in the overall evaluation agent set are collected according to hierarchical set and hierarchical belonging relationship to generate performance data set and local evaluation result set;

[0122] The performance data set records the performance data corresponding to each layer of network objects, while the local evaluation result set records the local evaluation results corresponding to each layer of network objects.

[0123] In this embodiment, constructing the hierarchical evaluation graph based on the hierarchical evaluation system, the set of evaluation agents, the performance data set, and the set of local evaluation results specifically includes:

[0124] For each evaluation agent in the overall set of evaluation agents, a corresponding evaluation node is established to obtain a set of evaluation nodes. Each evaluation node corresponds one-to-one with each evaluation agent. Each evaluation node is used to carry the hierarchical identifier, monitoring range, performance data and local evaluation results of the corresponding evaluation agent.

[0125] For each layer in the layer set, a corresponding virtual proxy node is established to obtain a set of virtual proxy nodes. Among them, the wireless link layer corresponds to the virtual proxy node of the wireless link layer, the wireless access node layer corresponds to the virtual proxy node of the wireless access node layer, the regional wireless network layer corresponds to the virtual proxy node of the regional wireless network layer, and the global wireless network layer corresponds to the virtual proxy node of the global wireless network layer.

[0126] Based on the evaluation node set and the virtual agent node set, a node set for the hierarchical evaluation graph is generated. Based on the hierarchical affiliation relationship, the correspondence matrix, the performance data set, and the local evaluation result set, a corresponding hierarchical index, node type identifier, and object affiliation identifier are assigned to each node in the node set. The node type identifier is used to distinguish between evaluation nodes and virtual agent nodes, and the object affiliation identifier is used to indicate the hierarchical network object affiliation relationship corresponding to each node.

[0127] Based on the set of intra-layer association relationships, the intra-layer association edges between each evaluation node within the same layer are determined, resulting in a set of intra-layer association edges. Among them, when two evaluation nodes correspond to the same level index and there is a same-level association relationship between the two evaluation nodes, an intra-layer association edge is established between the two evaluation nodes.

[0128] When two evaluation nodes have the same hierarchical index and there is no hierarchical association between the two evaluation nodes, no intra-level association edge is established between the two evaluation nodes.

[0129] Based on the set of inter-layer convergence relationships, inter-layer convergence edges between adjacent layers are determined, resulting in a set of inter-layer convergence edges. When the layer corresponding to the lower evaluation node is adjacent to the layer corresponding to the upper evaluation node, and the lower network object corresponding to the lower evaluation node has a convergence relationship with the upper network object corresponding to the upper evaluation node, an inter-layer convergence edge is established between the lower evaluation node and the upper evaluation node.

[0130] When the aggregation relationship does not exist, no inter-layer aggregation edge is established; and a set of proxy aggregation edges is established according to the aggregation relationship between the evaluation node of each level and the corresponding virtual proxy node, wherein each evaluation node in each level establishes a proxy aggregation edge with the virtual proxy node corresponding to its level.

[0131] Based on the result transmission requirements and correction transmission requirements between adjacent levels, an inter-layer feedback relationship is established, and an inter-layer feedback edge set is established based on the inter-layer feedback relationship. When an upper-level evaluation node needs to transmit evaluation results or correction information to a lower-level evaluation node, an inter-layer feedback edge is established between the upper-level evaluation node and the lower-level evaluation node.

[0132] When the upper-layer evaluation node does not need to pass evaluation results or correction information to the lower-layer evaluation node, no inter-layer feedback edge is established; and according to the hierarchical feedback requirements between virtual agent nodes at each level, a set of agent feedback edges is established, wherein the virtual agent node corresponding to the global wireless network layer establishes a feedback edge to the virtual agent node corresponding to the regional wireless network layer, the virtual agent node corresponding to the regional wireless network layer establishes a feedback edge to the virtual agent node corresponding to the wireless access node layer, and the virtual agent node corresponding to the wireless access node layer establishes a feedback edge to the virtual agent node corresponding to the wireless link layer.

[0133] Based on the performance data set and the local evaluation result set, the conflict judgment value between each pair of evaluation nodes is calculated to obtain the conflict judgment result. The conflict judgment value is obtained by weighted summation of the local performance score difference item, the local performance level difference item, and the abnormal state mark difference item. The local performance score difference item is the absolute value of the difference between the local performance scores of the two corresponding evaluation nodes at the same time. The local performance level difference item is the absolute value of the difference between the local performance levels of the two corresponding evaluation nodes at the same time. The abnormal state mark difference item is the absolute value of the difference between the abnormal state marks of the two corresponding evaluation nodes at the same time.

[0134] The local performance score difference item corresponds to the first conflict judgment weight, the local performance level difference item corresponds to the second conflict judgment weight, and the abnormal state mark difference item corresponds to the third conflict judgment weight. The sum of the first conflict judgment weight, the second conflict judgment weight, and the third conflict judgment weight is one.

[0135] When the conflict determination value is greater than or equal to the conflict determination threshold, it is determined that the two corresponding evaluation nodes meet the conflict determination conditions; when the conflict determination value is less than the conflict determination threshold, it is determined that the two corresponding evaluation nodes do not meet the conflict determination conditions.

[0136] Based on the conflict determination results of each evaluation node pair, determine the set of conflict node pairs and establish conflict verification relationships between each conflict node pair;

[0137] A set of conflict verification edges is established based on the set of conflict node pairs and the conflict verification relationship. When two evaluation nodes meet the conflict determination conditions, the two evaluation nodes are determined to constitute a conflict node pair, and a conflict verification edge is established between the two evaluation nodes. When the two evaluation nodes do not meet the conflict determination conditions, the two evaluation nodes are determined not to constitute a conflict node pair, and no conflict verification edge is established.

[0138] Based on the set of intra-layer associated edges, inter-layer convergence edges, proxy convergence edges, inter-layer feedback edges, proxy feedback edges, and conflict verification edges, generate the edge set of the hierarchical evaluation graph.

[0139] Based on the node set and edge set of the hierarchical evaluation graph, a hierarchical evaluation graph is constructed. The nodes in the hierarchical evaluation graph include the evaluation nodes corresponding to the evaluation agents at each level and the virtual agent nodes corresponding to each level. The edges in the hierarchical evaluation graph include intra-layer association edges, inter-layer convergence edges, inter-layer feedback edges, and conflict verification edges.

[0140] In this embodiment, the input embedding process is performed on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results, specifically including:

[0141] Obtain the performance data, local evaluation results, agent role information, hierarchical identification information, time status information, wireless network topology relationship information, and centrality information corresponding to each evaluation node in the hierarchical evaluation graph;

[0142] Based on the performance data set and the local evaluation result set, the performance statistical features and local evaluation features corresponding to each evaluation node are extracted to generate a performance evaluation input vector. The performance evaluation input vector includes the statistical values ​​of each performance indicator within a preset time window, the statistical values ​​of local performance scores, the statistical values ​​of local performance levels, and the statistical values ​​of abnormal state markers. Among them, the statistical values ​​of each performance indicator correspond to the performance indicator items embedded in the input of the participating nodes, the statistical values ​​of local performance scores correspond to the statistical results of local performance scores of the evaluation nodes within the preset time window, the statistical values ​​of local performance levels correspond to the statistical results of local performance levels of the evaluation nodes within the preset time window, and the statistical values ​​of abnormal state markers correspond to the statistical results of abnormal state markers of the evaluation nodes within the preset time window.

[0143] Based on the deployment attributes of each evaluation agent in the overall set of evaluation agents, an agent role vector is constructed for each evaluation node. The agent role vector includes the wireless link layer evaluation agent role marker, the wireless access node layer evaluation agent role marker, the regional wireless network layer evaluation agent role marker, and the global wireless network layer evaluation agent role marker, which are used to represent the role category of the evaluation agent to which the corresponding evaluation node belongs.

[0144] Based on the hierarchical set and the hierarchical index of each evaluation node, a hierarchical identifier vector corresponding to each evaluation node is constructed. The hierarchical identifier vector includes the wireless link layer identifier component, the wireless access node layer identifier component, the regional wireless network layer identifier component, and the global wireless network layer identifier component.

[0145] When an evaluation node belongs to the radio link layer, the radio link layer identifier component in the hierarchical identifier vector is set to one, and the other components are set to zero; when an evaluation node belongs to the radio access node layer, the radio access node layer identifier component in the hierarchical identifier vector is set to one, and the other components are set to zero; when an evaluation node belongs to the regional radio network layer, the regional radio network layer identifier component in the hierarchical identifier vector is set to one, and the other components are set to zero; when an evaluation node belongs to the global radio network layer, the global radio network layer identifier component in the hierarchical identifier vector is set to one, and the other components are set to zero.

[0146] Based on the time sampling results corresponding to each evaluation node in the standardized acquisition sequence, a time state vector for each evaluation node is constructed. The time state vector includes an absolute time index, a relative time offset, and a time series position encoding value. The absolute time index is used to represent the absolute time position of the corresponding evaluation node at the current moment, the relative time offset is used to represent the time offset of the corresponding evaluation node relative to the start time of the preset time window, and the time series position encoding value is used to represent the time series position of the corresponding evaluation node within the preset time window.

[0147] Based on the connection relationship between each evaluation node in the hierarchical evaluation diagram, a wireless network topology relationship vector is constructed for each evaluation node. The wireless network topology relationship vector includes the number of incoming edges, the number of outgoing edges, the total number of associated edges, and the shortest hop count to the corresponding virtual proxy node. The number of incoming edges represents the number of edges ending at the corresponding evaluation node, the number of outgoing edges represents the number of edges starting at the corresponding evaluation node, the total number of associated edges represents the total number of edges connected to the corresponding evaluation node, and the shortest hop count represents the shortest path length from the corresponding evaluation node to the virtual proxy node at this level.

[0148] The centrality vector of each evaluation node is calculated based on the graph structure of the hierarchical evaluation graph. The centrality vector includes degree centrality, betweenness centrality and proximity centrality. The degree centrality is determined by dividing the total number of associated edges of the corresponding evaluation node by the total number of nodes in the hierarchical evaluation graph minus one. The betweenness centrality is used to represent the degree of betweenness of the corresponding evaluation node on the shortest path of other nodes. The proximity centrality is used to represent the degree of proximity of the corresponding evaluation node to other nodes.

[0149] Based on the performance evaluation input vector, agent role vector, hierarchical identifier vector, time state vector, wireless network topology relationship vector, and centrality vector, construct the spliced ​​input vector for each evaluation node;

[0150] Linear mapping is performed on the concatenated input vector of each evaluation node to obtain the node representation of the corresponding evaluation node. The node representation is determined by mapping the node input embedding mapping relationship to the concatenated input vector and then superimposing the node input embedding bias term. The node representations corresponding to each evaluation node are then aggregated to generate a node representation set.

[0151] For each edge in the hierarchical evaluation graph, an edge attribute vector is constructed based on the edge relationship type, starting node level, target node level, edge direction, edge connection strength, and conflict determination value. The edge relationship type is used to distinguish intra-layer associated edges, inter-layer convergence edges, inter-layer feedback edges, and conflict checking edges. The edge direction is used to indicate the starting and ending directions of the edge. The edge connection strength is used to indicate the strength of the connection of the corresponding edge. The conflict determination value indicates the degree of conflict between the associated nodes of the corresponding edge.

[0152] Linear mapping is performed on the edge attribute vectors of each edge to obtain the edge attribute codes of the corresponding edges. The edge attribute codes are determined by mapping the edge attribute codes to the edge attribute vectors according to the edge attribute code mapping relationship and then superimposing the edge attribute code bias term. The edge attribute codes corresponding to each edge in the hierarchical evaluation graph are collected to generate the edge attribute code results.

[0153] In this embodiment, based on any two evaluation nodes in the hierarchical evaluation graph, a cross-level semantic distance encoding result is generated, and the conflict-aware bias result is obtained according to different level evaluation nodes, specifically including:

[0154] For any two evaluation nodes in the hierarchical evaluation graph, construct a set of node pairs for performing cross-level semantic distance encoding and conflict-aware bias generation based on the node representation set, edge attribute encoding results, performance data set, and local evaluation result set.

[0155] For any node pair in the node pair set, generate the topological association quantity of the node pair based on the shortest path length between the two corresponding evaluation nodes in the hierarchical evaluation graph.

[0156] Based on the hierarchical indexes corresponding to the two evaluation nodes in the node pair, generate the hierarchical span metric of the node pair. The hierarchical span metric is the absolute value of the difference between the hierarchical indexes of the two evaluation nodes.

[0157] Based on the performance evaluation input vectors of the two evaluation nodes corresponding to the node pair in the performance dataset, the index correlation quantity of the node pair is generated. The index correlation quantity is one minus the cosine similarity of the two performance evaluation input vectors. The cosine similarity is determined by the ratio of the inner product of the two performance evaluation input vectors to the product of the magnitudes of the two performance evaluation input vectors.

[0158] Based on the time state vectors of the two evaluation nodes corresponding to the node pair, the time synchronization deviation of the node pair is generated. The time synchronization deviation is defined as the absolute value of the difference between the relative time offsets of the two evaluation nodes at the current moment.

[0159] Based on the abnormal state label sequence of the two evaluation nodes corresponding to the node pair in the local evaluation result set, the abnormal propagation risk of the node pair is generated. The abnormal propagation risk is the average of the absolute values ​​of the difference between the abnormal state labels of the two evaluation nodes at each sampling time within the preset time window.

[0160] Based on the topological association quantity, hierarchical cross-sectionality quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity, the cross-hierarchical semantic distance value of the node pair is generated. The cross-hierarchical semantic distance value is obtained by multiplying the topological association quantity, hierarchical cross-sectionality quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity with the corresponding cross-hierarchical semantic distance fusion weights and then summing them. The sum of the cross-hierarchical semantic distance fusion weights is one.

[0161] Based on the cross-level semantic distance values, cross-level semantic distance codes for node pairs are generated. The cross-level semantic distance codes are obtained by taking the topological association quantity, level cross-rate quantity, indicator correlation quantity, time synchronization deviation quantity, anomaly propagation risk quantity, and cross-level semantic distance value as input vectors, and performing a linear mapping using the cross-level semantic distance code mapping matrix and the cross-level semantic distance code bias vector. The cross-level semantic distance codes corresponding to each node pair in the node pair set are then aggregated to generate the cross-level semantic distance code results.

[0162] For any node pair in the node pair set, based on the local performance scores of the two corresponding evaluation nodes in the local evaluation result set at the current time, the local performance score difference of the node pair is generated. The local performance score difference is defined as the absolute value of the difference between the local performance scores of the two evaluation nodes at the current time.

[0163] Based on the local performance score sequence of the two evaluation nodes in the local evaluation result set within the preset time window, a trend correlation quantity of the node pair is generated. The trend correlation quantity is one minus the correlation coefficient of the local performance score sequence of the two evaluation nodes within the preset time window. The correlation coefficient is determined by the sum of the products of the deviations between the local performance scores of the two evaluation nodes and the mean of the corresponding local performance scores, and the product of the square root of the sum of squares of the deviations of the two evaluation nodes.

[0164] Based on the time state vectors of the two evaluation nodes corresponding to the node pair, the time deviation of the node pair is generated. The time deviation is the absolute value of the difference between the time series position encoding values ​​of the two evaluation nodes at the current time.

[0165] The conflict degree information of node pairs is generated based on the local performance score difference, trend correlation, and time deviation. The conflict degree information is obtained by multiplying the local performance score difference, trend correlation, and time deviation by the corresponding conflict degree fusion weights and then summing them. The sum of each conflict degree fusion weight is one.

[0166] Based on the conflict degree information and the conflict determination value, a conflict perception bias is generated for the node pair. The conflict perception bias is obtained by multiplying the conflict degree information and the conflict determination value by the corresponding conflict perception bias fusion weights and then summing them. The sum of the fusion weights of each conflict perception bias is one.

[0167] The conflict degree information and conflict perception bias of each node pair in the node pair set are aggregated to generate conflict perception bias results.

[0168] In this embodiment, the node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are input into the improved Graphormer model, and multi-head attention computation is performed to obtain the evaluation node update representation set, which specifically includes:

[0169] The node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are input into the improved Graphormer model to construct the input node feature matrix of the improved Graphormer model. The input node feature matrix is ​​formed by arranging the node representations corresponding to each evaluation node in the evaluation node set in a preset order.

[0170] In the improved Graphormer model, a relation enhancement attention unit is constructed. For any node pair, a relation enhancement bias term is generated based on the corresponding edge attribute encoding result, cross-level semantic distance encoding result, and conflict-aware bias result. The relation enhancement bias term is obtained by weighted fusion of the edge attribute encoding mapping result, the cross-level semantic distance encoding mapping result, and the conflict-aware bias result. The relation enhancement bias term satisfies the following: the relation enhancement bias term is equal to the first fusion weight multiplied by the edge attribute encoding projection result, plus the second fusion weight multiplied by the cross-level semantic distance encoding projection result, plus the third fusion weight multiplied by the conflict-aware bias result.

[0171] In the improved Graphormer model, intra-layer local consistency feature extraction paths and inter-layer convergence feature extraction paths are constructed. Intra-layer local consistency masks and inter-layer convergence masks are generated based on the level indices of any two evaluation nodes. When any two evaluation nodes belong to the same level, the intra-layer local consistency mask is zero and the inter-layer convergence mask is negative infinity. When any two evaluation nodes belong to different levels, the intra-layer local consistency mask is negative infinity and the inter-layer convergence mask is zero.

[0172] In each encoding layer of the improved Graphormer model, query mapping, key mapping, and value mapping are performed on the feature matrix of the input node of the current encoding layer to obtain the query matrix, key matrix, and value matrix corresponding to each attention head. The query matrix corresponding to each attention head is obtained by multiplying the feature matrix of the input node of the current encoding layer with the corresponding query mapping matrix. The key matrix corresponding to each attention head is obtained by multiplying the feature matrix of the input node of the current encoding layer with the corresponding key mapping matrix. The value matrix corresponding to each attention head is obtained by multiplying the feature matrix of the input node of the current encoding layer with the corresponding value mapping matrix.

[0173] In the intra-layer local consistency feature extraction path, the intra-layer attention weight is calculated based on the dot product of the query vector and key vector of the corresponding evaluation node, the relation enhancement bias term, and the intra-layer local consistency mask. The corresponding value vector is then weighted and summed according to the intra-layer attention weight to obtain the intra-layer local consistency features corresponding to each evaluation node. The calculation of the intra-layer attention weight includes dividing the dot product of the query vector and key vector of the corresponding evaluation node by the square root of the current attention head feature dimension, adding it to the relation enhancement bias term and the intra-layer local consistency mask, and performing normalization processing on all candidate evaluation nodes.

[0174] In the inter-layer convergence feature extraction path, the inter-layer attention weights are calculated based on the dot product of the query vector and key vector of the corresponding evaluation node, the relation enhancement bias term, and the inter-layer convergence mask. The corresponding value vectors are then weighted and summed according to the inter-layer attention weights to obtain the inter-layer convergence features corresponding to each evaluation node. The calculation of the inter-layer attention weights includes dividing the dot product of the query vector and key vector of the corresponding evaluation node by the square root of the current attention head feature dimension, adding it to the relation enhancement bias term and the inter-layer convergence mask, and performing normalization processing on all candidate evaluation nodes.

[0175] The intra-layer local consistency features output by each attention head are concatenated to obtain the intra-layer local consistency representation;

[0176] The inter-layer convergence features output by each attention head are concatenated to obtain the inter-layer convergence representation;

[0177] In the improved Graphormer model, a dual-path fusion update unit is constructed. The intra-layer local consistency representation and the inter-layer convergence representation are concatenated and then input into the feature fusion mapping matrix to obtain the fusion representation. The fusion representation is obtained by concatenating the intra-layer local consistency representation and the inter-layer convergence representation, followed by linear mapping, bias superposition, and nonlinear mapping.

[0178] Based on the fusion representation, residual update and feedforward update are performed to obtain the output representation of each evaluation node in the current coding layer. The output representations corresponding to each evaluation node of the last coding layer of the improved Graphormer model are aggregated to generate a set of updated evaluation node representations. The residual update is to add the current coding layer input representation and the fusion representation and then perform normalization. The feedforward update is to input the normalized result into the feedforward mapping unit, add the feedforward mapping result and the normalized result, and then perform normalization again.

[0179] In this embodiment, a hierarchical virtual proxy node chain is used to aggregate the updated representation sets of the evaluation nodes step by step to obtain the global network performance evaluation results, specifically including:

[0180] Obtain the set of evaluation node update representations and extract the hierarchical index corresponding to each evaluation node. Based on the hierarchical index, classify the update representations of each evaluation node into the radio link layer, radio access node layer, regional radio network layer and global radio network layer respectively, and obtain the node update representation subsets corresponding to each layer.

[0181] In the hierarchical virtual agent node chain, for the virtual agent node corresponding to the wireless link layer, the updated representation subset of the node corresponding to the wireless link layer is received.

[0182] Based on the proxy aggregation edge relationship between each evaluation node and the virtual proxy node corresponding to the wireless link layer, the node update representation subset corresponding to the wireless link layer is weighted and aggregated to obtain the wireless link layer proxy aggregation result.

[0183] The wireless link layer proxy aggregation result is input into the virtual proxy node corresponding to the wireless access node layer, and then combined with the node update representation subset corresponding to the wireless access node layer for joint aggregation to obtain the wireless access node layer proxy aggregation result.

[0184] Among them, the wireless access node layer proxy aggregation result is obtained by weighted aggregation of the wireless link layer proxy aggregation result and the node update representation subset corresponding to the wireless access node layer;

[0185] The wireless access node layer proxy aggregation result is input into the virtual proxy node corresponding to the regional wireless network layer, and then combined with the node update representation subset corresponding to the regional wireless network layer for joint aggregation to obtain the regional wireless network layer proxy aggregation result.

[0186] Among them, the regional wireless network layer proxy aggregation result is obtained by weighted aggregation of the wireless access node layer proxy aggregation result and the node update representation subset corresponding to the regional wireless network layer;

[0187] The regional wireless network layer proxy aggregation result is input into the virtual proxy node corresponding to the global wireless network layer, and then combined with the node update representation subset corresponding to the global wireless network layer for joint aggregation to obtain the global wireless network layer proxy aggregation result.

[0188] Among them, the global wireless network layer proxy aggregation result is obtained by weighted aggregation of the regional wireless network layer proxy aggregation result and the node update representation subset corresponding to the global wireless network layer;

[0189] For any virtual proxy node at any level, the proxy aggregation weight of the updated representation of each input node in the level relative to the current virtual proxy node is calculated. The proxy aggregation weight is obtained by normalizing the correlation calculation result between the updated representation of the current input node and the state of the current virtual proxy node. The proxy aggregation result corresponding to the current level is obtained by weighted summation based on the updated representation of each input node and the corresponding proxy aggregation weight.

[0190] Based on the proxy aggregation results corresponding to each level in the hierarchical virtual proxy node chain, the system performs step-by-step transmission and aggregation in order from the lower level to the higher level to obtain the global performance representation of the wireless communication network. The global performance representation of the wireless communication network is jointly generated by the global wireless network layer proxy aggregation result and the proxy aggregation results of each lower level in the hierarchical virtual proxy node chain.

[0191] The global performance representation of the wireless communication network is input into the global evaluation output layer to obtain the global performance evaluation result of the network. The global performance evaluation result of the network includes at least the global performance score, the global performance level, and the global abnormal state label.

[0192] The global performance score is calculated based on the global performance representation of the wireless communication network. The global performance score is obtained by mapping the global performance representation of the wireless communication network through a scoring mapping matrix and superimposing a scoring bias. The global performance level is determined based on the relationship between the global performance score and a preset global level classification threshold.

[0193] The global anomaly status label is determined based on the relationship between the global performance score and the preset global anomaly determination threshold.

[0194] In this embodiment, detecting conflicts between the network global performance evaluation results and the evaluation itself, correcting the network global performance evaluation results, and generating resource control strategies or operating parameter optimization strategies specifically include:

[0195] Obtain the global network performance evaluation results and the proxy aggregation results corresponding to each level of virtual proxy nodes, and construct an evaluation conflict detection input set based on the evaluation difference between the global network performance evaluation results and the outputs of each level of virtual proxy nodes;

[0196] Based on the network global performance evaluation results, the deviation between the network global performance evaluation results and the outputs of the virtual proxy nodes corresponding to the wireless link layer, the wireless access node layer, the regional wireless network layer, and the global wireless network layer is calculated to obtain the evaluation conflict detection values ​​corresponding to each layer. The evaluation conflict detection values ​​corresponding to each layer are obtained by weighted fusion of the differences in global performance scores, global performance levels, and global abnormal state markings.

[0197] The existence of an evaluation conflict is determined based on the evaluation conflict detection value corresponding to each level. Specifically, when the evaluation conflict detection value corresponding to any level is greater than or equal to the preset conflict trigger threshold, an evaluation conflict is determined to exist, and the conflict verification process is triggered.

[0198] When the evaluation conflict detection value corresponding to each level is less than the preset conflict trigger threshold, it is determined that there is no evaluation conflict and the current global network performance evaluation result is maintained.

[0199] After the conflict verification process is triggered, the corresponding conflict node pair is located based on the hierarchical virtual agent node that generated the evaluation conflict, combined with the conflict perception bias result and the conflict verification edge, to obtain a set of conflict node pairs. Each conflict node pair in the set of conflict node pairs is determined by the conflict degree information, conflict judgment value and hierarchical association relationship between the corresponding evaluation nodes.

[0200] For each conflict node pair in the conflict node pair set, update the attention weight, edge relationship, and local evaluation result of the corresponding conflict node pair. The update of the attention weight is obtained by modifying the original attention weight with the conflict suppression coefficient. The update of the edge relationship includes adjusting at least one edge relationship among intra-layer association edge, inter-layer convergence edge, inter-layer feedback edge, and conflict verification edge. The update of the local evaluation result includes modifying at least one result among local performance score, local performance level, and abnormal state label.

[0201] Based on the updated attention weights, updated edge relationships, and updated local evaluation results, the corresponding edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are regenerated, and the hierarchical evaluation graph is reconstructed.

[0202] The updated hierarchical evaluation graph is re-input into the improved Graphormer model, and intra-layer local consistency feature extraction, inter-layer convergence feature extraction, and dual-path fusion update are re-executed to obtain the corrected set of evaluation node update representations.

[0203] Based on the revised evaluation node update representation set, the hierarchical aggregation is re-executed through the hierarchical virtual proxy node chain to generate the revised global performance representation of the wireless communication network, and the revised global network performance evaluation result is regenerated based on the revised global performance representation of the wireless communication network.

[0204] Based on the corrected global network performance evaluation results, a resource regulation strategy or an operation parameter optimization strategy for the wireless communication network is generated. The resource regulation strategy includes at least one or more of the following: resource block allocation adjustment strategy, bandwidth allocation adjustment strategy, and access control adjustment strategy. The operation parameter optimization strategy includes at least one or more of the following: handover parameter optimization strategy, scheduling parameter optimization strategy, and power control optimization strategy.

[0205] Example 1: To verify the feasibility of this invention in practice, it was applied to the operation evaluation and optimization scenario of a large-scale campus wireless communication network. This campus network simultaneously carries high-definition video backhaul, mobile inspection terminal access, industrial sensor data upload, and voice dispatch services. The network includes multiple wireless access nodes, several regional subnets, and a unified core scheduling unit. With the increase in service concurrency, problems such as increased link latency fluctuations, uneven load distribution between regions, decreased access success rate, fluctuating handover success rate, and enhanced local interference emerged during network operation. Existing methods typically monitor a single node or a single indicator, which can obtain local operational results but is difficult to reflect the transmission relationships between the link layer, access node layer, regional wireless network layer, and global wireless network layer, and also difficult to identify inconsistencies between local and global evaluation results. Therefore, during peak network service periods, although the throughput of some access nodes remains high, the overall service experience fluctuates significantly, and subsequent resource adjustments often lag, making it difficult to formulate targeted optimization strategies.

[0206] In this embodiment, the wireless communication network to be evaluated is first divided into layers. Wireless link objects are assigned to the wireless link layer, base stations, access units, and edge access devices are assigned to the wireless access node layer, access nodes with the same service coverage are aggregated into the regional wireless network layer, and the entire campus private network is used as the global wireless network layer. Subsequently, evaluation agents are deployed at each layer, each responsible for collecting data on latency, throughput, packet loss rate, jitter, bandwidth utilization, wireless retransmission rate, resource block utilization, channel quality, interference intensity, access success rate, handover success rate, and resource utilization for the corresponding wireless network object. The collected data forms a performance dataset, and the evaluation agents at each layer generate local evaluation results. Then, a hierarchical evaluation graph is constructed based on the hierarchical evaluation system, the set of evaluation agents, the performance dataset, and the local evaluation result set. This graph simultaneously expresses intra-layer correlations, inter-layer convergence relationships, inter-layer feedback relationships, and conflict verification relationships, enabling a unified correlation between local fluctuations at the link layer, load changes at the access layer, congestion trends at the regional layer, and service quality at the global layer.

[0207] In practical applications, performance data, local evaluation results, agent role information, hierarchical identification information, temporal state information, wireless network topology relationship information, and centrality information are input and embedded to obtain a set of node representations and edge attribute encoding results. Further, based on the topological association, hierarchical span, index correlation, time synchronization deviation, and anomaly propagation risk between any two evaluation nodes, cross-hierarchical semantic distance encoding results are generated. Conflict-aware bias results are generated based on the differences in local performance scores, trend correlation, and time deviation between evaluation nodes at different levels. Then, the set of node representations, edge attribute encoding results, cross-hierarchical semantic distance encoding results, and conflict-aware bias results are input into an improved Graphormer model. In the model, intra-layer local consistency features and inter-layer convergence features are extracted to obtain the set of updated representations for the evaluation nodes. Finally, a hierarchical virtual agent node chain is used to aggregate the updated representations at each level, outputting the global network performance evaluation results. When a threshold-exceeding conflict is detected between the global evaluation result and the output of the regional layer virtual proxy node, the system automatically triggers a conflict verification process. This process corrects the attention weights, edge relationships, and local evaluation results of the conflicting node pairs and regenerates resource control strategies or operational parameter optimization strategies. For example, it reallocates resource blocks for high-load access nodes, adjusts handover thresholds for handover boundary areas, and adjusts channel and bandwidth usage configurations for areas with high interference. In this way, the network operation and maintenance process, which previously relied solely on local monitoring, is transformed into a dynamic optimization process oriented towards multi-level collaborative perception, global evaluation, and conflict correction. This solves the problems of existing technologies struggling to uniformly handle cross-level correlations, performance conflicts, and insufficient accuracy in global evaluation.

[0208] To verify the effectiveness of this invention, the method of this invention was compared with the traditional single-layer indicator statistical evaluation method over several consecutive evaluation periods. The test objects were wireless communication networks within the same campus, with consistent service load, terminal scale, and network topology.

[0209] Table 1 Comparison of the Implementation Effects of Layered Evaluation and Optimization of Wireless Communication Networks in the Park

[0210] Indicator Categories Indicator Name Traditional single-level indicator statistical evaluation method Method of the present invention range of change Consistency of assessment Consistency rate between overall assessment and manual review 89.4% 96.8% An increase of 7.4 percentage points Consistency of assessment Post-conflict verification error assessment 8.6% 5.4% Decrease of 37.2% latency performance Average latency 48.7ms 34.2ms Decrease of 29.8% latency performance Delay Standard Deviation 12.6ms 7.1ms Decrease of 43.7% Transmission performance Average throughput 612.4Mbps 688.9Mbps Increased by 12.5% Transmission performance Peak throughput 731.6Mbps 804.3Mbps Improved by 9.9% reliability Average packet loss rate 1.86% 0.92% Decrease of 50.5% reliability Wireless retransmission rate 6.7% 3.8% Decrease of 43.3% stability Average jitter 9.8ms 5.4ms Decrease of 44.9% Resource status Bandwidth utilization 81.3% 72.6% Down 10.7% Resource status Resource block utilization 88.5% 79.4% Down 10.3% Resource status Resource utilization balance 0.71 0.84 An increase of 18.3% Wireless environment Average Channel Quality Indicators 12.4 15.1 An increase of 21.8% Wireless environment Average Interference Intensity -89.6dBm -94.1dBm Optimized by 4.5dB Access capabilities Access success rate 96.2% 98.7% Increased by 2.5 percentage points Switching capabilities Switching success rate 93.8% 97.1% An increase of 3.3 percentage points Conflict handling Percentage of re-fusion cycles triggered 0 18.6% Increase conflict verification capabilities Business Experience Number of times video upload stutters 118 times / cycle 69 times / cycle Decrease of 41.5% Business Experience Number of times the inspection terminal reconnects 95 times / cycle 60 times / cycle Down 36.8% Business Experience Voice dispatch service interruption rate 1.87% 1.24% Down 33.7%

[0211] As can be seen from the table above, this invention significantly improves upon traditional single-layer statistical evaluation methods in several key indicators of hierarchical evaluation and optimization of wireless communication network performance. Regarding global evaluation capability, the consistency rate between the global evaluation and manual verification achieved by the method of this invention reaches 96.8%, higher than the 89.4% of the traditional method, representing an improvement of 7.4 percentage points. Simultaneously, the evaluation error after conflict verification decreased from 8.6% to 5.4%, a reduction of 37.2%. This result demonstrates that by constructing a hierarchical evaluation system, a hierarchical evaluation graph, and introducing cross-level semantic distance coding and conflict-aware bias, this invention can more effectively handle the coupling relationship between local and global evaluation results, improving the ability of global performance evaluation results to characterize the actual operating state of the network.

[0212] In terms of network transmission performance and stability, the method of this invention also demonstrates good optimization effects. The average latency decreased from 48.7 milliseconds to 34.2 milliseconds, a reduction of 29.8%, and the latency standard deviation decreased from 12.6 milliseconds to 7.1 milliseconds, indicating improvements in both network latency level and latency fluctuation. The average throughput increased from 612.4 megabits per second to 688.9 megabits per second, an increase of 12.5%, and the peak throughput increased from 731.6 megabits per second to 804.3 megabits per second, indicating enhanced overall network transmission capacity. The average packet loss rate decreased from 1.86% to 0.92%, the wireless retransmission rate decreased from 6.7% to 3.8%, and the average jitter decreased from 9.8 milliseconds to 5.4 milliseconds, demonstrating that the method of this invention can effectively alleviate link instability and data retransmission problems. These results indicate that by using the improved Graphormer model to fuse intra-layer local consistency features and inter-layer convergence features, the performance change trends of wireless network objects at different layers can be identified more accurately, providing a valid basis for subsequent optimization.

[0213] The method of this invention also has positive effects on resource status, wireless environment, and service experience. Bandwidth utilization decreased from 81.3% to 72.6%, resource block utilization was optimized from 88.5% to 79.4%, and resource utilization balance improved from 0.71 to 0.84, indicating a more balanced resource allocation process. The average channel quality index improved from 12.4 to 15.1, and the average interference intensity was optimized from -89.6 dB / mW to -94.1 dB / mW, indicating an improved wireless environment. The access success rate increased from 96.2% to 98.7%, and the handover success rate increased from 93.8% to 97.1%, further demonstrating that this invention can improve the stability of the access and handover processes. Meanwhile, the number of video transmission stutters, the number of reconnection attempts at inspection terminals, and the voice dispatch service interruption rate decreased by 41.5%, 36.8%, and 33.7%, respectively, reflecting that while improving the accuracy of global network performance assessment, this invention also enhances the targeting of resource control strategies and operational parameter optimization strategies, thereby improving the actual service carrying effect.

[0214] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A hierarchical evaluation method for network performance based on multi-agent systems, characterized in that, Includes the following steps: The wireless communication network to be evaluated is divided into layers, a hierarchical evaluation system is constructed, and multiple evaluation agents are deployed at the corresponding layers to obtain a set of evaluation agents; Each evaluation agent in the evaluation agent set collects performance data of the corresponding wireless network object, generating a performance data set and a local evaluation result set. A hierarchical evaluation graph is constructed based on the hierarchical evaluation system, the set of evaluation agents, the set of performance data, and the set of local evaluation results. Input embedding processing is performed on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results; Based on any two evaluation nodes in the hierarchical evaluation graph, generate cross-level semantic distance encoding results, and obtain conflict-aware bias results according to evaluation nodes at different levels. The node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are input into the improved Graphormer model, and multi-head attention computation is performed to obtain the evaluation node update representation set. By using a hierarchical virtual agent node chain, the updated representation set of the evaluation nodes is aggregated level by level to obtain the global network performance evaluation result; The system detects conflicts between the network global performance evaluation results and the evaluation results, corrects the network global performance evaluation results, and generates resource control strategies or operating parameter optimization strategies.

2. The network performance hierarchical evaluation method based on a multi-agent system according to claim 1, characterized in that, The hierarchical evaluation system includes the wireless link layer, the wireless access node layer, the regional wireless network layer, and the global wireless network layer.

3. The hierarchical evaluation method for network performance based on a multi-agent system according to claim 1, characterized in that, The process of collecting performance data of the corresponding wireless network object from each evaluation agent in the evaluation agent set, and generating a performance data set and a local evaluation result set, specifically includes: The evaluation agents in the overall set of evaluation agents collect performance data on the network objects at the corresponding level according to the corresponding relationship matrix. Construct performance data vectors based on the hierarchy and business type of the hierarchical network objects; Time alignment and validity filtering are performed on each indicator component in the performance data vector to obtain a standardized collection sequence; For each indicator component in the standardized collection sequence, an indicator normalization result is constructed, and for negative performance indicators, a corresponding reverse normalization result is constructed. Based on the index weights corresponding to each performance index component, a local evaluation result is generated. The performance data vectors, standardized acquisition sequences, and local evaluation results output by each evaluation agent in the overall evaluation agent set are aggregated according to hierarchical sets and hierarchical affiliations to generate performance data sets and local evaluation result sets.

4. The network performance hierarchical evaluation method based on a multi-agent system according to claim 1, characterized in that, The construction of the hierarchical evaluation graph based on the hierarchical evaluation system, the set of evaluation agents, the set of performance data, and the set of local evaluation results specifically includes: For each evaluation agent in the overall set of evaluation agents, a corresponding evaluation node is established, resulting in a set of evaluation nodes; For each level in the hierarchy set, a corresponding virtual proxy node is established to obtain a set of virtual proxy nodes; Based on the set of evaluation nodes and the set of virtual agent nodes, a set of nodes for a hierarchical evaluation graph is generated, and a hierarchical index, node type identifier, and object ownership identifier are assigned to each node. Based on the set of intra-layer association relationships and the set of inter-layer convergence relationships, we obtain the set of intra-layer association edges and the set of inter-layer convergence edges; Based on the result transmission requirements and correction transmission requirements between adjacent levels, establish inter-layer feedback relationships, and establish an inter-layer feedback edge set based on the inter-layer feedback relationships; Based on the performance data set and the local evaluation result set, calculate the conflict determination value between each pair of evaluation nodes to obtain the conflict determination result. Based on the conflict determination results of each evaluation node pair, determine the set of conflict node pairs, establish conflict verification relationships between each conflict node pair, and establish a set of conflict verification edges. Based on the set of intra-layer associated edges, inter-layer convergence edges, proxy convergence edges, inter-layer feedback edges, proxy feedback edges, and conflict verification edges, generate the edge set of the hierarchical evaluation graph. Construct a hierarchical evaluation graph based on the set of nodes and the set of edges of the hierarchical evaluation graph.

5. The network performance hierarchical evaluation method based on a multi-agent system according to claim 1, characterized in that, The specific steps of performing input embedding processing on each evaluation node in the hierarchical evaluation graph to obtain the node representation set and edge attribute encoding results include: Obtain the hierarchical evaluation graph, and perform input embedding processing on each evaluation node in the hierarchical evaluation graph to generate performance evaluation input vector, agent role vector, hierarchical identifier vector, time state vector, wireless network topology relationship vector, and centrality vector; Construct a spliced ​​input vector based on the performance evaluation input vector, agent role vector, hierarchy identifier vector, time state vector, wireless network topology vector, and centrality vector; Linear mapping is performed on the concatenated input vectors of each evaluation node to obtain the node representation of the corresponding evaluation node, and a set of node representations is generated. For each edge in the hierarchical evaluation graph, construct an edge attribute vector based on the edge relationship type, starting node level, target node level, edge direction, edge connection strength, and conflict determination value. Linear mapping is performed on the edge attribute vectors of each edge to obtain the edge attribute codes of the corresponding edges, and the edge attribute code results are generated.

6. The network performance hierarchical evaluation method based on a multi-agent system according to claim 1, characterized in that, The process of generating cross-level semantic distance encoding results based on any two evaluation nodes in the hierarchical evaluation graph, and obtaining conflict-aware bias results according to different levels of evaluation nodes, specifically includes: For any two evaluation nodes in the hierarchical evaluation graph, construct a set of node pairs based on the node representation set, edge attribute encoding results, performance data set, and local evaluation result set; For any node pair in the set of node pairs, based on the hierarchical evaluation graph, generate the topological association quantity, hierarchical span quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity; Based on the topological association quantity, hierarchical cross-level quantity, indicator correlation quantity, time synchronization deviation quantity, and anomaly propagation risk quantity, a cross-level semantic distance value is generated. Based on the cross-level semantic distance values, a cross-level semantic distance code is generated, and the cross-level semantic distance codes are aggregated to generate a cross-level semantic distance coding result. For any node pair in the node pair set, generate a local performance score difference based on the local performance scores of the two corresponding evaluation nodes in the local evaluation result set at the current time. Based on the local performance score sequence of the two corresponding evaluation nodes in the local evaluation result set, a trend correlation quantity is generated, and based on the time state vector of the two corresponding evaluation nodes, a time deviation quantity is generated. Based on the difference in local performance scores, the correlation of trends, and the time deviation, conflict degree information is generated. Based on the conflict degree information and the conflict judgment value, a conflict perception bias is generated. The conflict degree information and conflict perception bias of each node pair in the node pair set are aggregated to generate conflict perception bias results.

7. The hierarchical evaluation method for network performance based on a multi-agent system according to claim 1, characterized in that, The process of inputting the node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results into the improved Graphormer model, performing multi-head attention computation, and obtaining the evaluation node update representation set specifically includes: Input the node representation set, edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results into the improved Graphormer model to construct the input node feature matrix; Construct a relation-enhancing attention unit, and for any node pair, generate a relation-enhancing bias term based on the edge attribute encoding result, the cross-level semantic distance encoding result, and the conflict-aware bias result; In the improved Graphormer model, we construct intra-layer local consistency feature extraction paths and inter-layer convergence feature extraction paths to generate intra-layer local consistency masks and inter-layer convergence masks. In each coding layer, query mapping, key mapping, and value mapping are performed on the feature matrix of the input node of the current coding layer to obtain the query matrix, key matrix, and value matrix, respectively. In the intra-layer local consistency feature extraction path, the intra-layer attention weights are calculated and weighted summation is performed to obtain the intra-layer local consistency features; In the inter-layer convergence feature extraction path, the inter-layer attention weights are calculated and weighted summation is performed to obtain the inter-layer convergence features; The intra-layer local consistency features output by each attention head are concatenated to obtain the intra-layer local consistency representation; The inter-layer convergence features output by each attention head are concatenated to obtain the inter-layer convergence representation; In the dual-path fusion update unit, the intra-layer local consistency representation and the inter-layer convergence representation are concatenated and then input into the feature fusion mapping matrix to obtain the fused representation; Based on the fusion representation, residual updates and feedforward updates are performed to obtain the output representation of each evaluation node in the current coding layer, and a set of updated representations for the evaluation nodes is generated.

8. The network performance hierarchical evaluation method based on a multi-agent system according to claim 1, characterized in that, The process of using a hierarchical virtual proxy node chain to aggregate the updated representation sets of evaluation nodes step by step to obtain the global network performance evaluation results specifically includes: Obtain the set of evaluation node update representations and extract the hierarchical index corresponding to each evaluation node to obtain a subset of node update representations; In the hierarchical virtual agent node chain, for the virtual agent node corresponding to the wireless link layer, the updated representation subset of the node corresponding to the wireless link layer is received and aggregated step by step to obtain the agent aggregation result of each layer. For any virtual proxy node at any level, calculate the proxy aggregation weight and obtain the proxy aggregation result for the current level; Based on the agent aggregation results corresponding to each level in the hierarchical virtual agent node chain, a global performance representation of the wireless communication network is obtained; The global performance representation of the wireless communication network is input into the global evaluation output layer to obtain the global performance evaluation result of the network.

9. The hierarchical evaluation method for network performance based on a multi-agent system according to claim 1, characterized in that, The process of detecting conflicts between the network global performance evaluation results and the evaluation results, correcting the network global performance evaluation results, and generating resource regulation strategies or operating parameter optimization strategies specifically includes: Obtain the global network performance evaluation results and the proxy aggregation results corresponding to virtual proxy nodes at each level, and construct the evaluation conflict detection input set; Based on the network global performance evaluation results, calculate the evaluation conflict detection value corresponding to each level; The existence of an assessment conflict is determined based on the assessment conflict detection value corresponding to each level. When an assessment conflict is determined to exist, the conflict verification process is triggered. Based on the hierarchical virtual agent node that generates the evaluation conflict, combined with the conflict perception bias result and the conflict verification edge, the corresponding conflict node pair is located to obtain the set of conflict node pairs. For each conflict node pair in the set of conflict node pairs, update the attention weight, edge relationship, and local evaluation result of the corresponding conflict node pair; Based on the updated attention weights, updated edge relationships, and updated local evaluation results, the corresponding edge attribute encoding results, cross-level semantic distance encoding results, and conflict-aware bias results are regenerated, and the hierarchical evaluation graph is reconstructed. The updated hierarchical evaluation graph is re-input into the improved Graphormer model to regenerate the corrected global network performance evaluation results. Based on the revised global network performance evaluation results, resource regulation strategies or operating parameter optimization strategies for wireless communication networks are generated.

Images