Dynamic graph neural network modeling method for spatiotemporal big data

Abstract

The application provides a dynamic graph neural network modeling method for space-time big data, and relates to the technical field of data processing.The method comprises the following steps: mapping network function entities into topological vertices and mapping topological correlation characteristics into weighted transmission links, driving topological reconstruction through a signaling event trigger sequence, and generating a communication network topological model; inputting the communication network topological model into a dynamic graph neural network, performing state feature space aggregation of a topological vertex neighborhood through a space-time feature extraction layer, and fusing time evolution dependence of a historical topological sequence to generate a network node space-time state tensor; based on the network node space-time state tensor, a particle swarm optimization algorithm is used to calculate a network-wide risk level quantization topological feature, and network resource strategy optimization is dynamically performed to suppress end-to-end risk conduction.The application improves the adaptability of the network to dynamic scenarios.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dynamic graph neural network modeling method for spatiotemporal big data. Background Technology

[0002] Communication networks are undergoing a critical evolution from static architecture to dynamic intelligence, with the core challenge being how to efficiently handle complex data coupled with spatiotemporal factors. Traditional methods, some of which rely on static topology modeling and empirical rule optimization, struggle to cope with the dynamic migration of network functional entities, multipath propagation of wireless signals, and the cascading effects of signaling events.

[0003] The ultra-dense deployment of 5G / 6G networks leads to frequent and dynamic changes in the spatial location of network functional entities (such as base stations and user terminals). At the same time, wireless signal propagation is affected by environmental factors such as terrain and weather, exhibiting strong spatiotemporal correlation. For example, in intelligent transportation scenarios, the high-speed movement of vehicles can trigger rapid switching of signal coverage areas. Some traditional static topology models cannot reflect these changes in a timely manner, resulting in resource allocation mismatch and end-to-end latency degradation. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a dynamic graph neural network modeling method for spatiotemporal big data, which improves the network's adaptability to dynamic scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a dynamic graph neural network modeling method for spatiotemporal big data, the method comprising:

[0007] Step 1: Real-time acquisition of multi-dimensional spatiotemporal data streams in the communication network, including spatial location sampling sets of network functional entities, topological association characteristics, and signaling event triggering sequences;

[0008] Step 2: Analyze the spatial location sampling set to extract the spatial distribution characteristics of network functional entities; perform wireless propagation modeling based on the spatial distribution characteristics, select the extreme value of the sampling density as the main coverage anchor point, and generate the main direction of bidirectional signal radiation; calculate the angle between the main directions of bidirectional signal radiation to divide the coverage area; dynamically deploy signal acquisition devices at the boundary of the area; obtain multipath propagation characteristics based on the time-domain activation sequence of the signal acquisition devices and calculate the channel propagation compensation coefficient;

[0009] Step 3: Integrate the channel propagation compensation coefficient and topology correlation characteristics to dynamically construct a communication network topology model. Map network functional entities to topology vertices and topology correlation characteristics to weighted transmission links. Then, drive topology reconstruction through signaling event trigger sequences to generate the communication network topology model.

[0010] Step 4: Input the communication network topology model into the dynamic graph neural network, perform state feature space aggregation of the topology vertex neighborhood through the spatiotemporal feature extraction layer, and fuse the temporal evolution dependency of the historical topology sequence to generate the spatiotemporal state tensor of the network nodes.

[0011] Step 5: Based on the spatiotemporal state tensor of network nodes, use the particle swarm optimization algorithm to calculate the quantitative topological representation of the risk level of the entire network, and dynamically execute network resource strategy optimization to suppress end-to-end risk transmission.

[0012] Furthermore, step 2 includes:

[0013] Step 21: Analyze the spatial location sampling set to calculate the spatial sampling point density distribution function, identify the local maxima of the density distribution function as the main coverage anchor points; calculate the signal propagation stability index distribution based on the main coverage anchor points, and select two complementary radiation directions that satisfy the preset threshold of stability index as the main bidirectional signal radiation directions;

[0014] Step 22: Measure the azimuth angle difference between the main directions of bidirectional signal radiation, construct a ring-shaped coverage area based on the azimuth angle difference, and deploy a cluster of signal acquisition devices in the boundary transition zone of the ring-shaped coverage area.

[0015] Step 23: Adjust the collaborative working cycle of the signal acquisition device cluster by setting the time-domain activation sequence, and extract multipath propagation features based on the spatial response characteristics of multipath reflection signals; generate channel propagation compensation coefficients by weighted fusion based on the time delay distribution discrete coefficient and attenuation path loss index of the multipath propagation features.

[0016] Furthermore, step 3 includes:

[0017] Step 31: Perform normalization processing on the channel propagation compensation coefficient to generate the link transmission reliability factor; parse the topological association characteristics to construct the initial adjacency matrix and inject the link transmission reliability factor into the corresponding matrix elements to generate the weighted adjacency matrix;

[0018] Step 32: Based on the weighted adjacency matrix, map the network functional entities to a set of topological vertices to construct the initial communication network topology model;

[0019] Step 33: Based on the initial communication network topology model, perform real-time topology reconstruction according to the signaling event trigger sequence. If the event type is node access, add a new vertex to the topology vertex set and expand the rows and columns in the weighted adjacency matrix; if the event type is link switching, update the weight value of the corresponding link in the weighted adjacency matrix; if the event type is node failure, delete the target vertex in the topology vertex set and remove the associated rows and columns from the weighted adjacency matrix.

[0020] Step 34: Combine the real-time reconstructed topology vertex set with the weighted adjacency matrix to generate a dynamic communication network topology model.

[0021] Furthermore, the communication network topology model is input into a dynamic graph neural network. A spatiotemporal feature extraction layer performs state feature space aggregation of the topology vertex neighborhoods and integrates the temporal evolution dependencies of historical topology sequences to generate spatiotemporal state tensors for network nodes, including:

[0022] Step 41: Input the dynamic communication network topology model into the spatiotemporal feature extraction layer of the dynamic graph neural network; determine the set of adjacent vertices for each topology vertex based on the weighted adjacency matrix in the dynamic communication network topology model; aggregate the state features of the adjacent vertex sets to generate the spatial feature vector of the topology vertices;

[0023] Step 42: Based on the spatial feature vector, extract the historical spatial feature sequence within the preset time window and generate time evolution dependent features through a temporal convolutional network;

[0024] Step 43: After concatenating the spatial feature vector with the temporal evolution-dependent feature vector, input the concatenation into the gated fusion unit to generate a spatiotemporal joint feature vector; organize all spatiotemporal joint feature vectors in vertex index order to obtain a three-dimensional tensor structure and generate the spatiotemporal state tensor of the network nodes.

[0025] Furthermore, based on the spatiotemporal state tensor of network nodes, a particle swarm optimization algorithm is used to calculate the quantified topological characteristics of the entire network risk level, and network resource strategy optimization is dynamically executed to suppress end-to-end risk transmission, including:

[0026] Step 51: Input the spatiotemporal state tensor of the network nodes into the particle swarm optimization algorithm, map each network node to a particle and initialize the particle position vector according to the feature value distribution, construct the end-to-end risk transmission assessment index, and calculate the risk transmission loss estimate for each particle position.

[0027] Step 52: Adjust the particle velocity vector based on the individual historical minimum loss position and the group minimum loss position, update the particle position vector based on the velocity vector and recalculate the risk transmission loss estimate to obtain the updated particle swarm state.

[0028] Step 53: Based on the updated particle swarm state, if the preset iteration termination threshold is reached, extract the minimum loss position vector of the swarm and perform position vector decoding to generate a quantitative topology representation of the risk level of the entire network; input the quantitative topology representation of the risk level into the resource scheduling decision-maker to generate a resource scheduling scheme.

[0029] Step 54: Execute the resource scheduling scheme and monitor the end-to-end risk transmission and attenuation trend, and adjust the particle swarm optimization parameters based on the monitoring results.

[0030] Furthermore, the spatial location sampling set is analyzed to calculate the spatial sampling point density distribution function, and the local maxima of the density distribution function are identified as the main coverage anchor points. Based on the main coverage anchor points, the signal propagation stability index distribution is calculated, and two complementary radiation directions that satisfy a preset threshold are selected as the main bidirectional signal radiation directions, including:

[0031] Step 221: Perform kernel density estimation on the spatial location sampling set to generate the probability density distribution surface of the spatial sampling points; detect the gradient change characteristics of the probability density distribution surface, and identify the extreme points where the gradient changes from positive to negative as local maxima.

[0032] Step 222: Select the location with the highest probability density value from the local maxima to determine the main coverage anchor point; divide the surrounding space into multiple spatial sectors at equal angles with the main coverage anchor point as the center; calculate the statistical dispersion of the reference signal received power in each spatial sector as a signal propagation stability index.

[0033] Step 223: Select spatial sectors with statistical dispersion lower than the preset stability threshold to obtain the sector center direction set, and calculate the azimuth difference of all direction pairs based on the set. Select the direction pair with the closest difference to the complementary state as the main direction of bidirectional signal radiation.

[0034] Furthermore, based on the weighted adjacency matrix, network functional entities are mapped to a set of topological vertices to construct an initial communication network topology model, including:

[0035] Step 321: Based on the row and column index identifier set of the weighted adjacency matrix, map each row and column index identifier to a topological vertex and assign vertex identifiers to generate a topological vertex set; traverse the non-zero elements of the weighted adjacency matrix with the topological vertex set as input and parse each non-zero element as a weighted transmission link, where the row index identifier is mapped to the source vertex identifier and the column index identifier is mapped to the destination vertex identifier, and the element value is mapped to the link weight value to generate a weighted transmission link set;

[0036] Step 322: Establish a graph structure vertex set based on the topological vertex set, and parse the topological edge set; establish a graph structure edge set based on the topological edge set and assign the link weight value to the weight attribute of the corresponding edge;

[0037] Step 323: Establish the connection relationship between vertices and edges based on the vertex set and weighted edge set of the graph structure, associate the source vertex identifier of the edge with the source vertex in the vertex set, and associate the destination vertex identifier with the destination vertex in the vertex set; generate the initial communication network topology model based on the vertex set, weighted edge set and vertex-edge connection relationship of the graph structure.

[0038] Furthermore, based on spatial feature vectors, historical spatial feature sequences within a preset time window are extracted and temporal evolution-dependent features are generated through a temporal convolutional network, including:

[0039] Step 421: Based on the timestamp index of the spatial feature vector, retrieve the historical spatial feature vector sequence within a preset time window from the spatiotemporal feature repository; perform time alignment processing on the historical spatial feature vector sequence;

[0040] Step 422: Input the time-aligned feature sequence into the first dilated convolutional layer of the temporal convolutional network. Perform local feature extraction by sliding the convolutional kernel of the first dilated convolutional layer along the time dimension to capture the short-term dependencies between adjacent time steps and obtain the short-term time-dependent feature sequence.

[0041] Step 423: The short-term time-dependent feature sequence is passed as input to the second dilated convolutional layer. The time receptive field is expanded by increasing the dilation coefficient of the second dilated convolutional layer, so that the convolutional kernel performs feature extraction across discontinuous time steps, captures the medium- and long-term dependencies across time steps, and obtains the medium- and long-term time-dependent feature sequence.

[0042] Step 424: Fuse the output features of the two dilated convolution layers and generate a time-evolution-dependent feature vector through time-dimensional pooling operation.

[0043] The above-described solution of the present invention has at least the following beneficial effects:

[0044] In this embodiment of the invention, by optimizing the main coverage anchor point and extracting multipath propagation features, selecting the extreme value of sampling density as the anchor point, dynamically deploying the boundary signal acquisition device, and combining the time-domain activation sequence to calculate the channel propagation compensation coefficient, spatial features such as multipath effects and changes in signal radiation direction can be accurately captured, thereby reducing channel modeling errors.

[0045] This invention achieves dynamic updates to the communication network topology model by triggering a sequence of signaling events to drive topology reconstruction. Compared to some traditional static topology models that cannot reflect the migration of network functional entities in real time, this method ensures that the topology model remains consistent with the actual network state. In dynamic scenarios, it reduces topology update latency and improves the network's adaptability to dynamic conditions. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the dynamic graph neural network modeling method for spatiotemporal big data provided in an embodiment of the present invention. Detailed Implementation

[0047] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0048] like Figure 1 As shown, embodiments of the present invention propose a dynamic graph neural network modeling method for spatiotemporal big data, the method comprising the following steps:

[0049] Step 1: Real-time acquisition of multi-dimensional spatiotemporal data streams in the communication network, including spatial location sampling sets of network functional entities, topological association characteristics, and signaling event triggering sequences;

[0050] Step 2: Analyze the spatial location sampling set to extract the spatial distribution characteristics of network functional entities; perform wireless propagation modeling based on the spatial distribution characteristics, select the extreme value of the sampling density as the main coverage anchor point, and generate the main direction of bidirectional signal radiation; calculate the angle between the main directions of bidirectional signal radiation to divide the coverage area; dynamically deploy signal acquisition devices at the boundary of the area; obtain multipath propagation characteristics based on the time-domain activation sequence of the signal acquisition devices and calculate the channel propagation compensation coefficient;

[0051] Step 3: Integrate the channel propagation compensation coefficient and topology correlation characteristics to dynamically construct a communication network topology model. Map network functional entities to topology vertices and topology correlation characteristics to weighted transmission links. Then, drive topology reconstruction through signaling event trigger sequences to generate the communication network topology model.

[0052] Step 4: Input the communication network topology model into the dynamic graph neural network, perform state feature space aggregation of the topology vertex neighborhood through the spatiotemporal feature extraction layer, and fuse the temporal evolution dependency of the historical topology sequence to generate the spatiotemporal state tensor of the network nodes.

[0053] Step 5: Based on the spatiotemporal state tensor of network nodes, use the particle swarm optimization algorithm to calculate the quantitative topological representation of the risk level of the entire network, and dynamically execute network resource strategy optimization to suppress end-to-end risk transmission.

[0054] In this embodiment of the invention, by optimizing the main coverage anchor point and extracting multipath propagation features, selecting the extreme value of sampling density as the anchor point, dynamically deploying the boundary signal acquisition device, and combining the time-domain activation sequence to calculate the channel propagation compensation coefficient, spatial features such as multipath effects and changes in signal radiation direction can be accurately captured, thereby reducing channel modeling errors.

[0055] This invention achieves dynamic updates to the communication network topology model by triggering a sequence of signaling events to drive topology reconstruction. Compared to some traditional static topology models that cannot reflect the migration of network functional entities in real time, this method ensures that the topology model remains consistent with the actual network state. In dynamic scenarios, it reduces topology update latency and improves the network's adaptability to dynamic conditions.

[0056] In a preferred embodiment of the present invention, step 2 involves parsing the spatial location sampling set to extract the spatial distribution characteristics of network functional entities; performing wireless propagation modeling based on the spatial distribution characteristics, selecting the extreme value of the sampling density as the main coverage anchor point, and generating the main direction of bidirectional signal radiation; calculating the angle between the main directions of bidirectional signal radiation to divide the coverage area; dynamically deploying signal acquisition devices at the area boundary; obtaining multipath propagation characteristics based on the time-domain activation sequence of the signal acquisition devices and calculating the channel propagation compensation coefficient, including:

[0057] Step 21: Analyze the spatial location sampling set to calculate the spatial sampling point density distribution function, identify the local maxima of the density distribution function as the main coverage anchor points; calculate the signal propagation stability index distribution based on the main coverage anchor points, and select two complementary radiation directions that satisfy the preset threshold of stability index as the main bidirectional signal radiation directions;

[0058] Step 22: Measure the azimuth angle difference between the main directions of bidirectional signal radiation, construct a ring-shaped coverage area based on the azimuth angle difference, and deploy a cluster of signal acquisition devices in the boundary transition zone of the ring-shaped coverage area.

[0059] Step 23: Adjust the collaborative working cycle of the signal acquisition device cluster by setting the time-domain activation sequence, and extract multipath propagation features based on the spatial response characteristics of multipath reflection signals; generate channel propagation compensation coefficients by weighted fusion based on the time delay distribution discrete coefficient and attenuation path loss index of the multipath propagation features.

[0060] In step 21 above, the average distance between each sampling point and the five nearest surrounding sampling points is calculated. If the deviation of a sampling point from the average distance exceeds 10%-15% of the average distance between sampling points, it is identified as an outlier and removed. The average distance between sampling points is calculated by averaging the pairwise distances of all sampling points.

[0061] 3D mesh generation standard: Based on the length, width, and height of the target space, divide it into cubic meshes with side lengths of 0.5 meters to 2 meters, ensuring that the number of meshes is not less than 500 (when the space scale is small, it can be appropriately reduced, but not less than 100); the mesh side length selection should meet the following requirements: 0.5 meters to 1 meter for indoor scenes, and 1 meter to 2 meters for outdoor open scenes.

[0062] Construction and implementation of inverse distance weighted interpolation model: When constructing the interpolation model, the center of the grid is used as the interpolation point, and the 8 grids closest to the interpolation point are selected as reference points. The weight of the reference point is inversely proportional to the distance (the closer the distance, the greater the weight). The interpolation radius is set to 3-5 grid side lengths to ensure that each interpolation point has sufficient reference data. In implementation, the sum of the products of the density values ​​of each reference point and the distance is calculated first, and then divided by the sum of the distances to obtain the density values ​​of the interpolation points, and finally a continuous density distribution function is generated.

[0063] Local extremum detection process: For each grid, detect the density value within a 3×3×3 grid range around it. If the density value of the current grid is more than 10% higher than the density values ​​of all neighboring grids, and the value is greater than twice the average density of the overall sampling area, it is marked as a candidate anchor point. The average density of the overall sampling area is calculated by the arithmetic mean of the density values ​​of all grids.

[0064] Calculate the straight-line distance between all candidate anchor points. If the distance between two anchor points is less than 5 meters, retain the anchor point with the higher density value; if the distance is greater than or equal to 5 meters, retain both. Repeat this process until there are no anchor points with a distance of less than 5 meters, forming the final main cover anchor point set.

[0065] Signal sample acquisition specifications: Set up a signal transmitter at each main coverage anchor point, and arrange signal receiving points in 16 directions: 0°, 22.5°, 45°...337.5°, with each receiving point 10-20 meters away from the anchor point; continuously collect data for 5-10 minutes at a sampling frequency of 1Hz-10Hz, and acquire 50-100 sets of signal strength data in each direction (ensuring that the amount of data meets statistical requirements).

[0066] The steps for calculating the coefficient of variation are as follows: First, calculate the arithmetic mean of the signal strength data in each direction. Then, calculate the sum of squares of the deviations between the data and the mean. Divide the sum by the data volume and take the square root to obtain the standard deviation. Finally, divide the standard deviation by the mean to obtain the coefficient of variation. The reciprocal of the coefficient of variation is the stability index. The threshold of 0.6-0.8 needs to be determined with reference to the requirements of the communication protocol. A higher value should be taken for scenarios sensitive to bit error rate, and a lower value should be taken for scenarios prioritizing coverage.

[0067] Selection of the main direction of bidirectional signal radiation: For the selected directions that meet the stability index, calculate the angle between any two directions (take the minimum angle value), and select combinations with angles between 150° and 180°; calculate the overlapping area of ​​the coverage areas of the two directions, and the ratio of the overlapping area to the coverage area of ​​a single direction must be ≤20% (the coverage area is calculated based on the boundary of the area where the signal strength attenuates to the threshold).

[0068] For directional combinations that meet the conditions, calculate the sum of the stability indices of the two directions, and select the combination with the largest sum as the two-way principal directions; if the sums are the same, select the combination with the included angle closer to 180°.

[0069] Step 22 above, measuring equipment and operation: Use a high-precision electronic compass (measurement accuracy ±0.5°), set up the equipment at the main coverage anchor point, align it with the signal receiving points in the two main radiation directions, and read the absolute azimuth angle value (accurate to 0.1°); repeat the measurement 3 times and take the average value to reduce measurement error.

[0070] Angle difference calculation rules: Calculate the difference between two absolute azimuth angles. If the difference is greater than 180°, subtract the difference from 360°. The final angle difference value is controlled within the range of 0°-180°, and the result is rounded to one decimal place.

[0071] Ring-shaped coverage area division and construction: Based on the signal transmission power and receiving sensitivity of the main coverage anchor point, the distance at which the signal strength attenuates to the communication threshold is determined through field testing. For indoor scenarios, the range is 50 meters to 200 meters, and for outdoor scenarios, the range is 200 meters to 500 meters as the coverage radius R.

[0072] Sector-shaped sub-region division: If the azimuth angle difference is α, calculate n=360° / α and round it to the nearest integer (e.g., if α=90° then n=4) to ensure that the angle deviation of each sub-region is ≤5°; take the main coverage anchor point as the origin and divide the ray boundary according to the sub-region angle to form a sector-shaped region.

[0073] The transition zone is defined as follows: The transition zone is an annular area with a radius of R1 = R × (1-10%) within the annular area and an outer radius of R2 = R (width is R × 5%-10%). The smaller width is used for indoor scenes and the larger width is used for outdoor scenes.

[0074] Device spacing calculation: Calculate the wavelength λ=c / f (where c is the speed of light) based on the signal operating frequency f. Set the device spacing to 0.5λ-2λ (e.g., for a 2.4GHz signal, λ≈0.125 meters, the spacing is 0.0625 meters-0.25 meters). Distribute devices evenly at the spacing within the transition zone of each sector sub-region, ensuring a quantity of ≥3. The device height should be consistent with the height of the anchor signal transmitter (error ±0.5 meters). Calibrate the verticality of the installation using a level.

[0075] Step 23 above, activation sequence generation: construct a periodic pulse sequence model, the period is determined based on the maximum multipath delay (obtained through previous testing, usually 0.1-1 seconds); the sequence duty cycle is 30%-50%, that is, the activation time accounts for 30%-50% of the period, and a smaller duty cycle is used in high interference scenarios; the sequence adopts binary encoding, and adjacent devices are assigned complementary code patterns (when one is active, the other is dormant).

[0076] A microprocessor is used as the sequence controller. The activation sequence code is pre-stored and activation instructions are sent to each device through a wireless synchronization signal (synchronization error ≤10μs) to ensure that the devices work alternately according to the preset time sequence and avoid co-channel interference.

[0077] The filter is constructed based on the signal operating frequency f0, with the center frequency set to f0 and the bandwidth taken as f0×10%-20% (e.g., f0=5GHz, bandwidth 500MHz-1GHz). The filter adopts a Butterworth structure with an order of 4-8 (to balance the filtering effect and complexity), and is implemented by adjusting the circuit component parameters (resistor and capacitor values ​​are calculated according to the filter design formula and then fine-tuned).

[0078] After acquiring 5-10 cycles of time-domain waveform data, out-of-band noise is removed using a filter. For the filtered data, the peak value of the waveform is detected by a sliding window (the window length is 1 / 10 of the signal period). Points where the peak value exceeds 3 times the average signal strength are determined to be the arrival time of the multipath signal, and the signal strength value at that point is recorded as the attenuation feature.

[0079] Calculation of the dispersion coefficient of time delay distribution: Statistically count the arrival times of all multipath signals and calculate the arithmetic mean arrival time; the difference between each arrival time and the mean is the deviation value, calculate the standard deviation of the deviation value, and then divide it by the mean arrival time to obtain the dispersion coefficient (value ranges from 0.1 to 0.5, the higher the degree of dispersion, the larger the coefficient).

[0080] Attenuation path loss index fitting: The propagation distance d is calculated by the GPS coordinates of the anchor point and the device, and the corresponding signal strength L is recorded. The relationship between L and d is fitted by the least squares method (L=k×log(d)+b), where k is the path loss index (value 2-4, usually 3-4 for indoor scenes and 2-3 for outdoor open scenes).

[0081] The weights are dynamically adjusted according to the scenario. In scenarios with severe multipath interference (discretion coefficient > 0.3), the weight of the dispersion coefficient is increased to 0.5-0.6. In scenarios dominated by distance attenuation (loss index > 3), the weight of the loss index is increased to 0.5-0.6. After the weights are determined, the weighted sum is achieved through an adder to obtain the compensation coefficient (0.8-1.2), which is then stored in the channel compensation module for signal correction.

[0082] In this embodiment of the invention, the quality of spatial sampling data is improved by outlier removal and precise interpolation. The main coverage anchor point is accurately identified by combining dual verification and deduplication rules. The bidirectional radiation direction is optimized through complementary screening and stability evaluation, reducing coverage blind spots and improving wireless resource utilization efficiency. High-precision angle measurement and scientific division rules enable precise coverage area division. Targeted transition zone deployment and spacing design ensure comprehensive boundary signal monitoring at a controllable cost, avoiding the resource waste or insufficient coverage problems of traditional division methods.

[0083] In a preferred embodiment of the present invention, step 3, which involves fusing channel propagation compensation coefficients and topology correlation characteristics to dynamically construct a communication network topology model, mapping network functional entities to topology vertices and topology correlation characteristics to weighted transmission links, and generating the communication network topology model by driving topology reconstruction through signaling event trigger sequences, includes:

[0084] Step 31: Perform normalization processing on the channel propagation compensation coefficient to generate the link transmission reliability factor; parse the topological association characteristics to construct the initial adjacency matrix and inject the link transmission reliability factor into the corresponding matrix elements to generate the weighted adjacency matrix;

[0085] Step 32: Based on the weighted adjacency matrix, map the network functional entities to a set of topological vertices to construct the initial communication network topology model;

[0086] Step 33: Based on the initial communication network topology model, perform real-time topology reconstruction according to the signaling event trigger sequence. If the event type is node access, add a new vertex to the topology vertex set and expand the rows and columns in the weighted adjacency matrix; if the event type is link switching, update the weight value of the corresponding link in the weighted adjacency matrix; if the event type is node failure, delete the target vertex in the topology vertex set and remove the associated rows and columns from the weighted adjacency matrix.

[0087] Step 34: Combine the real-time reconstructed topology vertex set with the weighted adjacency matrix to generate a dynamic communication network topology model.

[0088] Step 31 above, sample size determination and threshold setting: Statistical channel propagation compensation coefficient sample size. If the sample size is ≥30 groups, determine the minimum value (the minimum effective value below the first quartile) and the maximum value (the maximum effective value above the third quartile) by sorting. If the sample size is <30 groups, directly call the preset threshold (minimum 0.8, maximum 1.2). The preset value is calibrated based on historical data of the same type of scenario (error ≤ ±0.05).

[0089] Linear normalization operation procedure: Perform normalization calculation for each compensation coefficient x. First, calculate the difference between (x and minimum value), then divide by the range between (maximum value and minimum value) to obtain the median value in the range of 0-1. Multiply the median value by 0.5 and add 0.5 to convert it into a reliability factor of 0.5-1.0. For example, if x=1.0, minimum value=0.8, maximum value=1.2, then the difference=0.2, the range=0.4, the median value=0.5, and the reliability factor=0.5×0.5+0.5=0.75.

[0090] Boundary verification and correction mechanism: The value-by-value verification method is adopted to scan all generated reliability factors. If there is a factor < 0.5, it is forcibly replaced with 0.5; if it is > 1.0, it is forcibly replaced with 1.0. The correction record needs to be stored in the data quality log (including the original coefficient value, the corrected value and the reason for the correction).

[0091] Topology association characteristic parsing method: Entity relationships are obtained through three-layer parsing: ① Physical layer: Read the device MAC address binding list to identify wired connection entity pairs; ② Link layer: Collect the associated terminal list of wireless access points (APs) and extract the authenticated wireless connections; ③ Application layer: Analyze the communication protocol logs and filter entity pairs that have data interaction in the past 5 minutes; The union of the three types of relationships is taken as the direct communication relationship set, and temporary connections with communication interruption records (interruption duration > 10 seconds) are removed.

[0092] Count the number of valid entities N, and construct an N×N square matrix (use a two-dimensional array for storage when N≤1000, and a sparse matrix for storage when N>1000); for entities i and j (i≠j), if they are in the relation set, then A(i,j)=1, otherwise A(i,j)=0; the diagonal element A(i,i) is forcibly set to 0 through batch assignment instructions to avoid self-loop link interference; after the matrix is ​​constructed, perform redundancy check to ensure that there are no duplicates or false associations (association records must match the device hardware identifier).

[0093] Weighting and Symmetry Handling: Establish a link-factor lookup table, match the corresponding reliability factor according to the entity pair (i, j) index, and replace the element in the initial matrix where A(i,j)=1; for wireless links, compare the difference between A(i,j) and A(j,i) by matrix transpose. If the difference is > 0.1, calculate the arithmetic mean of the two ((A(i,j) + A(j,i)) / 2) and update it synchronously to both positions; if the difference is ≤ 0.1, keep the original value; for wired links, determine the directionality based on port attributes: bidirectional ports maintain symmetry, unidirectional ports (such as fiber optic transceivers) are marked with a directionality identifier, and symmetry is not forced.

[0094] Verify the proportion of non-zero elements (which should be consistent with the actual number of communication links, with an error of ≤5%) and the distribution of weight values ​​(more than 90% of elements should be within the range of 0.6-0.9). If the criteria are not met, backtrack to the compensation coefficient calculation stage to regenerate the factors.

[0095] In step 32 above, the vertex ID allocation rule is as follows: the encoding rule of "type prefix + sequence number" is adopted, such as "BS-001" for base stations and "UE-001" for terminals. The sequence number is continuously incremented from 001 to ensure the uniqueness of the ID (verified by the database deduplication mechanism); the mapping relationship between the ID and the adjacency matrix index is stored in the index table, and the index value = sequence number - 1 (such as "BS-001" corresponding to index 0), which supports O(1) time complexity search.

[0096] The attribute table contains required and optional fields: Required fields are ID, type (enumerated values: base station / terminal / repeater), location coordinates (latitude and longitude accurate to 0.0001°, indoor coordinates accurate to 0.1 meters), and activation time; Optional fields are maximum transmission rate (unit Mbps, value range 10-1000), supported frequency bands (e.g., 2.4GHz / 5GHz), and transmit power (unit dBm, value range 10-30); attribute values ​​must be read directly through the device interface (not manually entered) and their integrity verified by a checksum.

[0097] Initial topology model construction: An adjacency list storage model is adopted. Each vertex node contains an ID, attribute pointer, and a list of links (storing the IDs and weights of adjacent vertices); the link node contains a source vertex pointer, a target vertex pointer, a weight value, and a link type (wired / wireless); the data structure dynamically allocates memory through linked lists, supporting the dynamic addition and deletion of vertices and links.

[0098] Quantity verification: Number of vertices = adjacency matrix order, number of non-zero links = number of non-zero elements in the adjacency matrix / 2 (symmetric matrix); Weight verification: Traverse all link weights to ensure they are all within the range of 0.5-1.0, and the symmetry error of wireless link weights is ≤0.05; Attribute verification: Each vertex attribute table has no empty values ​​in the required fields, and the position coordinates are verified by map matching (deviation ≤5 meters); When verification fails, an alarm is triggered and the source of the error is located (such as duplicate vertex IDs or weight out of bounds).

[0099] In step 33 above, the signaling monitoring module is deployed to capture network layer signaling frames in real time (sampling frequency 100Hz) and identify the event type through the frame header type field.

[0100] Node access event: includes an "Association Request" field, carrying the hardware address of the new entity and the ID of the target entity requesting access; Link switching event: includes a "Switching Indication" field, carrying the source entity, target entity, and signal strength change (in dB); Node failure event: includes a "Disconnect Notification" field or three consecutive heartbeat timeouts (heartbeat cycle 1 second), carrying the ID of the failed entity and an error code; Event parsing delay is controlled within 50ms, and the parsing results are stored in the event queue (queue length ≤ 1000 entries).

[0101] The delay time from the completion of parsing a node access or failure event to the start of reconstruction is calculated. If it is greater than 1 second, accelerated processing is triggered (CPU resources are prioritized). If the delay of a link switching event is greater than 0.5 seconds, non-critical switching events (events with signal strength changes of less than 3dB) are ignored to ensure that core events are processed first.

[0102] New vertex addition process:

[0103] ID Allocation: Query the current maximum sequence number, the new sequence number = maximum sequence number + 1, and generate a new ID (e.g., if the current maximum is "UE-099", the new ID is "UE-100"); Attribute Acquisition: Obtain the entity type through the access request frame, collect the location coordinates through the GPS / indoor positioning module (temporary coordinates "0,0" are temporarily stored before positioning is completed), and extract the communication parameters from the device capability negotiation frame; Attribute Calibration: Correct the location coordinates within 30 seconds after access by measuring the distance (wireless signal time-of-flight method), reducing the error to within ±1 meter.

[0104] Adjacency matrix expansion operation: Add a row and a column to the end of the original matrix, and set the initial values ​​to 0 in batches using the memset function; locate the corresponding position of the new vertex in the matrix and the target entity according to the target entity ID in the association request, obtain the compensation coefficient through signal testing (average of 10 consecutive samples), and fill in the weight value after normalization (the default initial value is 0.7 for the first connection, and it is dynamically calibrated to the actual value within the next 3 minutes).

[0105] Weight update determination logic:

[0106] Location Link: Locate the element position (i, j) in the adjacency matrix using the source ID and target ID; Factor Calculation: Recalculate the compensation coefficient based on the signal strength change ΔS (dB): If ΔS ≥ 3dB, coefficient = original coefficient + 0.1 × (ΔS / 3); If ΔS ≤ -3dB, coefficient = original coefficient - 0.1 × (|ΔS| / 3), and normalize to generate a new factor; Difference Verification: If the absolute value of the new factor - the old factor > 0.1, perform the update; otherwise, skip. The update record must include the new and old factor values ​​and the reason for the change.

[0107] Boundary control: If the updated factor is <0.5, it is forced to be set to 0.5 and marked as "poor link quality"; if it is >1.0, it is forced to be set to 1.0 and marked as "excellent link quality". The marking information is synchronized to the link attribute table.

[0108] Vertex and matrix reduction operations:

[0109] Vertex deletion: Locate the invalid ID node in the topological vertex set linked list, remove the node by adjusting the pointer, and release memory space; Matrix reduction: Remove the row and column data corresponding to the invalid ID as a whole, shift the remaining rows up in the original order, and shift the remaining columns to the left in the original order to ensure that the new matrix index strictly corresponds to the remaining vertex ID (index value = sequence number - 1); Residual verification: Check whether there are isolated rows / columns (all zero elements) in the new matrix. If they exist, mark the corresponding vertex as an "isolated node" and trigger an alarm.

[0110] Step 34 above, dynamic topology model combination generation: A mapping relationship between the vertex set and the adjacency matrix is ​​established through ID indexing, generating an association table (containing vertex ID, matrix row index, and matrix column index), ensuring that each vertex uniquely corresponds to one row and one column in the matrix; link weights are matched through dual indexes: the source vertex ID is the row index, and the target vertex ID is the column index, with an index positioning error ≤0.1%; metadata generation specifications: timestamps accurate to milliseconds (format: YYYY-MM-DDHH:MM:SS.XXX); number of vertices = number of elements in the current vertex set; total number of links = number of non-zero elements in the adjacency matrix (symmetric matrices require deduplication counting); average reliability factor = sum of non-zero weights / total number of links, rounded to two decimal places; metadata is bound and stored with the main model.

[0111] Model integrity and consistency verification:

[0112] Traverse the vertex set, checking that each ID has a corresponding index in the association table, and that the number of rows / columns in the matrix equals the number of vertices; Link characteristic verification: Randomly select 20% of the wireless links and check that the difference between A(i,j) and A(j,i) is ≤0.05; Select 20% of the wired links and verify that the directional identifier is consistent with the actual port type; Weight distribution test: Calculate the standard deviation (σ=) of all non-zero weights. (where μ is the average value and n is the number of links). If σ > 0.2, generate an abnormal link list (links with weights < 0.6 or > 0.9) and prompt for manual calibration; the verification pass rate must be ≥ 95%, otherwise return to the reconstruction stage.

[0113] Log and Snapshot Management: Update logs are stored in CSV format, with fields including event ID, type, timestamp, involved vertex ID, involved link (source-target), and weight change value (old → new). Log files are split by day, retaining data from the most recent 3 days (≥72 hours). The snapshot function is triggered by a scheduled task, generating a binary image file of the model every hour. The storage path includes the timestamp (e.g., "snapshot_20250808_1200.bin"). Model rollback based on snapshots is supported (rollback time ≤30 seconds).

[0114] Full model verification trigger conditions: Real-time statistics of vertex number change rate = |current vertex number - last verified vertex number| / last verified vertex number. If > 20%, verification is triggered; Calculate the average change in link weight = Σ|new weight - old weight| / total number of links. If > 0.15, verification is triggered; The verification process includes three indicators: completeness, consistency, and weight rationality. The passing score for each indicator is ≥ 90 points (out of 100). Unqualified items require the generation of a rectification list.

[0115] In this embodiment of the invention, normalization processing (dynamic threshold + boundary correction) adapted to sample size ensures that the link transmission reliability factor accurately maps to the channel quality, improving the consistency of link evaluation; a three-layer parsing and matrix verification mechanism achieves comprehensive capture and redundancy elimination of topological association characteristics, laying a high-quality data foundation for the construction of the weighted matrix and ensuring the scenario adaptability of the matrix weights. The ID encoding rule of "type prefix + sequence number" and the mapping of the index table ensure that the vertex identifier is unique and traceable and the search efficiency reaches O(1); mandatory verification of the required fields of the attribute table and graphical adjacency list storage ensure that the vertex attributes are complete and accurate and the model supports dynamic addition and deletion; triple integrity verification (quantity / weight / attribute) avoids basic errors such as vertex-matrix mismatch and weight out-of-bounds errors from the source, improving the reliability of the initial model.

[0116] In a preferred embodiment of the present invention, step 4 involves inputting the communication network topology model into a dynamic graph neural network, performing state feature space aggregation of the topology vertex neighborhood through a spatiotemporal feature extraction layer, and fusing the temporal evolution dependencies of historical topology sequences to generate a spatiotemporal state tensor for network nodes, including:

[0117] Step 41: Input the dynamic communication network topology model into the spatiotemporal feature extraction layer of the dynamic graph neural network; determine the set of adjacent vertices for each topology vertex based on the weighted adjacency matrix in the dynamic communication network topology model; aggregate the state features of the adjacent vertex sets to generate the spatial feature vector of the topology vertices;

[0118] Step 42: Based on the spatial feature vector, extract the historical spatial feature sequence within the preset time window and generate time evolution dependent features through a temporal convolutional network;

[0119] Step 43: After concatenating the spatial feature vector with the temporal evolution-dependent feature vector, input the concatenation into the gated fusion unit to generate a spatiotemporal joint feature vector; organize all spatiotemporal joint feature vectors in vertex index order to obtain a three-dimensional tensor structure and generate the spatiotemporal state tensor of the network nodes.

[0120] In step 41 above, the dynamic communication network topology model (including the topology vertex set and weighted adjacency matrix) generated in step 34 is input into the spatiotemporal feature extraction layer of the dynamic graph neural network; the layer parameters are initialized as follows: the feature extraction dimension is set to 16-32 dimensions (16 dimensions when the number of vertex attributes is ≤8, and 32 dimensions when it is >8); the neighborhood aggregation radius threshold is set to 0.6 (vertices with a link reliability factor ≥0.6 are included in the neighborhood); the weight matrix is ​​initialized using a Xavier uniform distribution with a value range of [-0.5, 0.5]; the bias term is initialized to 0; after initialization, the parameters are validated to ensure that there are no null values ​​or outliers.

[0121] For each topological vertex e, traverse the non-zero elements in the e-th row of the weighted adjacency matrix, extract the column indices of elements with a value (link reliability factor) ≥ 0.6, and the corresponding vertices constitute the set of adjacent vertices of vertex e; if the number of adjacent vertices is < 3 (below the minimum neighborhood size), lower the threshold to 0.5 and re-filter; if still insufficient, supplement with the nearest vertex according to the straight-line distance of the position coordinates to ensure that each vertex has at least 3 adjacent vertices; the set of adjacent vertices is sorted in descending order of reliability factor, and the first 5-8 vertices (5 for indoor scenes and 8 for outdoor scenes) are retained as the effective neighborhood.

[0122] Vertex state feature extraction: Read attribute features from the topological vertex set, including location coordinates (latitude and longitude / indoor coordinates normalized to [0, 1]), maximum transmission rate (10-1000Mbps normalized to [0, 1]), mean signal strength (-100 to -50dBm normalized to [0, 1]), and activation duration (normalized to [0, 1] per hour), to form an initial feature vector (dimension consistent with the number of attributes).

[0123] Neighborhood feature weighted aggregation: For the effective neighboring vertices of vertex e, the features are aggregated by weighting according to the link reliability factor: the feature vector of each adjacent vertex o is multiplied by its link reliability factor (weight) with e, the sum is divided by the total weight to obtain the neighborhood aggregated feature; the initial feature of vertex e and the neighborhood aggregated feature are concatenated by dimension, and mapped to 16-32 dimensions by linear transformation (weight matrix) to generate spatial feature vector; the feature value truncation threshold is set to [-1, 1], and the value outside the range is forcibly corrected to the boundary value.

[0124] In step 42 above, the preset time window size is 5-10 time steps (time step length = topology model update cycle, 1 minute / step; 5 steps for high dynamic scenarios, 10 steps for low dynamic scenarios); the sliding window method is used to extract the spatial feature vector of each vertex in the most recent T time steps (T=5 or 10) to form a historical feature sequence (dimension is T×feature dimension); if the historical data of a vertex is less than T time steps (such as when a new vertex is added), the missing time steps are filled with the initial feature vector (filling ratio ≤30%, if it exceeds, it is marked as "new vertex" type).

[0125] Temporal Convolutional Network Construction and Training: A two-layer temporal convolutional network is constructed. The first layer has a kernel size of 3 (covering 3 consecutive time steps), and the number of output channels equals the feature dimension. The second layer has a kernel size of 2, and the number of output channels equals the feature dimension / 2. Each convolutional layer is followed by batch normalization (mean=0, variance=1) and a ReLU activation function (activation threshold=0, negative values ​​are set to 0). The network input is a historical feature sequence, and the output is a time-dependent feature (dimensionality=feature dimension / 2).

[0126] Training process: The topological model and corresponding performance metrics from the past 7 days are used as training data, and the training set and validation set are divided in a 7:3 ratio; the loss function is mean squared error (the sum of squared deviations between predicted and actual features); the optimizer uses stochastic gradient descent; the initial learning rate is 0.01, which decreases by 5% every 5 epochs; training is conducted for 30-50 epochs; the training stops if the validation set loss does not decrease for 3 consecutive epochs; an early stopping mechanism is used to prevent overfitting, and the model parameters with the minimum validation set loss are saved.

[0127] The historical spatial feature sequence is input into the trained temporal convolutional network. The convolutional layer captures local dependencies in the temporal dimension (such as feature trends and abrupt change points). After activation function and batch normalization, the output features are time evolution dependent. L2 normalization (magnitude = 1) is performed on the output features to enhance feature stability.

[0128] In step 43 above, the spatial feature vector (16-32 dimensions) is concatenated with the temporal evolution dependency features (8-16 dimensions) from step 42 according to their dimensions, forming a concatenated feature vector (24-48 dimensions); a gated fusion unit is constructed, including a reset gate and an update gate.

[0129] Reset gate: The reset weight is calculated using the sigmoid activation function (output [0, 1]) to control the proportion of historical time features retained. The parameters are determined through training (training data is the same as the temporal convolutional network). Update gate: The update weight is calculated using the sigmoid activation function to control the proportion of current spatial features fused with historical time features. The input of the gate unit is a concatenated feature vector, and the output is a spatiotemporal joint feature vector (16-32 dimensions, consistent with the dimension of the spatial feature vector).

[0130] The reset gate output is multiplied by the time evolution dependent feature to obtain the reset time feature; the reset time feature is added to the spatial feature vector and then activated by the tanh activation function (output [-1, 1]) to generate candidate features; the update gate output is multiplied by the candidate features and then multiplied by (1 - update gate output) and the spatial feature vector to obtain the spatiotemporal joint feature vector; the feature vector is truncated by a threshold (range [-1, 1]) to ensure numerical stability.

[0131] The spatiotemporal joint feature vectors of all vertices are organized according to the topological vertex set ID order (from smallest to largest) to form a three-dimensional tensor: the first dimension is the number of vertices (consistent with the number of vertices in the current topological model), the second dimension is the feature dimension (16-32), and the third dimension is the time step (fixed at 1, representing the current moment); after the tensor is generated, it is verified: the number of vertices, feature dimension, and time step match the preset values, and the feature values ​​are all within the range of [-1, 1]. After the verification is passed, a timestamp (accurate to the second) is marked.

[0132] Calculate the average variance of the tensor features. If the variance is less than 0.05 (feature distribution is concentrated), add small random perturbations (amplitude ≤ 0.01) through the feature enhancement module to improve the discriminative power. If there are all-zero feature vectors (outliers), replace them with the average feature vectors of the same type of vertices. The optimized tensor is stored in the feature database and can be called in subsequent network analysis.

[0133] In this embodiment of the invention, dynamic thresholding of adjacent vertices and weighted aggregation of features ensure that the spatial feature vector accurately reflects the local topological relationships and attribute differences of vertices. The neighborhood supplementation mechanism avoids feature loss under sparse topology, improving the completeness and representativeness of spatial features. A temporal convolutional network combined with a sliding window extracts historical features, effectively capturing temporal evolution patterns through optimized network parameters, and normalization enhances the stability of temporal features. The gated fusion unit adaptively balances the weights of spatial and temporal features through a dual-gating mechanism, avoiding the dominance of a single feature. The three-dimensional tensor structure is organized according to vertex order, ensuring the structured expression of spatiotemporal features. The tensor optimization mechanism further improves feature quality and usability.

[0134] In a preferred embodiment of the present invention, step 5, based on the spatiotemporal state tensor of network nodes, uses the particle swarm optimization algorithm to calculate the quantified topological characteristic of the overall network risk level, and dynamically executes network resource strategy optimization to suppress end-to-end risk transmission, includes:

[0135] Step 51: Input the spatiotemporal state tensor of the network nodes into the particle swarm optimization algorithm, map each network node to a particle and initialize the particle position vector according to the feature value distribution, construct the end-to-end risk transmission assessment index, and calculate the risk transmission loss estimate for each particle position.

[0136] Step 52: Adjust the particle velocity vector based on the individual historical minimum loss position and the group minimum loss position, update the particle position vector based on the velocity vector and recalculate the risk transmission loss estimate to obtain the updated particle swarm state.

[0137] Step 53: Based on the updated particle swarm state, if the preset iteration termination threshold is reached, extract the minimum loss position vector of the swarm and perform position vector decoding to generate a quantitative topology representation of the risk level of the entire network; input the quantitative topology representation of the risk level into the resource scheduling decision-maker to generate a resource scheduling scheme.

[0138] Step 54: Execute the resource scheduling scheme and monitor the end-to-end risk transmission and attenuation trend, and adjust the particle swarm optimization parameters based on the monitoring results.

[0139] In step 51 above, the spatiotemporal state tensor of the network nodes is used as the input data for the particle swarm optimization algorithm. Each network node corresponds to one particle in the algorithm, and the number of particles is consistent with the number of network nodes. The dimension of the particle position vector is set to the feature dimension of the spatiotemporal state tensor of the network nodes, with a value range of 16-32 (matching the feature dimension of the tensor). When initializing the particle position vector, a uniform sampling method is used based on the feature value distribution, and values ​​are randomly selected from the interval [-1, 1] to fill the vector, ensuring that the initial position covers the main area of ​​the feature space. An end-to-end risk transmission assessment index system is constructed, which includes three core indicators: node load rate (value range 0-1, determined by the number of nodes). The following metrics are used to calculate the risk propagation loss for each particle position: the ratio of current traffic to maximum carrying capacity, the link reliability attenuation coefficient (range 0-1, determined based on the real-time change rate of link weights in the weighted adjacency matrix), and the historical fault propagation probability (range 0-1, obtained by statistically analyzing the frequency of neighboring node failures caused by node failures within the past 72 hours). Weights are assigned to these metrics, with node load rate weighted at 0.4, link reliability attenuation coefficient weighted at 0.3, and historical fault propagation probability weighted at 0.3 (the sum of these weights is 1). A weighted summation is used to calculate the estimated risk propagation loss for each particle position, with the estimated value mapped to a quantitative score of 0-10 (a higher score indicates a greater likelihood of risk propagation).

[0140] In step 52 above, the initial range of the particle velocity vector is set to [-0.5, 0.5], consistent with the dimension of the position vector. When adjusting the velocity vector based on the individual's historical minimum loss position and the group's minimum loss position, three parameters are introduced: inertia weight, individual learning factor, and group learning factor. The initial value of the inertia weight is set to 0.8, which decreases linearly to 0.4 with the number of iterations (decreasing by 0.1 every 5 iterations), used to balance the particle's global exploration and local development capabilities. The individual learning factor and the group learning factor are both set to 1.8, controlling the intensity of the particle's learning towards its own historical best position and the group's best position, respectively. After the velocity vector is updated, boundary verification is required. If the velocity value exceeds the range of [-0.5, 0.5], it is forcibly corrected to the nearest boundary value. The particle position vector is adjusted according to the updated velocity vector. The new position = current position + velocity vector. Boundary verification is also performed to ensure that the position vector elements are within the range of [-1, 1]. The risk transmission loss estimate of the adjusted particle position is recalculated. The index weighted summation method in step 51 is repeated to record the individual historical minimum loss position of each particle (the position corresponding to the minimum estimate experienced by the current particle) and the group minimum loss position (the position corresponding to the minimum estimate among all particles), forming the updated particle swarm state.

[0141] Step 53 above, the preset iteration termination threshold includes two conditions: the number of iterations reaches 30 (for small-scale indoor networks) or 50 (for large-scale outdoor networks), or the change in the minimum loss estimate of the population is less than 0.01 (tending to stabilize) in 5 consecutive iterations; when either condition is met, the minimum loss position vector of the population is extracted and a decoding operation is performed. The decoding process maps each element of the position vector from the interval [-1, 1] to a risk level of 1-5 (-1 corresponds to level 1, 0 corresponds to level 3, 1 corresponds to level 5, and intermediate values ​​are converted linearly), and the risk levels are arranged in order of node index. A quantified topology representation of the network's risk level is generated, where level 1 represents extremely low risk and level 5 represents extremely high risk. This quantified topology representation is input into a resource scheduling decision-maker, which generates a resource scheduling plan based on a preset rule base: for high-risk nodes (levels 4-5), priority is given to allocating additional bandwidth resources (increasing current bandwidth by 10%-20%) and power resources (increasing transmit power by 3-5 dB); for medium-risk nodes (levels 2-3), basic resource configuration is maintained and resource utilization is monitored; for low-risk nodes (level 1), redundant resources can be appropriately reduced (reducing them by 5%-10%). After the plan is generated, the total resource constraints must be verified to ensure that the network's resource utilization rate does not exceed 90% after scheduling (reserve emergency resources).

[0142] In step 54 above, after executing the resource scheduling scheme, the network performance monitoring module collects end-to-end risk propagation data in real time, with a monitoring frequency of once every 5 minutes. The focus is on tracking changes in load rate of high-risk nodes, link reliability recovery, and the number of fault propagation events. A risk propagation attenuation trend index is calculated, which is the percentage difference between the average of the estimated risk loss from three consecutive monitoring sessions after scheduling and the average before scheduling. If the percentage difference is ≥20%, the attenuation trend is considered significant; if the percentage difference is <10%, the attenuation effect is considered poor. Based on the monitoring results, the particle swarm optimization parameters are adjusted: when the attenuation effect is poor, the inertia weight is increased to 0.9, and the individual learning factor is increased to 2.0 to enhance the particle's ability to explore new positions; when the attenuation trend is stable, the current parameters are maintained; when over-optimization leads to resource waste (resource utilization <60%), the inertia weight is reduced to 0.6, and the number of iterations is reduced to 20-30. The adjusted parameters are stored in the optimization parameter library as the initial configuration for the next round of particle swarm optimization.

[0143] In this embodiment of the invention, a multi-dimensional risk indicator system and particle mapping mechanism are used to achieve accurate quantification of network risk and reasonable construction of the initial optimization space, providing a reliable foundation for subsequent optimization. The speed adjustment strategy balances global exploration and local optimization capabilities, ensuring that the particle swarm converges efficiently to the optimal solution and improving the accuracy of risk assessment. The iterative termination mechanism and decoding rules make the risk level representation intuitive and interpretable, and the resource scheduling scheme is highly targeted, effectively achieving resource allocation in high-risk areas. The dynamic parameter adjustment mechanism enables the algorithm to adapt to changes in network state, continuously optimize the risk suppression effect, reduce resource waste, and improve the network's risk resistance.

Claims

1. A dynamic graph neural network modeling method for spatiotemporal big data, characterized in that, The method includes: Step 1: Real-time acquisition of multi-dimensional spatiotemporal data streams in the communication network, including spatial location sampling sets of network functional entities, topological association characteristics, and signaling event triggering sequences; Step 2: Analyze the spatial location sampling set to extract the spatial distribution characteristics of network functional entities; perform wireless propagation modeling based on the spatial distribution characteristics, select the extreme value of the sampling density as the main coverage anchor point, and generate the main direction of bidirectional signal radiation; calculate the angle between the main directions of bidirectional signal radiation to divide the coverage area; dynamically deploy signal acquisition devices at the boundary of the area; obtain multipath propagation characteristics based on the time-domain activation sequence of the signal acquisition devices and calculate the channel propagation compensation coefficient; Step 3: Integrate the channel propagation compensation coefficient and topology correlation characteristics to dynamically construct a communication network topology model. Map network functional entities to topology vertices and topology correlation characteristics to weighted transmission links. Then, drive topology reconstruction through signaling event trigger sequences to generate the communication network topology model. Step 4: Input the communication network topology model into the dynamic graph neural network, perform spatial aggregation of state features of the topology vertex neighborhood through the spatiotemporal feature extraction layer, and fuse the temporal evolution dependencies of historical topology sequences to generate the spatiotemporal state tensor of network nodes. This includes: Step 41: Input the dynamic communication network topology model into the spatiotemporal feature extraction layer of the dynamic graph neural network; determine the set of adjacent vertices for each topology vertex based on the weighted adjacency matrix in the dynamic communication network topology model; aggregate the state features of the adjacent vertex sets to generate spatial feature vectors of the topology vertices; Step 42: Based on the spatial feature vectors, extract the historical spatial feature sequences within a preset time window and generate temporal evolution dependency features through a temporal convolutional network; Step 43: Concatenate the spatial feature vectors and temporal evolution dependency features and input them into the gated fusion unit to generate spatiotemporal joint feature vectors; organize all spatiotemporal joint feature vectors according to vertex index order to obtain a three-dimensional tensor structure and generate the spatiotemporal state tensor of network nodes. Step 5: Based on the spatiotemporal state tensor of network nodes, use the particle swarm optimization algorithm to calculate the quantitative topological representation of the risk level of the entire network, and dynamically execute network resource strategy optimization to suppress end-to-end risk transmission.

2. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 1, step 2 includes: Step 21: Analyze the spatial location sampling set to calculate the spatial sampling point density distribution function, and identify the local maxima of the density distribution function as the main coverage anchor points; Based on the distribution of signal propagation stability indexes calculated from the main coverage anchor point, two complementary radiation directions whose stability indices meet the preset thresholds are selected as the main directions of bidirectional signal radiation. Step 22: Measure the azimuth angle difference between the main directions of bidirectional signal radiation, construct a ring-shaped coverage area based on the azimuth angle difference, and deploy a cluster of signal acquisition devices in the boundary transition zone of the ring-shaped coverage area. Step 23: Adjust the collaborative working cycle of the signal acquisition device cluster by setting the time-domain activation sequence, and extract multipath propagation features based on the spatial response characteristics of multipath reflection signals; generate channel propagation compensation coefficients by weighted fusion based on the time delay distribution discrete coefficient and attenuation path loss index of the multipath propagation features.

3. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 2, characterized in that, Step 3 includes: Step 31: Perform normalization processing on the channel propagation compensation coefficient to generate the link transmission reliability factor; parse the topological association characteristics to construct the initial adjacency matrix and inject the link transmission reliability factor into the corresponding matrix elements to generate the weighted adjacency matrix; Step 32: Based on the weighted adjacency matrix, map the network functional entities to a set of topological vertices to construct the initial communication network topology model; Step 33: Based on the initial communication network topology model, perform real-time topology reconstruction according to the signaling event trigger sequence. If the event type is node access, add a new vertex to the topology vertex set and expand the rows and columns in the weighted adjacency matrix; if the event type is link switching, update the weight value of the corresponding link in the weighted adjacency matrix; if the event type is node failure, delete the target vertex in the topology vertex set and remove the associated rows and columns from the weighted adjacency matrix. Step 34: Combine the real-time reconstructed topology vertex set with the weighted adjacency matrix to generate a dynamic communication network topology model.

4. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 3, characterized in that, Based on the spatiotemporal state tensor of network nodes, a particle swarm optimization algorithm is used to calculate the quantified topological characteristics of the entire network risk level. Network resource strategy optimization is then dynamically executed to suppress end-to-end risk propagation, including: Step 51: Input the spatiotemporal state tensor of the network nodes into the particle swarm optimization algorithm, map each network node to a particle and initialize the particle position vector according to the feature value distribution, construct the end-to-end risk transmission assessment index, and calculate the risk transmission loss estimate for each particle position. Step 52: Adjust the particle velocity vector based on the individual historical minimum loss position and the group minimum loss position, update the particle position vector based on the velocity vector and recalculate the risk transmission loss estimate to obtain the updated particle swarm state. Step 53: Based on the updated particle swarm state, if the preset iteration termination threshold is reached, extract the minimum loss position vector of the swarm and perform position vector decoding to generate a quantitative topology representation of the risk level of the entire network; input the quantitative topology representation of the risk level into the resource scheduling decision-maker to generate a resource scheduling scheme. Step 54: Execute the resource scheduling scheme and monitor the end-to-end risk transmission and attenuation trend, and adjust the particle swarm optimization parameters based on the monitoring results.

5. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 4, characterized in that, Step 21: Analyze the spatial location sampling set to calculate the spatial sampling point density distribution function, and identify the local maxima of the density distribution function as the main coverage anchor points; Based on the distribution of signal propagation stability indexes calculated from the main coverage anchor point, two complementary radiation directions whose stability indices meet preset thresholds are selected as the main bidirectional signal radiation directions, including: Step 211: Perform kernel density estimation on the spatial location sampling set to generate the probability density distribution surface of the spatial sampling points; detect the gradient change characteristics of the probability density distribution surface, and identify the extreme points where the gradient changes from positive to negative as local maxima. Step 212: Select the location with the highest probability density value from the local maxima to determine the main coverage anchor point; divide the surrounding space into multiple spatial sectors at equal angles with the main coverage anchor point as the center; calculate the statistical dispersion of the reference signal received power in each spatial sector as a signal propagation stability index. Step 213: Select spatial sectors with statistical dispersion lower than the preset stability threshold to obtain the sector center direction set, and calculate the azimuth difference of all direction pairs based on the set. Select the direction pair with the closest difference to the complementary state as the main direction of bidirectional signal radiation.

6. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 5, characterized in that, Step 32: Based on the weighted adjacency matrix, map network functional entities to a set of topological vertices to construct an initial communication network topology model, including: Step 321: Based on the row and column index identifier set of the weighted adjacency matrix, map each row and column index identifier to a topological vertex and assign vertex identifiers to generate a topological vertex set; traverse the non-zero elements of the weighted adjacency matrix with the topological vertex set as input and parse each non-zero element as a weighted transmission link, where the row index identifier is mapped to the source vertex identifier and the column index identifier is mapped to the destination vertex identifier, and the element value is mapped to the link weight value to generate a weighted transmission link set; Step 322: Establish a graph structure vertex set based on the topological vertex set, and parse the topological edge set; establish a graph structure edge set based on the topological edge set and assign the link weight value to the weight attribute of the corresponding edge; Step 323: Establish the connection relationship between vertices and edges based on the vertex set and weighted edge set of the graph structure, associate the source vertex identifier of the edge with the source vertex in the vertex set, and associate the destination vertex identifier with the destination vertex in the vertex set; generate the initial communication network topology model based on the vertex set, weighted edge set and vertex-edge connection relationship of the graph structure.

7. The dynamic graph neural network modeling method for spatiotemporal big data according to claim 6, characterized in that, Step 42: Based on the spatial feature vector, extract the historical spatial feature sequence within a preset time window and generate time evolution dependent features through a temporal convolutional network, including: Step 421: Based on the timestamp index of the spatial feature vector, retrieve the historical spatial feature vector sequence within a preset time window from the spatiotemporal feature repository; perform time alignment processing on the historical spatial feature vector sequence; Step 422: Input the time-aligned feature sequence into the first dilated convolutional layer of the temporal convolutional network. Perform local feature extraction by sliding the convolutional kernel of the first dilated convolutional layer along the time dimension to capture the short-term dependencies between adjacent time steps and obtain the short-term time-dependent feature sequence. Step 423: The short-term time-dependent feature sequence is passed as input to the second dilated convolutional layer. The time receptive field is expanded by increasing the dilation coefficient of the second dilated convolutional layer, so that the convolutional kernel performs feature extraction across discontinuous time steps, captures the medium- and long-term dependencies across time steps, and obtains the medium- and long-term time-dependent feature sequence. Step 424: Fuse the output features of the two dilated convolution layers and generate a time-evolution-dependent feature vector through time-dimensional pooling operation.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images