An optimization method for a high-performance 5G communication module

Abstract

This invention relates to the field of 5G communication technology, and more particularly to an optimization method for high-performance 5G communication modules. The method includes the following steps: collecting key operating parameters under multi-carrier concurrent conditions to construct a feature graph model reflecting the evolution of these parameters; mining state clustering patterns using a graph polarization spectrum clustering algorithm, and combining this with a multi-hop path optimization algorithm based on graph attention mechanisms to identify critical paths and core parameter nodes affecting communication performance; using the critical path policy graph as input, and combining it with the characteristics of the target communication scenario, generating an optimal parameter configuration strategy using a graph structure-aligned meta-learning algorithm; applying this strategy to an actual module to construct an adaptive control model integrating fuzzy inference and reinforcement learning, achieving dynamic evaluation of the strategy execution effect and fine-tuning of parameters. This invention not only improves the performance scheduling efficiency of 5G communication modules under complex network conditions but also significantly enhances the system's intelligence, generalization capabilities, and scenario adaptability.

Description

Technical Field

[0001] This invention relates to the field of 5G communication technology, and more particularly to an optimization method for a high-performance 5G communication module. Background Technology

[0002] With the rapid development of 5G, driven by its key characteristics such as massive MIMO, ultra-dense networking, and multi-carrier concurrency, it is widely used in high-reliability, low-latency scenarios such as smart manufacturing, connected vehicles, and telemedicine. As a core component for terminal access and data transmission, the performance optimization of 5G communication modules is crucial for ensuring system stability and throughput efficiency. With the widespread application of multi-carrier aggregation, ultra-dense deployment, and heterogeneous network environments, 5G communication modules often face problems such as redundant parameter configurations, low resource scheduling efficiency, and unpredictable performance fluctuations during operation. Traditional optimization methods often rely on rule-driven or static modeling, which struggles to achieve efficient generalization and dynamic scheduling in complex and ever-changing communication environments. The following problems still exist: In environments with multi-carrier concurrency, high-frequency interference, and rapid channel changes, the operating state of communication modules is highly complex, and traditional methods cannot effectively extract the potential cooperative relationships between parameters; most existing optimization strategies rely on static analysis, making it difficult to discover multi-hop paths and their key parameter nodes that significantly affect communication performance, thus limiting the potential for in-depth optimization of system performance; traditional optimization methods often rely on empirical rules or fixed models under specific communication scenarios, which cannot adapt to load changes or environmental switching, causing the strategies to fail in different scenarios. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides an optimization method for high-performance 5G communication modules. This method effectively extracts the collaborative relationships between the operating parameters of 5G communication modules, identifies critical paths and parameter nodes, and achieves efficient generalized scheduling and dynamic optimization across scenarios in environments with multi-carrier concurrency, high-frequency interference, and complex heterogeneous networks. This improves the performance scheduling efficiency of 5G communication modules under complex network conditions.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] An optimization method for a high-performance 5G communication module includes the following steps:

[0006] S1: Obtain the operating parameters of the 5G communication module in multi-carrier concurrent state;

[0007] S2: Based on the operating parameters, the state graph is clustered and divided using the graph polarization spectrum clustering algorithm to construct a feature graph model. In the graph, nodes represent the operating parameter states, and edge weights represent the degree of coordinated change between parameters.

[0008] S3: Based on the feature graph model, a multi-hop path optimization algorithm with graph attention mechanism is used to jointly model the node weights in the path and score the path, identify high-weight paths and key parameter nodes that affect communication performance, and generate a critical path strategy graph containing node configuration relationships and path control logic.

[0009] S4: Using the critical path strategy graph as input, based on the scene characteristics in a multi-communication environment, a meta-learning algorithm based on graph structure alignment is adopted to construct a strategy graph transfer model. Through the mapping function, the generalized scheduling and optimal transfer of parameter configuration under different load conditions are realized, and the optimal parameter configuration strategy is generated.

[0010] S5: Apply the optimal parameter configuration strategy to the actual operation of the 5G communication module, collect the real-time operating status of the module as a fuzzy input variable, construct a reinforced fuzzy inference network model, evaluate the deviation and adaptability of the current execution strategy through fuzzy inference rules combined with reinforcement learning reward mechanism, and fine-tune the execution parameters based on performance feedback.

[0011] Furthermore, the operating parameters include transmit power, intermodulation interference, crystal oscillator offset, voltage-current response, board temperature, and instantaneous throughput.

[0012] Furthermore, step S2 includes the following steps:

[0013] Based on the aforementioned operating parameters, a time-series state diagram reflecting the parameter evolution process is constructed using parameter normalization and time window segmentation strategies. Nodes represent parameter states in different time slices, and edge weights characterize the magnitude of collaborative changes between adjacent parameter states.

[0014] By calculating the correlation coefficient matrix for each parameter node pair in the time-series state diagram and combining it with the maximum information coefficient to enhance the nonlinear collaborative characteristics, a weighted multi-attribute state map is generated.

[0015] Based on the weighted multi-attribute state graph, the graph polarization spectral clustering algorithm is used to perform spectral decomposition and feature space mapping on the node set. By introducing polarization operators to adjust the spectral clustering boundary, the parameter states with high noise and weak coupling are clustered and the parameter state clustering labels are output.

[0016] Based on the parameter state clustering labels, the core nodes and edge weight distribution in each subgraph are extracted, a feature graph model containing node features, edge collaborative weights and clustering labels is constructed, and a structured feature graph model is output.

[0017] Furthermore, the formula for constructing the feature map model is as follows:

[0018] ;

[0019] in, This represents the strength of the co-evolution among the running states in the constructed feature map; and These represent the normalized runtime parameter vectors within the i-th and j-th time windows, respectively; Represents the squared Euclidean distance between the multi-parameter states of the i-th and j-th time slices; Indicates the maximum information coefficient; This indicates points in the parameter variation graph where there are local curvature or abnormally drastic changes. This represents the global variance between the states of the operating parameters; The normalized scaling factor represents the maximum information coefficient; , and These represent the weighting coefficients of the Euclidean distance term, the nonlinear correlation term, and the spectral perturbation term, respectively.

[0020] Furthermore, step S3 includes the following steps:

[0021] Based on the feature graph model, a multi-hop path optimization algorithm based on graph attention mechanism is used to jointly model the state features and edge collaboration relationships of nodes in the graph within the multi-hop adjacency range, and construct a multi-hop path structure with cross-level information perception capability.

[0022] In the multi-hop path structure, an attention mechanism is used to dynamically weight the differences in edge weight distribution and state features between nodes within different hop count ranges to generate a node attention distribution map.

[0023] Based on the node attention distribution map, all reachable multi-hop paths are scored. Taking into account the attention weight of each node in the path, the edge collaboration strength and the impact of the number of hops, a set of high-weight paths with significant path scores is selected.

[0024] Based on the set of high-weight paths, key parameter nodes in the paths are extracted, their positions and roles in the path structure and their dependencies on other nodes are analyzed, and a critical path strategy graph that integrates control logic and priority relationships is constructed.

[0025] Furthermore, the formula for the multi-hop path optimization algorithm is as follows:

[0026] ;

[0027] in, H represents the multi-hop path score from node p to node q, used to construct the critical path strategy graph; H represents the maximum number of hops limit, which controls the hierarchical depth of path modeling. This represents the global adjustment factor for the path at the h-th hop; This represents the set of h-th hop neighbors of the current node p; Indicates attention weight; This represents the edge collaboration weight, where the edge weight indicates the degree of collaborative change among the operating parameters; It represents instantaneous throughput and is used to measure the dynamic impact of a node on communication performance in the h-th hop layer; Indicates the number of hops in the path topology; This represents the standard deviation of the plate-level temperature of the nodes associated with node p in the h-th hop; This represents the hop count decay factor.

[0028] Furthermore, step S4 includes the following steps:

[0029] Based on the critical path strategy graph, extract high-weight path sequences, key parameter node configuration combinations, and control dependencies between paths to construct a graph structure sample set associated with specific communication scenarios, forming a source scenario knowledge graph.

[0030] In the target communication environment, by combining the characteristics of real-time network load level, interference source distribution and terminal connection density, a graph structure representation of the current scene is constructed, and scene labels and path control demand vectors are initialized.

[0031] A graph-structure aligned meta-learning algorithm is adopted to perform structural mapping between the knowledge graph in the source scene and the target scene graph, and to generate the optimal path control mapping function in the target scene through a fast parameter adaptation mechanism;

[0032] Based on the aforementioned control mapping function, the critical path strategy graph in the source scenario is mapped to the target communication scenario, realizing the structural migration of the key parameter configuration relationship and outputting the optimal parameter configuration strategy.

[0033] Furthermore, the formula for the optimal path control mapping function is as follows:

[0034] ;

[0035] in, Let N represent the optimal path control mapping function under the target communication scenario R; N represents the number of source scenario samples participating in the control learning. This represents the critical path parameter vector extracted from the nth source scene; This represents the expected optimal parameter configuration result for the nth sample in the target scenario; R represents the critical path strategy graph structure constructed in the nth source communication scenario; R represents the scenario graph structure in the target communication environment. This represents a path control mapping function; Represents the regularization term of function J; This represents the regularity coefficient.

[0036] Furthermore, the enhanced fuzzy inference network model is specifically based on the core parameter nodes in the critical path strategy graph as input dimensions, collecting state indicators such as transmit power offset, voltage and current fluctuations, crystal oscillator frequency offset and temperature changes, constructing a fuzzy input vector, and completing the fuzzification process through membership functions; combining the node dependency relationships in the strategy graph to generate a fuzzy rule base, and introducing a reinforcement learning mechanism to dynamically adjust the weights of fuzzy rules and the priority of execution strategies.

[0037] Furthermore, the reinforcement learning reward mechanism is specifically constructed based on the dynamic feedback of communication performance indicators, including using throughput improvement, bit error rate reduction and handover latency convergence rate as core reward factors, and generating real-time performance scores by combining the weighting function of each indicator.

[0038] The beneficial effects of this invention are as follows:

[0039] This invention uses a graph polarization spectrum clustering algorithm to cluster the operational states of 5G communication modules, effectively revealing the collaborative relationships between parameters and laying a structural foundation for subsequent optimization. This improves the interpretability and structural awareness of the internal states of the communication modules. A multi-hop path optimization algorithm utilizing graph attention mechanisms is used to jointly model the weights of parameter nodes in the path. This accurately identifies the paths and key nodes that have the greatest impact on communication performance, providing a highly reliable decision-making basis for policy generation and significantly improving the accuracy and controllability of path regulation. By introducing a meta-learning algorithm for graph structure alignment, a policy graph model that can be transferred across multiple communication environments is constructed. This solves the problems of poor policy adaptability and inability to generalize in traditional methods, achieving efficient transfer and dynamic scheduling of parameter configurations and improving the robustness and adaptability of the system in complex network environments. In actual deployment, a feedback optimization model jointly constructed based on fuzzy inference and reinforcement learning mechanisms can dynamically sense subtle changes in the communication state and adjust parameters in real time, avoiding performance degradation or policy failure, and improving the overall stability and continuous optimization capability of the system. This invention integrates intelligent technologies such as graph neural network computing, meta-learning, and reinforced fuzzy control to construct a closed-loop optimization system from state perception and policy generation to dynamic execution, which significantly enhances the high-performance operation capability of 5G communication modules in multi-carrier and multi-scenario environments. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating an optimization method for a high-performance 5G communication module according to the present invention.

[0041] Figure 2 This is a flowchart illustrating step S3 provided in an embodiment of the present invention.

[0042] Figure 3 This is a flowchart illustrating step S4 provided in an embodiment of the present invention. Detailed Implementation

[0043] Please see Figure 1-3 As shown, this invention relates to an optimization method for a high-performance 5G communication module.

[0044] Example

[0045] An optimization method for a high-performance 5G communication module includes the following steps:

[0046] S1: Obtain the operating parameters of the 5G communication module in multi-carrier concurrent state; the operating parameters include transmit power, intermodulation interference, crystal oscillator offset, voltage-current response, board temperature and instantaneous throughput.

[0047] In one embodiment, a 5G communication module is deployed on an experimental test platform, and a base station simulator supporting NR (New Radio) multi-carrier aggregation is configured. Multiple carriers are simultaneously activated via a scheduling tool. During module operation, the following parameter data are collected:

[0048] Transmit power: obtained by sampling through the power detection module of the RF transceiver chip;

[0049] Intermodulation interference: Real-time monitoring was conducted in the nonlinear mixing frequency band using a spectrum analyzer;

[0050] Crystal Oscillator Offset: The difference between the crystal oscillator frequency and the reference frequency is acquired using a high-precision clock offset measurement unit;

[0051] Voltage-current response: The response curves of supply voltage and current as the number of carriers enabled are collected by the PMU module (Power Management Unit);

[0052] Board-level temperature: The temperature distribution in key areas of the PCB is monitored using a thermistor array;

[0053] Instantaneous throughput: The actual amount of data transmitted per millisecond (TTI) is extracted through MAC layer scheduling feedback information.

[0054] The data period is 10ms, and the continuous collection time is 3 minutes, forming a high time resolution dataset of operating parameters.

[0055] S2: Based on the operating parameters, the state graph is clustered and divided using the graph polarization spectrum clustering algorithm to construct a feature graph model. In the graph, nodes represent the operating parameter states, and edge weights represent the degree of coordinated change between parameters.

[0056] Step S2 includes the following steps:

[0057] Based on the aforementioned operating parameters, a time-series state diagram reflecting the parameter evolution process is constructed using parameter normalization and time window segmentation strategies. Nodes represent parameter states in different time slices, and edge weights characterize the magnitude of collaborative changes between adjacent parameter states.

[0058] Specifically, all operating parameters are processed uniformly across intervals, such as normalizing various indicators to the [0, 1] interval; a maximum-minimum scaling method is used, and the timestamps corresponding to the original values ​​are retained for easy restoration. A sliding time window is set (e.g., a window of 2 seconds, sliding once every 500ms); the values ​​of each operating parameter at the current moment are extracted within each window to form a parameter state vector, which serves as a node in the graph. Each node in the graph represents the parameter state within a certain time window; adjacent time windows are connected by edges, and the initial edge weights are measured using the Euclidean distance of the parameter vectors (i.e., the smaller the difference, the larger the edge weight).

[0059] By calculating the correlation coefficient matrix for each parameter node pair in the time-series state diagram and combining it with the maximum information coefficient to enhance the nonlinear collaborative characteristics, a weighted multi-attribute state map is generated.

[0060] Specifically, for all operating parameters within each time window, a set of state vectors is constructed; the Pearson correlation coefficient (a measure of linear synchronization) between different parameter pairs is calculated; for example, current fluctuations and temperature rise trends are often positively correlated; and a parameter correlation matrix is ​​generated to provide a basis for the initial edge weights of the state diagram.

[0061] For parameter relationships that cannot be explained by linear functions (such as abrupt changes in crystal oscillator frequency offset and bit error rate), the maximum information coefficient (MIC) is used to measure their correlation. MIC can capture nonlinear patterns, such as step changes or exponential trends, and improve the ability to model the state of complex scenarios. Finally, the edge weight between each pair of nodes is coupled by linear correlation and nonlinear information entropy, and the edge is given multiple attribute labels.

[0062] Each node is bound to the following: a parameter status value vector (such as the 6 operating parameter values ​​at a certain moment); a timestamp tag (used for timing sorting); and the ID of the physical module to which it belongs (such as PA, LO, power management module).

[0063] Each edge is bound to the following: collaborative weight (calculated based on correlation and MIC); parameter type label pair (e.g., temperature-current); evolution direction label (whether it is in the rising or falling range of parameter change);

[0064] This ultimately forms a multidimensional weighted, time-series oriented state graph, laying a precise foundation for subsequent graph analysis.

[0065] Based on the weighted multi-attribute state graph, the graph polarization spectral clustering algorithm is used to perform spectral decomposition and feature space mapping on the node set. By introducing polarization operators to adjust the spectral clustering boundary, the parameter states with high noise and weak coupling are clustered and the parameter state clustering labels are output.

[0066] Specifically, the normalized Laplacian matrix is ​​used to perform feature spectrum decomposition on the entire graph; the first K principal component vectors of the graph are extracted, and the high-dimensional node states are projected into the low-dimensional spectral space; the Euclidean distance between the low-dimensional vectors reflects the structural similarity between the original states.

[0067] In traditional spectral clustering, cluster boundaries are easily blurred due to high noise interference. The system introduces a polarization operator to adjust the weights of cluster centers and boundaries. The polarization operator assigns lower participation weights to boundary nodes based on multiple indicators such as local node density, edge strength change rate, and state mutation index. For state nodes with strong noise, drastic fluctuations, but no stable pattern, they are preferentially separated into independent subclasses or subjected to soft clustering.

[0068] Each node is labeled with a precise clustering tag (e.g., Cluster-3, Cluster-7); nodes in the same state exhibit a typical change pattern (e.g., "low power + high temperature + frequency deviation increase"); the clustering results are visualized as multiple color-coded subgraphs to distinguish different operating behavior patterns and abnormal operating conditions.

[0069] Based on the parameter state clustering labels, the core nodes and edge weight distribution in each subgraph are extracted, a feature graph model containing node features, edge collaborative weights and clustering labels is constructed, and a structured feature graph model is output.

[0070] Specifically, in each cluster subgraph, representative core nodes are selected based on indicators such as degree centrality, mean edge weight, and time stability. These core nodes correspond to key control point parameters in actual operation, such as "temperature peak" and "power jump point".

[0071] The distribution of collaborative weights for all edges in each subgraph is statistically analyzed, including the following:

[0072] Mean edge weight: represents the overall coupling strength within this type of state;

[0073] Boundary weight variance: reflects the stability of the relationship between parameters within this type of state;

[0074] Maximum path weight: Identifies the most coupled path in the subgraph, which is used for subsequent path optimization strategies.

[0075] Each cluster subgraph is encapsulated as a structure, which includes the following:

[0076] Cluster labels (identify subgraph categories);

[0077] Core node indexes (such as nodes 3, 8, and 15);

[0078] Node vector (parameter values ​​+ module location information);

[0079] Edge connection matrix (records whether an edge exists and its weight);

[0080] All subgraphs are combined to form the final feature map model library, which serves as the input basis for subsequent graph attention mechanism modeling and path recognition.

[0081] Furthermore, the formula for constructing the feature map model is as follows:

[0082] ;

[0083] in, This represents the strength of the co-evolution among the running states in the constructed feature map; and These represent the normalized operating parameter vectors for the i-th and j-th time windows, respectively. The parameters include transmit power P, intermodulation interference I, and crystal oscillator offset. Voltage-current response V / I, plate temperature T, instantaneous throughput R; Represents the squared Euclidean distance between the multi-parameter states of the i-th and j-th time slices; This represents the maximum information coefficient, used to capture the degree of nonlinear coupling between state pairs; This indicates points in the parameter variation graph where there are local curvature or abnormally drastic changes. This represents the global variance between the states of the operating parameters; The normalized scaling factor represents the maximum information coefficient; , and These represent the weighting coefficients of the Euclidean distance term, the nonlinear correlation term, and the spectral perturbation term, respectively.

[0084] The specific calculation formula is as follows:

[0085] ;

[0086] in, and This indicates that the data is divided into discrete grids; This represents the mutual information of the joint grid.

[0087] S3: Based on the feature graph model, a multi-hop path optimization algorithm with graph attention mechanism is used to jointly model the node weights in the path and score the path, identify high-weight paths and key parameter nodes that affect communication performance, and generate a critical path strategy graph containing node configuration relationships and path control logic.

[0088] Step S3 includes the following steps:

[0089] Based on the feature graph model, a multi-hop path optimization algorithm based on graph attention mechanism is used to jointly model the state features and edge collaboration relationships of nodes in the graph within the multi-hop adjacency range, and construct a multi-hop path structure with cross-level information perception capability.

[0090] In the multi-hop path structure, an attention mechanism is used to dynamically weight the differences in edge weight distribution and state features between nodes within different hop count ranges to generate a node attention distribution map.

[0091] For each target node, the system extracts its 1-3 hop neighbor nodes from the feature graph model. The feature combination of each hop neighbor node includes: node state value (such as frequency shift, current anomaly, etc.); edge weight (cooperative change intensity) between the target node and the target node; hop number label (used to guide hierarchical sensitive modeling); all features are concatenated and input into the attention mechanism of the graph neural network to form the feature interaction vector between node pairs.

[0092] The system introduces a graph attention mechanism (such as the GATv2 structure): each target node calculates an attention coefficient for each of its neighboring nodes; the attention coefficient reflects the "weight of neighboring nodes in the target node's decision-making". Low-hop-count (e.g., 1-hop) neighbors are given higher initial weights, but 2-hop or 3-hop neighbors with high edge cooperation strength can also increase their weights; when a node's state changes drastically (e.g., a sudden increase in power or a drastic temperature fluctuation), its attention weight is amplified; the attention weights are updated and iterated over time, and the system refreshes the attention distribution of all nodes every 1 second.

[0093] Each node outputs a weight vector corresponding to the attention coefficient of its neighboring nodes; the system organizes the attention weights of all nodes in the entire graph in a matrix manner to form a node attention distribution map; this distribution map is not only used for subsequent path scoring, but also for analyzing clusters of communication performance bottlenecks.

[0094] Based on the node attention distribution map, all reachable multi-hop paths are scored. Taking into account the attention weight of each node in the path, the edge collaboration strength and the impact of the number of hops, a set of high-weight paths with significant path scores is selected.

[0095] Specifically, for all reachable paths (limited to 3 hops) originating from key nodes in the graph, the system calculates the following three key metrics in sequence:

[0096] Node average attention score: The average weight of all nodes in the path in the current attention graph;

[0097] Total edge collaboration strength: the sum of the collaboration weights of all edges in the path;

[0098] Hop count penalty factor: The longer the path, the larger this factor becomes, reducing the priority of redundant paths.

[0099] The system calculates a comprehensive path score based on a weighted combination of the three factors mentioned above. A high-scoring path indicates that it is highly important in the state evolution and that the nodes are closely coordinated; the shorter the path, the easier it is for real-time control.

[0100] For each key node (such as a node with abnormal indicators), the system starts from it and performs a depth-first path search (DFS); for each 2-3 hop path, the node set and edge set are extracted in sequence; based on the current attention distribution graph and edge weight graph, the score is calculated in real time; all path scores are uniformly sorted, and the top N paths (such as Top 10) are selected as the high-weight path set.

[0101] High-weight paths are stored in a graph structure, with nodes identified by numbers and edges stored by attributes. The path structure and score values ​​can be presented in a visual interface, which facilitates subsequent model debugging and policy rule optimization. The system periodically caches the set of high-weight paths and tracks their evolution trend to discover new key evolutionary paths.

[0102] Based on the set of high-weight paths, key parameter nodes in the paths are extracted, their positions and roles in the path structure and their dependencies on other nodes are analyzed, and a critical path strategy graph that integrates control logic and priority relationships is constructed.

[0103] Specifically, in each high-priority path, the system identifies the following role nodes:

[0104] Source node: The location where the state change first occurs, often a node where temperature or power parameters change abruptly;

[0105] Transmission node: An intermediate node connecting the source and the target. It has a high edge weight, and its state changes have a significant impact on the terminal.

[0106] Target node: A parameter node that exhibits significant changes in communication performance (such as decreased throughput and increased bit error rate);

[0107] The system records the above nodes as key parameter nodes and categorizes them into a "cause-transmission-result" model structure.

[0108] Graph traversal and causal chain analysis are performed on the path structure between key nodes to identify the following relationships: whether there are strong dependent paths (such as temperature rise always being accompanied by crystal oscillator drift); whether there are state feedback loops (such as parameter adjustment affecting itself in reverse); and generating control rules based on the analysis: such as "if A rises, it causes B to be abnormal, and B's abnormality affects C's performance", then the system generates a control logic chain: "monitor A, adjust B, protect C".

[0109] A strategy diagram consists of the following elements:

[0110] Node set: contains the unique identifier, parameter type, and logical role (cause, result) of key nodes;

[0111] Edge set: includes edge start and end nodes, control direction, and dependency level;

[0112] Path priority: Assigned high, medium, or low levels based on path scoring results;

[0113] Strategy tags: such as "power regulation type", "temperature control protection type", "frequency deviation recovery type", etc.;

[0114] The policy graph will serve as input for subsequent policy transfer models (such as meta-learning models) or execution engines to perform intelligent parameter scheduling.

[0115] Furthermore, the formula for the multi-hop path optimization algorithm is as follows:

[0116] ;

[0117] in, H represents the multi-hop path score from node p to node q, used to construct the critical path strategy graph; H represents the maximum number of hops limit, which controls the hierarchical depth of path modeling. This represents the global adjustment factor for the h-th hop, reflecting the importance of path aggregation in the current hop layer; This represents the set of h-th hop neighbors of the current node p; Indicates attention weight; This represents the edge collaboration weight, where the edge weight indicates the degree of collaborative change among the operating parameters; It represents instantaneous throughput and is used to measure the dynamic impact of a node on communication performance in the h-th hop layer; Indicates the number of hops in the path topology; This represents the standard deviation of the plate-level temperature of the nodes associated with node p in the h-th hop; This represents the hop count decay factor.

[0118] The calculation formula is as follows:

[0119] ;

[0120] in, This represents the instantaneous throughput of node k in the h-th time window; and This indicates adjacent time windows, used in conjunction with time window segmentation strategies;

[0121] The calculation formula is as follows:

[0122] ;

[0123] in, This represents the plate-level temperature sample value of the h-th hop neighbor node k; This represents the average plate-level temperature of the h-th hop neighbor of node p.

[0124] S4: Using the critical path strategy graph as input, based on the scene characteristics in a multi-communication environment, a meta-learning algorithm based on graph structure alignment is adopted to construct a strategy graph transfer model. Through the mapping function, the generalized scheduling and optimal transfer of parameter configuration under different load conditions are realized, and the optimal parameter configuration strategy is generated.

[0125] Step S4 includes the following steps:

[0126] Based on the critical path strategy graph, extract high-weight path sequences, key parameter node configuration combinations, and control dependencies between paths to construct a graph structure sample set associated with specific communication scenarios, forming a source scenario knowledge graph.

[0127] Specifically, the following core structural information is extracted from the critical path strategy graph generated by S3:

[0128] High-weighted path sequence: such as "transmit power fluctuation → crystal oscillator frequency deviation → throughput decrease";

[0129] Key parameter node configuration combinations: Node attributes include parameter value range, adjustment direction, and device module ID;

[0130] Inter-path control dependencies: For example, if two paths share a key node, they need to be jointly controlled.

[0131] The above structural information is organized into graph structure samples, where nodes represent parameter configuration entities and edges represent dependency control paths. Each sample is bound to a label of the current communication scenario, such as "high-interference urban area" or "medium-density rural area". The policy graph structure sample sets collected under multiple different typical communication environments are encapsulated into a source scenario knowledge graph library for subsequent alignment learning.

[0132] In the target communication environment, by combining the characteristics of real-time network load level, interference source distribution and terminal connection density, a graph structure representation of the current scene is constructed, and scene labels and path control demand vectors are initialized.

[0133] Specifically, based on the monitoring module deployed on the base station side, the system periodically collects the following communication environment indicators to form the input feature set of the target scenario:

[0134] Network load level: Calculated based on resource block (PRB) utilization, such as: idle <30%, medium 30~70%, overload >70%;

[0135] Interference source distribution: The number and intensity of interference sources are clustered by measuring the signal strength and SINR value of neighboring cells;

[0136] Terminal connection density: The number of UEs currently connected to the cell is counted, and the density level is identified by combining the UE's movement speed.

[0137] The process of constructing the target scene graph structure is as follows:

[0138] Node construction: One node for each type of environmental factor, for example:

[0139] Node A: Load level = High; Node B: Interference = Medium; Node C: Connection density = High.

[0140] Edge definition: Based on historical data analysis, add edges to represent the impact of interactions, such as how high load can easily lead to increased interference;

[0141] Node attributes: Each node carries information such as the corresponding factor's numerical range, trend of change, and mutation probability;

[0142] Image Labels: The entire image is assigned a unique scene label (e.g., "high load scene in dense urban areas") and scene objective (e.g., "improve spectrum utilization and reduce bit error rate").

[0143] Based on the current environmental analysis results and performance requirements, the system automatically generates a set of parameter adjustment priorities: such as frequency stability > power efficiency > temperature suppression; this priority vector will serve as the target guide for subsequent mapping functions to adjust the configuration of key nodes.

[0144] A graph-structure aligned meta-learning algorithm is adopted to perform structural mapping between the knowledge graph in the source scene and the target scene graph, and to generate the optimal path control mapping function in the target scene through a fast parameter adaptation mechanism;

[0145] Specifically, graph embedding algorithms (such as GraphSAGE or GINE) are used to generate graph structure vectors for the source and target scene graphs respectively; similarity calculations are performed on the graph structure vectors to measure the structural differences between the two graphs (such as differences in the number of nodes and edge connectivity); and a graph editing distance metric is introduced to quantify the minimum set of operations required to transform the source graph into the target graph.

[0146] Build cross-scenario migration strategies using the MAML algorithm framework:

[0147] Meta-task definition: Each source scene graph structure and its control result are defined as a training task;

[0148] Inner loop update: For each task, the policy model is trained within a few iterations, and the direction of parameter update is recorded;

[0149] Outer loop optimization: Learn how to initialize a policy network so that it can converge quickly to new tasks;

[0150] After training, the meta-model can accept any new scene graph structure as input and quickly adjust the policy network to generate an adaptive mapping function.

[0151] The final meta-learning model outputs a structure alignment mapping function, with the input being the structural features of the target scene graph and the output being the policy configuration adjustment instructions;

[0152] This function has the following capabilities: adjusting the range of node parameter values ​​in the source strategy graph based on the characteristics of the target scene nodes; modifying the control intensity and triggering conditions of the path based on the structure mapping results; and outputting a set of efficient path combinations adapted to the target scene.

[0153] Based on the aforementioned control mapping function, the critical path strategy graph in the source scenario is mapped to the target communication scenario, realizing the structural migration of the key parameter configuration relationship and outputting the optimal parameter configuration strategy.

[0154] Specifically, after the graph structure alignment is completed, the system automatically performs the following strategy graph migration operation:

[0155] Map the critical path in the source graph to the functional area in the target graph (e.g., project the "frequency compensation path" to the "interference-dense node subgraph"); update the parameter configuration range of the critical nodes (e.g., adjust the original power limit of 2025dBm to 1822dBm); dynamically replace low-fitness paths or nodes, and add supplementary paths based on the target scenario.

[0156] The optimal parameter configuration strategies output include:

[0157] Parameter dimension table: Lists the parameter names, optimal settings, adjustment ranges, and limit boundaries of key nodes;

[0158] Control logic diagram: Based on the path, control logic in the form of "IF-THEN" is generated, such as: IF High interference level & high connection density THEN Reduce transmit power by 2dBm;

[0159] Policy application conditions: The system automatically marks which load intervals, time periods, or device types the policy applies to;

[0160] Set up real-time monitoring and feedback. If the performance drops beyond the threshold after the strategy is executed, activate the alternative strategy or roll back the old strategy. Add a strategy scoring mechanism to dynamically evaluate the effect of the adjustment and use it for subsequent strategy updates and iterations.

[0161] After the strategy is pushed to the 5G communication module, the control engine dynamically adjusts the module parameters according to the control strategy; the module operation results are uploaded in real time, and the following data are fed back for strategy backtracking and optimization: throughput change trend; BER change curve; parameter response latency and stabilization time; the feedback data is compared and analyzed with the effect of the initial strategy for reference in the next fine-tuning of the meta-model.

[0162] Furthermore, the formula for the optimal path control mapping function is as follows:

[0163] ;

[0164] in, Let N represent the optimal path control mapping function under the target communication scenario R; N represents the number of source scenario samples participating in the control learning. This represents the critical path parameter vector extracted from the nth source scene; This represents the expected optimal parameter configuration result for the nth sample in the target scenario; R represents the critical path strategy graph structure constructed in the nth source communication scenario; R represents the scenario graph structure in the target communication environment. This represents a path control mapping function; Represents the regularization term of function J; This represents the regularity coefficient.

[0165] S5: Apply the optimal parameter configuration strategy to the actual operation of the 5G communication module, collect the real-time operating status of the module as a fuzzy input variable, construct a reinforced fuzzy inference network model, evaluate the deviation and adaptability of the current execution strategy through fuzzy inference rules combined with reinforcement learning reward mechanism, and fine-tune the execution parameters based on performance feedback.

[0166] Specifically, the enhanced fuzzy inference network model is based on the core parameter nodes in the critical path strategy graph as the input dimension, collecting state indicators such as transmit power offset, voltage and current fluctuation, crystal oscillator frequency offset and temperature change, constructing a fuzzy input vector, and completing the fuzzification process through membership functions; combining the node dependency relationships in the strategy graph to generate a fuzzy rule base, and introducing a reinforcement learning mechanism to dynamically adjust the weight of fuzzy rules and the priority of execution strategies;

[0167] The reinforcement learning reward mechanism is specifically constructed based on the dynamic feedback of communication performance indicators, including the throughput improvement value, the reduction of bit error rate and the handover delay convergence rate as core reward factors, and generating real-time performance scores by combining the weighting function of each indicator.

[0168] Specifically, the generated optimal parameter configuration strategy is deployed in the operation and control unit of the 5G communication module. After system startup, the module's operational status data under multi-carrier concurrent communication is collected in real time, mainly including:

[0169] Transmit power offset: Indicates the difference between the current transmit power and the target transmit power.

[0170] Voltage and current fluctuations: reflect the stability of the power supply system and capture instantaneous changes caused by environmental or load fluctuations.

[0171] Crystal frequency offset: Used to monitor the accuracy of the clock system and has a significant impact on frequency synchronization stability.

[0172] Temperature variation: Record the heat accumulation state of key components and assess their thermal sensitivity impact on module performance.

[0173] The four types of state indicators mentioned above are mapped to fuzzy input variables. Based on empirical rules or expert systems, membership functions (such as "three-stage" low-medium-high membership levels) are set to fuzzify each variable, forming a fuzzy input vector.

[0174] Based on the dependencies between the core parameter nodes in the critical path policy graph, a fuzzy rule base is constructed. Each rule describes the optimal control behavior under a specific combination of inputs, such as: "If the transmit power offset is high and the voltage fluctuation is medium, then reduce the power control weight." These rules serve as the foundation of the inference engine for real-time policy determination.

[0175] A reinforcement learning module based on temporal decision-making is embedded in the inference engine. This module continuously tracks the deviation between the fuzzy inference results and the actual performance. By comparing the system performance after the fuzzy output policy is executed with the expected benefits, it dynamically updates the weights of each fuzzy rule and adjusts the policy selection priority during the inference process.

[0176] Establish a performance feedback channel to collect three key communication performance indicators in real time after the system is running:

[0177] Throughput improvement value: Measures the degree to which the current strategy enhances data transmission capabilities.

[0178] Bit error rate reduction: reflects the level of improvement in link quality.

[0179] Handover delay convergence rate: Evaluate the response efficiency during multi-band handover or carrier aggregation.

[0180] Each indicator is combined with a weighted coefficient to generate a comprehensive performance score, which serves as an immediate reward signal in reinforcement learning, driving the policy model to converge toward a better direction.

[0181] The system compares the comprehensive performance score with the preset threshold. If there is a significant deviation in the current strategy execution effect, the module will trigger an adaptive parameter fine-tuning process to automatically adjust the core parameter configuration in the critical path until the system runs stably in the optimal state region, thus achieving closed-loop optimization control.

[0182] In summary, this invention, by introducing a high-temporal-resolution data acquisition mechanism and fusing monitoring of multi-source heterogeneous parameters (such as power, frequency offset, temperature, and current), constructs a state map of operating parameters under realistic and complex working conditions. This significantly improves the system's perception accuracy of dynamic communication states, providing solid data support for subsequent control. The joint mapping using the maximum information coefficient (MIC) and linear correlation indices not only enhances the ability to uncover nonlinear parameter coupling relationships but also achieves highly robust modeling of complex physical properties, effectively identifying key state change patterns that are difficult to discover using traditional linear methods.

[0183] This invention introduces a polarization operator to effectively filter out abnormal nodes and achieve soft clustering in the face of high noise and high disturbance states, ensuring higher stability and resolution of the state graph clustering results, and effectively supporting the accuracy and reliability of subsequent critical path construction. By combining graph neural networks with a multi-hop attention mechanism, the dependency modeling capability between state nodes across levels and hop counts is enhanced, effectively identifying core paths and key parameter nodes affecting communication performance, providing a structured and high-value policy graph for regulation strategy generation. By aligning the source scenario knowledge graph with the target communication environment's graph structure, and combining graph embedding and MAML meta-learning algorithms, the system possesses the ability to "adapt quickly with few samples," enabling rapid generation of adaptive optimal parameter configuration strategies under rapidly changing communication environments (such as carrier switching and sudden changes in network load).

[0184] This invention performs structural mapping and parameter transfer on the critical path strategy graph, effectively enabling the transfer and reuse of control knowledge across different scenarios and devices, reducing system development and optimization costs, and improving the engineering versatility and reusability of strategy design. It integrates fuzzy control and reinforcement learning to construct an adaptive inference network model, and uses communication performance feedback (throughput, bit error rate, switching latency) for dynamic parameter fine-tuning, achieving continuous optimization of the optimal strategy and online self-learning capabilities, ensuring the long-term performance stability and intelligent response capabilities of the system.

[0185] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An optimization method for a high-performance 5G communication module, characterized in that, The following steps are involved: S1: Obtain the operating parameters of the 5G communication module in multi-carrier concurrent state; S2: Based on the operating parameters, the state graph is clustered and divided using the graph polarization spectrum clustering algorithm to construct a feature graph model. In the graph, nodes represent the operating parameter states, and edge weights represent the degree of coordinated change between parameters. S3: Based on the feature graph model, a multi-hop path optimization algorithm with graph attention mechanism is used to jointly model the node weights in the path and score the path, identify high-weight paths and key parameter nodes that affect communication performance, and generate a critical path strategy graph containing node configuration relationships and path control logic. S4: Using the critical path strategy graph as input, based on the scene characteristics in a multi-communication environment, a meta-learning algorithm based on graph structure alignment is adopted to construct a strategy graph transfer model. Through the mapping function, the generalized scheduling and optimal transfer of parameter configuration under different load conditions are realized, and the optimal parameter configuration strategy is generated. S5: Apply the optimal parameter configuration strategy to the actual operation of the 5G communication module, collect the real-time operating status of the module as a fuzzy input variable, construct a reinforced fuzzy inference network model, evaluate the deviation and adaptability of the current execution strategy through fuzzy inference rules combined with reinforcement learning reward mechanism, and fine-tune the execution parameters based on performance feedback.

2. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, The operating parameters include transmit power, intermodulation interference, crystal oscillator offset, voltage-current response, board temperature, and instantaneous throughput.

3. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, Step S2 includes the following steps: Based on the aforementioned operating parameters, a time-series state diagram reflecting the parameter evolution process is constructed using parameter normalization and time window segmentation strategies. Nodes represent parameter states in different time slices, and edge weights characterize the magnitude of collaborative changes between adjacent parameter states. By calculating the correlation coefficient matrix for each parameter node pair in the time-series state diagram and combining it with the maximum information coefficient to enhance the nonlinear collaborative characteristics, a weighted multi-attribute state map is generated. Based on the weighted multi-attribute state graph, the graph polarization spectral clustering algorithm is used to perform spectral decomposition and feature space mapping on the node set. By introducing polarization operators to adjust the spectral clustering boundary, the parameter states with high noise and weak coupling are clustered and the parameter state clustering labels are output. Based on the parameter state clustering labels, the core nodes and edge weight distribution in each subgraph are extracted, a feature graph model containing node features, edge collaborative weights and clustering labels is constructed, and a structured feature graph model is output.

4. The optimization method for a high-performance 5G communication module according to claim 3, characterized in that, The formula for constructing the feature map model is as follows: ； in, This represents the strength of the co-evolution among the running states in the constructed feature map; and These represent the normalized runtime parameter vectors within the i-th and j-th time windows, respectively; Represents the squared Euclidean distance between the multi-parameter states of the i-th and j-th time slices; Indicates the maximum information coefficient; This indicates points in the parameter variation graph where there are local curvature or abnormally drastic changes. This represents the global variance between the states of the operating parameters; The normalized scaling factor represents the maximum information coefficient; , and These represent the weighting coefficients of the Euclidean distance term, the nonlinear correlation term, and the spectral perturbation term, respectively.

5. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, Step S3 includes the following steps: Based on the feature graph model, a multi-hop path optimization algorithm based on graph attention mechanism is used to jointly model the state features and edge collaboration relationships of nodes in the graph within the multi-hop adjacency range, and construct a multi-hop path structure with cross-level information perception capability. In the multi-hop path structure, an attention mechanism is used to dynamically weight the differences in edge weight distribution and state features between nodes within different hop count ranges to generate a node attention distribution map. Based on the node attention distribution map, all reachable multi-hop paths are scored. Taking into account the attention weight of each node in the path, the edge collaboration strength and the impact of the number of hops, a set of high-weight paths with significant path scores is selected. Based on the set of high-weight paths, key parameter nodes in the paths are extracted, their positions and roles in the path structure and their dependencies on other nodes are analyzed, and a critical path strategy graph that integrates control logic and priority relationships is constructed.

6. The optimization method for a high-performance 5G communication module according to claim 5, characterized in that, The formula for the multi-hop path optimization algorithm is as follows: ； in, H represents the multi-hop path score from node p to node q, used to construct the critical path strategy graph; H represents the maximum number of hops limit, which controls the hierarchical depth of path modeling. This represents the global adjustment factor for the path at the h-th hop; This represents the set of h-th hop neighbors of the current node p; Indicates attention weight; This represents the edge collaboration weight, where the edge weight indicates the degree of collaborative change among the operating parameters; It represents instantaneous throughput and is used to measure the dynamic impact of a node on communication performance in the h-th hop layer; Indicates the number of hops in the path topology; This represents the standard deviation of the plate-level temperature of the nodes associated with node p in the h-th hop; This represents the hop count decay factor.

7. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, Step S4 includes the following steps: Based on the critical path strategy graph, extract high-weight path sequences, key parameter node configuration combinations, and control dependencies between paths to construct a graph structure sample set associated with specific communication scenarios, forming a source scenario knowledge graph. In the target communication environment, by combining the characteristics of real-time network load level, interference source distribution and terminal connection density, a graph structure representation of the current scene is constructed, and scene labels and path control demand vectors are initialized. A graph-structure aligned meta-learning algorithm is adopted to perform structural mapping between the knowledge graph in the source scene and the target scene graph, and to generate the optimal path control mapping function in the target scene through a fast parameter adaptation mechanism; Based on the aforementioned control mapping function, the critical path strategy graph in the source scenario is mapped to the target communication scenario, realizing the structural migration of the key parameter configuration relationship and outputting the optimal parameter configuration strategy.

8. The optimization method for a high-performance 5G communication module according to claim 7, characterized in that, The formula for the optimal path control mapping function is as follows: ； in, Let N represent the optimal path control mapping function under the target communication scenario R; N represents the number of source scenario samples participating in the control learning. This represents the critical path parameter vector extracted from the nth source scene; This represents the expected optimal parameter configuration result for the nth sample in the target scenario; R represents the critical path strategy graph structure constructed in the nth source communication scenario; R represents the scenario graph structure in the target communication environment. This represents a path control mapping function; Represents the regularization term of function J; This represents the regularity coefficient.

9. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, The enhanced fuzzy inference network model is specifically based on the core parameter nodes in the critical path strategy graph as input dimensions, collecting state indicators such as transmit power offset, voltage and current fluctuations, crystal oscillator frequency offset and temperature changes, constructing a fuzzy input vector, and completing the fuzzification process through membership functions; combining the node dependencies in the strategy graph to generate a fuzzy rule base, and introducing a reinforcement learning mechanism to dynamically adjust the weights of fuzzy rules and the priority of execution strategies.

10. The optimization method for a high-performance 5G communication module according to claim 1, characterized in that, The reinforcement learning reward mechanism is specifically constructed based on the dynamic feedback of communication performance indicators, including the throughput improvement value, the reduction of bit error rate and the handover delay convergence rate as core reward factors, and generating real-time performance scores by combining the weighting function of each indicator.

Images