A path loss prediction method suitable for spatial non-stationarity

Abstract

This invention discloses a path loss prediction method applicable to spatial nonstationarity. Addressing the challenge that traditional empirical models struggle to characterize the near-field spherical wave effect, spatial nonstationarity, and multi-dimensional physical feature coupling in ultra-large-scale MIMO, this invention designs a CNN-Transformer hybrid network. First, XL-MIMO channel data is acquired and samples with ineffective propagation links are filtered out. A six-dimensional feature vector is constructed from spatial scale, orientation, and propagation state parameters, labeled with path loss. Local feature mapping is performed through one-dimensional convolution, maintaining the sequence length. High-order local features are combined with learnable position embeddings and fed into a multi-layer encoder for cross-dimensional dependency modeling. Subsequently, global pooling is performed to construct a multi-layer perceptron, which is trained using the Adam optimizer, preserving a low-error regression prediction model. This invention significantly reduces path loss prediction errors in complex spatial nonstationar scenarios and is suitable for coverage prediction and refined modeling.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication channel modeling and machine learning application technology, and in particular relates to a path loss prediction method suitable for spatial non-stationarity. Background Technology

[0002] In coverage prediction, link budgeting, and network planning of wireless communication systems, path loss is a crucial parameter describing large-scale signal fading. As communication systems evolve towards Ultra-Large-Scale Multiple-Input Multiple-Output (XL-MIMO), antenna array apertures increase dramatically, leading to channel propagation exhibiting near-field spherical wave effects and significant spatial non-stationary characteristics, distinct from traditional MIMO systems. In this complex spatial non-stationary propagation environment, path loss is no longer solely determined by propagation distance but is dynamically influenced by the coupling of multiple physical factors, including spatial location, propagation state, directional relationships, and multipath structure.

[0003] Existing empirical statistical path loss models typically use propagation distance as the main input variable, and their model structure is fixed with low computational complexity. However, under the non-stationary conditions of XL-MIMO space, these models have extremely limited ability to express the complex nonlinear relationship between path loss and multidimensional physical features. On the other hand, while traditional shallow data-driven models (such as random forests) can introduce some nonlinear fitting ability, they still have significant shortcomings in characterizing the deep dependencies between different input features. Furthermore, when conventional deep learning models face high costs for real-world data acquisition and limited sample size, directly relying on small samples for training often leads to poor model convergence stability and weak generalization ability.

[0004] Therefore, there is an urgent need for a path loss prediction method that can fully combine prior knowledge of the physical environment with the powerful deep representation capabilities of networks. This method should be uniquely designed to address the spatial non-stationary characteristics of XL-MIMO scenarios, enabling the model to accurately learn local feature combination mechanisms and deep dependencies across feature dimensions by utilizing multi-dimensional physical feature parameters as input. This would effectively reduce path loss prediction errors in complex spatial non-stationary propagation scenarios, providing precise support for intelligent planning of future mobile communication networks. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems that existing empirical path loss models are unable to characterize the multidimensional physical feature coupling relationship in complex XL-MIMO propagation scenarios and that shallow data-driven models have limited ability to express deep dependencies. This invention proposes a path loss prediction method suitable for spatial non-stationarity, which can be used to achieve regression prediction of path loss in complex propagation environments.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a path loss prediction method suitable for spatial non-stationarity, comprising the following steps:

[0007] S101: Data Acquisition and Preprocessing. Acquire raw XL-MIMO channel data generated based on ray tracing, and filter out samples with no effective propagation links by setting a path loss threshold to obtain a set of effective samples;

[0008] S102: Feature Extraction and Label Construction. The spatial scale, spatial orientation, and propagation state parameters from the effective sample set are read to form a six-dimensional physical feature vector, which serves as the input to the path loss regression prediction model. The corresponding true path loss values ​​from the effective sample set are used as supervision labels.

[0009] S103: Local Feature Mapping. The six-dimensional physical feature vector is input into a local feature mapping module consisting of two cascaded one-dimensional convolutional neural networks. The input features are converted into high-dimensional local features through one-dimensional convolution operations, and batch normalization is performed after each convolutional layer. At the same time, no pooling layer is set to maintain the length of the feature sequence.

[0010] S104: Global Encoding and Cross-Dimensional Modeling. The high-dimensional local features output by the local feature mapping module are added to the learnable location embedding matrix and fed into a multi-layer Transformer encoder to model cross-feature dimension dependencies through a multi-head self-attention mechanism;

[0011] S105: Global Pooling and Regression Output. Global average pooling is performed on the feature sequence output by the Transformer encoder to obtain a fixed-length global feature vector. This global feature vector is then input into the multilayer perceptron, which outputs a path loss prediction scalar value.

[0012] S106: Network Training and Parameter Optimization. Using the root mean square error as the loss function, the Adam optimizer and error backpropagation mechanism are used to batch train the overall prediction network composed of the local feature mapping module, Transformer encoder, and multilayer perception mechanism, and the model parameters with lower errors are saved through an early stopping control mechanism.

[0013] Specifically, in step S101, the path loss threshold is 250 dB, and samples with a path loss label greater than or equal to 250 dB are determined as samples without a valid propagation link and are removed.

[0014] Furthermore, in step S102, the six-dimensional physical feature vector includes the 3D transmission and reception distance, line-of-sight propagation marker, transmitter height, receiver height, horizontal azimuth angle, and elevation angle, which are used to characterize the link spatial scale, directional relationship, and obstruction conditions.

[0015] In one embodiment of this application, in step S103, the channels of the two-layer cascaded one-dimensional convolutional neural network are configured as 1-16 and 16-64, the kernel size is 3, the stride is 1, and batch normalization and ReLU nonlinear mapping are performed after the convolutional layers.

[0016] Furthermore, in step S104, the embedding dimension of the Transformer encoder is 64, the number of attention heads is 4, the number of encoder layers is 3, the feedforward layer dimension is 256, and the Dropout is 0.1.

[0017] Preferably, in step S105, the multilayer perceptron has a structure of 64-128-64-1, which is used to map the global feature vector after global average pooling into a path loss prediction scalar value.

[0018] The present invention has the following beneficial effects:

[0019] (1) The present invention uses the 3D transmission and reception distance, line-of-sight propagation marker, transmitter height, receiver height, horizontal azimuth angle and elevation angle as input features, so that the path loss prediction model can utilize spatial scale, directional relationship and occlusion status information with clear physical meaning, and reduce the modeling bias caused by relying solely on propagation distance for prediction.

[0020] (2) This invention uses a two-layer one-dimensional convolutional neural network to perform local feature mapping on six-dimensional physical features, and maintains the feature sequence length without setting a pooling layer. Then, it learns cross-feature dimension dependencies through learnable position embedding and Transformer encoder, so that adjacent feature combination expression and long-range feature dependency can be utilized in the same prediction network.

[0021] (3) Under the same test set settings, the root mean square error of the CNN-Transformer model is 4.546 dB, the mean absolute error is 3.338 dB, and the coefficient of determination is... The root mean square error (RMSE) was 0.873; compared to the CI model, the RMSE decreased from 9.77 dB to 4.546 dB; compared to the random forest model, the RMSE decreased from 5.453 dB to 4.546 dB. When the prediction error was less than 5 dB, the cumulative probability corresponding to the CNN-Transformer model was 78%; when the prediction error was less than 10 dB, the cumulative probability corresponding to the CNN-Transformer model was approximately 96%. These results demonstrate that this method can reduce path loss prediction error and improve the overall fit. Attached Figure Description

[0022] Figure 1This is a step-by-step illustration of a path loss prediction method applicable to spatial nonstationarity proposed in this invention;

[0023] Figure 2 This is a diagram of the overall network architecture of CNN-Transformer in the method of this invention;

[0024] Figure 3 This is an example diagram illustrating the model training convergence process in the method of this invention;

[0025] Figure 4 This is a diagram showing the prediction results of the CNN-Transformer model in the method of this invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0027] To verify the effectiveness and superiority of the method described in this invention in handling spatial non-stationary characteristics, this embodiment takes a new intermediate frequency (e.g., 15 GHz) XL-MIMO outdoor urban street scene as an example for specific verification.

[0028] like Figure 1 As shown, the present invention provides a path loss prediction method applicable to spatial nonstationarity. First, an effective path loss feature set is constructed based on ray tracing data, and then a six-dimensional physical feature input is formed. Path loss regression prediction is achieved through a CNN-Transformer hybrid network.

[0029] In step S101, the raw channel data is generated based on the DeepMIMO dataset. This dataset, based on WirelessInsite software, describes the direct, reflected, diffracted, and scattered propagation processes in an urban environment using ray tracing. The example uses an outdoor urban street scene with a center frequency of 15 GHz, a system bandwidth of 250 MHz, 3 base stations, a VMIMO array with 128 antenna elements at the base station, a single-antenna reception method at the user end, an antenna spacing parameter of 0.5, 1436 effective samples, and a maximum path count of 5. The simulation output includes parameters such as path loss, propagation distance, multipath power, multipath delay, path count, and line-of-sight markers.

[0030] Since the simulation data may contain samples with no effective propagation links or abnormal values, during the data preprocessing stage, samples with a path loss greater than or equal to 250 dB are identified as samples with no effective propagation links and are removed. The cleaned set of effective samples is denoted as:

[0031]

[0032] in, Indicates the first The input feature vector corresponding to each sample; Indicates the first The true value of the path loss label corresponding to each sample; This represents the total number of valid samples after cleaning.

[0033] In step S102, a six-dimensional input feature for path loss prediction is extracted from the cleaned valid samples. The six-dimensional input feature consists of the transmit / receive 3D distance. Line-of-sight propagation markers Launch terminal height Receiver height Horizontal azimuth and pitch angle Composition. Among them, The three-dimensional Euclidean distance between the transmitter and receiver is expressed in meters. The line-of-sight (LoS) propagation marker is set, while the non-line-of-sight (NLoS) condition is set to 0, and the LoS condition is set to 1. The height of the transmitter is in meters. The height of the receiving end is in meters (m). This is the horizontal azimuth angle, in rad. The pitch angle is expressed in rad. Inputs and outputs are written to a CSV file in a one-to-one correspondence according to the sample dimensions for subsequent model training and testing.

[0034] like Figure 2 As shown, in steps S103 to S105, the path loss prediction model adopts a serially fused CNN-Transformer structure. The input six-dimensional feature vector first enters the local feature mapping module, then enters the Transformer global modeling module, and finally outputs the path loss prediction result through global average pooling and multilayer perceptron.

[0035] In step S103, the local feature mapping module consists of two cascaded one-dimensional convolutional neural networks. The first one-dimensional convolutional layer has 1 input channel, 16 output channels, a kernel size of 3, and a stride of 1; the second one-dimensional convolutional layer has 16 input channels, 64 output channels, a kernel size of 3, and a stride of 1. This module is used to map the input physical features into a 64-dimensional local feature representation.

[0036] To accelerate convergence and reduce the impact of internal covariate shifts, a batch normalization layer is introduced after each convolutional layer. For layers containing... The calculation process for the current batch of each sample is as follows:

[0037]

[0038]

[0039]

[0040]

[0041] in, This represents the number of samples in the current batch. For the first in the current batch The feature matrix of each sample after one-dimensional convolution operation; and These are the mean and variance of the features in the current batch, respectively. To prevent the use of a pre-defined small constant with a denominator of zero; The standardized feature matrix; and These are the learnable scaling parameters and the learnable translation parameters, respectively. This is the output feature matrix after batch normalization. After batch normalization, a nonlinear mapping is introduced using the Rectified Linear Unit (ReLU) function.

[0042] No pooling layer is set in this local feature mapping module, so that the length of the feature sequence after passing through the one-dimensional convolutional neural network is reduced. In this embodiment, the length of the feature sequence remains unchanged. Equal to the input feature dimension, i.e. This is so that the Transformer encoder can use position encoding to preserve the positional information in the feature sequence.

[0043] In step S104, the feature tensor output by the local feature mapping module is added to the learnable location embedding matrix to obtain the combined feature tensor input to the Transformer encoder, whose expression is:

[0044]

[0045] in, This is the local feature tensor output by the local feature mapping module; The learnable location embedding matrix is ​​initialized with a random distribution and adaptively updated via backpropagation during training. This is the combined feature tensor input to the Transformer encoder.

[0046] The Transformer encoder has an embedding dimension of 64, 4 attention heads, 3 encoder layers, a feedforward layer dimension of 256, and a Dropout value of 0.1. With four attention heads, the model can model the relationships between 3D distance, height, azimuth, pitch, and line-of-sight states in parallel across different subspaces. The multi-head self-attention layer and feedforward neural network layer in the Transformer encoder, combined with layer normalization and residual connections, are used to maintain gradient propagation stability during deep network training.

[0047] In step S105, the feature sequence output by the Transformer encoder is input into the global average pooling layer, and averaged over the sequence length dimension to obtain a fixed-length global feature vector, the expression of which is:

[0048]

[0049] in, This indicates the position of the feature sequence output by the Transformer encoder in the sequence. The state vector at that location; The length of the feature sequence; This is the global feature vector. Then, The input is a multilayer perceptron with a structure of 64-128-64-1, and a Dropout mechanism is introduced in the fully connected layer. Finally, the output is a scalar value of the predicted path loss.

[0050] In step S106, model training is performed in batch mode. The sample data is divided into 80% training set and 20% test set, with a batch size of 64, a maximum training epoch count of 300, and an initial learning rate of 100 for the Adam optimizer. The number of early stopping wheels is 20.

[0051] The loss function uses the root mean square error, and its expression is:

[0052]

[0053] in, This refers to the number of samples in a single batch. For the first batch The true value of path loss for each sample; For the first batch The path loss prediction value for each sample. During training, the network parameters are updated through the error backpropagation mechanism; if the error in the training process does not improve for 20 consecutive rounds, the training is terminated early and the model parameters with lower error are saved.

[0054] like Figure 3As shown, the training error decreases rapidly in the initial stage and then gradually stabilizes, indicating that the CNN-Transformer model can converge on the sample data.

[0055] like Figure 4 As shown, on the test set, the predicted scatter points of the CNN-Transformer model are generally distributed near the ideal fitted line. To evaluate the path loss prediction results, mean absolute error, root mean square error, and coefficient of determination are used as evaluation metrics, and their calculation formulas are as follows:

[0056]

[0057]

[0058]

[0059] in, The total number of test samples; For the first The true value of path loss for each sample; For the first Predicted path loss values ​​for each sample; This is the arithmetic mean of the true path loss values ​​for the test samples.

[0060] According to the experimental results of the embodiments of the present invention, the root mean square error of the CNN-Transformer model is 4.546 dB, the mean absolute error is 3.338 dB, and the coefficient of determination is... The coefficient of determination is 0.873; the root mean square error of the CI model is 9.77 dB, the mean absolute error is 8.69 dB, and the coefficient of determination is... The root mean square error (RMSE) of the random forest model is 0.413, the mean absolute error (MAO) is 5.453 dB, and the coefficient of determination is 3.954 dB. The value is 0.817. These results demonstrate that the CNN-Transformer model can reduce path loss prediction error under this test setting.

[0061] Furthermore, the ablation experiments in the embodiments of the present invention show that, when only a CNN structure is used, the root mean square error is 4.906 dB, the mean absolute error is 3.569 dB, and the coefficient of determination is... The root mean square error is 0.852; when using only the Transformer structure, the root mean square error is 6.005 dB, the mean absolute error is 4.294 dB, and the coefficient of determination is... The result was 0.778; when using the CNN-Transformer hybrid structure, the three metrics were 4.546 dB, 3.338 dB, and 0.873, respectively. This indicates that one-dimensional convolutional local feature extraction and Transformer cross-feature dimension dependency modeling work together in the path loss prediction process.

[0062] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A path loss prediction method applicable to spatial nonstationarity, characterized in that, Includes the following steps: S101: Data Acquisition and Preprocessing. Acquire raw XL-MIMO channel data generated based on ray tracing, and filter out samples with no effective propagation links by setting a path loss threshold to obtain a set of effective samples; S102: Feature Extraction and Label Construction. The spatial scale, spatial orientation, and propagation state parameters from the effective sample set are read to form a six-dimensional physical feature vector, which serves as the input to the path loss regression prediction model. The corresponding true path loss values ​​from the effective sample set are used as supervision labels. S103: Local Feature Mapping. The six-dimensional physical feature vector is input into a local feature mapping module consisting of two cascaded one-dimensional convolutional neural networks. The input features are converted into high-dimensional local features through one-dimensional convolution operations. Batch normalization is performed after each convolutional layer, and no pooling layer is set to keep the feature sequence length unchanged. S104: Global Encoding and Cross-Dimensional Modeling. The high-dimensional local features output by the local feature mapping module are added to the learnable location embedding matrix and fed into a multi-layer Transformer encoder to model cross-feature dimension dependencies through a multi-head self-attention mechanism; S105: Global Pooling and Regression Output. Global average pooling is performed on the feature sequence output by the Transformer encoder to obtain a fixed-length global feature vector. This global feature vector is then input into the multilayer perceptron, which outputs a path loss prediction scalar value. S106: Network Training and Parameter Optimization. Using the root mean square error as the loss function, the Adam optimizer and error backpropagation mechanism are used to batch train the overall prediction network composed of the local feature mapping module, Transformer encoder, and multilayer perception mechanism, and the model parameters with lower errors are saved through an early stopping control mechanism.

2. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S101, the path loss threshold is 250 dB. Samples with a path loss label greater than or equal to 250 dB are determined to be samples without a valid propagation link and are removed. The cleaned set of valid samples is represented as follows: in, Indicates the first The input feature vector corresponding to each sample; Indicates the first The true value of the path loss label corresponding to each sample; This represents the total number of valid samples after cleaning.

3. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S102, the six-dimensional feature vector includes the 3D distance between the transmitter and receiver. Line-of-sight propagation markers Launch terminal height Receiver height Horizontal azimuth and pitch angle ;in, The three-dimensional Euclidean distance between the transmitter and receiver is expressed in meters. This is a line-of-sight propagation marker; the value is 0 for non-line-of-sight conditions and 1 for line-of-sight conditions. The transmitter height is in meters (m). The height of the receiver is in meters (m). This is the horizontal azimuth angle, in rad. The pitch angle is expressed in rad.

4. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S103, the first one-dimensional convolutional layer of the two-layer cascaded one-dimensional convolutional neural network has 1 input channel, 16 output channels, a kernel size of 3, and a stride of 1; the second one-dimensional convolutional layer has 16 input channels, 64 output channels, a kernel size of 3, and a stride of 1; and a nonlinear mapping is performed after the convolutional layers using the linear rectified function ReLU.

5. The path loss prediction method applicable to spatial nonstationarity according to claim 4, characterized in that: The batch normalization process performs the following calculations on the one-dimensional convolutional output features within the current batch: in, This represents the number of samples in the current batch. For the first in the current batch The feature matrix of each sample after one-dimensional convolution operation; and These are the mean and variance of the features in the current batch, respectively. To prevent the use of a pre-defined small constant with a denominator of zero; The standardized feature matrix; and These are the learnable scaling parameters and the learnable translation parameters, respectively. This is the output feature matrix after batch normalization.

6. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S104, the learnable location embedding matrix is ​​initialized to a random distribution and updated through backpropagation during network training; the expression for adding the output features of the local feature mapping module to the learnable location embedding matrix is: in, This is the local feature tensor output by the local feature mapping module; For learnable position embedding matrix; The input is the combined feature tensor of the Transformer encoder; the Transformer encoder has an embedding dimension of 64, an attention head of 4, an encoder layer of 3, a feedforward layer dimension of 256, and a Dropout of 0.

1.

7. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S105, the global average pooling layer compresses the feature sequence output by the Transformer encoder into a fixed-length global feature vector, the expression of which is: in, This indicates the position of the feature sequence output by the Transformer encoder in the sequence. The state vector at that location; The length of the feature sequence; The global feature vector is represented by the multilayer perceptron, which has a structure of 64-128-64-1.

8. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: In step S106, the formula for calculating the root mean square error is: in, The number of samples used in error calculation; For the first The true value of path loss for each sample; For the first The predicted path loss values ​​for each sample; the batch size during training is 64, and the initial learning rate of the Adam optimizer is... The maximum number of training rounds is 300, the number of early stop rounds is 20, the training set ratio is 80%, and the test set ratio is 20%.

9. The path loss prediction method applicable to spatial nonstationarity according to claim 1, characterized in that: The original XL-MIMO channel data was generated based on the DeepMIMO ray tracing dataset. The simulation scenario was an outdoor urban street scene with a center frequency of 15 GHz, a system bandwidth of 250 MHz, 3 base stations, a base station array size of 128 antenna elements, a single antenna receiving method at the user end, an antenna spacing parameter of 0.5, 1436 effective samples, and a maximum number of paths of 5. The simulation output included path loss, propagation distance, multipath power, multipath delay, number of paths, and line-of-sight markers.

Images