Channel modeling method and system based on multipath generation mechanism of different types of structures

Abstract

The present invention discloses a channel modeling method and system based on the multipath generation mechanism of different types of structures, which relates to the field of channel modeling. The method includes: constructing a data set; performing data enhancement and grouping on the data set; constructing a deep learning model; training the deep learning model using the data set after data enhancement and data grouping; retraining the deep learning model using the grouped data; pruning the trained deep learning model to obtain a lightweight model; inputting the relative position information of the structure, transmitter, and receiver and the structure category label into the lightweight model to obtain a multipath signal propagation parameter set; and then constructing a railway scene millimeter wave time-varying air-to-ground channel model. The above scheme in the present invention can realize the accurate modeling and efficient prediction of the millimeter wave time-varying air-to-ground channel in the railway scene, and provide the necessary theoretical basis for the design and performance optimization of the millimeter wave mobile communication system of unmanned aerial vehicles in the railway scene.

Description

Technical Field

[0001] The present invention relates to the technical field of channel modeling, and in particular to a channel modeling method and system based on a multipath generation mechanism of different types of structures. Background Art

[0002] Existing modeling methods for air-to-ground channels can be roughly classified into two categories, namely deterministic channel modeling methods and statistical channel modeling methods. Compared with the traditional ground channel model, since the height of the UAV is much higher than the ground transceiver, the number of scatterers around it is much less than that of the ground transceiver; the distribution of scatterers between the two is very different, which will lead to different channel time-space-frequency domain characteristics. Therefore, the existing traditional ground channel model and its parameters cannot be directly applied to the UAV air-to-ground channel.

[0003] Deterministic channel modeling method: The establishment of the deterministic model of the air-to-ground channel of UAV is generally based on the electromagnetic field and electromagnetic wave propagation theory and geometric optics theory. Using the antenna model, the three-dimensional propagation environment model and the scattering model of the structure therein, according to the radio wave propagation mechanism, the multipath signals of each antenna unit at continuous time are calculated one by one to construct a spatial consistency channel model. Based on the fine and accurate three-dimensional modeling of the propagation environment, the deterministic channel model has high accuracy. In recent years, with the improvement of computer memory capacity and computing power, the deterministic channel modeling method has gained more and more attention in the air-to-ground channel modeling, and the more representative one is the ray tracing method. In the literature, ray tracing is used to model the air-to-ground channel of UAV in the 28GHz and 60GHz frequency bands, urban, suburban and marine environments. This type of channel model can simulate the dynamic characteristics of the channel delay domain and three-dimensional angle domain, as well as the change law of multipath distribution and structure over time. However, the channel ray tracing simulator currently used has not been calibrated by channel measurement data, and its accuracy needs to be further verified. It is worth noting that the deterministic channel model has limited application in the design and evaluation of UAV communication systems due to its high complexity and large amount of calculation, and it is difficult to promote.

[0004] Statistical channel modeling method: The statistical air-ground channel model is established by extracting channel characteristic parameters from channel measurement data, combining statistical probability distribution, and constructing a geometry-based stochastic channel model (GBSCM) and a non-geometrical stochastic channel model (NGSCM). This type of channel model does not pursue the accurate characterization of channel propagation characteristics in specific scenarios and specified locations, but is effective and fast in characterizing the general channel characteristics of a typical propagation environment. The NGSCM model was established in the literature, and its time-varying fast fading is described by a Markov process, switching between Ricean and Loo distributions according to the aircraft altitude; but this model is only applicable to narrowband and single-antenna communication systems. GBSCMs use point scatterers to represent the environment and simplify the radio wave propagation environment. For example, the structures around the ground station and the drone are modeled as point scatterers on the surface of a cylinder or a sphere. Starting from the geometric concept and driven by theoretical derivation, a regular-shaped GBSCM is constructed, and a closed-form expression of the channel space-time domain correlation is formed. However, in the actual environment, the distribution of scatterers mainly depends on the distribution and shape of the surface structures, and tends to be irregular and clustered. There are significant differences in the scattering characteristics of different structures, and the complex amplitude modeling of multipath signals should be different. The 3rd Generation Partnership Project (3GPP) provides a cluster structure channel model for drones as user terminals in LTE cellular communication networks. However, this model is aimed at the air-to-ground channel in the 2GHz frequency band, and its applicability in the millimeter wave frequency band needs further verification. In addition, under mobile conditions, the model cannot accurately describe the evolution of multipath propagation parameters over time, and the accuracy of the model has encountered bottlenecks. The currently published UAV air-to-ground GBSCM and NGSCM related work, although it can well describe the channel statistical characteristics of a typical propagation environment, is difficult to ensure the spatial consistency of the channel under the condition of UAV movement, resulting in a bottleneck in the accuracy improvement of the air-to-ground channel model.

[0005] In summary, although there is a millimeter-wave time-varying air-to-ground channel model, the accuracy of the model needs to be further verified due to the lack of actual channel measurement data. In addition, the deterministic air-to-ground channel modeling method can ensure the accuracy of the modeled time-varying channel under the premise of measurement verification of the simulator, but the large amount of calculation and high memory usage make it difficult to promote and apply in the link-level and system-level simulation of air-to-ground communication systems; the statistical channel modeling method is effective and fast in characterizing the general statistical characteristics of the channel in a type of propagation environment, but under mobile conditions, the spatial consistency of the constructed channel cannot be guaranteed, and the accuracy improvement of the model encounters a bottleneck. For the construction of the millimeter-wave time-varying air-to-ground channel model, how to achieve high-precision and low-complexity channel modeling while ensuring its spatial consistency is a challenging problem and the key to the promotion and application of the model, which needs to be further studied.

[0006] Existing work has trained neural networks in two steps to construct a time-varying air-ground channel model based on air-ground channel data generated by ray tracing simulation: the first step is to predict the communication link status (line-of-sight / non-line-of-sight) based on the transceiver location information, and the second step is to input the communication link status and location information to generate multipath propagation parameters, including PL, delay, and arrival and departure angles. The proposed channel model can "play back" the air-ground channel data of a specific scene, and well fits the specific distribution characteristics and correlation of the air-ground channel data with a certain position relationship in a specific scene; however, it does not deeply explore and learn the abstract statistical characteristics of the causal relationship between the propagation environment and the channel propagation parameters, and the generalization ability and robustness of the proposed model need to be further improved. However, building an air-ground channel model based on a neural network may bring opportunities for solving the above challenging problems.

[0007] Therefore, the present invention uses deep learning methods, based on the actual measurement data of millimeter-wave time-varying air-to-ground channels in railway scenarios, to analyze the complex mapping relationship between time-varying channels and propagation environments, and deeply explore the key features of radio wave propagation mechanisms in railway scenarios. Subsequently, a method for constructing a deep learning model of multipaths of different types of structures in railway air-to-ground propagation environments is proposed, which includes data enhancement and model structured pruning methods to reduce model complexity and calculation time, and improve the generalization ability of the model. The constructed lightweight model will be applied to the accurate modeling and efficient prediction of millimeter-wave time-varying air-to-ground channels in railway scenarios, providing the necessary theoretical basis for the design and performance optimization of millimeter-wave mobile communication systems for unmanned aerial vehicles in railway scenarios. Summary of the invention

[0008] The purpose of the present invention is to provide a channel modeling method and system based on the multipath generation mechanism of different types of structures, to achieve accurate modeling and efficient prediction of millimeter-wave time-varying air-to-ground channels in railway scenarios, and to provide the necessary theoretical basis for the design and performance optimization of millimeter-wave mobile communication systems for unmanned aerial vehicles in railway scenarios.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] In a first aspect, the present invention provides a channel modeling method based on a multipath generation mechanism of different types of structures, comprising:

[0011] Constructing a data set; the data set includes: relative position information of the structure, the transmitter, and the receiver, and a structure category label;

[0012] Performing data augmentation and data grouping on the data set;

[0013] Build deep learning models;

[0014] Training the deep learning model using the data set after data enhancement and data grouping;

[0015] Retraining the deep learning model using the grouped data;

[0016] Prune the trained deep learning model to obtain a lightweight model;

[0017] The relative position information of the structure, transmitter, and receiver as well as the structure category label are input into the lightweight model to obtain a set of multipath signal propagation parameters;

[0018] Based on the multipath signal propagation parameters, a millimeter-wave time-varying air-to-ground channel model for railway scenarios is constructed.

[0019] Optionally, performing data enhancement on the data set specifically includes:

[0020] The plane formed by the structure, the transmitter and the receiver is rotated and translated so that the transmitter and the receiver are located on the X-axis and are symmetrical about the origin.

[0021] Optionally, grouping the data set specifically includes:

[0022] The data with the same structure category label are grouped together to examine the impact of different relative positions of the transmitter, receiver and structure on multipath propagation parameters; then the category label is changed to continue to examine the impact of different relative positions, so as to achieve effective learning of the geometric characteristics of the multipath of the structure;

[0023] The data with similar relative positions of transmitters, receivers and structures are grouped together to investigate the influence of different structure category labels on multipath propagation parameters. Then, the relative position combination is changed to investigate the influence of different structure category labels. Effective learning of the scattering characteristics of multipath of structure action is achieved.

[0024] Optionally, pruning the trained deep learning model to obtain a lightweight model specifically includes the following steps:

[0025] Based on graph theory, the spectral clustering method is used to decompose the trained deep learning model to obtain multiple different neuron sets.

[0026] Optionally, based on graph theory, a spectral clustering method is used to decompose the trained deep learning model to obtain multiple different neuron sets, which specifically includes the following steps:

[0027] Consider each neuron collection as a “module”;

[0028] Represent the trained deep learning model as a weighted undirected graph, with all neurons considered as points in space;

[0029] If two neurons are in adjacent layers, there will be an undirected edge between them, and the weight of the edge will be equal to the absolute value of the weight of the connection between the two neurons; if the two neurons are not in adjacent layers, that is, there is no connection, the weight of the edge is 0;

[0030] Constructing a weight matrix of a deep learning model based on the weights;

[0031] Decomposing deep learning using a spectral clustering algorithm based on the weight matrix to obtain multiple modules;

[0032] Through damage experiments, the importance of different modules is evaluated;

[0033] Guiding structured pruning of deep learning models based on the importance of different modules.

[0034] Optionally, the structured pruning of the deep learning model guided by the importance of different modules specifically includes the following steps:

[0035] For each module, set all weights passed into its neurons to 0 and keep the rest of the deep learning model unchanged;

[0036] Evaluate the prediction accuracy of the damaged deep learning model, determine the impact of pruning on the deep learning model based on the loss function, and get the importance ranking of each module;

[0037] Prune the modules in order from low to high importance according to the current segmentation status;

[0038] Fine-tune the pruned deep learning model and evaluate the prediction accuracy of the deep learning model;

[0039] When the prediction accuracy of the fine-tuned deep learning model falls below the set threshold, pruning stops.

[0040] In a second aspect, the present invention provides a channel modeling system based on a multipath generation mechanism of different types of structures, the system comprising:

[0041] A data set construction module, used to construct a data set; the data set includes: relative position information of a structure, a transmitter, and a receiver, and a structure category label;

[0042] A data enhancement and grouping module, used for performing data enhancement and data grouping on the data set;

[0043] Deep learning model building module, used to build deep learning models;

[0044] A first training module, used to train the deep learning model using the data set after data enhancement and data grouping;

[0045] A second training module, used to retrain the deep learning model using the grouped data;

[0046] The pruning module is used to prune the trained deep learning model to obtain a lightweight model;

[0047] A multipath signal propagation parameter set determination module is used to input the relative position information of the structure, the transmitter, and the receiver and the structure category label into the lightweight model to obtain a multipath signal propagation parameter set;

[0048] The channel model construction module is used to construct a millimeter wave time-varying air-to-ground channel model for a railway scenario based on the multipath signal propagation parameters.

[0049] In a third aspect, the present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned channel modeling method based on the multipath generation mechanism of different types of structures.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned channel modeling method based on the multipath generation mechanism of different types of structures.

[0051] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0052] The present invention proposes a deep learning model construction method, data enhancement and model structured pruning method for multipaths of different types of structures in a railway air-ground propagation environment, and establishes a generalized lightweight model that can support spatially consistent high-precision, low-complexity millimeter-wave time-varying air-ground channel modeling. In order to achieve spatially consistent high-precision, low-complexity channel model construction, the present invention first studies the data enhancement method based on the radio wave propagation mechanism with the help of the good self-learning ability and nonlinear fitting ability of deep learning to prevent the model from learning "pseudo-relationships" and fundamentally guarantee the robustness of the model; then studies the construction method of the multipath deep learning model of the structure, and combines the multipath distance theory to establish a loss measure, and through supervised learning, uses the layer-by-layer greedy method to train the model, studies the parameter assignment and initial structure adjustment strategy of the model to improve the model training efficiency; studies the model convergence evaluation method; based on the graph theory, the training model is "moduled", and the measurement criteria of the importance of each module are studied. Based on this, a structured pruning method of the model is proposed to reduce the complexity of the model and improve the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0054] Figure 1 A flow chart of a channel modeling method based on a multipath generation mechanism of different types of structures provided by the present invention;

[0055] Figure 2 A schematic diagram of the structure of the deep learning model of multipath of structural action provided by the present invention;

[0056] Figure 3 A schematic diagram of the modular structure of the deep learning model provided by the present invention;

[0057] Figure 4 A schematic diagram of the structure of a channel modeling system based on the multipath generation mechanism of different types of structures provided by the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] The purpose of the present invention is to provide a channel modeling method and system based on the multipath generation mechanism of different types of structures, to achieve accurate modeling and efficient prediction of millimeter-wave time-varying air-to-ground channels in railway scenarios, and to provide the necessary theoretical basis for the design and performance optimization of millimeter-wave mobile communication systems for unmanned aerial vehicles in railway scenarios.

[0060] In the millimeter wave frequency band, the modeling of multipath of structures is generally based on geometric optics, and the model is highly accurate; however, it requires fine data to accurately characterize the structure, and the computational complexity is high, making it difficult to promote and apply in link-level and system-level simulations of air-to-ground communication systems. Therefore, based on deep learning, this paper studies the deep learning model modeling method of multipath of different types of structures, and studies data enhancement and model structured pruning methods to improve the generalization ability of the model; creates a lightweight model for accurate channel construction. The details are as follows:

[0061] First, based on the mechanism of radio wave propagation, we study the data enhancement method for training multipath models of different types of structures, explore the standardization method of input transmitter, receiver and structure position information, and design a data grouping strategy in supervised training based on causal theory;

[0062] Second, we study the construction method of deep learning models for multipaths of different types of structures, establish a loss function based on multipath distance theory, and study the model's parameter assignment and initial structure adjustment strategy in supervised training, as well as the model's convergence evaluation method;

[0063] Third, based on graph theory and with the help of spectral clustering methods, the trained model is decomposed into different "modules". Through damage experiments, the importance of modules and their measurement methods are studied. Based on this, the structured pruning method of the trained model is studied. For the lightweight model after pruning, the functions of its different modules are studied to make the model interpretable.

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Embodiment 1

[0066] Figure 1 The flow chart of the channel modeling method based on the multipath generation mechanism of different types of structures provided by the present invention is as follows: Figure 1 As shown, the method in the present invention comprises:

[0067] Step 1: Construct a data set; the data set includes: relative position information of the structure, transmitter, and receiver, and a structure category label.

[0068] Based on the channel measurement of railway scenes and the analysis of radio wave propagation mechanism, it can be seen that the scattering characteristics of the structures in the railway scenes and the relative position of the structures and the transceivers are the fundamental factors affecting the generation of channel propagation parameters. Therefore, deep learning is used for modeling That is, it simulates the propagation parameters of multipaths affected by the structure; its input data is: the relative position information of the structure, transmitter, and receiver, as well as the structure category label; the output data is a set of propagation parameters of a certain number of multipaths.

[0069] Step 2: Perform data augmentation and data grouping on the dataset.

[0070] The positions of the transmitter, receiver and structure are arbitrary, but the relative position relationship of the three points mainly affects the propagation parameters. Before data input, the plane formed by the three points is rotated and translated so that the transmitter and receiver are located on the X-axis and symmetrical about the origin to avoid interference with the model geometric feature learning caused by the arbitrary position distribution of the transmitter, receiver and structure.

[0071] In addition, in order to effectively learn the geometric characteristics and scattering characteristics of multipath of the structure, it is planned to group the data for learning. According to the influence rules and action mechanisms of different types of structures on multipath, the groups are as follows:

[0072] The data with the same structure category label are grouped together to examine the impact of different relative positions of the transmitter, receiver and structure on multipath propagation parameters; then the category label is changed to continue to examine the impact of different relative positions, so as to achieve effective learning of the geometric characteristics of the multipath of the structure;

[0073] The data with similar relative positions of transmitters, receivers and structures are grouped together to investigate the influence of different structure category labels on multipath propagation parameters. Then, the relative position combination is changed to investigate the influence of different structure category labels. Effective learning of the scattering characteristics of multipath of structure action is achieved.

[0074] Step 3: Build a deep learning model.

[0075] See also Figure 2 , Figure 2 is the deep learning network used for model building, whose input is the rotated and translated transmitter position X' Tx , receiver position X' Rx , the centroid position of the structure And the structure category label n, n = 1, ..., N, where N represents the N types of structures in the railway scene; the output is the multipath signal propagation parameter Ω related to the structure l , in, Represented by the structure Θ kThe modeled multipath propagation parameters do not include Doppler frequency shift, which is mainly determined by the moving speed and carrier wavelength. When the three-dimensional arrival angle and departure angle of the multipath have been modeled by the model, combined with the moving speed and carrier wavelength of the transceiver in the communication system, the multipath Doppler frequency shift can be modeled.

[0076] Step 4: Train the deep learning model using the data set after data enhancement and data grouping.

[0077] Step 5: Retrain the deep learning model using the grouped data.

[0078] The training process is a supervised learning process. It should be noted that the number of multipaths affected by the structure is not fixed, but considering the sparse multipath of the millimeter wave air-to-ground channel, the number of multipaths affected by the structure is not large. Defined as the maximum number of multipaths that the structure acts on; for insufficient multipaths In all cases, the corresponding output position is filled with "0" value. The loss function is defined based on multipath distance theory and cross entropy.

[0079] During the training process, the model is first trained with complete data using the layer-by-layer greedy training method; then, the grouped data is used to enable the model to achieve reinforcement learning of the geometric characteristics and scattering characteristics of the structure. By training the model, the model parameter assignment and initial structure adjustment strategy are studied to speed up the training. In addition, with the help of optimization theory and nonlinear function approximation theory, the model convergence evaluation method is studied to provide feedback to guide the training of the multipath model of the structure.

[0080] Step 6: Prune the trained deep learning model to obtain a lightweight model.

[0081] For the training of multipath models with different types of structures, it is difficult to determine the structure of the model (deep learning network level, structure, etc.), so a sufficiently large network is usually selected; at the same time, there will be parameter redundancy in the training to adjust the plasticity of the model and achieve rapid optimization of the model. However, after the training is completed, during the inference process, the model needs to be properly pruned to remove redundant nodes to make the model more efficient and mature, so as to reduce complexity and calculation time and promote model application; at the same time, proper pruning can also avoid overfitting, explore the model structure, and improve the generalization performance of the model. In the present invention, the compression of the deep learning model is achieved by structured pruning.

[0082] Typically, trained deep learning models are modular and can be divided into different sets of neurons, where each set is highly connected internally and the connections between sets are low. Based on graph theory, a spectral clustering method is proposed to decompose the trained network and regard each set as a "module". The trained network is represented as a weighted undirected graph, and all neurons are regarded as points in space. If two neurons are in adjacent layers, there will be an undirected edge between them, and the weight of the edge will be equal to the absolute value of the connection weight between the two neurons; if there is no connection, the weight of the edge is 0. The weight matrix of the trained network is thus defined, and the "module" decomposition of the network is achieved with the help of the spectral clustering algorithm. Figure 3 Illustration of the modular structure of a deep learning model.

[0083] Furthermore, through damage experiments, the importance of different modules is evaluated to guide the structured pruning of the model. For each module, all weights passed to its neurons are set to 0, and the rest of the network is kept unchanged; the prediction accuracy of the damaged network is evaluated, the impact of pruning on the network is judged based on the loss function, and the measurement criteria for evaluating the importance of modules are studied. Based on this, the pruning order will be carried out from low to high according to the importance of the modules in the current segmentation state. After pruning, the model is fine-tuned and the model prediction accuracy is evaluated; when the prediction accuracy of the fine-tuned model is lower than the set threshold, pruning is stopped. The pruned lightweight model has a clear modular structure, based on which the functions of each module can be understood, providing an effective way to interpret deep learning models.

[0084] Step 7: Input the relative position information of the structure, transmitter, and receiver as well as the structure category label into the lightweight model to obtain a set of multipath signal propagation parameters.

[0085] Step 8: Based on the multipath signal propagation parameters, a millimeter-wave time-varying air-ground channel model for a railway scenario is constructed.

[0086] Embodiment 2

[0087] In order to execute the method corresponding to the above embodiment 1 to achieve the corresponding functions and technical effects, a channel modeling system based on the multipath generation mechanism of different types of structures is provided below. Figure 4 , the system in the present invention comprises:

[0088] The data set construction module 201 is used to construct a data set; the data set includes: relative position information of the structure, the transmitter, and the receiver, and a structure category label;

[0089] A data enhancement and grouping module 202, configured to perform data enhancement and data grouping on the data set;

[0090] A deep learning model construction module 203, used to construct a deep learning model;

[0091] A first training module 204, configured to train the deep learning model using the data set after data enhancement and data grouping;

[0092] A second training module 205 is used to retrain the deep learning model using the grouped data;

[0093] A pruning module 206 is used to prune the trained deep learning model to obtain a lightweight model;

[0094] A multipath signal propagation parameter set determination module 207 is used to input the relative position information of the structure, the transmitter, and the receiver and the structure category label into the lightweight model to obtain a multipath signal propagation parameter set;

[0095] The channel model construction module 208 is used to construct a millimeter-wave time-varying air-to-ground channel model for a railway scenario based on the multipath signal propagation parameters.

[0096] Embodiment 3

[0097] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the channel modeling method based on the multipath generation mechanism of different types of structures provided in embodiment 1.

[0098] Embodiment 4

[0099] Based on the description of the third embodiment, the fourth embodiment of the present invention provides a storage medium on which a computer program is stored. The computer program can be executed by a processor to implement the channel modeling method of the first embodiment based on the multipath generation mechanism of different types of structures.

[0100] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0101] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A channel modeling method based on the multipath generation mechanism of different types of structures, characterized in that: include: Build a dataset; The data set includes: relative position information of the structure, transmitter, and receiver, and a structure category label; Data enhancement and data grouping are performed on the data set; data enhancement of the data set specifically includes: rotating and translating the plane composed of the structure, the transmitter and the receiver so that the transmitter and the receiver are located on the X-axis and are symmetrical about the origin; data grouping of the data set specifically includes: grouping data with the same structure category label as one group, examining the influence of different relative positions of transmitters, receivers and structures on multipath propagation parameters; then changing the category label, continuing to examine the influence of different relative positions; so as to achieve effective learning of geometric features of multipath of structure effect; grouping data with similar relative positions of transmitters, receivers and structures as one group, examining the influence of different structure category labels on multipath propagation parameters; then changing the relative position combination, examining the influence of different structure category labels; achieving effective learning of scattering features of multipath of structure effect; Constructing a deep learning model, specifically including: inputting the transmitter position, receiver position, structure centroid position and structure category label after rotation and translation, and outputting the multipath signal propagation parameters related to the structure; the modeled multipath propagation parameters do not include Doppler frequency shift, which is determined by the moving speed and carrier wavelength; when the three-dimensional arrival angle and departure angle of the multipath have been modeled by the model, combined with the moving speed and carrier wavelength of the transceiver in the communication system, the multipath Doppler frequency shift is modeled; Training the deep learning model using the data set after data enhancement and data grouping; The deep learning model is trained again using the grouped data, and the training process is a supervised learning process; the number of multipaths acted on by the structure is not fixed, and considering that the multipaths of the millimeter wave air-to-ground channel are sparse, the number of multipaths acted on by the structure is not large; L k Defined as the maximum number of multipaths that the structure acts on; for multipaths less than L k In any case, the corresponding output position is filled with a "0" value; the loss function is defined based on the multipath distance theory and cross entropy; in the training process, the layer-by-layer greedy training method is used to first train the deep learning model with complete data; then, the grouped data is used to enable the deep learning model to implement reinforcement learning on the geometric characteristics and scattering characteristics of the structure; by training the deep learning model, the parameter assignment and initial structure adjustment strategy of the deep learning model are studied to speed up the training speed; through the optimization theory and nonlinear function approximation theory, the convergence evaluation method of the deep learning model is studied to provide feedback to guide the training of the multipath model of the structure; The trained deep learning model is pruned to obtain a lightweight model, which specifically includes the following steps: based on graph theory, the trained deep learning model is decomposed by using a spectral clustering method to obtain multiple different neuron sets, which specifically includes the following steps: each neuron set is regarded as a "module"; the trained deep learning model is represented as a weighted undirected graph, and all neurons are regarded as points in space; if two neurons are in adjacent layers, there will be an undirected edge between them, and the weight of the edge will be equal to the absolute value of the connection weight between the two neurons; if the two neurons are not in adjacent layers, that is, there is no connection, the weight of the edge is 0; a weight matrix of the deep learning model is constructed based on the weights; based on the weight matrix, a spectral clustering algorithm is used to calculate the weight matrix; The method decomposes the deep learning model into multiple modules; evaluates the importance of different modules through damage experiments; guides the structured pruning of the deep learning model based on the importance of different modules, which specifically includes the following steps: for each module, set all weights passed into its neurons to 0, and keep the rest of the deep learning model unchanged; evaluate the prediction accuracy of the damaged deep learning model, judge the impact of pruning on the deep learning model based on the loss function, and get the importance ranking of each module; prune the modules in order from low to high according to the importance of the modules in the current segmentation state; fine-tune the pruned deep learning model and evaluate the prediction accuracy of the deep learning model; stop pruning when the prediction accuracy of the fine-tuned deep learning model is lower than the set threshold; The relative position information of the structure, transmitter, and receiver as well as the structure category label are input into the lightweight model to obtain a set of multipath signal propagation parameters; Based on the multipath signal propagation parameters, a millimeter-wave time-varying air-to-ground channel model for railway scenarios is constructed.

2. A channel modeling system based on the multipath generation mechanism of different types of structures, characterized in that: The system comprises: A data set construction module, used to construct a data set; the data set includes: relative position information of a structure, a transmitter, and a receiver, and a structure category label; The data enhancement and grouping module is used to perform data enhancement and data grouping on the data set; the data enhancement of the data set specifically includes: rotating and translating the plane composed of the structure, the transmitter and the receiver so that the transmitter and the receiver are located on the X-axis and are symmetrical about the origin; the data grouping of the data set specifically includes: grouping data with the same structure category label, investigating the influence of different relative positions of transmitters, receivers and structures on multipath propagation parameters; then changing the category label, continuing to investigate the influence of different relative positions; so as to achieve effective learning of geometric features of multipath of structure effect; grouping data with similar relative positions of transmitters, receivers and structures, investigating the influence of different structure category labels on multipath propagation parameters; then changing the relative position combination, investigating the influence of different structure category labels; achieving effective learning of scattering features of multipath of structure effect; The deep learning model construction module is used to construct the deep learning model, which specifically includes: the input is the transmitter position, receiver position, structure centroid position and structure category label after rotation and translation, and the output is the multipath signal propagation parameters related to the structure; the modeled multipath propagation parameters do not include Doppler frequency shift, which is determined by the moving speed and carrier wavelength; when the three-dimensional arrival angle and departure angle of the multipath have been modeled by the model, combined with the moving speed and carrier wavelength of the transceiver in the communication system, the multipath Doppler frequency shift is modeled; A first training module, used to train the deep learning model using the data set after data enhancement and data grouping; The second training module is used to retrain the deep learning model using the grouped data, and the training process is a supervised learning process; the number of multipaths acted on by the structure is not fixed. Considering that the multipaths of the millimeter wave air-to-ground channel are sparse, the number of multipaths acted on by the structure is not large; L k Defined as the maximum number of multipaths that the structure acts on; for multipaths less than L k In any case, the corresponding output position is filled with a "0" value; the loss function is defined based on the multipath distance theory and cross entropy; in the training process, the layer-by-layer greedy training method is used to first train the deep learning model with complete data; then, the grouped data is used to enable the deep learning model to implement reinforcement learning on the geometric characteristics and scattering characteristics of the structure; by training the deep learning model, the parameter assignment and initial structure adjustment strategy of the deep learning model are studied to speed up the training speed; through the optimization theory and nonlinear function approximation theory, the convergence evaluation method of the deep learning model is studied to provide feedback to guide the training of the multipath model of the structure; The pruning module is used to prune the trained deep learning model to obtain a lightweight model, which specifically includes the following steps: based on graph theory, the trained deep learning model is decomposed by using the spectral clustering method to obtain multiple different neuron sets, which specifically includes the following steps: each neuron set is regarded as a "module"; the trained deep learning model is represented as a weighted undirected graph, and all neurons are regarded as points in space; if two neurons are in adjacent layers, there will be an undirected edge between them, and the weight of the edge will be equal to the absolute value of the connection weight between the two neurons; if the two neurons are not in adjacent layers, that is, there is no connection, the weight of the edge is 0; based on the weights, a weight matrix of the deep learning model is constructed; based on the weight matrix, a spectral clustering method is used to calculate the weight matrix of the deep learning model; based on the spectral clustering method, a weight matrix of the deep learning model is constructed by using the spectral clustering method .... The clustering algorithm decomposes the deep learning model to obtain multiple modules; the importance of different modules is evaluated through damage experiments; the structured pruning of the deep learning model is guided based on the importance of different modules, which specifically includes the following steps: for each module, all weights passed to its neurons are set to 0, and the rest of the deep learning model is kept unchanged; the prediction accuracy of the damaged deep learning model is evaluated, and the impact of pruning on the deep learning model is judged based on the loss function to obtain the importance ranking of each module; the modules are pruned in order from low to high importance in the current segmentation state; the pruned deep learning model is fine-tuned and the prediction accuracy of the deep learning model is evaluated; when the prediction accuracy of the fine-tuned deep learning model is lower than the set threshold, the pruning is stopped; A multipath signal propagation parameter set determination module is used to input the relative position information of the structure, the transmitter, and the receiver and the structure category label into the lightweight model to obtain a multipath signal propagation parameter set; The channel model construction module is used to construct a millimeter wave time-varying air-to-ground channel model for a railway scenario based on the multipath signal propagation parameters.

3. An electronic device, characterized in that: It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the channel modeling method based on the multipath generation mechanism of different types of structures as described in claim 1.

4. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, the channel modeling method based on the multipath generation mechanism of different types of structures as claimed in claim 1 is implemented.

Images