Non-uniform evaporation atmospheric waveguide channel modeling method

Abstract

The invention provides a non-uniform evaporation atmospheric waveguide channel modeling method, and relates to the technical field of atmospheric evaporation waveguide channel modeling, and the method comprises the steps: obtaining spatial refractive index distribution through calculation based on meteorological parameters and related weather parameter continuous distribution before and after generation; calculating by using an equation according to the spatial refractive index distribution to obtain a relation between a ray trajectory and a ray angle; constructing an evaporation waveguide prediction model by using a sequence feature coding module and a constructed channel model decoding module; and training the evaporation waveguide prediction model by using the relation between the ray trajectory and the ray angle to obtain a trained evaporation waveguide prediction model, and completing modeling of the non-uniform evaporation atmospheric waveguide channel. According to the method, the problem that high-precision and high-efficiency channel modeling is difficult to carry out in a horizontal non-uniform evaporation waveguide environment is solved.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of atmospheric evaporation duct channel modeling, in particular to a non-uniform evaporation atmospheric duct channel modeling method. BACKGROUND

[0002] Evaporation duct is a special marine atmospheric refraction phenomenon that can capture electromagnetic wave propagation within a certain angle, realize over-the-horizon, low-loss transmission, and thus has important application value in over-the-horizon radar, long-distance communication and other national defense and civilian fields. The core modeling problem is to accurately obtain the influence of the abnormal channel on the amplitude, time delay and other characteristics of electromagnetic wave propagation. However, the atmospheric duct environment is affected by multiple parameters such as temperature, humidity, air pressure, etc., and presents a highly complex and time-varying characteristic, making it particularly difficult to accurately obtain the channel response.

[0003] At present, evaporation duct channel modeling is in a rapid development stage. The traditional atmospheric duct channel modeling research mainly relies on two types of approaches: numerical calculation and statistical analysis. In terms of numerical calculation, the parabolic equation method (PE) simplifies the wave equation into a parabolic partial differential equation for solving, which can effectively handle complex boundary and refractive index conditions, and is a classic method for calculating path loss (PL); the ray tracing method (RT) is based on geometric optics and Snell's law, which traces the ray trajectory by solving the eikonal equation, and then analyzes the multipath time delay and angle of arrival (AOA) of the channel. However, such numerical methods have a sharp increase in computational complexity when simulating large-scale propagation scenarios, and the time consumption is significant, making it difficult to meet the real-time requirements of real communication environments and other applications with strict real-time requirements. In terms of statistical analysis, in order to improve efficiency, relevant scholars have adopted schemes such as least squares fitting model coefficients, Kullback-Leibler divergence to determine shadow fading distribution, etc. In addition, some studies have also conducted statistical modeling for specific sea areas. Such methods have relatively small computational load, but the accuracy and universality of the model are strongly dependent on the sample data of a specific sea area, and lack the ability to adapt to different regions and different weather conditions, resulting in insufficient generalization and accuracy. SUMMARY

[0004] In view of the above problems in the prior art, the non-uniform evaporation atmospheric duct channel modeling method provided by the present application solves the problem of difficult high-precision and high-efficiency channel modeling in a horizontally non-uniform evaporation duct environment.

[0005] In order to achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows: a non-uniform evaporation atmospheric duct channel modeling method, comprising: S1: based on meteorological parameters, generating a continuous distribution of weather parameters related in front and back, and obtaining a spatial refractive index distribution by calculation; S2: According to the spatial refractive index distribution, the equation is used for calculation to obtain the ray trajectory and the ray angle relationship; S3: A sequence feature coding module and a channel model decoding module are used to construct an evaporation waveguide prediction model; S4: The ray trajectory and the ray angle relationship are used to train the evaporation waveguide prediction model to obtain a trained evaporation waveguide prediction model, and the modeling of the non-uniform evaporation atmospheric waveguide channel is completed; wherein the trained evaporation waveguide prediction model is used to analyze the meteorological parameters to obtain an evaporation waveguide prediction result.

[0006] Further, the expression of the spatial refractive index distribution is: ; wherein, represents the spatial refractive index distribution, represents the air temperature, represents the atmospheric pressure, and represent constants, represents the water vapor pressure.

[0007] Further, the expression of the ray trajectory and the ray angle relationship is: ; ; ; ; wherein, represents the refractive index distribution, represents the horizontal index of the sampling point, represents the height index of the sampling point, represents the total number of height sampling points, represents the radius of the earth, represents the height of the ray trajectory, represents the ray angle, represents the height change, represents the angle change, represents the horizontal coordinate of the ray.

[0008] Further, the evaporation waveguide prediction model comprises: a sequence feature coding module, configured to extract key features affecting electromagnetic wave propagation from a horizontally non-uniform environmental parameter sequence to obtain a global feature vector; a channel model decoding module, configured to restore the global feature vector to obtain an evaporation waveguide prediction result.

[0009] Furthermore, the sequence feature encoding module includes a multi-layer Transformer encoder architecture, which extracts key features affecting electromagnetic wave propagation from a horizontally non-uniform environmental parameter sequence and establishes correlations between different locations. The sequence feature encoding module includes: an input feature mapping layer, a sequence enhancement preprocessing layer, a distance embedding layer, a position encoding layer, a multi-layer Transformer encoder, a multi-head self-attention mechanism, a feedforward neural network, residual connections and layer normalization, and a feature aggregation layer. Input Feature Mapping Layer: A two-layer fully connected network is used to map the 8-dimensional environmental parameters (temperature, humidity, wind speed, air pressure, transmit power, frequency, antenna height, distance) of each sampling point to a 256-dimensional Transformer hidden feature space. The specific structure is: 8-dimensional input to a 128-dimensional intermediate layer to a 256-dimensional output layer; The sequence enhancement preprocessing layer is used to enhance the local features of the input sequence using one-dimensional convolution, while preserving the original information through residual connections. The distance embedding layer is used to learn a 64-dimensional distance embedding vector for each position in the sequence, mapping position indices to distance features through a trainable embedding table; The positional coding layer is used to add absolute positional information to the sequence using sine-cosine positional coding. The coding dimension is 256, and the maximum sequence length is 200. The multi-layer Transformer encoder is constructed by stacking 6 Transformer encoder layers. Each layer contains: a multi-head self-attention mechanism: 8 attention heads, each with a dimension of 32, using scaled dot product attention. For each sample point in the sequence, the correlation weight with all other points is calculated to achieve global information interaction. The multi-head design enables the network to pay attention to different types of spatial correlations at the same time. Feedforward neural networks employ a structure that first increases dimensionality and then decreases it, using the GELU activation function to enhance the network's nonlinear expressive power. Residual connections and layer normalization are used in each sublayer; The feature aggregation layer employs a dual aggregation strategy, combining global average pooling and attention pooling. Global average pooling extracts overall features, while attention pooling highlights important positions using learned weights. The two types of features are then fused into the final global feature vector through a linear layer.

[0010] Furthermore, the channel model decoding module includes: a multi-layer feature projection network, a multi-scale sequence feature mapper, cross-scale feature fusion, a residual refinement module, and a final output mapping layer; The multi-layer feature projection network uses a three-layer fully connected network to progressively expand the one-dimensional global features: from 256 dimensions to 1024 dimensions to 512 dimensions to 4096 dimensions, and finally reshapes them into a 4×4 feature map with 256 channels, which serves as the initial seed for the generation of the two-dimensional field. A multi-scale sequence feature mapper is designed with two different scales to map 101-dimensional sequence features into 16×16 and 32×32 two-dimensional feature maps. Each mapper includes: sequence-to-space rearrangement: expanding 101 sequence points to 121 points through linear interpolation, reshaping into an 11×11 spatial layout; convolutional upsampling: upsampling the 11×11 feature map to the target size using transposed convolutions; adaptive pooling: precisely adjusting to the 16×16 or 32×32 size; and a progressive upsampling network: employing 5 levels of transposed convolutional upsampling: Level 1: 4×4 to 8×8, Level 2: 8×8 to 16×16 and fused with sequence features, Level 3: 16×16 to 32×32 and fused with sequence features, and Level 4: 32×32 to 64×64. Each level uses grouping normalization, GELU activation, and Dropout2d regularization. Cross-scale feature fusion involves concatenating and fusing upsampled features with corresponding sequence mapping features at 16×16 and 32×32 scales. This cross-scale fusion mechanism ensures that spatial information of the sequence is preserved during decoding, avoiding information loss. The residual refinement module uses three residual blocks to refine features at a 64×64 resolution. Each residual block contains two 3×3 convolutional layers, grouping normalization, GELU activation, and Dropout. The residual structure enables the network to finely adjust local details while preserving the large-scale propagation trend. The output mapping layer uses a three-layer convolutional network to map the multi-channel feature map to single-channel path loss values: from 24 channels to 16 channels to 8 channels to 1 channel. Finally, it uses bilinear interpolation to accurately upsample to a resolution of 101×101.

[0011] Furthermore, the expression for the loss function used in training the evaporating waveguide prediction model is as follows: ; in, Represents the loss function. Indicates the number of sampling points. This indicates the weight of the corresponding sampling point. Represents the predictive channel model for the sampling points. Let i represent the actual channel model of the sampling point, where i represents the i-th sampling point.

[0012] The beneficial effects of the present invention are as follows: The present invention provides a non-uniform evaporating atmospheric waveguide channel modeling method. According to the spatial refractive index distribution, the equation is used to calculate the relationship between the ray trajectory and the ray angle. The evaporating waveguide prediction model is trained using the relationship between the ray trajectory and the ray angle to obtain the trained evaporating waveguide prediction model and complete the modeling of the non-uniform evaporating atmospheric waveguide channel. (1) Spatial environment features are extracted by the global attention mechanism based on Transformer. The features are restored to a high-resolution two-dimensional field distribution by using a transposed convolutional network based on a CNN decoder. (2) The weather distribution related before and after is generated by Markov chain, and the propagation of electromagnetic waves in horizontal non-uniform conditions is simulated by segmented ray tracing. (3) The end-to-end evaporating waveguide channel modeling method directly takes weather parameters and transceiver parameters as input and outputs the final channel model. There is no need to perform "ray tracing-loss calculation" step by step. The output channel model is complete, the model is more practical, the modeling ability of long-distance spatial dependence is better, the characterization of horizontal non-uniformity is more realistic and the error is smaller, the modeling accuracy is higher, the channel model is more complete, the physical credibility is higher, the applicability is wider, and the computational efficiency is significantly improved. Attached Figure Description

[0013] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart of a non-uniform evaporating atmospheric waveguide channel modeling method according to some embodiments of this specification. Figure 2 This is an exemplary schematic diagram of the refractive index of an evaporating waveguide according to some embodiments of this specification; Figure 3 This is an exemplary schematic diagram of a non-uniform evaporation waveguide ray tracing trajectory according to some embodiments of this specification; Figure 4 This is an exemplary schematic diagram of an evaporating waveguide prediction model architecture shown in some embodiments of this specification; Figure 5 This is an exemplary schematic diagram of prediction results on a test set according to some embodiments of this specification. Detailed Implementation

[0014] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0015] Example Figure 1 This is an exemplary flowchart illustrating a non-uniform evaporating atmospheric waveguide channel modeling method according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0016] In some embodiments, the processor can first generate a horizontally non-uniform atmospheric parameter distribution that conforms to real physical laws through a Markov chain, combine this with an evaporation waveguide model to predict the spatial refractive index distribution, and then use piecewise ray tracing technology to generate an accurate channel response containing loss and time delay information. The proposed deep learning network with a self-attention mechanism-enhanced encoding-decoding structure can effectively capture the long-range dependencies between atmospheric parameters at different spatial locations along the propagation path, and can progressively restore global features to a high-resolution two-dimensional path loss field distribution, achieving end-to-end atmospheric waveguide channel modeling.

[0017] S1: Based on meteorological parameters, generate a continuous distribution of related weather parameters before and after, and obtain the spatial refractive index distribution through calculation.

[0018] Meteorological parameters include temperature, humidity, wind speed, air pressure, transmission power, frequency, antenna height, and distance.

[0019] Spatial refractive index distribution is data that reflects the relationship between refractive index and weather parameters.

[0020] In some embodiments, the processor can generate a continuous distribution of related weather parameters based on given meteorological parameters using Markov chain technology, and calculate the spatial refractive index distribution using an evaporation waveguide model; wherein, in the radio wave environment, the following is taken: , .

[0021] In some embodiments, such as Figure 2 The results of horizontal position sampling shown are used to randomly generate initial weather parameters based on the atmospheric parameter range of common atmospheric waveguides. Then, a Markov chain is used to generate a continuous atmospheric distribution in space, and the refractive index distribution in the entire space is obtained by the evaporation waveguide model.

[0022] In some embodiments, the expression for the spatial refractive index distribution is: ; in, Represents the spatial refractive index distribution. Indicates temperature, Indicates atmospheric pressure. and Represents a constant. This indicates water vapor pressure.

[0023] S2: Based on the spatial refractive index distribution, the relationship between the ray trajectory and the ray angle is obtained by using the equation.

[0024] The relationship between the ray trajectory and the ray angle is the discrete form of Snell's law in a spherically layered atmosphere.

[0025] In some embodiments, the processor can utilize the continuous distribution of refractive index in the modeling space, solve the Eckner equation by applying Snell's law, and comprehensively consider the influence of the spherical layered atmosphere and the curvature of the Earth. We then derive the discrete form of Snell's law in the spherical layered atmosphere, namely the relationship between the ray trajectory and the ray angle.

[0026] In some embodiments, the expression for the relationship between the ray trajectory and the ray angle is: ; ; ; ; in, Indicates the refractive index distribution. Indicates the horizontal index of the sampling point. Indicates the sampling point height index. This indicates the total number of height sampling points. Represents the Earth's radius. Indicates the height of the ray trajectory. Indicates the angle of the ray. Indicates changes in altitude. Indicates a change in angle. This represents the horizontal coordinate of the ray.

[0027] In some embodiments, the processor may take into account that the angle change in the evaporation waveguide effect is often small, i.e. and Smaller, through approximation and The original formula is simplified, and through partitioned iteration, the final relationship between the ray trajectory and ray angle is obtained. Compared to traditional algorithms that approximate... , only the refractive index of the overall area is equivalently processed through the waveguide strength and waveguide height. In this invention, the spatial refractive index in the first step is modeled, and through the differential approximation method, specifically, the atmospheric refractive index matrix is set The element in the i-th row (vertical direction, corresponding to height) and the j-th column (horizontal direction) in is calculated approximately by the difference method. When i = 1, forward difference is used; when i = rows, backward difference is used; when 1 < i < rows, central difference is used. This can effectively approximate the refractive index gradient distribution in a non-uniform environment at a limited resolution. After obtaining the overall refractive index gradient distribution, the ray propagation situation in the entire space can be obtained by sampling and iterating within a certain angle based on the relationship between the ray trajectory and the refractive index, as Figure 3 shown.

[0028] S3: Use the sequence feature encoding module and the constructed channel model decoding module to construct an evaporation duct prediction model.

[0029] The evaporation duct prediction model adopts a network architecture of Transformer encoder + CNN decoder to realize the prediction of the two-dimensional path loss field from the environmental parameters of the spatial sequence. The encoder part uses the global attention mechanism of Transformer to extract the spatial environmental features, and the decoder part uses the transposed convolutional network to restore the features to a high-resolution two-dimensional field distribution.

[0030] In some embodiments, as Figure 4 shown, the evaporation duct prediction model includes: a sequence feature encoding module, which is used to extract the key features affecting the electromagnetic wave propagation from the horizontally non-uniform environmental parameter sequence to obtain a global feature vector; a constructed channel model decoding module, which is used to restore the global feature vector to obtain the evaporation duct prediction result.

[0031] The sequence feature encoding module (TransformerEncoder) adopts a multi-layer Transformer encoder architecture to extract the key features affecting the electromagnetic wave propagation from the horizontally non-uniform environmental parameter sequence and establish the correlation relationship between different positions. The sequence feature encoding module includes: an input feature mapping layer, a sequence enhancement preprocessing layer, a distance embedding layer, a position encoding layer, a multi-layer Transformer encoder, a multi-head self-attention mechanism, a feed-forward neural network, a residual connection and layer normalization, and a feature aggregation layer.

[0032] Input Feature Mapping Layer: A two-layer fully connected network is used to map the 8-dimensional environmental parameters (temperature, humidity, wind speed, air pressure, transmit power, frequency, antenna height, and distance) of each sampling point to a 256-dimensional Transformer hidden feature space. The specific structure is: 8-dimensional input → 128-dimensional intermediate layer (LayerNorm + GELU activation + 0.1 Dropout) → 256-dimensional output layer (LayerNorm + Dropout). This two-layer design enhances feature representation capabilities, layer normalization stabilizes feature distribution, and Dropout prevents overfitting.

[0033] Sequence Enhancement Preprocessing Layer: This layer uses one-dimensional convolutions (kernel size 3, padding 1) to enhance local features of the input sequence, preserving original information through residual connections. This layer captures local correlations between adjacent sampling points, providing richer local context for subsequent global attention mechanisms.

[0034] Distance embedding layer: Learns a 64-dimensional distance embedding vector for each position in the sequence, mapping position indices to distance features through a trainable embedding table. After concatenating the distance embeddings with the input features, spatial location information is fused through a linear projection layer (320-dimensional → 256-dimensional), enabling the network to distinguish environmental parameters at different distances.

[0035] Positional encoding layer: Sine-cosine positional encoding is used to add absolute positional information to the sequence. The encoding dimension is 256, the maximum sequence length is 200, and it is fused with the input features through addition, enabling the network to perceive the positional relationships in the sequence.

[0036] Multi-layer Transformer encoder: Composed of 6 stacked Transformer encoder layers, each layer includes: Multi-head self-attention mechanism: 8 attention heads, each with a dimension of 32 (256 / 8), using scaled dot product attention. For each sample point in the sequence, the relevance weights with all other points are calculated to achieve global information interaction. The multi-head design enables the network to simultaneously focus on different types of spatial correlations (such as local changes at close range and large-scale patterns at long range).

[0037] Feedforward Neural Network: It adopts a structure of first increasing the dimensionality (256→1024) and then decreasing the dimensionality (1024→256), and uses the GELU activation function to enhance the nonlinear expressive power of the network.

[0038] Residual connections and layer normalization: Each sublayer uses residual connections and layer normalization to ensure the training stability of deep networks.

[0039] Feature aggregation layer: A dual aggregation strategy is employed, combining global average pooling and attention pooling. Global average pooling extracts overall features, while attention pooling highlights important positions using learned weights. The two types of features are fused into the final global feature vector through a linear layer (512-dimensional → 256-dimensional).

[0040] The channel model decoding module restores the 256-dimensional global environment features extracted by the encoder into a 101×101 distance-height two-dimensional path loss distribution. The channel model decoding module includes: a multi-layer feature projection network, a multi-scale sequence feature mapper, cross-scale feature fusion, a residual refinement module, and a final output mapping layer.

[0041] Multilayer Feature Projection Network: A three-layer fully connected network is used to progressively expand the one-dimensional global features: 256-dimensional → 1024-dimensional (LayerNorm + GELU + Dropout) → 512-dimensional (LayerNorm + GELU) → 4096-dimensional (256×4×4). Finally, it is reshaped into a 256-channel 4×4 feature map, serving as the initial "seed" for generating the two-dimensional field.

[0042] Multi-scale Sequence Feature Mapper: Two feature mappers of different scales are designed to map 101-dimensional sequence features into 16×16 and 32×32 two-dimensional feature maps. Each mapper includes: Sequence-to-space rearrangement: The 101 sequence points are expanded to 121 points through linear interpolation, reshaping them into an 11×11 spatial layout; Convolutional upsampling: The 11×11 feature map is upsampled to the target size using transposed convolutions; Adaptive pooling: Precisely adjusted to 16×16 or 32×32 size; Stepwise upsampling network: A 5-stage transposed convolutional upsampling network is used: Stage 1: 4×4→8×8 (256→128 channels, convolution...) The first stage uses a 4×4 kernel with a step size of 2. The second stage uses 8×8 to 16×16 (128 to 96 channels) + sequence feature fusion (96 + 32 to 64 channels). The third stage uses 16×16 to 32×32 (64 to 48 channels) + sequence feature fusion (48 + 32 to 32 channels). The fourth stage uses 32×32 to 64×64 (32 to 24 channels). Each stage uses grouping normalization, GELU activation, and Dropout2d regularization.

[0043] Cross-scale feature fusion: At the 16×16 and 32×32 scales, upsampled features are concatenated and fused with the corresponding sequence mapping features. This cross-scale fusion mechanism ensures that the spatial information of the sequence is preserved during decoding, avoiding information loss.

[0044] Residual Refinement Module: Feature refinement is performed using three residual blocks at a 64×64 resolution. Each residual block contains two 3×3 convolutional layers, grouped normalization, GELU activation, and Dropout. The residual structure enables the network to finely adjust local details (such as waveguide layer boundaries, interference fringes, etc.) while preserving large-scale propagation trends.

[0045] The final output mapping layer uses a three-layer convolutional network to map the multi-channel feature maps to single-channel path loss values: 24 channels → 16 channels → 8 channels → 1 channel. Finally, bilinear interpolation is used to accurately upsample to a 101×101 resolution.

[0046] The evaporation waveguide prediction results reflect the two-dimensional path loss distribution of 101×101 distance-height.

[0047] S4: Using the relationship between ray trajectory and ray angle, the evaporation waveguide prediction model is trained to obtain a trained evaporation waveguide prediction model, thus completing the modeling of the non-uniform evaporation atmospheric waveguide channel; the trained evaporation waveguide prediction model is used to analyze meteorological parameters and obtain evaporation waveguide prediction results.

[0048] In some embodiments, the processor may employ an adaptive learning rate adjustment strategy, setting the batch size to 16, the training epoch to 100, and using the gradient descent-based Adam optimization algorithm to train the model. Addressing the electromagnetic propagation characteristics of non-uniform waveguides, particularly the long signal propagation distance and significant intensity variations within the waveguide layer, this invention specifically improves the traditional mean square error loss function by introducing a weighted mean square error loss function to strengthen the physical constraints on key waveguide regions during training.

[0049] In some embodiments, the expression for the loss function used in training the evaporating waveguide prediction model is: ; in, Represents the loss function. Indicates the number of sampling points. This indicates the weight of the corresponding sampling point. Represents the predictive channel model for the sampling points. Let i represent the actual channel model of the sampling point, where i represents the i-th sampling point.

[0050] In some embodiments, the processor can set the weight distribution according to the height range of the evaporating waveguide: the weight of sampling points within the waveguide layer height is set to 5, and the weight of sampling points outside the waveguide layer is set to 1. This weight configuration mechanism can effectively enhance the model's ability to learn the propagation characteristics of the waveguide core region during training, highlighting the influence of key physical regions while maintaining the numerical stability of the overall training process.

[0051] In some embodiments, the processor can use a segmented ray tracing method to accurately model the propagation behavior of electromagnetic waves throughout the space and combine it with a path loss model to calculate the time delay and loss values ​​of each propagation path. Finally, these values ​​are synthesized into a complete channel impulse response dataset, which serves as a true label for model learning.

[0052] In some embodiments, the processor can take the weather parameters and transmitter parameters from step 1 as input, including atmospheric temperature, sea surface temperature, atmospheric pressure, relative humidity, wind speed, transmitter and receiver altitudes, and signal frequency, respectively. It then uses a deep learning network to predict spatial losses and receiver location delays, and calculates a weighted loss function by comparing the predicted results with the actual values. The loss function is as follows: .

[0053] like Figure 5 The results and analysis of the test set samples shown demonstrate that the proposed method can accurately predict path loss and receiver delay in non-uniform evaporating waveguide environments, verifying that the invention has high prediction accuracy and reliability in complex waveguide channel modeling.

[0054] In some embodiments of this specification, a non-uniform evaporating atmospheric waveguide channel modeling method is provided. Based on the spatial refractive index distribution, the equation is used to calculate the relationship between the ray trajectory and the ray angle. The evaporating waveguide prediction model is trained using the relationship between the ray trajectory and the ray angle to obtain the trained evaporating waveguide prediction model, thus completing the modeling of the non-uniform evaporating atmospheric waveguide channel. (1) Spatial environment features are extracted by a global attention mechanism based on Transformer, and the features are restored to a high-resolution two-dimensional field distribution by using a transposed convolutional network based on a CNN decoder; (2) Weather distributions related to the preceding and following phases are generated by Markov chains, and the propagation of electromagnetic waves in horizontal non-uniform conditions is simulated by segmented ray tracing; (3) The end-to-end evaporating waveguide channel modeling method directly takes weather parameters and transceiver parameters as inputs and outputs the final channel model. There is no need to perform "ray tracing-loss calculation" step by step. The output channel model is complete, the model is more practical, the modeling ability of long-distance spatial dependence is better, the characterization of horizontal non-uniformity is more realistic and the error is smaller, the modeling accuracy is higher, the channel model is more complete, the physical credibility is higher, the applicability is wider, and the computational efficiency is significantly improved.

Claims

1. A method for modeling a non-uniform evaporating atmospheric waveguide channel, characterized in that, include: S1: Based on meteorological parameters, generate a continuous distribution of related weather parameters before and after, and obtain the spatial refractive index distribution through calculation; S2: Based on the spatial refractive index distribution, the relationship between the ray trajectory and the ray angle is obtained by using the equation. S3: Construct an evaporating waveguide prediction model using the sequence feature encoding module and the channel model decoding module; S4: Using the relationship between ray trajectory and ray angle, the evaporation waveguide prediction model is trained to obtain a trained evaporation waveguide prediction model, thus completing the modeling of the non-uniform evaporation atmospheric waveguide channel; the trained evaporation waveguide prediction model is used to analyze meteorological parameters and obtain evaporation waveguide prediction results.

2. The non-uniform evaporating atmospheric waveguide channel modeling method according to claim 1, characterized in that, The expression for the spatial refractive index distribution is: ； in, Represents the spatial refractive index distribution. Indicates temperature, Indicates atmospheric pressure. and Represents a constant. This indicates water vapor pressure.

3. The non-uniform evaporating atmospheric waveguide channel modeling method according to claim 1, characterized in that, The expression for the relationship between the ray trajectory and the ray angle is: ； ； ； ； in, Indicates the refractive index distribution. Indicates the horizontal index of the sampling point. Indicates the sampling point height index. This indicates the total number of height sampling points. Represents the Earth's radius. Indicates the height of the ray trajectory. Indicates the angle of the ray. Indicates changes in altitude. Indicates a change in angle. This represents the horizontal coordinate of the ray.

4. The non-uniform evaporating atmospheric waveguide channel modeling method according to claim 1, characterized in that, The evaporation waveguide prediction model includes: The sequence feature encoding module is used to extract key features affecting electromagnetic wave propagation from a horizontally non-uniform environmental parameter sequence to obtain a global feature vector. A channel model decoding module is constructed to reconstruct the global feature vector and obtain the evaporation waveguide prediction results.

5. The non-uniform evaporating atmospheric waveguide channel modeling method according to claim 4, characterized in that, The sequence feature encoding module includes a multi-layer Transformer encoder architecture, which extracts key features affecting electromagnetic wave propagation from a horizontally non-uniform environmental parameter sequence and establishes correlations between different locations. The sequence feature encoding module includes: an input feature mapping layer, a sequence enhancement preprocessing layer, a distance embedding layer, a position encoding layer, a multi-layer Transformer encoder, a multi-head self-attention mechanism, a feedforward neural network, residual connections and layer normalization, and a feature aggregation layer. Input Feature Mapping Layer: A two-layer fully connected network is used to map the 8-dimensional environmental parameters (temperature, humidity, wind speed, air pressure, transmit power, frequency, antenna height, distance) of each sampling point to a 256-dimensional Transformer hidden feature space. The specific structure is: 8-dimensional input to a 128-dimensional intermediate layer to a 256-dimensional output layer; The sequence enhancement preprocessing layer is used to enhance the local features of the input sequence using one-dimensional convolution, while preserving the original information through residual connections. The distance embedding layer is used to learn a 64-dimensional distance embedding vector for each position in the sequence, mapping position indices to distance features through a trainable embedding table; The positional coding layer is used to add absolute positional information to the sequence using sine-cosine positional coding. The coding dimension is 256, and the maximum sequence length is 200. The multi-layer Transformer encoder is constructed by stacking 6 Transformer encoder layers. Each layer contains: a multi-head self-attention mechanism: 8 attention heads, each with a dimension of 32, using scaled dot product attention. For each sample point in the sequence, the correlation weight with all other points is calculated to achieve global information interaction. The multi-head design enables the network to pay attention to different types of spatial correlations at the same time. Feedforward neural networks employ a structure that first increases dimensionality and then decreases it, using the GELU activation function to enhance the network's nonlinear expressive power. Residual connections and layer normalization are used in each sublayer; The feature aggregation layer employs a dual aggregation strategy, combining global average pooling and attention pooling. Global average pooling extracts overall features, while attention pooling highlights important positions using learned weights. The two types of features are then fused into the final global feature vector through a linear layer.

6. The non-uniform evaporating atmospheric waveguide channel modeling method according to claim 4, characterized in that, The channel model decoding module includes: a multi-layer feature projection network, a multi-scale sequence feature mapper, a cross-scale feature fusion, a residual refinement module, and a final output mapping layer; The multi-layer feature projection network uses a three-layer fully connected network to progressively expand the one-dimensional global features: from 256 dimensions to 1024 dimensions to 512 dimensions to 4096 dimensions, and finally reshapes them into a 4×4 feature map with 256 channels, which serves as the initial seed for the generation of the two-dimensional field. A multi-scale sequence feature mapper is designed with two different scales to map 101-dimensional sequence features into 16×16 and 32×32 two-dimensional feature maps. Each mapper includes: sequence-to-space rearrangement: expanding 101 sequence points to 121 points through linear interpolation, reshaping into an 11×11 spatial layout; convolutional upsampling: upsampling the 11×11 feature map to the target size using transposed convolutions; adaptive pooling: precisely adjusting to the 16×16 or 32×32 size; and a progressive upsampling network: employing 5 levels of transposed convolutional upsampling: Level 1: 4×4 to 8×8, Level 2: 8×8 to 16×16 and fused with sequence features, Level 3: 16×16 to 32×32 and fused with sequence features, and Level 4: 32×32 to 64×64. Each level uses grouping normalization, GELU activation, and Dropout2d regularization. Cross-scale feature fusion involves concatenating and fusing upsampled features with corresponding sequence mapping features at 16×16 and 32×32 scales. This cross-scale fusion mechanism ensures that spatial information of the sequence is preserved during decoding, avoiding information loss. The residual refinement module uses three residual blocks to refine features at a 64×64 resolution. Each residual block contains two 3×3 convolutional layers, grouping normalization, GELU activation, and Dropout. The residual structure enables the network to finely adjust local details while preserving the large-scale propagation trend. The output mapping layer uses a three-layer convolutional network to map the multi-channel feature map to single-channel path loss values: from 24 channels to 16 channels to 8 channels to 1 channel. Finally, it uses bilinear interpolation to accurately upsample to a resolution of 101×101.

7. The method for modeling non-uniform evaporating atmospheric waveguide channels according to claim 1, characterized in that, The expression for the loss function used in training the evaporating waveguide prediction model is as follows: ； in, Represents the loss function. Indicates the number of sampling points. This indicates the weight of the corresponding sampling point. Represents the predictive channel model for the sampling points. Let i represent the actual channel model of the sampling point, where i represents the i-th sampling point.

Images