Heterogeneous graph neural network-based Internet of Vehicles path loss prediction method and system

Abstract

The invention belongs to the related technical field of wireless communication, and particularly relates to an Internet of Vehicles path loss prediction method and system based on a heterogeneous graph neural network, and the method comprises the steps: taking a vehicle and a selected receiver as nodes, taking the multi-dimensional physical driving characteristics of the vehicle and the receiver, and taking the spatial correlation between the receivers as edges to construct a heterogeneous graph; and predicting the heterogeneous graph by using the trained multi-modal mixed attention graph neural network model to obtain an Internet of Vehicles path loss prediction result. The heterogeneous graph not only integrates geometric physical characteristics and static environment characteristics of a vehicle and a receiver, but also integrates dynamic environment characteristics and microscopic path profile characteristics, so that the model can effectively capture channel violent changes caused by high-speed movement of the vehicle, rapid changes of surrounding shelters and a close-range microscopic sheltering effect; and the prediction robustness of the model in extreme dynamic and complex microenvironments is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and in particular relates to a method and system for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In Vehicle-to-Everything (V2X) communication scenarios, especially driven by the demand for high-reliability, low-latency communication, achieving high-accuracy path loss prediction is a crucial foundation for ensuring the performance and safety of future V2X systems. High-accuracy path loss prediction effectively supports wireless resource management, dynamic routing, communication reliability assurance, and mobility management, and has profound and far-reaching significance for improving autonomous driving safety, traffic efficiency, and user experience.

[0004] However, in high-frequency millimeter-wave, high-dynamic vehicular network communication environments, channel path loss prediction still faces significant complexity and technical challenges. First, there is the extreme dynamism of the environment. In real-world vehicular network scenarios, the transmitting end (vehicle) often moves at high speeds, while dynamic obstructions around the communication link (such as other fast-moving vehicles, pedestrians, trees, etc.) are constantly changing. This high dynamism causes the wireless channel characteristics to change drastically within milliseconds, making it difficult for traditional prediction models that rely on assumptions of a relatively static environment to capture and adapt to this real-time dynamism. Second, there is the complexity of the close proximity of the transmitting and receiving ends and the significant impact of the micro-environment. When the communication distance is short, the propagation characteristics of millimeter-wave signals become extremely sensitive to micro-environmental factors. For example, the branches and leaves of a tree, a parked bus, or even the local obstruction and reflection effects of the vehicle's own structure can significantly affect the signal, with an impact far exceeding that of macroscopic building obstructions typically considered by traditional models. Existing models are severely inadequate in characterizing these close-range micro-effects and fail to effectively utilize fine-grained perceptual information from modalities such as vision. Finally, there are the challenges of high-frequency signal sensitivity and non-line-of-sight (NLoS) prediction. Millimeter-wave signals have extremely poor penetration capabilities and are highly sensitive to any obstacle. In NLoS communication scenarios, signal loss is enormous and its attenuation pattern is complex and variable, posing significant difficulties for accurate modeling and prediction. Traditional models based on geometric optics or ray tracing (RT) often fail to meet the accuracy and real-time requirements of future intelligent vehicle connectivity when dealing with the complex multipath propagation and high sensitivity of millimeter waves.

[0005] Fortunately, future intelligent vehicles will be equipped with a variety of advanced communication and sensing devices, such as cameras, LiDAR, millimeter-wave radar, and diverse communication modules. These devices can independently perceive their surroundings or exchange information, thus enabling the intelligent fusion of multimodal perception and communication through a "machine synesthesia" framework. Through the intelligent fusion of multimodal perception and communication information, vehicles can acquire more comprehensive and in-depth environmental insights from heterogeneous data, providing unprecedented information support for complex tasks such as path loss prediction.

[0006] Existing research and technologies suffer from the following limitations, failing to effectively support dynamic channel prediction tasks in high-frequency vehicle-to-everything (V2X) networks during the 6G era, particularly in multimodal sensing information fusion. First, they rely excessively on static geographic information system (GIS) maps and empirical models. Most traditional path loss prediction methods assume a relatively static environment, making it difficult to effectively integrate real-time multimodal sensing data. Therefore, they cannot perceive and adapt to real-time dynamic changes and microscopic occlusion effects, nor can they capture the fine propagation mechanisms of millimeter-wave signals in complex multipath environments. Second, they lack the perception and fusion of real-time dynamic information and the microscopic environment. Existing models generally cannot effectively integrate real-time vehicle-mounted sensing data for deep fusion, thus failing to detect obstacles not marked on the map or those appearing instantaneously (such as vehicles suddenly cutting into lanes or fast-moving pedestrians). This results in significant prediction errors in highly dynamic NLoS scenarios, failing to utilize the complementary information provided by multimodal data. Furthermore, the generalization ability of traditional models relying on specific environmental information is generally limited. Most data-driven channel models are typically trained and tested in specific geographical scenarios or with limited datasets, generally exhibiting insufficient generalization ability across scenarios and geographical regions, making them difficult to deploy and apply in diverse real-world intelligent vehicle-to-everything (V2X) environments. Finally, traditional statistical models and path loss prediction models based on single-modal information struggle to efficiently handle complex, multi-source heterogeneous data. With the explosive growth of V2X sensing data, existing models face bottlenecks in the effective fusion of multi-source heterogeneous data (such as macro-maps, micro-trajectories, and real-time vision), high-dimensional feature extraction, and complex nonlinear relationship modeling, making it difficult to learn deep-seated channel propagation patterns from massive, heterogeneous data.

[0007] In summary, existing technologies cannot effectively address the challenges of extreme environmental dynamics, close-range microscopic complexity, and high-frequency signal sensitivity. Especially in the context of the intelligent fusion of multimodal perception and communication, there is an urgent need for a channel path loss prediction technology that deeply understands complex environments and possesses high accuracy and strong generalization capabilities to meet the stringent requirements of future intelligent transportation and autonomous driving for wireless communication. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting path loss in vehicle networking based on heterogeneous graph neural networks, which enables real-time and high-precision prediction of path loss, effectively identifies communication interruption events, and provides solid communication support for autonomous driving and intelligent transportation.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks, comprising: Using the vehicle and the selected receiver as nodes, a heterogeneous graph is constructed with the multidimensional physical driving characteristics of the vehicle and the receiver, as well as the spatial correlation between receivers, as edges; wherein, the multidimensional physical driving characteristics include the geometric physical characteristics, static environmental characteristics, micro-path profile characteristics, and dynamic environmental characteristics of the vehicle and the receiver. The heterogeneous graph is predicted using a trained multimodal hybrid attention graph neural network model to obtain the vehicle network path loss prediction result; The multimodal hybrid attention map neural network model processes the heterogeneous graph as follows: Extract the node features of the nodes and the edge features of the edges in the heterogeneous graph and embed them; A multi-layered heterogeneous graph Transformer layer is used to update node features, and the updated node features are extracted from different depths and fused to form fused node features. By embedding and concatenating node features and edge features, link-level features are learned through a feature interaction layer, and the link-level features are used to predict the path loss of the vehicle network.

[0010] Secondly, the present invention provides a vehicle network path loss prediction system based on heterogeneous graph neural networks, comprising: The heterogeneous graph construction module is configured to: construct a heterogeneous graph using the vehicle and the selected receiver as nodes, and using the multidimensional physical driving features of the vehicle and the receiver, as well as the spatial correlation between receivers, as edges; wherein, the multidimensional physical driving features include the geometric physical features, static environmental features, microscopic path profile features, and dynamic environmental features of the vehicle and the receiver. The prediction module is configured to: use a trained multimodal hybrid attention graph neural network model to predict the heterogeneous graph and obtain the vehicle network path loss prediction result; The multimodal hybrid attention map neural network model processes the heterogeneous graph as follows: Extract the node features of the nodes and the edge features of the edges in the heterogeneous graph and embed them; A multi-layered heterogeneous graph Transformer layer is used to update node features, and the updated node features are extracted from different depths and fused to form fused node features. By embedding and concatenating node features and edge features, link-level features are learned through a feature interaction layer, and the link-level features are used to predict the path loss of the vehicle network.

[0011] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0012] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0013] The above one or more technical solutions have the following beneficial effects: In this invention, the heterogeneous graph not only integrates the geometric and physical features and static environmental features of the vehicle and receiver, but also integrates dynamic environmental features and micro-path profile features. This enables the model to effectively capture the drastic channel changes caused by the high-speed movement of the vehicle, the rapid changes of surrounding obstructions, and the effect of close-range micro-obstruction, greatly improving the model's prediction robustness in extreme dynamic and complex micro-environments.

[0014] In this invention, multiple layers of heterogeneous graph Transformer layers are stacked, and each layer innovatively integrates local graph attention and global self-attention mechanisms. Through a learnable gating fusion unit, the outputs of these two attention mechanisms are dynamically balanced and combined to achieve a deep understanding and comprehensive modeling of near-field physical interactions and far-field spatial context, effectively addressing the sensitivity of millimeter-wave signals to obstacles and complex propagation environments.

[0015] In this invention, the training process is driven by minimizing a composite optimization objective that combines the MAE of uninterrupted samples with the F1 score, and an adaptive learning rate scheduler and an early stopping mechanism are used to achieve the stability and optimal performance of model training.

[0016] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of the vehicle network path loss prediction method based on heterogeneous graph neural network in Embodiment 1 of the present invention. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0021] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] Example 1 This embodiment discloses a path loss prediction method for vehicular networks based on heterogeneous graph neural networks, aiming to solve the challenge of real-time, high-precision prediction of path loss in complex urban environments such as 6G high-frequency millimeter wave (28 GHz) communication between vehicles and infrastructure. Specifically, this embodiment constructs a multimodal, physically driven heterogeneous graph structure and utilizes an innovative hybrid attention graph neural network architecture to achieve a deep understanding and accurate prediction of dynamic channel characteristics. The following is a detailed description of the vehicle network path loss prediction method based on heterogeneous graph neural networks proposed in this embodiment: Step 1: Construction of high-fidelity multimodal heterogeneous graph data.

[0023] This embodiment proposes a high-fidelity multimodal heterogeneous graph data construction method that integrates macroscopic map information, microscopic path profile information, dynamic vehicle trajectory information, and real-time vehicle visual perception information into a unified heterogeneous graph structure.

[0024] By constructing nodes in the graph, including vehicle nodes and selected receiver nodes, their features are normalized coordinates and type identifiers. The features of vehicle nodes are innovatively enhanced using real-time in-vehicle visual semantic features. Edges in the graph are constructed, primarily including vehicle-to-receiver (Ego-Rx) edges and receiver-to-receiver (Rx-Rx) edges. For Ego-Rx edges, this embodiment innovatively extracts a 13-dimensional, physically driven, multi-dimensional edge feature system. This system not only covers core geometric and physical features such as distance and frequency, but also deeply describes macroscopic static environmental features (e.g., the proportion of buildings and trees), microscopic path profile features (e.g., the type and distance of the first obstacle, the maximum continuous occlusion length, and the number of independent obstacles), and dynamic environmental features (e.g., the number of dynamic occluders and the distance to the nearest dynamic obstacle), thereby accurately characterizing the all-around physical propagation characteristics of the communication link. Through intelligent sampling of the original data list and allocation of difficulty-aware sampling weights based on the path loss difficulty distribution, the model's focus on complex scenarios during training is optimized.

[0025] The specific steps are as follows: Step 11: Loading raw data and selecting snapshot samples.

[0026] First, this embodiment loads a static environment map (such as a building mask) from a preset data source. Tree cover Geographical set of millimeter-wave receivers (RSUs) , and the affine transformation matrix used for coordinate transformation and pixel-meter scale Simultaneously, load the raw data list containing snapshots from various times, where each entry... Snapshot information was recorded.

[0027] Secondly, based on configuration parameters, such as sampling rate The original data list is sparsified to select a subset of snapshot samples to participate in graph construction. .

[0028] Finally, for the training set, this embodiment calculates the path loss difficulty distribution for each snapshot sample based on the pre-computed distribution. Calculate and assign difficulty-aware sampling weights The aim is to ensure that more challenging scenarios receive greater attention during model training. Difficulty-aware sampling weights. It can be used for sample extraction during the data loading phase (with the option to use weighted sampling instead of random sampling), or introduced as weights in the subsequent loss function calculation phase to ensure that more challenging or specific types of scenarios receive greater attention during model training, thereby improving the model's generalization ability and prediction accuracy for extreme cases. Difficulty-aware sampling weights The calculation comprehensively considers the extreme ratios and quantiles (such as the P50 value) of the path loss values ​​within the snapshot, i.e. ,in, The aim is to increase the sampling frequency for more challenging scenarios, thereby improving the model's generalization ability and prediction accuracy for extreme cases.

[0029] in, Indicates the first The difficulty-aware sampling weights for each snapshot sample. These weights are used to quantify the importance of the current snapshot during training, guiding the model to focus more on challenging scenarios; Indicates the index of the snapshot sample currently being processed; Indicates the first In a snapshot sample, the proportion of links with path loss values ​​in the extremely high range (e.g., above a certain extremely high quantile, such as the 99th quantile) reflects the frequency of extreme communication interruption events within that snapshot. Indicates the first The 50th percentile (i.e., median) of all link path loss values ​​in a snapshot sample reflects the overall level of path loss within that snapshot. This represents a set of predefined threshold parameters, such as multiple boundary points used to divide the difficulty level of path loss, and is used to classify snapshots into different difficulty levels (e.g., "easy", "medium", "hard" and "extreme") based on their extreme_ratio and P50 values. This represents a mapping function that maps based on a snapshot. , Values ​​and presets This is then categorized into a difficulty level (e.g., 0-3), and this difficulty level is further mapped to a weight. For example, a strategy of inverse class frequency weighting is often used, in which the weight of the difficulty level is proportional to the inverse of the number of samples in that difficulty level, so that rare but difficult samples get a higher sampling probability.

[0030] Step 12: Coordinate transformation and receiver node selection.

[0031] First, this embodiment performs a unified transformation on all physical coordinates. For any physical coordinate... Using a preloaded affine transformation matrix Convert it to pixel coordinates , represented as ,in It is Transformation matrix, pixel-to-meter scale Used to convert between pixel distance and physical distance.

[0032] Secondly, for each snapshot, given the current physical location of the vehicle (transmitter). This embodiment uses all available receiver sets. Filter out a specified number The most recent receiver is the set of valid receivers for the current snapshot. This filtering process ensures that the constructed graph focuses on the RSU nodes most relevant to vehicle communication.

[0033] Finally, verify all the transformed vehicle pixel coordinates. and selected receiver pixel coordinates All are within the effective range of the static environment map, that is and ,in and These represent the width and height of the map, respectively.

[0034] Step 13: Graph node feature construction.

[0035] First, this embodiment constructs the set of nodes in the graph. The node primarily consists of one ego-vehicle node. and Selected Receiver Nodes .

[0036] Secondly, for the self-driving node Assign its transformed 2D pixel coordinates And a binary identifier representing that it is a vehicle. .

[0037] Finally, for each receiver node Assign its transformed 2D pixel coordinates And a binary identifier indicating that it is a receiver. .

[0038] It is worth noting that this embodiment innovatively does not directly store the pre-computed RGB image embedding as node features, but instead stores the file path of the original RGB image. Storing these as graph metadata allows for on-demand loading and processing during model training or inference, thereby improving the flexibility and efficiency of data processing. Node feature vectors It can be represented as .

[0039] Step 14: Physics-driven multidimensional edge feature extraction and graph connection construction.

[0040] The graph constructed in this embodiment is a heterogeneous graph, containing two main types of edges: vehicle-to-receiver (Ego-Rx) edges and receiver-to-receiver (Rx-Rx) edges. Each edge carries a set of physically driven multidimensional feature systems to comprehensively describe the physical propagation environment of the communication link.

[0041] Step 14.1: Construction and feature extraction of the Ego-Rx edge.

[0042] First, for the self-driving node With each selected receiver node In this embodiment, a directed edge is constructed between them. .

[0043] Secondly, for each Ego-Rx edge, extract its corresponding multidimensional physics-driven features. The multidimensional physics-driven feature extraction process includes: Core geometry and physical features: Calculating the physical distance between the vehicle and the receiver And logarithmic distance and record the communication frequency. .

[0044] Macroscopic static environment characteristics: The straight-line propagation path from the vehicle to the receiver is drawn on a pixel map. Statistics on the path blocked by building membranes and tree cover The percentage of pixels covered. For example, the percentage of buildings. ,in The number of pixels on the path, and the proportion of trees. Calculate in a similar manner.

[0045] Microscopic path profile features: This embodiment innovatively extracts fine microscopic obstacle information through pixel-level path analysis. First, it identifies... Up by and Formed joint obstacle mask Then, the type of the first obstacle on the straight-line propagation path from the car to the receiver is calculated on the pixel map. (0: None, 1: Tree, 2: Building) and their distance to the vehicle The physical length of the largest continuous obstruction on the straight propagation path. By identifying The length of the connected components along the path and converted to physical distance; and the number of independent obstacles traversed along the path. Through the Connected components are labeled along the path (e.g., using the scipy.ndimage.label function). These features significantly improve the model's understanding of the attenuation mechanism of millimeter-wave signals in complex non-line-of-sight (NLoS) environments.

[0046] Dynamic environment characteristics: This embodiment utilizes real-time trajectory data of all other vehicles in the current snapshot. Calculate the shortest distance from these dynamic obstacles to the Ego-Rx line segment. Identify locations within a specified radius of the connecting line segment. Other vehicles within the range are considered dynamic obstructions, and the number of dynamic obstructions is extracted. ,Right now ,in The indicator function; and the minimum distance to the nearest dynamic occluder. These features enable the model to dynamically sense and adapt to the impact of instantaneously moving obstacles on the channel.

[0047] Specifically, by iterating through all other vehicles in the current snapshot. For each other vehicle Calculate its distance to its own vehicle With receiver connecting segments shortest distance If this shortest distance is less than the preset dynamic obstacle detection radius... If so, the vehicle is considered a dynamic occlusion, and the indicator function is used. Return 1; Finally, sum the number of all dynamic occlusions that meet the criteria to obtain... .

[0048] in, This indicates the total number of dynamic obstacles (i.e., other vehicles) identified near the current vehicle-to-receiver (Ego-Rx) communication link. : Iterate through the indices of all non-autonomous vehicles. : indicates an indicator function; if the condition within the parentheses is true, then The value is 1; if the condition is false, the value is 0. Indicates the first Real-time location of other vehicles To the vehicle-receiver connection segment The shortest distance, which can be a physical distance (meters) or a pixel distance, depending on the coordinate system and units used. Indicates the first The real-time physical location (e.g., two-dimensional coordinates) of other vehicles. These location information come from vehicle trajectory data. Indicates the connection of the vehicle's pixel coordinates and receiver pixel coordinates The straight line segments are geometric lines defined on the pixel map. This represents a preset dynamic obstacle detection radius or threshold distance (usually in meters or pixels). If the shortest distance from other vehicles to the vehicle-receiver line segment is less than this radius... If so, it is considered that the vehicle constitutes a dynamic obstruction to the current link.

[0049] Edge type identifier: Marks this edge as a vehicle-receiver type. .

[0050] Finally, all 13 extracted dimensional features (as defined by EdgeFeature) are combined into the attribute tensor of the Ego-Rx edge. .

[0051] Step 14.2: Construction and feature extraction of receiver-to-receiver (Rx-Rx) edges.

[0052] First, to capture the spatial correlation between receivers (i.e., between RSUs), this embodiment constructs additional edges between selected receiver nodes. For each receiver... Select by K-Nearest Neighbor (K-NN) search (e.g., using KDTree). The nearest receiver As its neighbor, and construct bidirectional connection edges. and .

[0053] Secondly, for each Rx-Rx edge, extract its corresponding features. This mainly includes the physical distance between receivers. And logarithmic distance .

[0054] Finally, label the edge as receiver-receiver type. And combine them into the attribute tensor of the Rx-Rx edge.

[0055] Step 15: Path loss label binding and data standardization.

[0056] First, measure the actual millimeter-wave path loss corresponding to each Ego-Rx edge. As a label for this heterogeneous graph .

[0057] Secondly, this embodiment standardizes the constructed heterogeneous graph data. During the training phase, the following is calculated for all training samples: Node coordinates: normalized to ,Right now , .

[0058] Edge features: Edge features for continuous variables Z-score standardization is performed, that is ,in and These are the mean and standard deviation of the corresponding features on the training set.

[0059] Path loss label Z-score standardization is performed, i.e. ,in and These are the mean and standard deviation of the path loss labels on the training set.

[0060] Save these statistics for later use.

[0061] Finally, all the constructed graph data (including standardized node features, edge indices, edge features, path loss labels, and RGB image paths) are saved in a lightweight format to a specified directory for efficient loading during subsequent model training and evaluation.

[0062] Step 2: Multimodal hybrid attention map neural network model that integrates visual semantics.

[0063] This embodiment designs a novel multimodal hybrid attention graph neural network model for end-to-end prediction of path loss.

[0064] Node and edge feature encoding: Independent encoders are used to map initial node features (including enhanced vehicle visual semantic features) and edge features to a unified latent space.

[0065] Heterogeneous Graph Transformer Layers: The model stacks multiple `GraphTransformerLayer` layers, each innovatively integrating local graph attention (such as GATv2Conv) with global self-attention mechanisms. Through a learnable gated fusion unit, the outputs of these two attention mechanisms are dynamically balanced and combined to achieve a deep understanding and comprehensive modeling of near-field physical interactions and far-field spatial context, effectively addressing the sensitivity of millimeter-wave signals to obstacles and complex propagation environments.

[0066] Multi-scale feature fusion and link-level feature learning: After learning multiple layers of GNN, node features are extracted from different depths (such as intermediate and last layers) and fused to form rich node representations. Subsequently, the fused features of the vehicle node, receiver node, and Ego-Rx edge, along with the vehicle's visual semantic features, are concatenated and learned through a feature interaction layer to obtain link-level features.

[0067] The specific steps are as follows: Step 21: RGB image encoder construction and preprocessing.

[0068] First, this embodiment constructs an RGB image encoder to extract high-level visual semantic features from images from an in-vehicle camera. This encoder is based on a pre-trained MobileNetV2 model and uses its feature extraction component, rgb_feature_extractor.

[0069] Secondly, the RGB encoder is selectively fine-tuned according to the configuration parameter `finetune_rgb_encoder`. If no fine-tuning is performed, all parameters of the encoder are frozen and it is used only as a static feature extractor. If fine-tuning is performed, several layers at the end of the encoder (specified by `num_finetune_layers`) are unfrozen, allowing parameter updates during training to better adapt to the vehicle-to-everything (V2X) image data.

[0070] Finally, during the forward propagation of the model, the RGB image data (of shape ) in the input batch are processed. First, feature maps are extracted using an RGB feature extractor. Then, they are transformed into fixed-dimensional embedding vectors through an adaptive average pooling layer and a flattening operation (rgb_pooler). These embedding vectors are then further processed through a sequence of linear layers, layer normalization, GELU activation function, and Dropout (rgb_processor) to obtain the final visual semantic features. .

[0071] Step 22: Encoding of node and edge features.

[0072] First, for each node in the heterogeneous graph Its original node feature vector First, the nodes pass through a node encoder, which consists of a linear layer, layer normalization, GELU activation function, and Dropout, mapping them to a high-dimensional hidden space to obtain the initial node embeddings. .

[0073] Secondly, for the EgoVehicle node, if it is configured to use RGB image features (use_rgb is true), then the visual semantic features extracted in step 21 will be used. Integrate it into its initial node embedding. Specifically, for each vehicle node in the graph... Its corresponding embedding Will be combined with visual semantic features The index_add_ operation is performed to inject the vehicle's own visual perception information into the graph structure.

[0074] Finally, for each edge in the graph Its original edge feature vector After passing through an edge encoder, which also consists of a linear layer, layer normalization, GELU activation function, and Dropout, the edge feature embedding is mapped to the hidden space to obtain the edge feature embedding. .

[0075] Step 23: Stacking of multimodal attention map neural network layers.

[0076] First, this embodiment stacks multiple layers (num_layers) of GraphTransformerLayer modules to perform message passing and feature learning on the graph. Each GraphTransformerLayer receives the embedding of the current node. The edge index of the graph is used as input.

[0077] Secondly, within each GraphTransformerLayer, node features Simultaneously, it is processed through two independent attention mechanisms: Local graph attention: Implemented using the GATv2Conv layer, which can perceive the features and edge attributes of neighboring nodes and aggregate local information through an attention mechanism. For nodes... and his neighbors Features of local attention aggregation It can be represented as ,node Local features By all its neighbor nodes The sum is obtained by weighting and summing the features of each neighbor node. Before summing, each neighbor node... Features It will undergo a linear transformation (by the weight matrix) and bias vector (implementation), then multiplied by the neighbor node. For nodes Attention weights This mechanism allows the model to selectively aggregate key information from different neighbors, thereby capturing fine patterns of near-field physical interactions.

[0078] in These are the attention weights calculated by GATv2Conv; Represents a node The updated local feature vector is obtained by aggregating its neighbor information through a local graph attention mechanism. This represents the target node in the graph, i.e., the node whose updated features are being calculated. Represents a node One of its neighboring nodes. Represents a node The set of neighboring nodes. This represents the attention weights calculated by the GATv2Conv layer. These weights quantify the neighbor nodes. For the target node Information aggregation contribution. The computation is usually based on nodes and The features and edge attributes (if the edge attributes are integrated into the attention calculation of GATv2Conv) are used. This represents a learnable weight matrix used for neighboring nodes. The features are linearly transformed. In GATv2Conv, this matrix maps the features of neighboring nodes to a shared feature space for aggregation. Representing neighboring nodes The current feature vector. This represents a learnable bias vector, which is also used to perform linear transformations on the features of neighboring nodes.

[0079] Global self-attention: Employs the standard Transformer self-attention mechanism, through a query matrix. Key matrix Value matrix Perform global interaction with all nodes.

[0080] For each target node in the graph Its global features It is calculated through the following steps: First, the target node Features It is converted into a query vector Meanwhile, the feature matrices of all nodes in the graph Converted into a key vector matrix Sum value vector matrix The query vector is multiplied by all key vectors to obtain an unnormalized attention score, which is then divided by... Scaling is performed; then, the Softmax function is applied to these scores to obtain the ratio of each node to the target node. Attention weights are assigned; finally, these attention weights are applied to the value vector matrix. The target node is obtained by weighted summation. global features This mechanism enables nodes It can capture information from all other nodes in the graph, thereby understanding the far-field spatial context and global dependencies. node Features after global self-attention It can be represented as:

[0081] in It is the dimension of the key vector; Represents a node The updated global feature vector is obtained by aggregating information from all nodes in the graph through a global self-attention mechanism. This represents the target node in the graph, i.e., the node whose updated features are being calculated. Indicates the use of nodes Features Transformed into a learnable weight matrix for query vectors. Represents the target node The current feature vector. This represents the feature matrix used to represent all nodes. Learnable weight matrix converted into key vector matrix. This represents a matrix composed of the eigenvectors of all nodes in the graph. If the graph contains... There are n nodes, and the feature dimension of each node is 1. ,but . This represents the transpose of all key vector matrices after a linear transformation. Indicates passage The dimension of the transformed key vector, in the denominator It is a scaling factor used to prevent the dot product result from becoming too large and to stabilize training. It is the Softmax activation function, used to normalize the attention score into a probability distribution. This represents the feature matrix used to represent all nodes. Learnable weight matrix converted into a value vector matrix. This represents a matrix consisting of all value vectors after a linear transformation.

[0082] Hybrid Attention Fusion and Gating Mechanism: This embodiment innovatively utilizes learnable fusion weight parameters. (After Softmax normalization) Output the local graph attention. With global self-attention output By performing weighted fusion, we obtain Subsequently, a gating mechanism is introduced, utilizing the residual connections based on the characteristics of the current node. With fusion features Together they generate a gating signal Finally, the updated node features are output. This gating mechanism allows the model to adaptively balance the original information and the fused information.

[0083] Finally, each GraphTransformerLayer is followed by a feedforward network (FFN) and layer normalization operations to further enhance feature representation capabilities. Through multi-layer stacking, the model is able to capture the complex multi-scale spatial dependencies of the channel environment.

[0084] Step 24: Node feature fusion and prediction.

[0085] First, after completing message passing for all GraphTransformerLayer layers, this embodiment extracts node features from the middle layer (the output of layers with indices approximately half the total number of layers in the stacked GraphTransformerLayer sequence) and the last layer, obtaining shallow features respectively. and deep features These two features are fused through concatenation and feature fusion to obtain the final fused node features. This multi-scale feature fusion strategy can take into account both local details and global contextual information.

[0086] Secondly, to predict path loss, the model needs to perform feature aggregation on the vehicle-receiver (Ego-Rx) edges. In this embodiment, all Ego-Rx edges are filtered out using the EdgeFeature.IS_EGO_RX edge mask, and the fused node features of the source node (vehicle) and target node (receiver) corresponding to these edges, as well as the corresponding encoded edge features, are extracted. Furthermore, the visual semantic features generated for the vehicle nodes in step 21 are... (If it exists) it is also aggregated to the corresponding Ego-Rx edge.

[0087] Here, `EdgeFeature.IS_EGO_RX` is an enumeration or constant defined in the code. It represents a specific index in the edge feature vector, used to identify whether the edge is a vehicle-to-receiver (Ego-Rx) type edge. `EdgeFeature.IS_EGO_RX` is actually a dimension index of the edge feature vector. During the graph construction phase, when constructing an Ego-Rx edge, the dimension containing `EdgeFeature.IS_EGO_RX` in its corresponding edge feature vector is set to... When constructing receiver-to-receiver (Rx-Rx) edges, this dimension is set to... .

[0088] Therefore, the filtering process is achieved by checking whether the dimension corresponding to the vehicle-receiver edge type identifier in the feature vector of each edge is 1.

[0089] Specifically, when processing graph data, the model obtains the feature matrix `edge_attr` for all edges. Each row of this matrix corresponds to an edge, and each column corresponds to an edge feature dimension. By accessing the column corresponding to `EdgeFeature.IS_EGO_RX` in `edge_attr`, a boolean tensor (or mask) can be obtained. If the value of this column is... If it is, then it corresponds to `True`; if it is... If the value is `False`, then the result is `False`. Applying this Boolean mask to the edge index `edge_index` of the graph allows for precise filtering of all edges belonging to the Ego-Rx type.

[0090] Finally, these features (source node features, target node features, Ego-Rx edge features, and vehicle visual features) are concatenated to form a comprehensive interaction feature vector. This vector is then fed into a multilayer perceptron (feature_interaction_layer) for deep interaction learning, ultimately yielding link-level features for prediction. .

[0091] Here, Ego-Rx edge features refer to the feature embeddings that have been encoded by the edge encoder and belong only to the vehicle-to-receiver (Ego-Rx) type edges. .

[0092] Step 25: Path loss regression and classification prediction.

[0093] First, link-level features It is input into two parallel prediction heads: Regression head: This is a multilayer perceptron used to directly predict the standardized path loss value. The output is a scalar regression prediction. .

[0094] Classification head: This is a multilayer perceptron used to predict the category of path loss (e.g., whether it is extremely high loss). The output is the log-odds of classification. .

[0095] Secondly, during the training phase, the model simultaneously optimizes the regression loss (e.g., mean squared error) and the classification loss (e.g., binary cross-entropy loss or multi-class cross-entropy loss) to improve the accuracy and robustness of predictions.

[0096] Finally, during the inference phase, you can choose to use either regression predictions or classification results based on your specific needs. Regression predictions It will be done through denormalization (using data saved during the training phase). and Convert back to the original path loss scale, i.e. .

[0097] This embodiment employs a composite loss function and a refined training strategy to optimize the model.

[0098] Dual-head prediction mechanism: The model simultaneously outputs regression predictions of path loss and classification predictions indicating communication interruption.

[0099] Advanced hybrid loss function: This loss function includes: FocalLoss and class balancing weights are used for classification tasks to effectively handle the sparsity of interruption events and class imbalance.

[0100] For the weighted HuberLoss regression task, it innovatively introduces a static weighting mechanism (giving higher weights to "simple" samples with low path loss) and a dynamic weighting mechanism (giving extra weights to the Top-K% samples with the largest prediction errors), thereby guiding the model to learn common scenarios more accurately and correct extreme prediction biases.

[0101] Composite optimization objective and adaptive learning rate scheduling: The training process is driven by minimizing a composite optimization objective that combines the MAE of uninterrupted samples and the F1 score, and adopts an adaptive learning rate scheduler (such as `ReduceLROnPlateau`) and an early stopping mechanism to achieve model training stability and optimal performance.

[0102] This implementation describes in detail the design principles of the composite loss function and the entire process of model training.

[0103] Step 31: Initialization and configuration of the loss function.

[0104] First, the loss function receives a normalized statistic during initialization. (including mean path loss) and standard deviation Define the path loss threshold for communication interruption. (For example ), and the weight parameters of each loss (regression_weight, classification_weight).

[0105] Secondly, to address the characteristics of path loss distribution, this embodiment configures additional weighting parameters: focal_gamma for FocalLoss, top_k_percent and top_k_weight for hard samples in dynamic weighted regression loss, and innovatively introduces low_loss_weight and low_loss_threshold for "simple" samples with low static weighted path loss.

[0106] Finally, HuberLoss is used internally within the loss function as the basis for the regressor sub-loss to improve robustness to outliers.

[0107] Step 32: Calculate classification loss.

[0108] First, the model outputs the classification log odds. Compared with the target value of the original path loss The true interruption label of the judgment (when The value is 1 if it is true and 0 otherwise.

[0109] Secondly, this embodiment uses FocalLoss to address the common class imbalance problem and the excessive number of "easily classifiable" samples in classification tasks. FocalLoss is calculated as follows: ,in It is the probability that the model predicts the true class. It is the focal_gamma parameter.

[0110] For each sample in the batch Its classification loss It can be represented as:

[0111] Final classification loss For all samples in the batch Average value: ; in, The final calculated classification loss is the average of the weighted Focal Loss of all samples in the batch. This indicates the total number of samples in the current training batch. This indicates the index of the sample in the current batch. The model represents the samples The output is the log-odds of the inactive path loss classification. Indicates sample The actual binary interrupt tag, whose value is based on the original path loss target value. Judgment: When A value of 1 indicates an interrupt, while a value of 0 indicates no interruption. The model represents the samples The predicted probability of the positive class (interruption), i.e. . The model represents the samples The probability of predicting the true class. If (If the real category is interruption), then .if (If the real category is non-interrupted), then . This represents the gamma parameter of Focal Loss, used to control the degree of weighted suppression for easily classified samples. When At that time, the model will pay more attention to samples that are difficult to classify. Indicates sample The category balance weight, if (If the real category is non-interrupted), then ,if (If the real category is interruption), then . This indicates the number of non-disruptive samples in the current batch. This indicates the number of interrupted samples in the current batch.

[0112] Finally, to further balance the contributions of interrupted and uninterrupted samples, this embodiment also dynamically calculates a class balancing weight. ,in and These represent the number of non-disruptive and disruptive samples in the batch, respectively. This weight is applied to FocalLoss to form a balanced FocalLoss. This is to ensure that the model does not favor non-disruptive samples that have a larger number of predictions.

[0113] Step 33: Calculate regression loss.

[0114] First, only for non-disruptive samples (i.e. The regression loss is calculated using the sample data. The standardized regression prediction output by the model is then used to calculate the regression loss. Standardized real path loss target Compare them.

[0115] Secondly, this embodiment uses HuberLoss as the basic regression loss. It combines the sum of squares and the linear part of the prediction error, thereby reducing the sensitivity to outliers when the error is large.

[0116] Furthermore, to address the potential issue of insufficient generalization ability of the model for "simple" samples with low path loss, this invention innovatively introduces a static weighting mechanism: for the original path loss target... Below the preset threshold The sample is given an additional weight. This move aims to encourage models to learn more accurately common and simple propagation scenarios.

[0117] For all non-disruptive samples in the batch (i.e. (of the sample), its regression loss It can be represented as:

[0118] Final regression loss For all non-disruptive samples in the batch Average value:

[0119] Among them, the Huber Loss function Defined as:

[0120] In this implementation, default .

[0121] in, The final calculated regression loss is the average of the weighted Huber Loss for all non-disruptive samples in the batch. This indicates the total number of non-disruptive samples in the current training batch. This indicates the index of the non-interrupted sample in the current batch. The model represents the samples The output is the standardized regression prediction value. Indicates sample The standardized real path loss target value. Indicates sample The prediction error. Indicates sample The static weighting factor (i.e., `low_loss_weight` in the code), if the sample Original path loss target (i.e., considered a "simple" sample), then ;otherwise, . Define the path loss threshold (in dB) for "simple" samples. Give extra weight to “simple” samples. : Indicates a sample The dynamic weighting factor (i.e., `top_k_weight` in the code). If the sample Prediction error If the error belongs to one of the largest `top_k_percent` proportions in the batch (i.e., it is considered a "hard" sample), then ;otherwise, . Define the proportion of samples with the largest errors that should be focused on in regression loss (e.g., (Representing the top 20%). Give extra weight to the Top-K% of samples with the largest errors. These are parameters for Huber Loss, controlling the transition point between the squared error and linear error components. The default value is usually [value missing]. .

[0122] Finally, this embodiment also employs a dynamic weighting mechanism to focus on the most difficult samples: among the non-disruptive samples, the top_k_percent proportion of samples with the largest prediction errors are identified, and the HuberLoss of these samples is multiplied by an additional weight. This allows the model to dynamically focus on the links with the least accurate predictions, thereby correcting outlier predictions and improving overall accuracy. The final weighted HuberLoss is the regressor sub-loss. .

[0123] Step 34: Calculation and weighted fusion of total loss.

[0124] First, total losses Weighted classification loss and weighted regression loss Linear combination ,in and These are the regression loss weight and the classification loss weight, respectively. These two weights are usually floating-point values ​​set by the user in the training configuration, used to control the relative importance of the regression task and the classification task in the total loss function.

[0125] Finally, during the backpropagation process, the total loss is... It is used to calculate the gradient of the model parameters and update them. If any loss value is detected to be NaN or Inf, the current update will be stopped to avoid training failure.

[0126] The model training and optimization process in this embodiment aims to achieve the best balance between prediction accuracy and interruption recognition capability through iterative optimization.

[0127] Step 41: Environment preparation and data loading.

[0128] First, set a random seed to ensure the reproducibility of the experiment and specify the computing device (GPU) used for training.

[0129] Secondly, the training and validation datasets are loaded according to the configuration, and normalized statistics of node, edge features, and path losses are calculated and obtained based on the training set. These statistics are used for subsequent model standardization and loss calculation.

[0130] Finally, a data loader is created to organize the dataset into batches for training and evaluation. The training data loader uses a standard random shuffle strategy.

[0131] Step 42: Instantiate the model, optimizer, and scheduler.

[0132] First, instantiate the EnhancedPathLossGraphTransformer model based on the parameters defined in the configuration file (such as hidden dimension, number of attention heads, number of layers, whether to use RGB features, whether to fine-tune the RGB encoder, etc.).

[0133] Secondly, the optimizer is instantiated; this embodiment uses the AdamW optimizer. To achieve finer control, if the model is configured for fine-tuning the RGB encoder, a separate learning rate is set for the parameters of the RGB encoder, typically by multiplying an `rgb_lr_factor` by the base learning rate to achieve differentiated learning rate updates. The remaining model parameters are optimized using the base learning rate.

[0134] Finally, a learning rate scheduler is instantiated; this embodiment uses the ReduceLROnPlateau scheduler. This scheduler dynamically adjusts the learning rate based on changes in a predefined optimization objective (such as MAE or F1 score) on the validation set. When the optimization objective does not improve within a specified patience period, the learning rate will decrease by a factor.

[0135] Step 43: Single-round training (Epoch) process.

[0136] First, at the start of each training epoch, the model is set to training mode.

[0137] Secondly, iterate through each batch in the training data loader: The data is moved to the designated device (GPU).

[0138] The optimizer gradient is cleared to zero.

[0139] Perform forward propagation of the model to obtain regression predictions and classification log odds.

[0140] This invention supports the use of Automatic Mixed Precision (AMP) to perform forward propagation and loss calculation in the torch.cuda.amp.autocast context, thereby accelerating training and reducing GPU memory usage.

[0141] Calculate the total loss using the composite loss function. and its components.

[0142] Perform backpropagation and calculate the gradient.

[0143] Perform gradient clipping to prevent gradient explosion.

[0144] Update the model parameters.

[0145] Finally, the loss for each batch is accumulated and recorded to calculate the average training loss for this round.

[0146] Step 44: Model evaluation and optimization target update.

[0147] First, at each evaluation interval or at the end of training, the model is set to evaluation mode.

[0148] Secondly, iterate through all batches in the validation data loader, perform model forward propagation, and collect all prediction results (regression predictions, classification log odds) and true labels.

[0149] Next, based on the collected prediction results and true labels, a series of detailed performance metrics are calculated, including but not limited to: total mean absolute error (MAE), MAE of non-outage samples (mae_non_outage), root mean square error (RMSE), R2 score, standardized mean absolute error (NMAE), standardized root mean square error (NMSE), as well as classification accuracy, precision, recall and F1 score.

[0150] Then, this embodiment innovatively introduces a composite optimization objective. The goal is to balance regression accuracy and classification performance simultaneously. It is defined as follows: ,in It is the mean absolute error of the uninterrupted samples. It is the F1 score for interrupted classification. It is a configurable balance weight (W_f1_balance). The model will minimize... Optimize for the target.

[0151] Next, Passed to the learning rate scheduler to adjust the learning rate.

[0152] Finally, if the current If the value is better than the historical best value, the best model is updated and the early stopping counter is reset; otherwise, the early stopping counter is incremented. Training terminates early when the counter reaches early_stopping_patience.

[0153] Step 45: Best model saving and final evaluation.

[0154] First, during training, whenever the model achieves its best performance on the composite optimization objective on the validation set, information such as the model's parameter state (model_state_dict), optimizer state, and normalized statistics will be saved as a checkpoint file.

[0155] Finally, after training, load the saved best model and perform a final comprehensive evaluation on the validation set. Generate detailed performance reports and visualizations (such as scatter plots of predictions versus true values, error distribution histograms, etc.) to comprehensively analyze the model's performance and generalization ability.

[0156] In this embodiment, thanks to multimodal data fusion, a physically driven multidimensional edge feature system, and a hybrid attention graph neural network architecture that integrates visual semantics, the present invention achieves a prediction accuracy for high-frequency millimeter-wave path loss far exceeding that of traditional methods in complex urban environments. Especially in NLoS scenarios, by integrating real-time perceived visual semantics, it can detect transient occlusions not present on the map, thereby reducing prediction errors.

[0157] In this embodiment, by integrating dynamic environmental features and micro-path profile features, and combining them with real-time vehicle perception data, the model can effectively capture the drastic channel changes caused by the high-speed movement of the transmitter, rapid changes in surrounding obstructions, and close-range micro-obstruction effects, which greatly improves the model's prediction robustness in extreme dynamic and complex micro-environments.

[0158] In this embodiment, by constructing a strictly geographically isolated high-fidelity dataset and employing difficulty-aware sampling and composite optimization objectives, the model can maintain its performance level when migrated to a completely new geographical scene, demonstrating strong cross-scene generalization ability and providing a reliable guarantee for actual deployment.

[0159] In this embodiment, the model, through a classification prediction head trained in parallel, can accurately identify communication interruption events where path loss exceeds a specific threshold, providing key information for the vehicle-to-everything (V2X) system to perform resource scheduling and path planning in advance, thereby ensuring the reliability and security of communication services.

[0160] In this embodiment, node features are no longer pre-calculated with RGB embedding, but rather images are loaded and processed in real time. The modular heterogeneous graph construction method effectively reduces data storage requirements and improves the flexibility and efficiency of the data processing flow.

[0161] Example 2 The purpose of this embodiment is to provide a vehicle network path loss prediction system based on heterogeneous graph neural networks, including: The heterogeneous graph construction module is configured to: construct a heterogeneous graph using the vehicle and the selected receiver as nodes, and using the multidimensional physical driving features of the vehicle and the receiver, as well as the spatial correlation between receivers, as edges; wherein, the multidimensional physical driving features include the geometric physical features, static environmental features, microscopic path profile features, and dynamic environmental features of the vehicle and the receiver. The prediction module is configured to: use a trained multimodal hybrid attention graph neural network model to predict the heterogeneous graph and obtain the vehicle network path loss prediction result; The multimodal hybrid attention map neural network model processes the heterogeneous graph as follows: Extract the node features of the nodes and the edge features of the edges in the heterogeneous graph and embed them; A multi-layered heterogeneous graph Transformer layer is used to update node features, and the updated node features are extracted from different depths and fused to form fused node features. By embedding and concatenating node features and edge features, link-level features are learned through a feature interaction layer, and the link-level features are used to predict the path loss of the vehicle network.

[0162] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0163] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0164] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0165] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0166] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0167] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0168] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0169] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0170] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0171] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0172] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks, characterized in that, include: Using the vehicle and the selected receiver as nodes, a heterogeneous graph is constructed with the multidimensional physical driving characteristics of the vehicle and the receiver, as well as the spatial correlation between receivers, as edges; wherein, the multidimensional physical driving characteristics include the geometric physical characteristics, static environmental characteristics, micro-path profile characteristics, and dynamic environmental characteristics of the vehicle and the receiver. The heterogeneous graph is predicted using a trained multimodal hybrid attention graph neural network model to obtain the vehicle network path loss prediction result; The multimodal hybrid attention map neural network model processes the heterogeneous graph as follows: Extract the node features of the nodes and the edge features of the edges in the heterogeneous graph and embed them; A multi-layered heterogeneous graph Transformer layer is used to update node features, and the updated node features are extracted from different depths and fused to form fused node features. By embedding and concatenating node features and edge features, link-level features are learned through a feature interaction layer, and the link-level features are used to predict the path loss of the vehicle network.

2. The method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1, characterized in that, Using the vehicle and the selected receiver as nodes, and the multi-dimensional physical driving characteristics of the vehicle and the receiver, as well as the spatial correlation between receivers, a heterogeneous graph is constructed as edges, specifically: Using the vehicle and receiver as nodes, the converted two-dimensional pixel coordinates and corresponding binary identifiers are assigned to the vehicle node and receiver node, respectively. A heterogeneous graph is constructed using the multidimensional physical driving characteristics of the vehicle and the receiver as edges, and the spatial correlation between receivers as edges.

3. A method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1 or 2, characterized in that, The extraction of the multidimensional physical driving features specifically involves: Geometric physical features are constructed based on the physical distance, logarithmic distance, and communication frequency between the vehicle and the receiver; Static environment features are constructed based on the proportion of pixels covered by building and tree masks along the straight-line propagation path from the vehicle to the receiver in the pixel map. Determine the joint obstacle mask formed by building mask and tree mask on the straight propagation path from the vehicle to the receiver in the pixel map, calculate the first obstacle type and the physical length of the largest continuous occlusion on the straight propagation path, and construct the micro path profile features by combining the number of independent obstacles passed through the straight propagation path. Calculate the shortest distance from dynamic obstacles to the line segment connecting the vehicle and the receiver based on the real-time trajectory data of all other vehicles in the current snapshot, identify other vehicles within a specified radius of the line segment as dynamic obstructions, extract the number of dynamic obstructions, and construct dynamic environment features.

4. A method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1 or 2, characterized in that, Extracting the spatial correlation between receivers, specifically: For each receiver, selection is performed using K-nearest neighbor search. The nearest receiver is selected as the neighbor of the corresponding receiver, and a bidirectional connection edge is constructed. For each receiver-to-receiver edge, extract the physical distance and logarithmic distance between the receivers as the corresponding features.

5. The method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1, characterized in that, A multi-layered heterogeneous graph Transformer is used to update node features. Updated node features are extracted from different depths and fused to form fused node features. Specifically: In each heterogeneous graph Transformer layer, the features of the current node are processed through local graph attention mechanism and global self-attention mechanism, respectively. By using learnable fusion weight parameters, the local graph attention output and the global self-attention output are weighted and fused to obtain fused features; A gating mechanism is introduced, which uses the residual connection and fusion features of the current node features to generate a gating signal, and obtains the updated node features. After completing the message passing of the heterogeneous graph Transformer layer, node features are extracted from different depths to obtain shallow features and deep features respectively. The shallow features and deep features are then fused to obtain the final fused node features.

6. The method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1, characterized in that, The multimodal hybrid attention map neural network model is trained using a total loss function composed of a classification loss function and a regression loss function; The classification loss function adopts the FocalLoss loss function and introduces dynamically calculated class balance weights. The regression loss function adopts the HuberLoss loss function and introduces a static weighting mechanism and a dynamic weighting mechanism. The static weighting mechanism assigns extra weights to samples whose original path loss target is lower than a preset threshold. The dynamic weighting mechanism identifies the top n samples with the largest prediction errors among the non-interrupted samples and assigns them extra weights.

7. The method for predicting path loss in vehicle-to-everything (V2X) networks based on heterogeneous graph neural networks as described in claim 1, characterized in that, When optimizing and evaluating the trained multimodal hybrid attention map neural network model, a composite optimization objective that balances regression accuracy and classification performance is introduced. The composite optimization objective consists of the mean absolute error of uninterrupted samples and the F1 score of interrupted classification, so as to minimize the composite optimization objective to optimize the multimodal hybrid attention map neural network model.

8. A vehicle network path loss prediction system based on heterogeneous graph neural networks, characterized in that, include: The heterogeneous graph construction module is configured to: construct a heterogeneous graph using the vehicle and the selected receiver as nodes, and using the multidimensional physical driving features of the vehicle and the receiver, as well as the spatial correlation between receivers, as edges; wherein, the multidimensional physical driving features include the geometric physical features, static environmental features, microscopic path profile features, and dynamic environmental features of the vehicle and the receiver. The prediction module is configured to: use a trained multimodal hybrid attention graph neural network model to predict the heterogeneous graph and obtain the vehicle network path loss prediction result; The multimodal hybrid attention map neural network model processes the heterogeneous graph as follows: Extract the node features of the nodes and the edge features of the edges in the heterogeneous graph and embed them; A multi-layered heterogeneous graph Transformer layer is used to update node features, and the updated node features are extracted from different depths and fused to form fused node features. By embedding and concatenating node features and edge features, link-level features are learned through a feature interaction layer, and the link-level features are used to predict the path loss of the vehicle network.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-7.

Images