Path loss prediction method and device based on multi-dimensional wireless environment characteristics

Abstract

The embodiment of the invention provides a path loss prediction method and device based on multi-dimensional wireless environment characteristics. The method comprises the following steps: determining scene data of a plurality of different scenes; obtaining path loss data corresponding to different scene data based on the scene data of different scenes; constructing a multi-dimensional wireless environment feature; performing data alignment on the path loss data and the multi-dimensional wireless environment characteristics, and constructing a target data set of a large-scale diversified scene; constructing a preset neural network architecture of an encoder-decoder based on jump connection, constructing a training set by using the target data set, training the preset neural network architecture by using the training set, and constructing a prediction model; the preset neural network architecture comprises a front feature processing module, a rear feature processing module, a codec module and a channel attention module. According to the scheme of the invention, a process for performing path loss prediction and constructing the prediction model based on the multi-dimensional wireless environment characteristics is provided, and the prediction precision and generalization ability of the model are improved.

Description

Technical Field

[0001] The present application relates to the technical field of wireless channel prediction, and in particular to a path loss prediction method and device based on multi-dimensional wireless environment characteristics. Background Art

[0002] In the sixth-generation (6G) mobile communication technology era, diverse scenarios such as high-speed mobility, dense urban environments, and complex industrial IoT scenarios are constantly emerging, placing unprecedentedly stringent demands on the performance, flexibility, and adaptability of communication systems. Traditional offline statistical paradigms are gradually exposing their limitations in the face of various dynamically changing environments, making it difficult to meet the real-time, high-precision communication requirements of 6G networks. To address these challenges, digital twin channel (DTC) technology has emerged. By constructing a bidirectional mapping between the physical environment and the digital space online, DTC can accurately characterize the entire process of channel fading and changes, providing a new development direction for 6G communication systems. The key to achieving this new paradigm lies in fast, large-scale channel prediction, of which path loss prediction (PLP) is the core component.

[0003] While traditional ray tracing (RT) technology can calculate highly accurate channel fading information, its extremely high computational complexity prevents it from meeting the real-time requirements of 6G network applications. In recent years, machine learning-based channel prediction methods have provided new insights into this area. These methods incorporate environmental information and use neural networks to learn the mapping between the environment and the channel to achieve path loss prediction. However, the inevitable model training overhead limits their rapid deployment and widespread use in practical applications.

[0004] In addition, current path loss prediction methods based on machine learning face two major challenges. On the one hand, the existing wireless environment features (WEF) extraction method is difficult to fully and accurately characterize complex and changing scene information, resulting in limited improvement in prediction performance; on the other hand, the model's zero-sample generalization performance is insufficient in different scenarios, making it difficult to adapt to the needs of diverse 6G network scenarios. The propagation characteristics of wireless signals are highly dependent on the environment, and the prior input information of the model directly determines the ability of the neural network to learn channel characteristics and generalize to new scenarios. Therefore, effectively extracting various basic environmental attributes, such as distance, occlusion, and electromagnetic characteristics, and using them to more comprehensively characterize the environment, has become the key to improving PLP prediction accuracy and generalization capabilities.

[0005] In summary, in order to achieve fast and generalizable large-scale path loss prediction in 6G systems, a new large-scale path loss prediction method with scenario generalization capability is urgently needed to meet the strict requirements of communication systems under the diverse scenario requirements of 6G networks and promote the development and application of 6G technology. Summary of the Invention

[0006] At least one embodiment of the present application provides a path loss prediction method and apparatus based on multi-dimensional wireless environment characteristics, which are used to solve the problems existing in the above-mentioned existing path loss prediction methods.

[0007] In order to solve the above technical problems, this application is implemented as follows:

[0008] In a first aspect, an embodiment of the present application provides a path loss prediction method based on multi-dimensional wireless environment characteristics, including:

[0009] Determine scenario data for multiple different scenarios;

[0010] Based on the scenario data of different scenarios, path loss data corresponding to the different scenario data is obtained;

[0011] Construct multi-dimensional wireless environment characteristics;

[0012] Aligning the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios;

[0013] Construct a preset neural network architecture of an encoder-decoder based on jump connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; the preset neural network architecture includes: pre- and post-feature processing modules, an encoder-decoder module, and a channel attention module.

[0014] Optionally, after building the prediction model, the method further includes:

[0015] The accuracy and generalization of the prediction model are evaluated, and path loss prediction is performed based on the evaluated prediction model.

[0016] Optionally, the configuration information of the path loss data includes at least one of a center frequency, a transmission order, a diffraction order, a transmission power, a receiving power threshold, a number of paths retained by each receiving antenna, a ray tracing simulation propagation model, a transmitting antenna, a receiving antenna, a transmitting antenna position, and a receiving antenna position;

[0017] The path loss data and the multi-dimensional wireless environment characteristics are aligned to construct a target data set for large-scale and diverse scenarios.

[0018] Optionally, the multi-dimensional wireless environment feature includes at least one of the following:

[0019] Calculate the distance factor between the transmitting antenna and the receiving antenna based on the Euclidean distance;

[0020] Free space path loss factor;

[0021] Transmitting antenna location factors;

[0022] Electromagnetic characteristics factors;

[0023] Obstruction factors of communication links in line-of-sight or non-line-of-sight situations;

[0024] Height map factors of buildings;

[0025] The height difference between the transmitter and receiver and the building;

[0026] Building outline factors;

[0027] The three-dimensional coordinate difference factor between the transmitting antenna and the receiving antenna;

[0028] Factors that damage buildings.

[0029] Optionally, the pre-processing and post-processing feature processing modules are used for pre-processing and post-processing; the pre-processing is used to align the feature dimensions of the multi-source heterogeneous input features in the training data based on the input training data, and the features after the feature dimension alignment are input to the codec module; the post-processing is used to perform residual processing and convolution mapping processing on the feature map output by the codec module to predict the true path loss map;

[0030] The codec module is used to receive the features after the feature dimension alignment, input the features after the feature dimension alignment into the first convolution block for preliminary feature fusion to obtain fused features, the encoder in the codec module extracts deep-level environmental features based on the fused features, and the decoder in the codec module outputs the feature map based on the deep-level environmental features; the encoder includes four convolution blocks and four maximum pooling layers; the decoder includes four upsampling layers and four convolution blocks, and there is a skip connection between the feature map of the i-th layer of the encoder and the feature map of the j-th layer of the decoder, where j = the total number of layers of the codec module - 1;

[0031] The channel attention module is arranged between each of the upsampling layers and the convolution block in each of the encoders; the channel attention module is used to compress the feature map to be processed in the codec module along the spatial dimension, and use the fully connected layer to generate normalized weights, and assign the normalized weights to key features.

[0032] Optionally, evaluating the accuracy and generalizability of the prediction model includes:

[0033] Based on the prediction model, determining a predicted path loss value of the prediction model;

[0034] Get the actual path loss value;

[0035] Determining a root mean square error between the actual path loss value and the predicted path loss value;

[0036] If the root mean square error is within a first preset range, determining that the accuracy of the prediction model is passed; otherwise, performing an operation of retraining the prediction model;

[0037] If the accuracy of the prediction model is passed, determining a prediction value of the prediction model in a preset scenario, and evaluating the generalization of the prediction model based on the prediction model;

[0038] When the predicted value is within a second preset range, it is determined that the generalization of the prediction model is passed; otherwise, an operation of retraining the prediction model is performed.

[0039] In a second aspect, an embodiment of the present application provides a path loss prediction device based on multi-dimensional wireless environment characteristics, including:

[0040] A first determining module, configured to determine scene data of a plurality of different scenes;

[0041] A first acquisition module is used to acquire path loss data corresponding to different scenario data based on scenario data of different scenarios;

[0042] A first building module is used to build multi-dimensional wireless environment features;

[0043] A second construction module is configured to align the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios;

[0044] The third construction module is used to construct a preset neural network architecture of an encoder-decoder based on jump connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; the preset neural network architecture includes: pre- and post-feature processing modules, an encoder-decoder module, and a channel attention module.

[0045] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the steps of the method described in any one of the first aspects are implemented.

[0046] In a fourth aspect, an embodiment of the present application provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of the method described in any one of the first aspects.

[0047] Compared with the prior art, the path loss prediction method and device based on multi-dimensional wireless environment characteristics provided in the embodiments of the present application determine scene data of multiple different scenarios; based on the scene data of different scenarios, obtain path loss data corresponding to the different scene data; construct multi-dimensional wireless environment characteristics; align the path loss data and the multi-dimensional wireless environment characteristics to construct a target data set for large-scale diversified scenarios; construct a preset neural network architecture of an encoder-decoder based on jump connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; the preset neural network architecture includes: pre- and post-feature processing modules, a codec module, and a channel attention module. The solution of the present application constructs a cross-scenario data set by fusing geographic spatial data with electromagnetic propagation features, and designs a deep learning model including an attention mechanism. It provides a process for path loss prediction and construction of a prediction model based on multi-dimensional wireless environment characteristics, thereby improving the accuracy of path loss prediction and enhancing the prediction accuracy and generalization ability in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present application. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0049] Figure 1 A schematic diagram of a flow chart of a path loss prediction method based on multi-dimensional wireless environment characteristics provided in an embodiment of the present application;

[0050] Figure 2 A diagram showing scene data provided in an embodiment of the present application;

[0051] Figure 3 A schematic diagram of the overall process of the path loss prediction method provided in an embodiment of the present application;

[0052] Figure 4 A schematic diagram of multi-dimensional wireless environment characteristics provided by an embodiment of the present application;

[0053] Figure 5 A schematic diagram of the structure of a preset neural network architecture provided in an embodiment of the present application;

[0054] Figure 6 A schematic diagram comparing the accuracy of prediction results provided by the embodiments of the present application with those of the prior art;

[0055] Figure 7 A schematic diagram comparing the generalization of prediction results provided by the embodiments of the present application with those of the prior art;

[0056] Figure 8 This is a structural diagram of a path loss prediction device based on multi-dimensional wireless environment characteristics according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] The terms "first", "second", etc. in this application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same type, and do not limit the number of objects, for example, the first object can be one or more. In addition, "or" in this application represents at least one of the connected objects. For example, "A or B" covers three options, namely, Option 1: including A but not including B; Option 2: including B but not including A; Option 3: including both A and B. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.

[0058] The term "indication" in this application can be either a direct indication (or explicit indication) or an indirect indication (or implicit indication). A direct indication can be understood as the sender explicitly informing the receiver of specific information, the operation to be performed, or the requested result, etc. in the instruction sent; an indirect indication can be understood as the receiver determining the corresponding information based on the instruction sent by the sender, or making a judgment and determining the operation to be performed or the requested result, etc. based on the judgment result.

[0059] As described in the background, conventional ray tracing algorithms in the prior art use electromagnetic wave propagation to calculate path loss, but this method suffers from high computational overhead and is unable to meet the requirements of 6G online applications. Deep learning-based channel prediction methods have limited environmental feature extraction and lack consideration of more comprehensive, multi-scale environmental features to describe the essential properties of a scene. Furthermore, these methods are trained on specific scenarios and are difficult to generalize to unseen scenarios. Therefore, conventional deep learning algorithms currently focus on predictions for known scenarios, and no research has yet been conducted on how to directly migrate to new scenarios. These algorithms are unable to meet the requirements of 6G networks in multiple scenarios and complex environments. Specifically, while conventional ray tracing algorithms can calculate path loss using electromagnetic wave propagation, they suffer from high computational overhead and are unable to meet the requirements of 6G online applications. Deep learning-based channel prediction methods suffer from limitations in environmental feature extraction, a lack of multi-scale feature description, difficulty generalizing to new scenarios, and the inability to directly migrate. Both methods struggle to adapt to the multiple scenarios and complex environments of 6G networks. To address at least one of these issues, the present invention provides a path loss prediction method and apparatus based on multi-dimensional wireless environment features. This method can mitigate or avoid these issues, improve prediction accuracy and generalization to new scenarios, and reduce model training overhead.

[0060] Please refer to Figure 1 , an embodiment of the present application provides a path loss prediction method based on multi-dimensional wireless environment characteristics, including:

[0061] Step 11, determining scene data for a plurality of different scenes;

[0062] Step 12: Based on the scenario data of different scenarios, path loss (PL) data corresponding to the different scenario data is obtained.

[0063] In an embodiment of the present application, scene data of multiple different scenes are obtained, and the scene data can be determined based on real map data. For example, in step 11, real map data is obtained, a scene model is generated, and the scene model is used to collect full-scene channel data and environmental characteristics of different transmission (Transmit, Tx) positions, to construct a target data set for large-scale and diverse scenes, and provide a data basis for the subsequent training of the prediction model. Specifically, real map data is obtained worldwide, covering diverse geographic information, including but not limited to buildings of different heights, different numbers, and different distributions. Scene data construction is as follows: Figure 2As shown, each scene is 300m long and 200m wide, and scenario data containing 21 different scenarios are constructed, such as represented by the first data set or the second data set. The first data set includes 18 scenario data, and 18 training and testing scenarios are used to evaluate the performance of model training and testing at specific sites. The second data set includes 3 new scenarios for evaluating the generalization performance of the model.

[0064] Based on the scenario data determined above, channel data for the corresponding scenarios was collected in a variety of scenarios. 300 samples were generated for each scenario. The transmit position of each sample was different, and the receive (Receive, Rx) positions were evenly distributed in a 300m*200m grid with a grid resolution of 5m*5m. This ultimately resulted in a large-scale path loss dataset.

[0065] Optionally, the configuration information of the path loss data includes:

[0066] At least one of the following information: center frequency, transmission order, diffraction order, transmission power, receiving power threshold, number of paths retained by each receiving antenna, ray tracing simulation propagation model, transmitting antenna, receiving antenna, transmitting antenna position and receiving antenna position.

[0067] It's worth noting that any data acquisition method consistent with this method and process can be used, including but not limited to actual measurements, RT platform simulation, and other open-source datasets. Taking RT platform simulation as an example, the specific RT simulation data for the constructed dataset is shown in Table 1 below. Both the transmitter and receiver use a single antenna. Here, RT refers to ray tracing, and RT platform refers to the channel simulation platform.

[0068] Table 1: Contents of path loss data configuration information

[0069]

[0070] Step 13: construct multi-dimensional wireless environment features (MWEF).

[0071] Furthermore, extracting effective MWEFs is crucial for achieving better prediction performance and zero-shot learning capabilities. Starting from the perspective of electromagnetic propagation, we analyze environmental factors related to electromagnetic propagation characteristics in the scene and construct multidimensional wireless environment features based on these factors. Based on these multidimensional wireless environment features, we unify and batch-process different samples from diverse scenarios, calculate and save the corresponding MWEFs, and form a structured MWEF dataset.

[0072] Step 14: align the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios.

[0073] In this embodiment, multidimensional wireless environment features are a set of multidimensional features that are extracted from a comprehensive set of essential scene attributes to describe the impact of the environment on channel characteristics during wireless signal propagation. The target dataset is a multidimensional wireless environment feature-channel dataset constructed from path loss data and multidimensional wireless environment features for a large variety of scenarios.

[0074] The characteristic dimensions of the multi-dimensional wireless environment characteristics (MWEF) here include: spatial geometric characteristics, such as building height distribution and scatterer density; electromagnetic propagation characteristics, such as free space loss, multipath reflection intensity, and obstacle shielding coefficient.

[0075] Specifically, the path loss data of the different scenarios generated above are aligned with the generated MWEF datasets, and finally a multi-dimensional wireless environment feature-path loss dataset is constructed. The multi-dimensional wireless environment feature-path loss dataset is the target dataset, where each sample contains a set of MWEFs and a corresponding path loss map, where MWEF includes the following types: FSPL, distance, occlusion, three-dimensional coordinate difference, penetration, Tx position, electromagnetic characteristics, height, height difference from Tx, height difference from Rx, and contour features. The distribution, number, and height of scatterers in different scenarios are different to ensure that the dataset has sufficient diversity and provide a data basis for the subsequent training of the prediction model. The overall process of predicting the path loss map is as follows: Figure 3 shown.

[0076] This application can obtain a real global map to obtain diversified geographic data, that is, building scatterers of different heights, different distributions, and different numbers. According to the determined diversified scenarios, channel data is collected in different scenarios to construct a large-scale path loss data set. From the perspective of electromagnetic propagation, the electromagnetic propagation factors in the scene are analyzed, MWEF (such as distance, obstruction, electromagnetic characteristics), etc. are extracted, and features are extracted based on MWEF to construct a multi-dimensional wireless environment feature data set. Furthermore, the path loss data obtained from different scenarios are aligned with the obtained multi-dimensional wireless environment features to construct a multi-dimensional wireless environment feature-path loss data set.

[0077] Step 15: construct a preset neural network architecture of an encoder-decoder based on jump connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; the preset neural network architecture includes: pre- and post-feature processing modules, an encoder-decoder module, and a channel attention module.

[0078] It should be explained that the target dataset includes: a first sub-target dataset constructed based on the first dataset, i.e., based on 18 training and test scenarios; and a second sub-target dataset constructed based on the second dataset, i.e., based on three "new scenarios." The first sub-target dataset includes the training dataset described above.

[0079] In an embodiment of the present application, an encoder-decoder architecture based on skip connection is first constructed. This step involves designing a neural network with a specific structure for processing multi-scale wireless environment features. The core components of the architecture include: an encoder, which is responsible for gradually abstracting and reducing the dimensionality of the input multi-scale wireless environment features. The dimensions in the feature map are corresponding to multiple pooling layers, and then the features of the corresponding feature map are extracted using multiple convolution blocks, while increasing the number of channels. This process enables the network to capture features at different levels. The decoder is opposite to the encoder. The decoder gradually restores the spatial dimensions of the feature map through upsampling layers and convolution layers, and maps the features of different levels back to the size of the original input. The skip connection directly connects the feature map of a certain layer in the encoder to the input of the corresponding layer in the decoder. For example, the feature map of the first layer of the encoder is connected to the last layer of the decoder to help recover the detail information discarded by the encoder. This design effectively solves the problem of information loss in deep networks.

[0080] The target dataset (i.e., the first target sub-dataset) is then used to construct a training set and train the model. This step involves the complete process of data processing and model training: dataset partitioning. The collected first target sub-dataset is divided into a training set and a test set according to a preset ratio (e.g., 7:3 or 8:2). The training set is used for model learning, and the test set is used to evaluate the model's accuracy.

[0081] The prediction model of this application uses environmental information represented by MWEF and processes it based on the encoder-decoder backbone network. It introduces skip connections and channel attention mechanisms to optimize feature extraction and reconstruction. Finally, it fuses features through a residual connection structure and outputs path loss prediction results, which improves prediction accuracy and generalization ability in new scenarios and reduces model training overhead.

[0082] It should be noted that during the training process, the prediction model of this application is based on zero-sample learning for "new scenarios", which enhances the model's zero-sample learning ability. Therefore, the prediction model is used to evaluate the generalization of the prediction model's "new scenarios", so that predictions can be made directly in new scenarios without retraining, and it has the best zero-sample generalization performance in different scenarios.

[0083] This application aims to overcome the shortcomings of existing path loss prediction methods by providing a high-precision, generalizable path loss prediction method based on multi-dimensional wireless environment characteristics. This method extracts environmental characteristics such as distance, obstruction, and electromagnetic properties and combines them with a deep learning network to predict high-precision path loss maps. The method also provides verification methods and model performance.

[0084] Optionally, the electromagnetic propagation factor includes at least one of the following:

[0085] Calculate the distance factor between the transmitting antenna and the receiving antenna based on the Euclidean distance;

[0086] Free space path loss factor;

[0087] Transmitting antenna location factors;

[0088] Electromagnetic characteristics factors;

[0089] Obstruction factors of communication links in line-of-sight or non-line-of-sight situations;

[0090] Height map factors of buildings;

[0091] The height difference between the transmitting and receiving ends and the building; that is, the height difference between the transmitting antenna and the building, and the height difference between the receiving antenna and the building;

[0092] Building outline factors;

[0093] The three-dimensional coordinate difference factor between the transmitting antenna and the receiving antenna;

[0094] Factors that damage buildings.

[0095] In the embodiments of the present application, the distance between Tx and Rx is a key factor in modeling path loss (PL). Due to free space attenuation, signal strength generally decreases with increasing distance. The Euclidean distance can be expressed as:

[0096] Formula (1); where (x Rx ,y Rx ,z Rx ) and (x Tx ,y Tx ,z Tx ) are the three-dimensional coordinates of the receiver and transmitter, that is, the three-dimensional coordinates of the transmitting antenna and the receiving antenna.

[0097] In the Free Space Path Loss (FSPL) factor, FSPL quantifies the signal power versus distance in an ideal free space environment without any obstacles. The FSPL equation is defined as:

[0098] Formula (2); where d is the distance, f is the signal frequency, and c is the speed of light.

[0099] Transmitting antenna position factor, this feature is represented by the position of Tx in the map. The element at the Tx index is set to 1, and all other elements are set to 0 to emphasize the spatial position of Tx. It is defined as: Formula (3); where (x, y) represents the coordinates of the point in the map, (x Tx ,y Tx ) represents the two-dimensional coordinate of Tx.

[0100] Electromagnetic properties: This feature is used to describe the electromagnetic properties of materials, such as walls and windows, which directly affect the reflection, scattering and absorption of signals. Electromagnetic properties are defined as: Formula (4); electromagnetic properties include dielectric constant and conductivity. In formula (4), M(x,y) represents the material type at point (x,y), e concrete and e glass represent the electromagnetic parameters of concrete and glass respectively.

[0101] Blockage factors: Non-line of sight (NLOS) conditions can cause reflections, scattering, etc., which can increase PL. The blockage characteristics of a communication link in line of sight (LOS) or NLOS conditions are defined as: Formula (5); where Path(x,y) represents the propagation path of Tx and Rx at point (x,y), L LOS and L NLOS Indicates LOS and NLOS areas respectively.

[0102] The height map factor is used to indicate that taller buildings may cause various propagation effects such as reflection and scattering, which may lead to increased PL. The height map is used to provide a bird's-eye view of the height of buildings in the scene. This feature is defined as: H map (x,y)=max i (I((x,y)∈A i )·h i ), formula (6); where I((x,y)∈A i ) represents the indicator function that is equal to 1 when the point (x, y) is within the area of building i, h i represents the height of building i, A i represents the area occupied by building i on the two-dimensional plane.

[0103] The height difference factor is used to represent the height difference between Tx / Rx and the building. The height difference map feature is defined as: ΔHTx (x,y)=H map (x,y)-z Tx ΔH Rx (x,y)=H map (x,y)-z Rx , formula (7); where ΔH Tx (x,y) and ΔH Rx (x,y) represents the height difference map feature.

[0104] The building outline factor is used to represent the boundary outline of the building and is defined as:

[0105] Formula (8); where Represents the edge boundary of the i-th building on the two-dimensional plane.

[0106] The 3D coordinate difference factor is used to represent the 3D coordinate difference between Tx and Rx, and is used to provide spatial information about signal propagation. This characteristic is defined as:

[0107] Δ3DCD(x,y,z)=(x Rx ,y Rx ,z Rx )-(x Tx ,y Tx ,z Tx ), formula (9); where Δ(x, y, z) represents the coordinate difference between Rx and Tx in three-dimensional space

[0108] The loss factor is used to represent the fact that signals inside buildings typically experience higher attenuation due to absorption by walls. The loss factor is defined as:

[0109]

[0110] For example, based on the defined MWEF, a representative scenario is selected for illustration, such as Figure 4 shown.

[0111] Optionally, the pre-processing and post-processing feature processing modules are used for pre-processing and post-processing; the pre-processing is used to align the feature dimensions of the multi-source heterogeneous input features in the training data based on the input training data, and the features after the feature dimension alignment are input to the codec module; the post-processing is used to perform residual processing and convolution mapping processing on the feature map output by the codec module to predict the true path loss map;

[0112] The codec module is used to receive the features after the feature dimension alignment, input the features after the feature dimension alignment into the first convolution block for preliminary feature fusion to obtain fused features, the encoder in the codec module extracts deep-level environmental features based on the fused features, and the decoder in the codec module outputs the feature map based on the deep-level environmental features; the encoder includes four convolution blocks and four maximum pooling layers; the decoder includes four upsampling layers and four convolution blocks, and there is a skip connection between the feature map of the i-th layer of the encoder and the feature map of the j-th layer of the decoder, where j = the total number of layers of the codec module - 1;

[0113] The channel attention module is arranged between each of the upsampling layers and the convolution block in each of the encoders; the channel attention module is used to compress the feature map to be processed in the codec module along the spatial dimension, and use the fully connected layer to generate normalized weights, and assign the normalized weights to key features.

[0114] It should be noted that the above training data is the target data set determined in this application.

[0115] It should also be noted that the feature map that needs to be processed in the codec module is not equivalent to the feature map output by the codec module. The codec module may involve multiple feature maps during the processing process. Therefore, the feature map that needs to be processed in the codec module is the key feature map that requires weight processing in the entire processing process.

[0116] In this embodiment, the pre- and post-feature processing modules are used for data preprocessing and result optimization. These modules are responsible for both data preparation and optimization, and are crucial for the overall architecture. Pre-processing is used to align and fuse heterogeneous features from multiple sources; the input features are heterogeneous data from multiple sources, potentially with varying dimensions, scales, and physical meanings.

[0117] During the feature dimension alignment process for multi-source heterogeneous input features in the training data, features from different sources can be mapped to a unified spatial coordinate system. Through feature concatenation, multi-source features are integrated into a first feature with consistent dimensions. For example, building height (meters) and distance from a base station (kilometers) are converted into a unified numerical feature vector. Post-processing is used for skip connections to achieve multi-scale feature fusion. This pre- and post-feature processing module can fuse shallow features that retain original details in the encoder with the prediction results output by the decoder. Shallow features provide spatial details, while deep features capture global propagation patterns. The combination of the two can generate more accurate multi-scale prediction results.

[0118] Reference Figure 5As shown, the training data is "MWEF". In view of the characteristics of multi-source heterogeneous environment features in wireless communication scenarios, the pre- and post-feature processing modules realize feature alignment and feature fusion, and adjust the dimension to output the path loss map. The input "MWEF" data consists of two parts: input 1 (input1) (B×8×240×160) and input 2 (input2) (B×6×240×160), where input1 is adjusted to B×8×240×160 by bilinear interpolation upsampling (Upsample), and then spliced with input2 along the channel dimension ( Figure 5 The "cat" on the left) forms a B×14×240×160 feature map, which is the first feature. In the downstream feature processing stage, the decoder output (B×64×240×160) is skipped and connected to the feature map (B×14×240×160) ( Figure 5 The "Skip Connection" in

[15] optimizes the flow of deep information through residual connections and maps it to the target output B×1×60×40. This design not only aligns the output resolution but also alleviates the vanishing gradient problem in deep networks through residual connections, improving the model's adaptability to complex environmental features.

[0119] The codec module is used for feature extraction and spatial reconstruction. The codec module is an encoder-decoder network with skip connections, which constitutes the core architecture of the MWEF-PLP network to achieve accurate path loss prediction. Figure 5 As shown in the figure, the first feature is converted to B×64×240×160 through the first convolution block (containing two convolution layers: kernel size (kernel_size) = 5, padding (padding) = 2, stride (stride) = 1, with batch normalization and ReLU activation function), and the feature is initially fused and sent to the encoder. The encoder uses four convolution blocks (which can be used for Figure 5 The decoder uses four upsampling layers ( Figure 5The spatial feature map is gradually reconstructed using "Upsampling" (using bilinear interpolation) and four convolutional blocks. However, this compression process can lead to information loss, forming a bottleneck and degrading prediction quality. To address this, MWEF-PLP transfers the features of the corresponding encoder layer through a channel attention module and incorporates skip connections between the encoder and decoder. Skip connections enable the decoder to directly access the multi-scale features of the encoder, effectively propagating fine spatial details and rich hierarchical contextual information to higher resolution layers.

[0120] The aforementioned skip connections are a key mechanism for achieving multi-scale feature fusion. For example, the feature map of the i-th layer of the encoder (with dimensions Hi*Wi*Ci) is directly concatenated or added to the input of the j-th layer of the decoder (where j corresponds to i, and j = total number of layers - i). Specifically, the feature map of the first layer of the encoder is fused with the feature map of the penultimate layer of the decoder to provide additional detailed information.

[0121] Furthermore, the channel attention module of this application is positioned within a neural network. This module is typically embedded between multiple layers of the encoder-decoder module. For example, within the encoder, it is applied after each downsampling block to enhance the encoder's ability to extract key features. Within the skip connection, it weights encoder features before feature fusion to ensure more valuable information is passed to the decoder. Within the decoder, it is applied after the upsampling block to optimize feature representation during reconstruction.

[0122] In this application, since the importance of multi-scale wireless environment features varies, a deep neural network architecture, such as SE-Net (Squeeze-and-Excitation Network), is introduced as a channel attention mechanism to dynamically adjust the weights of multi-scale features. Specifically, the spatial dimension of the environmental feature map is compressed by global average pooling, and then passed through two fully connected layers and normalized weights are generated. Dynamic weights are used to emphasize key features while suppressing irrelevant features, thereby improving prediction performance. Figure 5 As shown, the channel attention module is set between each of the upsampling layers and the convolution blocks in each of the encoders, i.e. Figure 5 The "Squeeze-and-Excitation Network" in

[15] is set between the upsampling layer of the decoder and the convolutional block in each encoder through skip connections.

[0123] Specifically, the present application completes the pre- and post-feature processing modules and the codec backbone network, and introduces a channel attention mechanism. The overall network MWEF-PLP of the preset neural network architecture of the present application consists of three key modules: pre- and post-feature processing modules, codec modules and channel attention modules. The overall architecture adopts an encoder-decoder backbone network based on jump connections, with MWEF as input, covering information such as Tx position, occlusion status, electromagnetic parameters, and wear characteristics. Through feature compression and reconstruction, and extracting multi-scale features, it can realize the prediction of path loss maps in all scenarios. The pre- and post-feature processing modules realize input feature alignment and fusion, and optimize the output PL map. The channel attention mechanism dynamically adjusts the multi-scale feature weights to highlight key high-dimensional environmental features.

[0124] Optionally, after step 15, the method further includes:

[0125] The accuracy and generalization of the prediction model are evaluated, and path loss prediction is performed based on the evaluated prediction model.

[0126] In an embodiment of the present application, this step is a process of evaluating the accuracy and generalization of the prediction model. For example, the accuracy of the model can be evaluated by comparing it with the existing advanced path loss prediction method; the trained model is tested in a new scenario and compared with the path loss prediction method and multiple feature combination methods of the prior art to evaluate the generalization of the model. The present application utilizes the evaluated prediction model to achieve path loss prediction in different scenarios. The prediction model constructed in the present application can predict new scenarios, thereby improving the zero-sample learning ability of the model. Here, the new scenario is a scenario that is different from the training and testing in the prediction model. For example, in the first scenario of training and testing, there are at least two features, and the at least two features can be determined based on sample data; the new scenario refers to a new application scenario that includes the at least two features, but the data corresponding to the at least two features is not obtained from the sample data, and the data content is completely different from that in the first scenario.

[0127] Optionally, the above-mentioned evaluation of the accuracy and generalization of the prediction model includes:

[0128] Based on the prediction model, determining a predicted path loss value of the prediction model;

[0129] Get the actual path loss value;

[0130] Determining a root mean square error between the actual path loss value and the predicted path loss value;

[0131] If the root mean square error is within a first preset range, determining that the accuracy of the prediction model is passed; otherwise, performing an operation of retraining the prediction model;

[0132] If the accuracy of the prediction model is passed, determining a prediction value of the prediction model in a preset scenario, and evaluating the generalization of the prediction model based on the prediction model;

[0133] When the predicted value is within a second preset range, it is determined that the generalization of the prediction model is passed; otherwise, an operation of retraining the prediction model is performed.

[0134] It should be noted that the evaluation accuracy can be evaluated using the test set in the first target sub-dataset mentioned above; the evaluation generalization can be evaluated using the second target sub-dataset mentioned above.

[0135] In the present embodiment, the accuracy and generalization of the proposed prediction model are evaluated. In this step, the accuracy and generalization of the prediction model are evaluated separately. The specific steps are as follows: the accuracy of the model is evaluated by comparing the root mean square error (RMSE) of the actual path loss value and the path loss value predicted by the neural network model.

[0136] Furthermore, the present application sets a first preset range and a second preset range. The two prediction ranges are respectively to use the output results of the prediction model alone to determine whether the prediction model meets the accuracy and generalization required by the present application. If the root mean square error is within the first preset range, the accuracy of the prediction model is determined to be passed, otherwise, the operation of retraining the prediction model is performed; if the accuracy of the prediction model is passed, the prediction value of the prediction model in the preset scenario is determined, and the generalization of the prediction model is evaluated based on the prediction model; the preset scenario is the "new scenario" defined above in the present application. If the prediction value is within the second preset range, the generalization of the prediction model is determined to be passed, otherwise, the operation of retraining the prediction model is performed.

[0137] This application can also use the root mean square error to compare with related technologies to determine whether the prediction model meets the accuracy requirements. For example, the data sets obtained from the 18 scenarios in the first category are divided into training sets and test sets in a ratio of 8:2. The training set is used to train the neural network, that is, to train the prediction model, and the test set is used to evaluate the performance of the proposed algorithm and the accuracy of the RadioNet and PMNet algorithms on the test data set. The PMNet and RadioUNet models only consider Tx location and city map information (penetration characteristics) to predict the path loss map. The specific training configuration of the proposed method MWEF-PLP is shown in Table 2 below.

[0138] Table 2:

[0139] Model MWEF-PLP Learning rate <![CDATA[8×10 -4 ]]> LR lambda,step size 0.9,20 Batch size 16 Optimizer Adam epoch number 80

[0140] All the above algorithms were trained for 80 epochs. 80 epochs means that the model was trained for 80 times and the MSE loss function was used. Figure 6 As shown in the figure, the proposed method achieved the lowest NMSE during network model training, reducing it to approximately 7dB. PMNet is a model designed based on dilated convolutions and skip connections, while RadioUNet is a model designed based on UNet. Compared with the above two methods, MWEF-PLP achieved gains of 31.7% and 69.2%, respectively. This demonstrates that MWEF plays a significant role in improving PLP accuracy.

[0141] Furthermore, the performance of the trained model was tested in "new scenarios" to verify its generalization. Furthermore, to evaluate the model's generalization ability, three new scenarios were created to obtain a dataset containing 900 samples, enabling zero-shot learning for path loss prediction. As shown in Table 3, the generalization ability of MWEF-PLP using individual wireless environment features was investigated. The results show that different features achieve varying prediction accuracy. Specifically, distance, FSPL, electromagnetic characteristics, 3D coordinate difference, Tx position, and occlusion features perform poorly. In contrast, other wireless environment features (WEFs) demonstrate some generalization ability in new scenarios, with the penetration feature achieving an RMSE of 16.53 dB. Because different scenarios exhibit fundamental differences in building size, distribution, and number, considering that location information alone does not significantly improve the model's generalization ability across different scenarios, four location-related features (LRFs)—distance, FSPL, 3D coordinate difference, and Tx position—are combined as a baseline. On this basis, its generalization ability is further studied by comparing its performance after incorporating other WEFs.

[0142] Table 3:

[0143]

[0144] Reference Figure 7As shown, LRF combined with other WEFs exhibits varying generalization capabilities. Specifically, combinations involving height, loss, and the height difference between Tx and Rx outperform the generalization performance of a single WEF, with the fusion including loss achieving an RMSE of 9.98 dB. In contrast, features such as occlusion, electromagnetic properties, and profiles showed no significant improvement, with some even exhibiting performance degradation. Notably, the RMSE of the MWEF-PLP method was 8.03 dB, a 54.11% improvement over the PMNet method, demonstrating the effectiveness of combining MWEFs with model design. This demonstrates that the mapping relationship between MWEFs and PLs is crucial for accurately characterizing channel propagation. In future work, extracting more comprehensive and effective WEFs and optimizing model algorithms will further improve the accuracy and generalization capability of PLP.

[0145] In summary, compared with the existing technology, this application has the following advantages: This method analyzes from the perspective of electromagnetic propagation characteristics and proposes multi-dimensional wireless environment features to capture various basic environmental attributes, such as distance, occlusion, and electromagnetic characteristics. It aims to comprehensively characterize the environmental model to achieve more accurate prediction of path loss, which is significantly better than the baseline method. This method introduces skip connections and channel attention mechanisms to achieve the extraction and optimization of multi-scale wireless environment features, and performs efficient and accurate path loss prediction for different scenarios. This method designs multi-dimensional wireless environment features to achieve path loss prediction, enhances the model's zero-shot learning ability, and has the best generalization performance in different scenarios compared to other baseline methods.

[0146] The above describes various methods of the embodiments of the present application. The following further provides apparatuses for implementing the above methods.

[0147] Please refer to Figure 8 The embodiment of the present application further provides a path loss prediction device based on multi-dimensional wireless environment characteristics, including:

[0148] A first determining module 81 is used to determine scene data of multiple different scenes;

[0149] A first acquisition module 82 is configured to acquire path loss data corresponding to different scenario data based on scenario data of different scenarios;

[0150] A first building module 83 is used to build multi-dimensional wireless environment features;

[0151] A second construction module 84 is configured to align the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios;

[0152] The third construction module 85 is used to construct a preset neural network architecture of an encoder-decoder based on jump connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; the preset neural network architecture includes: pre- and post-feature processing modules, an encoder-decoder module and a channel attention module.

[0153] Optionally, the path loss prediction device based on multi-dimensional wireless environment characteristics in the embodiment of the present application further includes:

[0154] The evaluation module is used to evaluate the accuracy and generalization of the prediction model and perform path loss prediction based on the evaluated prediction model.

[0155] Optionally, the configuration information of the path loss data includes at least one of the center frequency, transmission order, diffraction order, transmission power, receiving power threshold, the number of paths retained by each receiving antenna, ray tracing simulation propagation model, transmitting antenna, receiving antenna, transmitting antenna position and receiving antenna position.

[0156] Optionally, the multi-dimensional wireless environment feature includes at least one of the following:

[0157] Calculate the distance factor between the transmitting antenna and the receiving antenna based on the Euclidean distance;

[0158] Free space path loss factor;

[0159] Transmitting antenna location factors;

[0160] Electromagnetic characteristics factors;

[0161] Obstruction factors of communication links in line-of-sight or non-line-of-sight situations;

[0162] Height map factors of buildings;

[0163] The height difference between the transmitter and receiver and the building;

[0164] Building outline factors;

[0165] The three-dimensional coordinate difference factor between the transmitting antenna and the receiving antenna;

[0166] Factors that damage buildings.

[0167] Optionally, the pre-processing and post-processing feature processing modules are used for pre-processing and post-processing; the pre-processing is used to align the feature dimensions of the multi-source heterogeneous input features in the training data based on the input training data, and the features after the feature dimension alignment are input to the codec module; the post-processing is used to perform residual processing and convolution mapping processing on the feature map output by the codec module to predict the true path loss map;

[0168] The codec module is used to receive the features after the feature dimension alignment, input the features after the feature dimension alignment into the first convolution block for preliminary feature fusion to obtain fused features, the encoder in the codec module extracts deep-level environmental features based on the fused features, and the decoder in the codec module outputs the feature map based on the deep-level environmental features; the encoder includes four convolution blocks and four maximum pooling layers; the decoder includes four upsampling layers and four convolution blocks, and there is a skip connection between the feature map of the i-th layer of the encoder and the feature map of the j-th layer of the decoder, where j = the total number of layers of the codec module - 1;

[0169] The channel attention module is arranged between each of the upsampling layers and the convolution block in each of the encoders; the channel attention module is used to compress the feature map to be processed in the codec module along the spatial dimension, and use the fully connected layer to generate normalized weights, and assign the normalized weights to key features.

[0170] Optional assessment modules include:

[0171] A first determining unit, configured to determine a predicted path loss value of the prediction model based on the prediction model;

[0172] A first acquiring unit, configured to acquire a true path loss value;

[0173] A second determining unit, configured to determine a root mean square error between the actual path loss value and the predicted path loss value;

[0174] a first processing unit, configured to determine that the accuracy of the prediction model passes when the root mean square error is within a first preset range, and otherwise, perform an operation of retraining the prediction model;

[0175] a second processing unit, configured to, when the accuracy of the prediction model is passed, determine a prediction value of the prediction model in a preset scenario, and evaluate the generalization of the prediction model based on the prediction model;

[0176] The third processing unit is configured to determine that the generalization of the prediction model is passed when the prediction value is within a second preset range, and otherwise, perform an operation of retraining the prediction model.

[0177] It should be noted that the device in this embodiment is a device corresponding to the above-mentioned method, and the implementation methods in the above-mentioned embodiments are all applicable to the embodiments of this device and can achieve the same technical effects. The above-mentioned device provided in the embodiment of this application can implement all the method steps implemented in the above-mentioned method embodiment and can achieve the same technical effects. The parts and beneficial effects of this embodiment that are the same as those in the method embodiment will not be specifically described here.

[0178] The present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the various processes of the above-mentioned path loss prediction method based on multi-dimensional wireless environment characteristics are implemented, and the same technical effects are achieved. To avoid repetition, the details are not described here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0179] An embodiment of the present application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, the various processes of the above-mentioned path loss prediction method embodiment based on multi-dimensional wireless environment characteristics are implemented, and the same technical effect can be achieved. To avoid repetition, they will not be repeated here.

[0180] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0181] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), including a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) to execute the methods described in each embodiment of the present application.

[0182] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A path loss prediction method based on multi-dimensional wireless environment characteristics, characterized in that: include: Determine scenario data for multiple different scenarios; Based on the scenario data of different scenarios, path loss data corresponding to the different scenario data is obtained; Construct multi-dimensional wireless environment characteristics; Aligning the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios; Constructing a preset neural network architecture of an encoder-decoder based on skip connections, using the target data set to construct a training set, and using the training set to train the preset neural network architecture to construct a prediction model; The preset neural network architecture includes: pre- and post-feature processing modules, a codec module, and a channel attention module.

2. The method according to claim 1, characterized in that After building the prediction model, the method further includes: The accuracy and generalization of the prediction model are evaluated, and path loss prediction is performed based on the evaluated prediction model.

3. The method according to claim 1, characterized in that The configuration information of the path loss data includes at least one of the center frequency, transmission order, diffraction order, transmission power, receiving power threshold, the number of paths retained by each receiving antenna, ray tracing simulation propagation model, transmitting antenna, receiving antenna, transmitting antenna position and receiving antenna position.

4. The method according to claim 1, wherein The multi-dimensional wireless environment characteristics include at least one of the following: Calculate the distance factor between the transmitting antenna and the receiving antenna based on the Euclidean distance; Free space path loss factor; Transmitting antenna location factors; Electromagnetic characteristics factors; Obstruction factors of communication links in line-of-sight or non-line-of-sight situations; Height map factors of buildings; The height difference between the transmitter and receiver and the building; Building outline factors; The three-dimensional coordinate difference factor between the transmitting antenna and the receiving antenna; Factors that damage buildings.

5. The method according to claim 1, characterized in that The pre-processing and post-processing modules are used for pre-processing and post-processing; the pre-processing is used to align the feature dimensions of the multi-source heterogeneous input features in the training data based on the input training data, and the features after the feature dimension alignment are input to the codec module; the post-processing is used to perform residual processing and convolution mapping processing on the feature map output by the codec module to predict the true path loss map; The codec module is used to receive the features after the feature dimension alignment, input the features after the feature dimension alignment into the first convolution block for preliminary feature fusion to obtain fused features, the encoder in the codec module extracts deep-level environmental features based on the fused features, and the decoder in the codec module outputs the feature map based on the deep-level environmental features; the encoder includes four convolution blocks and four maximum pooling layers; the decoder includes four upsampling layers and four convolution blocks, and there is a skip connection between the feature map of the i-th layer of the encoder and the feature map of the j-th layer of the decoder, where j = the total number of layers of the codec module - 1; The channel attention module is arranged between each of the upsampling layers and the convolution block in each of the encoders; the channel attention module is used to compress the feature map to be processed in the codec module along the spatial dimension, and use the fully connected layer to generate normalized weights, and assign the normalized weights to key features.

6. The method according to claim 2, characterized in that Evaluate the accuracy and generalization of the predictive model, including: Based on the prediction model, determining a predicted path loss value of the prediction model; Get the actual path loss value; Determining a root mean square error between the actual path loss value and the predicted path loss value; If the root mean square error is within a first preset range, determining that the accuracy of the prediction model is passed; otherwise, performing an operation of retraining the prediction model; If the accuracy of the prediction model is passed, determining a prediction value of the prediction model in a preset scenario, and evaluating the generalization of the prediction model based on the prediction model; When the predicted value is within a second preset range, it is determined that the generalization of the prediction model is passed; otherwise, an operation of retraining the prediction model is performed.

7. A path loss prediction device based on multi-dimensional wireless environment characteristics, characterized in that: include: A first determining module, configured to determine scene data of a plurality of different scenes; A first acquisition module is used to acquire path loss data corresponding to different scenario data based on scenario data of different scenarios; A first building module is used to build multi-dimensional wireless environment features; A second construction module is configured to align the path loss data with the multi-dimensional wireless environment characteristics to construct a target data set for large-scale and diverse scenarios; A third construction module is used to construct a preset neural network architecture of an encoder-decoder based on skip connections, use the target data set to construct a training set, and use the training set to train the preset neural network architecture to construct a prediction model; The preset neural network architecture includes: pre- and post-feature processing modules, a codec module, and a channel attention module.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

9. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps of the method according to any one of claims 1 to 6.

Images