A 4, 5g signal three-dimensional propagation modeling method based on a graph neural network

Abstract

This invention discloses a method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks, comprising the following steps: S1, acquiring three-dimensional environmental data, signal source parameters, and user measurement data to establish a three-dimensional environmental model and generate training and validation sample sets; S2, performing propagation path modeling on the three-dimensional environmental model, calculating direct, reflection, and diffraction features and aligning them with the samples; S3, constructing a joint model of graph neural network and three-dimensional convolution, completing initial training to obtain baseline parameters; S4, establishing an elastic weight consolidation library based on validation results; S5, performing fine-tuning using incremental data to obtain updated parameters; S6, generating signal prediction and three-dimensional distribution maps; S7, iteratively updating and outputting the final model based on error results. This invention achieves high-precision modeling and adaptive optimization of 4G and 5G signals in complex three-dimensional space, improving prediction accuracy and model stability, and possessing the advantages of scalability and rapid convergence.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and artificial intelligence modeling technology, and in particular to a method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks. Background Technology

[0002] Existing wireless communication signal propagation modeling methods are mostly based on ray tracing, empirical formulas, or statistical fitting to simulate signal propagation in buildings, streets, and indoor environments. These methods typically assume simplified environmental structures and homogeneous materials, making it difficult to accurately reflect the complex reflection, diffraction, and occlusion characteristics in three-dimensional space. When signals propagate in multi-story buildings or complex terrain, the prediction results often show significant deviations, especially in 4G and 5G high-frequency scenarios, where this error is more pronounced. Traditional methods are highly dependent on field measurement data and lack a dynamic update mechanism after model construction, making it impossible to adaptively correct based on actual measurement data during network operation.

[0003] In recent years, deep learning methods have been introduced into signal propagation modeling. However, most model structures are still based on convolutional neural networks or multilayer perceptrons, and feature extraction is limited to local spatial grids. They are insufficient in describing the topological propagation relationships of signals between spatial nodes and fail to reflect the spatial correlations in complex scenes. Furthermore, these models typically use iterative parameter updates directly during incremental learning, lacking an assessment of the importance of historical parameters. This leads to the problem of forgetting old knowledge and parameter drift when training on new data, making the model's accuracy unstable over long-term learning.

[0004] Therefore, how to provide a method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a three-dimensional propagation modeling method for 4G and 5G signals based on graph neural networks. This invention combines graph neural networks and three-dimensional convolutional neural networks for signal propagation feature learning, and introduces an elastic weight consolidation mechanism to achieve stable incremental training of the model. It can accurately depict the signal propagation laws in complex spatial environments and has the advantages of high prediction accuracy, stable model updates, and strong adaptability.

[0006] A method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to an embodiment of the present invention includes the following steps: S1. Acquire 3D environment data, signal source parameter data and user measurement data, establish a 3D environment model and material parameter library, and generate training sample set and verification sample set; S2. Perform propagation path modeling on the 3D environment model, calculate the propagation feature set of direct, reflected, and diffracted light, and align it with the training sample set to form an aligned training sample set; S3. Construct a deep learning prediction model, using a multi-task regression structure of graph neural network and three-dimensional convolutional neural network, and use an aligned training sample set for initial training to obtain the baseline model parameter vector and baseline prediction results. S4. Use the verification sample set to evaluate the benchmark prediction results, calculate the parameter importance vector based on the historical training sample set, form an elastic weight consolidation parameter library, and set the benchmark model parameter vector as the reference parameter vector. S5. Receive incremental user measurement data, construct an incremental sample set based on the deep learning prediction model and the propagation feature set, perform incremental fine-tuning based on the reference parameter vector and elastic weights to obtain the updated model parameter vector and incremental prediction results. S6. Use the updated model parameter vector to infer the deep learning prediction model, generate signal strength prediction results and three-dimensional coverage distribution map, and calculate the error evaluation results. S7. Update the parameter importance vector and elastic weights to consolidate the parameter library based on the error evaluation results. If the convergence condition is not met, return to S5. If the convergence condition is met, output the deep learning prediction model and the three-dimensional coverage distribution map.

[0007] Optionally, S1 specifically includes: S11. Acquire three-dimensional environment data, establish a three-dimensional environment model, establish a material parameter library and set dielectric constant, conductivity and reflection coefficient; S12. Obtain the signal source parameter set and user measurement data set, record the antenna position, transmission frequency, angle, transmission power, time stamp, building identifier, room identifier, RSRP, and SINR; S13. Construct training and validation sample sets based on user measurement data sets and ray tracing results.

[0008] Optionally, S2 specifically includes: S21. Identify the signal propagation path in the three-dimensional environment model and determine the direct path, reflection path, and diffraction path of the signal from the signal source to the user's location. S22. Calculate the signal propagation characteristics based on different path types, consider the influence of spatial obstacles and material parameters, and generate a set of propagation characteristics that includes spatial location information and signal change characteristics. S23. Align the propagation feature set with the training sample set in terms of spatial location and time label to generate an aligned training sample set for training the deep learning prediction model.

[0009] Optionally, S3 specifically includes: S31. Construct a deep learning prediction model using a joint structure of graph neural network and three-dimensional convolutional neural network; S32. Input the three-dimensional environment data, signal source parameter data and user measurement data into the graph neural network and the three-dimensional convolutional neural network. The graph neural network outputs the node embedding representation, and the three-dimensional convolutional neural network outputs the local spatial feature tensor. S33. Fuse node embedding representation with local spatial feature tensor, perform feature concatenation and attention weighting, map to joint spatial feature representation, construct multi-task regression head, the first regression head is used to output signal strength, the second regression head is used to output signal-to-noise ratio; S34. Use the loss function to calculate the error between the predicted value and the true value, and use the weighted sum of the mean square error and the mean absolute error as the loss function. S35. Introduce a regularization mechanism during training, and use random deactivation and L2 weight decay. S36. Select the AdamW optimizer with weight decay for forward and backward propagation, update the deep learning prediction model parameter vector until the loss function value reaches the convergence condition, complete the initial training, and output the baseline model parameter vector and baseline prediction results. S37. After training convergence, export and save the baseline model parameter vector, generate baseline prediction results and corresponding evaluation metrics based on the validation sample set, and record logs and version information.

[0010] Optionally, S4 specifically includes: S41. Perform error analysis and performance evaluation on the benchmark prediction results based on the validation sample set; S42. Using the feature distribution of the training sample set and the parameter vector of the benchmark model, perform statistical analysis on the parameter sensitivity of the deep learning prediction model, calculate the degree of influence of each deep learning prediction model parameter on the prediction performance, and generate a parameter importance vector. S43. Establish an elastic weight consolidation parameter library based on parameter importance vector, and store and bind parameter importance with corresponding deep learning prediction model parameters; S44. Mark the baseline model parameter vector as the reference parameter vector and establish an index relationship in the parameter storage unit.

[0011] Optionally, S5 specifically includes: S51. Receive incremental user measurement data from the network operation environment. The incremental user measurement data includes RSRP and SINR observations in the latest time window, user location information, building and room identifiers, and performs time and space matching with signal source parameter data. After removing abnormal measurement points, a valid incremental dataset is formed. S52. Jointly encode the effective incremental dataset with the 3D environment model and propagation feature set, and generate an incremental sample set based on the corresponding time, spatial location and material parameters. The input data dimension is consistent with the aligned training sample set. S53. Load the reference parameter vector and the elastic weight consolidation parameter library, use the reference parameter vector as the initial weight, and use the parameter importance vector in the elastic weight consolidation parameter library as the basis for regularization constraints to limit the update range of key weights. S54. In the deep learning prediction model, the AdamW optimizer with weight decay is used for incremental training. The incremental sample set is input into the graph neural network and the three-dimensional convolutional neural network structure, forward propagation is performed to calculate the prediction output, and back propagation is performed based on the prediction error to update the parameters of the deep learning prediction model. S55. In the incremental training process, random deactivation and L2 regularization strategies are introduced, and elastic weights are combined to consolidate the regularization terms to jointly constrain parameter updates. S56. The prediction error of the incremental sample set is calculated using the loss function, and the weighted sum of the errors of signal strength and signal-to-noise ratio is used as the total loss. S57. Monitor the loss function during continuous iteration. When the total loss decrease rate is lower than the set threshold after several consecutive training rounds, determine convergence, export the updated model parameter vector and generate incremental prediction results. S58. Write the updated model parameter vector and incremental prediction results to the storage module, and record the corresponding training time, data batch number and optimization log.

[0012] Optionally, the application of elastic weight consolidation in S53 specifically includes: S531. Read the reference parameter values ​​and parameter importance weights of each deep learning prediction model parameter from the elastic weight consolidation parameter library, set the update range of each parameter to be within the range of ±0.05 standard deviation of the reference parameter value, and use the standard deviation range as the parameter adjustment boundary for this round of incremental training. S532. During incremental training, calculate the difference between the updated parameter value and the reference parameter value for each parameter, square the difference, multiply it by the parameter importance weight, and then sum it with the error value output by the loss function to form the total loss value. S533. Input the total loss value into the optimizer to calculate the gradient direction, adjust the model parameter update magnitude according to the weighting result, generate a new parameter vector, output the updated deep learning prediction model parameter set, and write the prediction model parameter set into the storage module.

[0013] Optionally, S6 specifically includes: S61. Load and update the model parameter vector into the deep learning prediction model. Input the 3D environment data, signal source parameter data and user measurement data into the model. Perform forward inference in the graph neural network and 3D convolutional neural network structure to generate the signal intensity prediction results and signal-to-noise ratio prediction results for each spatial location. S62. Map the signal strength prediction results to the spatial coordinates of the three-dimensional environment model, project the prediction values ​​onto the three-dimensional scene, and generate a signal coverage distribution map and spatial signal strength visualization results. S63. Compare the prediction results with the actual values ​​in the user's measurement data to calculate the mean absolute error, mean square error and relative error, and generate a comprehensive error assessment result. S64. After the error assessment is completed, store the signal strength prediction results, signal-to-noise ratio prediction results, three-dimensional coverage distribution map and error assessment results as the output dataset.

[0014] Optionally, S7 specifically includes: S71. Calculate the updated value of the parameter importance vector based on the error assessment results, adjust the importance of the corresponding parameters in the elastic weight consolidation parameter library, and record the update log. S72. Set convergence conditions, including determining convergence if the total loss decrease rate is lower than a preset threshold in multiple consecutive training rounds and if the mean absolute error and mean square error are stable within the target range. S73. Determine if the convergence condition is met. If not, return to S5 to execute the next round of incremental training. If it is met, export the final deep learning prediction model, the updated parameter importance vector, and the three-dimensional coverage distribution map as the final output.

[0015] The beneficial effects of this invention are: (1) This invention constructs a deep learning prediction model by combining graph neural networks and three-dimensional convolutional neural networks, thereby achieving simultaneous modeling of spatial topological relationships and local features in the signal propagation path. This model can accurately reflect the propagation characteristics of signals in complex three-dimensional environments and improve the prediction accuracy in multi-story buildings, buildings with different materials, and mixed indoor and outdoor scenarios.

[0016] (2) The present invention introduces an elastic weight consolidation mechanism, establishes a parameter importance vector and parameter library, and constrains key weights during the incremental training of the model, so that the model can maintain the stability of the learned knowledge when new measurement data is introduced, avoid catastrophic forgetting problem, and improve the model's continuous learning ability and long-term convergence performance.

[0017] (3) This invention utilizes a multi-task regression structure to synchronously output signal strength and signal-to-noise ratio results, and combines error evaluation feedback for dynamic optimization, so that the model achieves a balance between prediction accuracy, robustness and generalization ability, and has high engineering practicality and scalability. It can be used in scenarios such as wireless network planning, coverage optimization and signal simulation analysis. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows a method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks proposed in this invention. Figure 2 This is a schematic diagram of the fusion structure of graph neural network and three-dimensional convolutional neural network in the 4G and 5G signal three-dimensional propagation modeling method based on graph neural network proposed in this invention; Figure 3 This is a flowchart of the elastic weight consolidation mechanism of a three-dimensional propagation modeling method for 4G and 5G signals based on graph neural networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-3 A method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks includes the following steps: S1. Acquire 3D environment data, signal source parameter data and user measurement data, establish a 3D environment model and material parameter library, and generate training sample set and verification sample set; S2. Perform propagation path modeling on the 3D environment model, calculate the propagation feature set of direct, reflected, and diffracted light, and align it with the training sample set to form an aligned training sample set; S3. Construct a deep learning prediction model, using a multi-task regression structure of graph neural network and three-dimensional convolutional neural network, and use an aligned training sample set for initial training to obtain the baseline model parameter vector and baseline prediction results. S4. Use the verification sample set to evaluate the benchmark prediction results, calculate the parameter importance vector based on the historical training sample set, form an elastic weight consolidation parameter library, and set the benchmark model parameter vector as the reference parameter vector. S5. Receive incremental user measurement data, construct an incremental sample set based on the deep learning prediction model and the propagation feature set, perform incremental fine-tuning based on the reference parameter vector and elastic weights to obtain the updated model parameter vector and incremental prediction results. S6. Use the updated model parameter vector to infer the deep learning prediction model, generate signal strength prediction results and three-dimensional coverage distribution map, and calculate the error evaluation results. S7. Update the parameter importance vector and elastic weights to consolidate the parameter library based on the error evaluation results. If the convergence condition is not met, return to S5. If the convergence condition is met, output the deep learning prediction model and the three-dimensional coverage distribution map.

[0021] In this embodiment, S1 specifically includes: S11. Acquire three-dimensional environment data, establish a three-dimensional environment model, establish a material parameter library and set dielectric constant, conductivity and reflection coefficient; S12. Obtain the signal source parameter set and user measurement data set, record the antenna position, transmission frequency, angle, transmission power, time stamp, building identifier, room identifier, RSRP, and SINR; S13. Construct training and validation sample sets based on user measurement data sets and ray tracing results.

[0022] In this embodiment, the data preparation in S1 specifically includes: 3D environmental data: a set of 3D coordinate data for buildings, rooms, roads, and terrain. Each data point consists of three-dimensional coordinates. express.

[0023] Signal source parameter data: Antenna base station set Each antenna ,in Antenna position, For the antenna base station transmission frequency, For the antenna base station angle, This represents the antenna's transmit power.

[0024] User measurement data: User signal data set Each user ,in Indicates the time stamp of the user signal. Indicates the user's time The user's building or room Indicates time Real-time user signal data, such as RSRP and SINR, are used to correct and train deep learning prediction models.

[0025] In this embodiment, the three-dimensional spatial modeling in S1 specifically includes: Spatial coordinates of 3D environmental data Accurate three-dimensional spatial models can be established using point cloud data, CAD or BIM models. Assign electromagnetic parameters, including dielectric constant, to different materials in the scene (such as concrete, glass, and metal). conductivity Reflection coefficient .

[0026] In this embodiment, S2 specifically includes: S21. Identify the signal propagation path in the three-dimensional environment model and determine the direct path, reflection path, and diffraction path of the signal from the signal source to the user's location. S22. Calculate the signal propagation characteristics based on different path types, consider the influence of spatial obstacles and material parameters, and generate a set of propagation characteristics that includes spatial location information and signal change characteristics. S23. Align the propagation feature set with the training sample set in terms of spatial location and time label to generate an aligned training sample set for training the deep learning prediction model.

[0027] In this embodiment, S22 specifically includes: S221. Distance calculation, if the transmitter position is... The receiving user location is Then the Euclidean distance between the two points is: ; S222, Path Loss: ; in, This is the total distance between the user and the connected antenna. Path loss at reference distance This is the path loss index.

[0028] S223, Received power calculation: ; in, Transmit power is the signal power emitted by the antenna. It is path loss. This is the transmitting antenna gain, which indicates the transmitting antenna's radiation capability in a certain direction. It is the antenna gain of the receiver, which indicates the receiver's reception capability in a certain direction.

[0029] S224. Map the calculated signal intensity onto the 3D spatial scene network to generate a heat map: ; in, Indicates three-dimensional position and time Signal strength below, Material parameters include dielectric constant. conductivity Reflection coefficient .

[0030] In this embodiment, S3 specifically includes: S31. Construct a deep learning prediction model using a joint structure of graph neural network and three-dimensional convolutional neural network; S32. Input the three-dimensional environment data, signal source parameter data and user measurement data into the graph neural network and the three-dimensional convolutional neural network. The graph neural network outputs the node embedding representation, and the three-dimensional convolutional neural network outputs the local spatial feature tensor. S33. Fuse node embedding representation with local spatial feature tensor, perform feature concatenation and attention weighting, map to joint spatial feature representation, construct multi-task regression head, the first regression head is used to output signal strength, the second regression head is used to output signal-to-noise ratio; S34. Use the loss function to calculate the error between the predicted value and the true value, and use the weighted sum of the mean square error and the mean absolute error as the loss function. S35. During the training process, a regularization mechanism is introduced, a random deactivation method is used to suppress overfitting, and L2 weight decay is used to limit the parameter amplitude. S36. Using the AdamW optimizer with weight decay, perform forward and backward propagation to update the deep learning prediction model parameter vector until the loss function value reaches the convergence condition, completing the initial training, and outputting the baseline model parameter vector and baseline prediction results; the initial learning rate of the AdamW optimizer is set to... Cosine annealing scheduling is employed; the convergence condition is that the rate of change of the validation set loss function value is less than 1% for five consecutive rounds. S37. After training convergence, export and save the baseline model parameter vector, generate baseline prediction results and corresponding evaluation metrics based on the validation sample set, and record logs and version information.

[0031] In this embodiment, the regularization mechanism in step 35 and the AdamW optimizer optimization process in step 36 specifically include: S351. In the joint structure of graph neural network and three-dimensional convolutional neural network, a random deactivation unit is set up. At the output end of each hidden layer, some neurons are randomly blocked according to a preset deactivation probability. A dynamic activation structure is formed in each forward propagation to reduce the degree of feature coupling. S352. Apply L2 regularization constraints to all trainable weight parameters, calculate the sum of squares of each parameter and multiply it by the regularization coefficient to form a weight decay term, and add it to the loss function. S353. Select the AdamW optimizer with weight decay mechanism to perform parameter updates, and set the initial learning rate to... ,use , momentum parameters and The numerical stability term; S354. During the training process, a cosine annealing learning rate scheduling strategy is introduced. The current learning rate is calculated based on the iteration rounds and the total number of training rounds, so that the learning rate gradually decreases as the training progresses. S355. In each round of backpropagation, the optimizer calculates the gradient components corresponding to the main loss term and the regularization term respectively, keeps the gradient zero for inactive nodes and updates the parameters for effective nodes. S356. During continuous training, record the total loss and the ratio of the regularization term. When the rate of change of the total loss and the rate of change of the learning rate decay curve are both less than 5%, the training process is considered to have converged, and the optimized model parameter vector is output.

[0032] In this embodiment, the construction of the deep learning prediction model specifically includes: Input construction: User at time The three-dimensional spatial coordinates below Signal source parameters Including antenna base station transmission frequency Antenna base station angle Antenna base station transmit power Environmental characteristics Including dielectric constant conductivity Reflection coefficient Building obstruction coefficient .

[0033] Model Selection and Design: Graph neural networks, particularly graph attention networks, are employed to capture channel correlations among spatially adjacent users. Simultaneously, a three-dimensional convolutional neural network is introduced to extract local spatial features of signal strength from a three-dimensional grid. The model uses a multi-task regression approach to output predicted values, including signal strength. , and signal-to-noise ratio (SINR).

[0034] In this embodiment, S4 specifically includes: S41. Perform error analysis and performance evaluation on the benchmark prediction results based on the validation sample set; S42. Using the feature distribution of the training sample set and the parameter vector of the benchmark model, perform statistical analysis on the parameter sensitivity of the deep learning prediction model, calculate the degree of influence of each deep learning prediction model parameter on the prediction performance, and generate a parameter importance vector; in this embodiment, the parameter importance vector is calculated based on the Fisher information matrix. S43. Establish an elastic weight consolidation parameter library based on parameter importance vector, and store and bind parameter importance with corresponding deep learning prediction model parameters; S44. Mark the baseline model parameter vector as the reference parameter vector and establish an index relationship in the parameter storage unit.

[0035] In this embodiment, S5 specifically includes: S51. Receive incremental user measurement data from the network operation environment. The incremental user measurement data includes RSRP and SINR observations in the latest time window, user location information, building and room identifiers, and performs time and space matching with signal source parameter data. After removing abnormal measurement points, a valid incremental dataset is formed. S52. Jointly encode the effective incremental dataset with the 3D environment model and propagation feature set, and generate an incremental sample set based on the corresponding time, spatial location and material parameters. The input data dimension is consistent with the aligned training sample set. S53. Load the reference parameter vector and the elastic weight consolidation parameter library, use the reference parameter vector as the initial weight, and use the parameter importance vector in the elastic weight consolidation parameter library as the basis for regularization constraints to limit the update range of key weights. S54. In the deep learning prediction model, the AdamW optimizer with weight decay is used for incremental training. The incremental sample set is input into the graph neural network and the three-dimensional convolutional neural network structure, forward propagation is performed to calculate the prediction output, and back propagation is performed based on the prediction error to update the parameters of the deep learning prediction model. S55. In the incremental training process, random deactivation and L2 regularization strategies are introduced, and elastic weights are combined to consolidate the regularization terms to jointly constrain parameter updates. S56. The prediction error of the incremental sample set is calculated using the loss function, and the weighted sum of the errors of signal strength and signal-to-noise ratio is used as the total loss. S57. During continuous iteration, monitor the loss function. When the total loss decrease rate is lower than the set threshold after several consecutive training rounds, convergence is determined, the updated model parameter vector is exported, and incremental prediction results are generated. In this embodiment, the threshold is set to 0.001. S58. Write the updated model parameter vector and incremental prediction results to the storage module, and record the corresponding training time, data batch number and optimization log.

[0036] In this embodiment, the application of elastic weight consolidation in S53 specifically includes: S531. Read the reference parameter values ​​and parameter importance weights of each deep learning prediction model parameter from the elastic weight consolidation parameter library, set the update range of each parameter to be within the range of ±0.05 standard deviation of the reference parameter value, and use the standard deviation range as the parameter adjustment boundary for this round of incremental training. S532. During incremental training, calculate the difference between the updated parameter value and the reference parameter value for each parameter, square the difference, multiply it by the parameter importance weight, and then sum it with the error value output by the loss function to form the total loss value. S533. Input the total loss value into the optimizer to calculate the gradient direction, adjust the model parameter update magnitude according to the weighting result, generate a new parameter vector, output the updated deep learning prediction model parameter set, and write the prediction model parameter set into the storage module.

[0037] In this embodiment, S6 specifically includes: S61. Load and update the model parameter vector into the deep learning prediction model. Input the 3D environment data, signal source parameter data and user measurement data into the model. Perform forward inference in the graph neural network and 3D convolutional neural network structure to generate the signal intensity prediction results and signal-to-noise ratio prediction results for each spatial location. S62. Map the signal strength prediction results to the spatial coordinates of the three-dimensional environment model, project the prediction values ​​onto the three-dimensional scene, and generate a signal coverage distribution map and spatial signal strength visualization results. S63. Compare the prediction results with the actual values ​​in the user's measurement data to calculate the mean absolute error, mean square error and relative error, and generate a comprehensive error assessment result. S64. After the error assessment is completed, store the signal strength prediction results, signal-to-noise ratio prediction results, three-dimensional coverage distribution map and error assessment results as the output dataset.

[0038] In this embodiment, S7 specifically includes: S71. Calculate the updated value of the parameter importance vector based on the error assessment results, adjust the importance of the corresponding parameters in the elastic weight consolidation parameter library, and record the update log. S72. Set convergence conditions, including determining convergence if the total loss decrease rate is lower than a preset threshold and the mean absolute error and mean square error are stable within the target range in multiple consecutive training rounds; in this embodiment, the total loss is less than 0.5% for 5 consecutive rounds, the mean absolute error is less than 0.1dB, and the mean square error is less than 1.5dB2. S73. Determine if the convergence condition is met. If not, return to S5 to execute the next round of incremental training. If it is met, export the final deep learning prediction model, the updated parameter importance vector, and the three-dimensional coverage distribution map as the final output.

[0039] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical multi-story building and outdoor mixed scenario. This scenario includes several high-rise buildings, outdoor roads, green belts, and some open areas. The signal source is arranged at transmitting antennas at different heights, and the user measurement points are distributed on multiple floors and in different spatial locations, forming a complex three-dimensional propagation environment.

[0040] In this scenario, an environmental model is first established based on 3D spatial data, assigning electromagnetic properties such as dielectric constant, conductivity, and reflection coefficient to walls, floors, and windows of different materials. Then, training and validation sample sets are generated by combining user-measured signal data. The proposed graph neural network and 3D convolutional neural network joint structure is used for signal propagation feature modeling. Initial training is performed using the AdamW optimizer with weighted decay, with an initial learning rate of 1×10⁻³, and a cosine annealing scheduling strategy is introduced to achieve adaptive learning rate reduction. Training stops when the continuous training loss decreases below 0.5% and the validation set error stabilizes, obtaining the baseline model parameter vector.

[0041] After model training is complete, an elastic weight consolidation mechanism is introduced. By calculating the parameter importance vector, the range of weight updates during incremental learning is constrained, ensuring model stability when receiving new measurement data. Subsequently, new user measurement data from different time periods are input into the model for incremental fine-tuning. Verification shows that the model convergence speed is improved during the incremental training phase, and the model no longer exhibits the phenomenon of forgetting old knowledge after introducing new data.

[0042] To demonstrate the beneficial effects of the method of this invention, the prediction performance of three methods was compared: the traditional ray tracing method, the model using only a three-dimensional convolutional neural network, and the graph neural network fusion model proposed in this invention. The comparison metrics included the mean absolute error of signal strength prediction, the mean square error, the signal-to-noise ratio prediction error, and the number of convergence rounds.

[0043] Table 1: Comparison of Model Prediction Performance

[0044] As shown in Table 1, the method of this invention reduces the mean absolute error in signal strength prediction by approximately 69.8% compared to the traditional ray tracing method and by approximately 41.5% compared to the simple 3D convolution model; the mean squared error is reduced by approximately 40%, and the signal-to-noise ratio prediction error also decreases from 1.25 dB to 0.78 dB. Notably, the performance degradation rate of the method of this invention after introducing incremental data is only 3.1%, while the performance degradation of the traditional deep learning model exceeds 14% during incremental training, indicating that the elastic weight consolidation mechanism effectively prevents parameter drift and forgetting problems. The model parameter stability is improved to 96.4%, and the number of convergence rounds is reduced to 120, which is approximately 33% faster than the control group.

[0045] The model of this invention exhibits a significantly more uniform spatial prediction error distribution in multi-story and complex indoor-outdoor overlapping scenarios, with the standard deviation of the prediction residual decreasing from 1.18 dB to 0.63 dB. The alignment accuracy between the 3D coverage visualization results and the measured data reaches over 95%. These results demonstrate that the spatial topological feature modeling capability of graph neural networks can effectively compensate for the shortcomings of 3D convolutional networks in local spatial feature extraction.

[0046] The above experiments demonstrate that the present invention has the following advantages in practical applications: This invention significantly outperforms traditional methods in terms of the accuracy of three-dimensional modeling of signal propagation, and is able to capture reflection and diffraction behavior in complex spaces; The model of this invention has adaptive incremental learning capability, which can dynamically update parameters according to new measurement data and maintain the long-term stability of the model. It has higher training efficiency, and the model structure maintains high accuracy while having better generalization and convergence, making it suitable for widespread use in various indoor and outdoor scenarios.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks, characterized in that, Includes the following steps: S1. Acquire 3D environment data, signal source parameter data and user measurement data, establish a 3D environment model and material parameter library, and generate training sample set and verification sample set; S2. Perform propagation path modeling on the 3D environment model, calculate the propagation feature set of direct, reflected, and diffracted light, and align it with the training sample set to form an aligned training sample set; S3. Construct a deep learning prediction model, using a multi-task regression structure of graph neural network and three-dimensional convolutional neural network, and use an aligned training sample set for initial training to obtain the baseline model parameter vector and baseline prediction results. S4. Use the verification sample set to evaluate the benchmark prediction results, calculate the parameter importance vector based on the historical training sample set, form an elastic weight consolidation parameter library, and set the benchmark model parameter vector as the reference parameter vector. S5. Receive incremental user measurement data, construct an incremental sample set based on the deep learning prediction model and the propagation feature set, perform incremental fine-tuning based on the reference parameter vector and elastic weights to obtain the updated model parameter vector and incremental prediction results. S6. Use the updated model parameter vector to infer the deep learning prediction model, generate signal strength prediction results and three-dimensional coverage distribution map, and calculate the error evaluation results. S7. Update the parameter importance vector and elastic weights to consolidate the parameter library based on the error evaluation results. If the convergence condition is not met, return to S5. If the convergence condition is met, output the deep learning prediction model and the three-dimensional coverage distribution map.

2. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 1, characterized in that, S1 specifically includes: S11. Acquire three-dimensional environment data, establish a three-dimensional environment model, establish a material parameter library and set dielectric constant, conductivity and reflection coefficient; S12. Obtain the signal source parameter set and user measurement data set, record the antenna position, transmission frequency, angle, transmission power, time stamp, building identifier, room identifier, RSRP, and SINR; S13. Construct training and validation sample sets based on user measurement data sets and ray tracing results.

3. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 2, characterized in that, S2 specifically includes: S21. Identify the signal propagation path in the three-dimensional environment model and determine the direct path, reflection path, and diffraction path of the signal from the signal source to the user's location. S22. Calculate the signal propagation characteristics based on different path types, consider the influence of spatial obstacles and material parameters, and generate a set of propagation characteristics that includes spatial location information and signal change characteristics. S23. Align the propagation feature set with the training sample set in terms of spatial location and time label to generate an aligned training sample set for training the deep learning prediction model.

4. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 3, characterized in that, S3 specifically includes: S31. Construct a deep learning prediction model using a joint structure of graph neural network and three-dimensional convolutional neural network; S32. Input the three-dimensional environment data, signal source parameter data and user measurement data into the graph neural network and the three-dimensional convolutional neural network. The graph neural network outputs the node embedding representation, and the three-dimensional convolutional neural network outputs the local spatial feature tensor. S33. Fuse node embedding representation with local spatial feature tensor, perform feature concatenation and attention weighting, map to joint spatial feature representation, construct multi-task regression head, the first regression head is used to output signal strength, the second regression head is used to output signal-to-noise ratio; S34. Use the loss function to calculate the error between the predicted value and the true value, and use the weighted sum of the mean square error and the mean absolute error as the loss function. S35. Introduce a regularization mechanism during training, and use random deactivation and L2 weight decay. S36. Select the AdamW optimizer with weight decay for forward and backward propagation, update the deep learning prediction model parameter vector until the loss function value reaches the convergence condition, complete the initial training, and output the baseline model parameter vector and baseline prediction results. S37. After training convergence, export and save the baseline model parameter vector, generate baseline prediction results and corresponding evaluation metrics based on the validation sample set, and record logs and version information.

5. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 4, characterized in that, S4 specifically includes: S41. Perform error analysis and performance evaluation on the benchmark prediction results based on the validation sample set; S42. Using the feature distribution of the training sample set and the parameter vector of the benchmark model, perform statistical analysis on the parameter sensitivity of the deep learning prediction model, calculate the degree of influence of each deep learning prediction model parameter on the prediction performance, and generate a parameter importance vector. S43. Establish an elastic weight consolidation parameter library based on parameter importance vector, and store and bind parameter importance with corresponding deep learning prediction model parameters; S44. Mark the baseline model parameter vector as the reference parameter vector and establish an index relationship in the parameter storage unit.

6. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 5, characterized in that, S5 specifically includes: S51. Receive incremental user measurement data from the network operation environment. The incremental user measurement data includes RSRP and SINR observations in the latest time window, user location information, building and room identifiers, and performs time and space matching with signal source parameter data. After removing abnormal measurement points, a valid incremental dataset is formed. S52. Jointly encode the effective incremental dataset with the 3D environment model and propagation feature set, and generate an incremental sample set based on the corresponding time, spatial location and material parameters. The input data dimension is consistent with the aligned training sample set. S53. Load the reference parameter vector and the elastic weight consolidation parameter library, use the reference parameter vector as the initial weight, and use the parameter importance vector in the elastic weight consolidation parameter library as the basis for regularization constraints to limit the update range of key weights. S54. In the deep learning prediction model, the AdamW optimizer with weight decay is used for incremental training. The incremental sample set is input into the graph neural network and the three-dimensional convolutional neural network structure, forward propagation is performed to calculate the prediction output, and back propagation is performed based on the prediction error to update the parameters of the deep learning prediction model. S55. In the incremental training process, random deactivation and L2 regularization strategies are introduced, and elastic weights are combined to consolidate the regularization terms to jointly constrain parameter updates. S56. The prediction error of the incremental sample set is calculated using the loss function, and the weighted sum of the errors of signal strength and signal-to-noise ratio is used as the total loss. S57. Monitor the loss function during continuous iteration. When the total loss decrease rate is lower than the set threshold after several consecutive training rounds, determine convergence, export the updated model parameter vector and generate incremental prediction results. S58. Write the updated model parameter vector and incremental prediction results to the storage module, and record the corresponding training time, data batch number and optimization log.

7. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 6, characterized in that, The specific application of elastic weight consolidation in S53 includes: S531. Read the reference parameter values ​​and parameter importance weights of each deep learning prediction model parameter from the elastic weight consolidation parameter library, set the update range of each parameter to be within the range of ±0.05 standard deviation of the reference parameter value, and use the standard deviation range as the parameter adjustment boundary for this round of incremental training. S532. During incremental training, calculate the difference between the updated parameter value and the reference parameter value for each parameter, square the difference, multiply it by the parameter importance weight, and then sum it with the error value output by the loss function to form the total loss value. S533. Input the total loss value into the optimizer to calculate the gradient direction, adjust the model parameter update magnitude according to the weighting result, generate a new parameter vector, output the updated deep learning prediction model parameter set, and write the prediction model parameter set into the storage module.

8. The method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 6, characterized in that, S6 specifically includes: S61. Load and update the model parameter vector into the deep learning prediction model. Input the 3D environment data, signal source parameter data and user measurement data into the model. Perform forward inference in the graph neural network and 3D convolutional neural network structure to generate the signal intensity prediction results and signal-to-noise ratio prediction results for each spatial location. S62. Map the signal strength prediction results to the spatial coordinates of the three-dimensional environment model, project the prediction values ​​onto the three-dimensional scene, and generate a signal coverage distribution map and spatial signal strength visualization results. S63. Compare the prediction results with the actual values ​​in the user's measurement data to calculate the mean absolute error, mean square error and relative error, and generate a comprehensive error assessment result. S64. After the error assessment is completed, store the signal strength prediction results, signal-to-noise ratio prediction results, three-dimensional coverage distribution map and error assessment results as the output dataset.

9. A method for three-dimensional propagation modeling of 4G and 5G signals based on graph neural networks according to claim 8, characterized in that, Specifically, S7 includes: S71. Calculate the updated value of the parameter importance vector based on the error assessment results, adjust the importance of the corresponding parameters in the elastic weight consolidation parameter library, and record the update log. S72. Set convergence conditions, including determining convergence if the total loss decrease rate is lower than a preset threshold in several consecutive training rounds and if the mean absolute error and mean square error are stable within the target range. S73. Determine if the convergence condition is met. If not, return to S5 to execute the next round of incremental training. If it is met, export the final deep learning prediction model, the updated parameter importance vector, and the three-dimensional coverage distribution map as the final output.

Images