Model training and positioning method

Abstract

The invention discloses a model training and positioning method. The model training method comprises the following steps: acquiring signal strength indexes and position information of a plurality of distributed base stations; the distributed base stations are mapped into nodes in the graph structure, and signal strength indexes and position information are constructed to serve as feature representation of the nodes; fusing the features of the nodes in different modes to obtain fused feature representation of the nodes; aggregating the fused feature representation of the node and the fused feature representation of the neighbor node of the node to update the feature representation of the node; determining predicted position information of terminal equipment in the target cell according to the updated node features; and training the graph neural network model, and obtaining a trained graph neural network model under the condition that a preset stop condition is met. According to the method and the device, the technical problem that the position of the terminal equipment is difficult to accurately determine under a non-line-of-sight condition and a multipath propagation environment due to limited positioning precision and insufficient robustness of a related positioning technology under a complex environment is solved.

Description

Technical Field

[0001] This application relates to the fields of wireless communication and positioning technology, and more specifically, to a model training and positioning method. Background Technology

[0002] Relevant wireless positioning methods mainly include triangulation based on base station signals, RSSI (Received Signal Strength Indication) positioning, and fingerprint positioning. Taking 5G triangulation as an example, it measures the RSSI and Time of Arrival (TOA / TDOA) of the target device's signal using multiple base stations, and then uses a triangulation algorithm to calculate the target's location. However, these methods typically require the target device to establish a good line-of-sight connection with multiple base stations; otherwise, it will be affected by non-line-of-sight multipath attenuation, leading to a decrease in positioning accuracy. In addition, base station triangulation relies on accurate time synchronization and multi-site deployment, which results in high deployment costs in complex environments.

[0003] RSSI-based fingerprint localization methods require extensive prior measurements in the target area to build an environment-specific fingerprint database, and the use of machine learning models (such as deep neural networks) to compare and locate the fingerprint using real-time RSSI. Although deep learning fingerprint localization can effectively integrate multiple features, its drawbacks are also quite obvious: on the one hand, it requires a large amount of manual measurement data to build fingerprint maps, which is labor-intensive and costly; on the other hand, environmental changes (such as personnel movement or changes in furniture arrangement) can lead to changes in the signal environment, making the fingerprint model prone to failure and resulting in insufficient localization robustness.

[0004] In summary, the accuracy and robustness of relevant positioning technologies are limited by environmental complexity and linear model assumptions, making it difficult to fully explore the spatial correlation characteristics between multiple base stations.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a model training and localization method to at least solve the technical problem that it is difficult to accurately determine the location of terminal devices in non-line-of-sight and multipath propagation environments due to the limited positioning accuracy and insufficient robustness of related positioning technologies in complex environments.

[0007] According to one aspect of this application, a model training method is provided, comprising: an acquisition step: acquiring signal strength indicators and location information of multiple distributed base stations in a target cell; a graph construction step: mapping the distributed base stations as nodes in a graph structure, constructing signal strength indicators and location information as feature representations of the nodes, and determining the distance between any two distributed base stations corresponding to any two nodes as the edge attribute of the edge connecting any two nodes, wherein the signal strength indicators and location information are features of different modalities; a cross-modal fusion step: fusing the features of different modalities of the nodes in the graph structure based on a cross-modal attention mechanism of a graph neural network model to obtain the fused feature representation of the nodes; a feature aggregation step: aggregating the fused feature representation of the nodes with the fused feature representations of the nodes' neighboring nodes according to the edge attributes to update the feature representation of the nodes; a location prediction step: determining the predicted location information of terminal devices in the target cell based on the updated node features; and a model training step: training a graph neural network model using supervised learning, and obtaining a trained graph neural network model when a preset stopping condition is met.

[0008] Optionally, a cross-modal attention mechanism based on a graph neural network model fuses the features of nodes in different modalities in the graph structure to obtain a fused feature representation of the node. This includes: using a first linear transformation layer to linearly transform the signal strength index of the node to obtain a signal strength feature vector; using a second linear transformation layer to linearly transform the position information of the node to obtain a position coordinate feature vector; using the signal strength feature vector as a query vector and the position coordinate feature vector as a key vector and a value vector, calculating the dot product of the query vector and the key vector to obtain an attention score; normalizing the attention score to obtain cross-modal attention weights, and using the cross-modal attention weights to perform a weighted summation of the value vectors to obtain cross-modal fused features; and performing a residual connection between the cross-modal fused features and the signal strength feature vector, and processing the residual connection result through a nonlinear activation function to obtain a fused feature representation of the node.

[0009] Optionally, based on edge attributes, the fused feature representation of a node is aggregated with the fused feature representations of its neighboring nodes to update the node's feature representation. This includes: concatenating the fused feature representation of the node with the fused feature representations of its neighboring nodes to obtain a concatenated feature representation; calculating the attention coefficients between the node and its neighboring nodes using an attention calculation function; normalizing the attention coefficients to obtain normalized attention weights; weighted summing the fused feature representations of neighboring nodes using the normalized attention weights to obtain the neighbor aggregated features; and combining the neighbor aggregated features with the fused feature representation of the node, and processing the combination result using a nonlinear activation function to obtain the updated feature representation of the node.

[0010] Optionally, based on the updated node features, the predicted location information of the terminal device in the target cell is determined, including: performing dimensional transformation and feature compression on the updated feature representation of the node using a first fully connected layer to obtain the node feature representation; calculating the importance weight of the node feature representation using a second fully connected layer to obtain the initial contribution weight of the node to the positioning terminal device, wherein the output dimension of the second fully connected layer is a single-dimensional scalar; normalizing the initial contribution weight to obtain the contribution weight; weighting and summing the node's location information with the corresponding contribution weight to obtain the preliminary predicted location of the terminal device; and performing nonlinear transformation and coordinate mapping on the preliminary predicted location using a third fully connected layer to obtain the predicted location information of the terminal device, wherein the third fully connected layer has a two-dimensional output and includes a nonlinear activation function to map the preliminary location to the target coordinate space.

[0011] Optionally, distributed base stations are mapped to nodes in a graph structure. Signal strength indicators and location information are constructed as feature representations of nodes, and the distance between distributed base stations corresponding to any two nodes is determined as an edge attribute of the edge connecting the two nodes. This includes: creating node identifiers corresponding to distributed base stations and establishing a mapping relationship from distributed base stations to the graph structure based on the node identifiers; constructing a signal strength feature vector based on the preprocessed reference signal received power value and constructing a location coordinate vector based on the two-dimensional coordinates of the distributed base stations; concatenating the signal strength feature vector and the location coordinate vector to obtain the feature representation of the node; and based on the mapping relationship, identifying two distributed base stations corresponding to two nodes connected by an undirected edge and calculating the Euclidean distance between the two distributed base stations, determining the Euclidean distance as an edge attribute of the undirected edge.

[0012] Optionally, a supervised learning approach is used to train the graph neural network model. Upon meeting a preset stopping condition, a fully trained graph neural network model is obtained. This includes: calculating the mean squared error loss between the predicted location information and the corresponding real location information; calculating the gradient of the trainable parameters of the graph neural network model based on the mean squared error loss, and updating the trainable parameters of the graph neural network model using an optimizer based on the gradient; evaluating the model performance in each training epoch on a validation dataset and recording the validation loss value; dynamically adjusting the learner's learning rate parameter according to a learning rate decay strategy; determining that the model performance of the graph neural network model will no longer improve if the validation loss value does not decrease in multiple consecutive training epochs; and terminating the training process to obtain a fully trained graph neural network model.

[0013] Optionally, the learning rate parameter of the optimizer is dynamically adjusted according to the learning rate decay strategy, including: initializing the initial learning rate parameter of the optimizer and obtaining the learning rate decay trigger condition; after each training epoch is completed, obtaining the current training epoch number and the corresponding validation loss value; based on the current training epoch number and validation loss value, determining whether the learning rate decay trigger condition is met; if the learning rate decay trigger condition is met, calculating the decay of the current learning rate parameter according to the preset learning rate decay strategy to obtain the updated learning rate parameter.

[0014] According to another aspect of this application, a positioning method is also provided, comprising: acquiring signal strength indicators and location information of multiple distributed base stations in a target cell; processing the signal strength indicators and location information of the multiple distributed base stations using a graph neural network model to obtain predicted location information of a terminal device in the target cell, wherein the graph neural network model is obtained by training using the above-mentioned model training method.

[0015] According to another aspect of this application, a model training apparatus is also provided, comprising: an acquisition module for acquiring signal strength indicators and location information of multiple distributed base stations in a target cell; a graph construction module for mapping the distributed base stations as nodes in a graph structure, constructing signal strength indicators and location information as feature representations of the nodes, and determining the distance between any two distributed base stations corresponding to any two nodes as the edge attribute of the edge connecting the two nodes, wherein the signal strength indicators and location information are features of different modalities; a fusion module for fusing the features of different modalities of the nodes in the graph structure based on the cross-modal attention mechanism of the graph neural network model to obtain the fused feature representation of the nodes; an aggregation module for aggregating the fused feature representation of the nodes with the fused feature representations of the nodes' neighboring nodes according to the edge attribute to update the feature representation of the nodes; a prediction module for determining the predicted location information of terminal devices in the target cell based on the updated node features; and a training module for training the graph neural network model using supervised learning, and obtaining the trained graph neural network model when a preset stopping condition is met.

[0016] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above model training method.

[0017] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program executes the above-described model training method during runtime.

[0018] According to another aspect of this application, a computer program is also provided, wherein the computer program, when executed by a processor, implements the above-described model training method.

[0019] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, which implements the above-described model training method when executed by a processor.

[0020] In this application, the following steps are employed: Acquisition Step: Acquiring signal strength indicators and location information of multiple distributed base stations in the target cell; Graph Construction Step: Mapping distributed base stations as nodes in a graph structure, constructing signal strength indicators and location information as feature representations of nodes, and determining the distance between any two distributed base stations as the edge attribute connecting any two nodes, where signal strength indicators and location information are features of different modalities; Cross-Modal Fusion Step: Fusion of features of different modalities of nodes in the graph structure based on a cross-modal attention mechanism using a graph neural network model to obtain the fused feature representation of the node; Feature Aggregation Step: Based on the edge attributes, merging the fused feature representation of the node with the fused feature representations of the node's neighboring nodes. The algorithm performs aggregation to update the feature representation of nodes; the location prediction step: based on the updated node features, the predicted location information of the terminal device in the target cell is determined; the model training step: a graph neural network model is trained using supervised learning, and the trained graph neural network model is obtained when the preset stopping conditions are met. By combining deep learning algorithms with graph structures, the positioning accuracy and robustness are improved, thereby achieving the technical effect of accurately determining the location of terminal devices under complex and changing environmental conditions. This solves the technical problem that it is difficult to accurately determine the location of terminal devices under non-line-of-sight and multipath propagation environments due to the limited positioning accuracy and insufficient robustness of related positioning technologies in complex environments. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 This is a flowchart of a model training method according to an embodiment of this application;

[0023] Figure 2 This is a flowchart of another model training method according to an embodiment of this application;

[0024] Figure 3 This is a flowchart of a data acquisition and preprocessing method according to an embodiment of this application;

[0025] Figure 4 This is an architecture diagram of an artificial intelligence model according to an embodiment of this application;

[0026] Figure 5 This is a structural diagram of a model training device according to an embodiment of this application;

[0027] Figure 6 This is a hardware structure block diagram of a computer terminal for a model training method according to an embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] According to an embodiment of this application, a method embodiment for model training is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0031] Figure 1 This is a flowchart of a model training method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0032] Step S101, Acquisition Step: Acquire the signal strength indicators and location information of multiple distributed base stations in the target cell.

[0033] Distributed base stations include, but are not limited to: distributed radio frequency units. Signal strength indicators include, but are not limited to: reference signal received power (RSRP). Location information includes, but is not limited to: two-dimensional coordinates in the real world.

[0034] Step S102, Graph Construction Step: Map the distributed base stations as nodes in the graph structure, construct signal strength index and location information as feature representations of the nodes, and determine the distance between any two distributed base stations as the edge attribute of the edge connecting any two nodes, where signal strength index and location information are features of different modes.

[0035] In this process, signal strength indicators and location information are constructed as feature representations of nodes, which means that signal strength indicators and location information are determined as feature representations of nodes.

[0036] In step S102, each distributed base station is converted and mapped to an independent node in the graph structure. The signal strength index and the location information of each base station are set as the feature representation of the created node. Here, the signal strength index and location information are the signal mode characteristics and location mode characteristics of the node, respectively, which together constitute the node's attribute set. During the graph structure construction process, the distance value between the distributed base stations corresponding to two nodes is calculated and determined. The distance value is an attribute of the edge connecting these two nodes; that is, the edge attribute is equal to the Euclidean distance between the coordinates of the two base stations. By assigning specific distance values ​​to the edges, the graph structure can intuitively and quantitatively reflect the spatial layout and interrelationships of the distributed base stations, injecting spatial dimension knowledge into the learning process of the graph neural network.

[0037] Step S103, cross-modal fusion step: The cross-modal attention mechanism based on the graph neural network model fuses the features of different modalities of nodes in the graph structure to obtain the fused feature representation of the nodes.

[0038] In step S103, the goal of the cross-modal attention mechanism is to process the dual-modal features carried by each node in the graph structure: signal strength mode and location information mode. Specifically, the cross-modal attention mechanism extracts and analyzes the signal strength mode and location information mode on the node respectively, and realizes dynamic matching and weighted fusion of signal and spatial information by calculating attention weights. The core of the fusion process is that the cross-modal attention mechanism can adaptively adjust and balance the contribution of different modal information, ensuring that the node feature representation includes both accurate signal strength details and relevant spatial location clues. Through the above mechanism, the flexibility of deep learning and the spatial perception capability of graph neural networks are combined, and the generated node fusion feature representation is richer and more comprehensive, which can better characterize the complex signal propagation characteristics in the base station environment and provide high-quality feature input for subsequent positioning tasks. After processing by the cross-modal attention mechanism, the feature representation of each node is upgraded, including not only the original signal strength mode and location information mode, but also the high-level features generated by intelligent fusion between the two modes, thereby improving the performance and accuracy of graph neural networks for positioning tasks under multipath effects and non-line-of-sight propagation challenges.

[0039] Step S104, Feature aggregation step: Based on the edge attributes, the fused feature representation of the node is aggregated with the fused feature representation of the node's neighboring nodes to update the node's feature representation.

[0040] In step S104, the distance between the distributed base stations corresponding to the nodes is determined as the weighting criterion for aggregating node features. Specifically, the fused feature representation of each node and the corresponding feature representations of all its neighboring nodes are aggregated, where neighboring nodes refer to other nodes in the graph that are directly connected to the current node through edges. The core of the aggregation process is to adjust the importance of neighboring node features based on edge attributes (i.e., distance), ensuring that the features of nodes that are closer in distance occupy a larger weight in the aggregation, thereby reflecting the impact of spatial proximity on signal propagation and positioning accuracy. The above strategy fully considers the spatial relationships between nodes, making information fusion more consistent with the signal propagation laws of the physical world and improving the accuracy of the positioning model.

[0041] Furthermore, based on the aggregated node feature update mechanism, the feature representation of each node is iteratively updated. Through multiple rounds of aggregation and updating, the feature representation of each node gradually integrates information from all nodes across the entire graph, forming a more comprehensive and accurate description of the signal environment features.

[0042] Step S105, Location Prediction Step: Based on the updated node characteristics, determine the predicted location information of the terminal devices in the target cell.

[0043] In step S105, the updated node features are input into the fully connected regression layer of the graph neural network for processing. The fully connected regression layer decodes the updated features, converting them into two-dimensional location information predicted for the target terminal device, thus achieving intelligent prediction of the terminal device's coordinates. Through deep learning of multimodal fusion features using the graph neural network, the combined information of signal strength and spatial location can be effectively utilized to output more accurate positioning results.

[0044] Step S106, Model Training Step: Train the graph neural network model using supervised learning. Under the condition that the preset stopping condition is met, the trained graph neural network model is obtained.

[0045] When the model training meets the preset stopping conditions, such as reaching the maximum number of iterations or the loss on the validation set falling below a certain threshold, the training process is declared complete. At this point, the trained graph neural network model can effectively utilize signal strength indicators and spatial location information to output accurate predictions of the terminal device's location. The trained model possesses higher positioning accuracy and environmental robustness, enabling stable operation in complex indoor and outdoor scenarios, providing reliable positioning services for terminal devices.

[0046] The above steps involve: an acquisition step, obtaining signal strength indicators and location information of multiple distributed base stations in the target cell; a graph construction step, mapping the distributed base stations as nodes in a graph structure, constructing signal strength indicators and location information as feature representations of the nodes, and determining the distance between any two distributed base stations as the edge attribute connecting the two nodes, where signal strength indicators and location information are features of different modalities; a cross-modal fusion step, fusing the features of different modalities of nodes in the graph structure based on a cross-modal attention mechanism of a graph neural network model to obtain the fused feature representation of the nodes; a feature aggregation step, aggregating the fused feature representation of the nodes with the fused feature representations of the nodes' neighboring nodes according to the edge attributes to update the node's feature representation; a location prediction step, determining the predicted location information of the terminal device in the target cell based on the updated node features; and a model training step, training the graph neural network model using supervised learning, obtaining the trained graph neural network model when a preset stopping condition is met. By combining deep learning algorithms with graph structures, the positioning accuracy and robustness are improved, thus achieving the technical effect of accurately determining the location of terminal devices under complex and changing environmental conditions.

[0047] The following are Figure 1 The steps shown are illustrated and explained by way of example.

[0048] According to some optional embodiments of this application, the cross-modal attention mechanism based on a graph neural network model fuses the features of different modalities of nodes in a graph structure to obtain the fused feature representation of the nodes. This can be achieved by the following method: A first linear transformation layer is used to linearly transform the signal strength index of the nodes to obtain a signal strength feature vector; a second linear transformation layer is used to linearly transform the position information of the nodes to obtain a position coordinate feature vector; using the signal strength feature vector as the query vector and the position coordinate feature vector as the key vector and value vector, the dot product of the query vector and the key vector is calculated to obtain an attention score; the attention score is normalized to obtain cross-modal attention weights, and the value vectors are weighted and summed using the cross-modal attention weights to obtain cross-modal fused features; the cross-modal fused features are residually connected to the signal strength feature vector, and the residual connection result is processed by a nonlinear activation function to obtain the fused feature representation of the nodes.

[0049] In this embodiment, the first step is to apply a first linear transformation layer to the signal strength index of the node. The task of this linear transformation is to convert the original RSRP signal strength data into a signal strength feature vector, facilitating subsequent processing and feature fusion in the neural network layers. The same linear transformation operation is then performed on the node's location information using a second linear transformation layer. The aim is to convert the coordinate information into a location coordinate feature vector, enabling the location information to participate in the subsequent cross-modal attention mechanism in a structured vector form.

[0050] Secondly, the signal strength feature vector is used as the query vector, and the location coordinate feature vector is used as the key and value vectors. The attention score is determined by calculating the dot product between the query vector and the key vector. The attention score measures the degree of correlation between the signal strength features and the location coordinate features.

[0051] The attention scores are then normalized, specifically by using the softmax function to ensure that the sum of all attention scores on the key vector of a node equals 1. The generated normalized attention weights can represent the degree of importance that signal strength features place on different dimensions of location information, providing dynamic weight allocation for cross-modal feature fusion.

[0052] The position coordinate feature vector, i.e., the value vector, is then weighted and summed again using weights. This weighted summation operation is used to generate cross-modal fusion features. The modal fusion features integrate signal strength and spatial location information, and intelligent weighting and fusion of different modal features are achieved through an attention mechanism.

[0053] Finally, the cross-modal fused features are residually concatenated with the original signal intensity feature vector. Adding fused features enhances the representational power of the signal intensity feature vector while preserving the original signal intensity information and preventing information loss. Nonlinear activation functions (such as ReLU or SiLU) are used to process the residual concatenation results to ensure nonlinear transformation of node feature representations within the deep learning model, thereby improving the complexity of feature representation and the model's localization accuracy.

[0054] According to some alternative embodiments of this application, the fused feature representation of a node is aggregated with the fused feature representations of its neighboring nodes based on edge attributes to update the node's feature representation. This can be achieved through the following method: concatenating the fused feature representation of the node with the fused feature representations of its neighboring nodes to obtain a concatenated feature representation; calculating the attention coefficients between the node and its neighboring nodes using an attention calculation function; normalizing the attention coefficients to obtain normalized attention weights; weighting the fused feature representations of neighboring nodes using the normalized attention weights to obtain the neighbor aggregated features; and combining the neighbor aggregated features with the fused feature representation of the node, and processing the combination result using a nonlinear activation function to obtain the updated feature representation of the node.

[0055] In this embodiment, the fused feature representation of a node is first concatenated with the fused feature representations of all its neighboring nodes. This concatenation operation forms the concatenated feature representation, which includes both the node's own depth features and features from the surrounding environment. Next, an attention calculation function is introduced to compute the concatenated feature representation and the edge attributes of the edges connecting the node and its neighbors. The purpose of this calculation is to quantify the degree of attention a node pays to the features of different neighboring nodes, thus obtaining the attention coefficient between the node and its neighbors.

[0056] Then, the attention coefficients are normalized by using the softmax function to normalize the attention coefficient matrix of each node, generating normalized attention weights. These normalized attention weights ensure that nodes can allocate resources reasonably when aggregating neighbor features, giving greater weight to neighbors that are closer or whose features are more relevant, thereby improving the effectiveness and relevance of feature aggregation.

[0057] We use normalized attention weights to perform a weighted summation of the fused feature representations of neighbor nodes. The summation operation produces a neighbor aggregate feature, which is a summary of a node's information about its surrounding environment, where the contribution ratio of each neighbor node is determined by the normalized attention weights.

[0058] Finally, the fused feature representations of neighbor aggregation and the node itself are combined, which can be done by concatenation. A non-linear activation function is then used to process the combined features to increase their expressive power and the model's non-linear learning ability. The output of the non-linear activation function is the updated feature representation of the node, which integrates comprehensive information from the node itself and its surrounding environment, providing richer and more accurate feature support for location prediction.

[0059] In some optional embodiments of this application, determining the predicted location information of a terminal device in a target cell based on the updated node features can be achieved through the following methods: A first fully connected layer is used to perform dimensional transformation and feature compression on the updated feature representation of the node to obtain a node feature representation; a second fully connected layer is used to calculate the importance weight of the node feature representation to obtain the initial contribution weight of the node to the positioning terminal device, wherein the output dimension of the second fully connected layer is a single-dimensional scalar; the initial contribution weight is normalized to obtain the contribution weight; the location information of the node and the corresponding contribution weight are weighted and summed to obtain the preliminary predicted location of the terminal device; a third fully connected layer is used to perform nonlinear transformation and coordinate mapping on the preliminary predicted location to obtain the predicted location information of the terminal device, wherein the third fully connected layer has a two-dimensional output and includes a nonlinear activation function to map the preliminary location to the target coordinate space.

[0060] In this embodiment, a first fully connected layer is first applied to process the updated feature representation of the nodes, aiming to extract the core information of the node feature representation through dimensionality transformation and feature compression. The first fully connected layer can capture the complex relationships within the node features, while compressing the feature dimensions to reduce computational complexity, making the node feature representation more compact and information-rich.

[0061] Secondly, the second fully connected layer is used to calculate the importance weights of the compressed node feature representations. The core of the importance weight calculation is to determine the initial contribution weight of each node to the positioning terminal device. The output of the second fully connected layer is a single-dimensional scalar, which means that each node will receive a value representing its influence on the estimated location of the terminal device. Through such weight calculation, the relative importance of each node in the positioning process can be quantified, providing a basis for subsequent weighted summation.

[0062] The weights are then normalized, ensuring that the sum of the contribution weights of all nodes equals 1. This normalization result, the contribution weight, ensures a balanced allocation of resources during neighbor node feature aggregation, preventing any single node's features from excessively dominating location prediction and thus improving the fairness and accuracy of the localization process.

[0063] The normalized contribution weights are then weighted and summed with the node location information to obtain the preliminary predicted location of the terminal device. This summation process is essentially a weighted average, where the weights reflect the correlation between node features and the localization task, and the location information reflects the actual coordinates of the nodes in the spatial coordinate system. By weighting the node location information with contribution weights, the contributions of all nodes can be comprehensively considered.

[0064] Finally, a third fully connected layer is used to perform a nonlinear transformation and coordinate mapping on the preliminary predicted location. This third fully connected layer has a two-dimensional output, including a nonlinear activation function, such as sigmoid or tanh, used to map the preliminary predicted location to the target coordinate space, i.e., the actual geographic coordinate system of the terminal device. This mapping process not only adjusts the numerical range of the preliminary location prediction, making it conform to the scale of the target coordinate space, but also enhances the nonlinear relationship between the predicted coordinates and node features through nonlinear transformation, making the predicted location closer to reality and improving the accuracy and reliability of the positioning.

[0065] As some optional embodiments of this application, the distributed base stations are mapped as nodes in a graph structure, signal strength indicators and location information are constructed as feature representations of the nodes, and the distance between any two distributed base stations corresponding to any two nodes is determined as the edge attribute of the edge connecting any two nodes. This can be achieved by the following method: creating node identifiers corresponding to the distributed base stations, and establishing a mapping relationship from the distributed base stations to the graph structure based on the node identifiers; constructing a signal strength feature vector based on the preprocessed reference signal received power value, and constructing a location coordinate vector based on the two-dimensional coordinates of the distributed base stations; concatenating the signal strength feature vector and the location coordinate vector to obtain the feature representation of the node; based on the mapping relationship, identifying the two distributed base stations corresponding to the two nodes connected by the undirected edge, calculating the Euclidean distance between the two distributed base stations, and determining the Euclidean distance as the edge attribute of the undirected edge.

[0066] In this embodiment, a unique node identifier must first be created and assigned to each distributed base station to establish a one-to-one mapping relationship between physical base stations and nodes in the graph structure. This mapping relationship ensures that the characteristics and information of each base station are accurately reflected on the corresponding nodes in the graph.

[0067] Secondly, a signal strength feature vector is constructed based on the preprocessed Reference Signal Received Power (RSRP) value. This feature vector includes the signal strength information received by the base station and is a crucial indicator of signal quality in the positioning model. Simultaneously, a location coordinate vector is constructed based on the two-dimensional coordinates of each distributed base station.

[0068] The signal strength feature vector and the location coordinate vector are then concatenated to generate the feature representation of the node. The concatenation operation integrates the base station's signal strength and physical location information into a single feature vector, ensuring that the graph neural network considers both signal quality and spatial location when processing node features.

[0069] Finally, based on the previously established mapping relationship, the two nodes connected by any two undirected edges are identified, and the two distributed base stations represented by these nodes are determined. The Euclidean distance between these two base stations is calculated. This Euclidean distance is a direct metric derived from the base station coordinate information and reflects the spatial proximity between the base stations. The Euclidean distance is assigned as an edge attribute of this undirected edge to characterize the spatial relationship between nodes. This edge attribute enables the graph neural network to learn distance-based association features between base stations, thereby improving the performance and accuracy of the localization model.

[0070] In some optional embodiments of this application, a supervised learning method is used to train the graph neural network model. The trained graph neural network model is obtained when a preset stopping condition is met. This can be achieved through the following methods: calculating the mean squared error loss between the predicted location information and the corresponding real location information; calculating the gradient of the trainable parameters of the graph neural network model based on the mean squared error loss value, and updating the trainable parameters of the graph neural network model using an optimizer according to the gradient; evaluating the model performance of the training epochs on a validation dataset and recording the validation loss value; dynamically adjusting the learner's learning rate parameter according to a learning rate decay strategy; determining that the model performance of the graph neural network model will no longer improve when the validation loss value has not decreased for multiple consecutive training epochs; terminating the training process when the model performance of the graph neural network model will no longer improve, thus obtaining the trained graph neural network model.

[0071] Specifically, dynamically adjusting the learner's learning rate parameter according to the learning rate decay strategy can be achieved as follows: initialize the optimizer's initial learning rate parameter and obtain the learning rate decay trigger condition; after each training epoch, obtain the current training epoch number and the corresponding validation loss value; based on the current training epoch number and validation loss value, determine whether the learning rate decay trigger condition is met; if the learning rate decay trigger condition is met, perform decay calculation on the current learning rate parameter according to the preset learning rate decay strategy to obtain the updated learning rate parameter.

[0072] In this embodiment, the error between the model-predicted terminal device location information and the actual location information is first calculated. This error can be measured using the mean squared error (MSE) loss value. The MSE quantifies the average squared distance between the model-predicted location and the actual location, providing direct feedback on the model's positioning accuracy. Specifically, for each training sample, the model-predicted two-dimensional coordinate position is subtracted from the corresponding known actual location to obtain an error vector. The sum of the squares of this error vector is then calculated, and finally averaged to obtain the MSE loss value.

[0073] Then, based on the calculated mean squared error loss value, the gradients of all trainable parameters (such as the weight matrix and bias vector) in the graph neural network are calculated using the backpropagation algorithm. The gradient is a multi-dimensional vector that indicates the direction and rate of change of the loss function at the current parameter value, providing guidance for parameter adjustment. The optimizer (e.g., Adam, SGD) uses this gradient information to update the model parameters according to a preset learning rate parameter, aiming to minimize the mean squared error loss value, i.e., improve the model's localization accuracy.

[0074] During training, in addition to training the model on the training set, it is also necessary to evaluate the model's performance on a separate validation dataset to ensure its generalization ability. For example, this can be done after each training epoch by calculating the mean squared error loss of the model on the validation set, recording and monitoring changes in the validation loss. Recording the validation loss helps determine whether the model is overfitting the training data and whether the model's performance is improving.

[0075] To accelerate model convergence and prevent overfitting, this embodiment introduces a learning rate decay strategy. This strategy dynamically adjusts the optimizer's learning rate parameter based on the training progress, gradually decreasing it from a higher initial value to a more stable value, enabling the model to converge to the optimal state faster and more smoothly. Learning rate decay methods include exponential decay and step decay, with the appropriate decay mode selected based on the specific training conditions.

[0076] During training, the trend of the validation loss value is monitored. When the validation loss value no longer decreases significantly after several consecutive training epochs, it means that the model's performance on the validation set has reached saturation and there is no longer significant room for improvement. At this point, the training process is terminated to avoid unnecessary waste of computational resources and the risk of overfitting. The resulting trained graph neural network model is the version with the best performance on the validation set.

[0077] This application also provides another model training method, which specifically includes the following steps:

[0078] Step S1: Collect RSRP values ​​for multiple channels from multiple distributed radio units (RUs) and combine them with the known two-dimensional coordinate information of each RU. Normalize the collected data and filter out invalid or outlier values ​​(such as signals with RSRP=0). Then, construct a graph structure from the processed RSRP data and the spatial coordinates of the corresponding RUs, with each RU corresponding to a node in the graph. Node characteristics include the RSRP values ​​of all channels under that RU and the coordinate information of that RU. A fully connected graph is constructed so that there is an edge between any two RU nodes; this edge attribute can be set based on the Euclidean distance between RUs (i.e., edge attribute = distance) to reflect the spatial relationship.

[0079] Specifically, such as Figure 2 and Figure 3 As shown, multiple distributed radio frequency units (RF units) are first deployed in the target area, and the multi-channel RSRP value of the signal received by each RU from the target device is measured. Simultaneously, the position of each RU in a two-dimensional coordinate system is acquired. After data acquisition, preprocessing is performed: the RSRP vector of each RU is normalized; invalid values ​​are removed; and valid RSRP values ​​are combined with the corresponding RU coordinates to form node features. Graph construction: Each RU is treated as a node in the graph, and an edge connects any two nodes, with the edge attribute being the distance between the two RUs (which can be pre-calculated). This results in a complete graph with attributes, where node features include signal mode (RSRP) and location information modes.

[0080] Step S2 employs a graph neural network (GNN) as the base model. For each node in the graph, its RSRP features and location information are input into the network as two modalities. A cross-modal attention mechanism is introduced to extract features from the RSRP signal modality and coordinate modality respectively, and the information from different modalities is dynamically fused using attention weights. This effectively aligns RSRP and spatial location information, improving feature representation capabilities. After fusion, GNN layers such as graph attention layers are used to aggregate information about nodes and their neighbors, further mining the spatial correlation features of the entire graph. Finally, after passing through a fully connected regression layer, the output of the graph neural network is mapped to the two-dimensional location information (such as latitude and longitude coordinates) of the target device.

[0081] When using AI modeling, such as Figure 4 As shown, preliminary feature extraction for the RSRP and coordinate modes is first performed through several fully connected or convolutional layers, and then fed into a cross-modal attention layer. In this layer, attention weight matrices from the RSRP and coordinate modes are calculated, and the modal information is weighted and fused. The fused features are then used as node features in the graph neural network and input into the graph attention network layer. In the GAT layer, the features of each node and all its neighboring nodes are aggregated and updated according to the edge weights (inverse distance) in the graph, iterating layer by layer to obtain the final node representation. Finally, the representations of all nodes are summarized in a fully connected regression layer, and the predicted 2D coordinates of the target device are decoded through the output layer. To improve positioning accuracy, strategies such as residual connections and multi-layer fusion can be introduced to enhance feature representation.

[0082] Step S3: Train the above graph neural network model using supervised learning. Mean squared error (MSE) is used as the loss function to evaluate the deviation between the predicted and actual positions. Simultaneously, optimization strategies such as dynamic learning rate decay and early stopping mechanisms are employed to accelerate convergence and prevent overfitting. The model obtained after offline training can output the two-dimensional coordinate position of the target device.

[0083] During the model training phase, supervised training was performed using collected multi-scene, multi-location measurement datasets. A mean squared error loss function was set to minimize the Euclidean distance difference between the predicted and calibrated locations. The Adam optimizer was used during training, with an initial learning rate of 0.001, which was gradually decreased as the validation error did not decrease. An early stopping mechanism was employed to prevent overfitting. After training, the model's localization error was evaluated. The results show that, compared to traditional triangulation and conventional fingerprint algorithms, the method provided in this application can significantly reduce localization error in complex multipath environments.

[0084] In summary, the above steps have the following advantages compared to related technologies:

[0085] Enhanced feature extraction and modality alignment capabilities: By introducing a cross-modal attention mechanism, information is extracted from two different modalities—RSRP signal strength and base station coordinates—and dynamically weighted and fused. This mechanism enables the model to adaptively adjust the attention allocation between different modalities, thereby more accurately aligning signal features with spatial features and improving positioning accuracy.

[0086] This invention utilizes a graph structure to simulate the spatial relationships between RUs: Multiple RUs are treated as nodes in a graph, with pairwise relationships described by fully connected edges. Edge attributes are set based on geometric distance, fully leveraging the ability of GNNs to automatically extract node and edge features. This graph structure more intuitively expresses the geometric relationships between multiple base stations / RF units, enabling the positioning model to comprehensively consider the joint information of all RUs, rather than a simple binary comparison, thereby improving the overall system performance.

[0087] Flexible handling of complex multipath attenuation features: Compared to traditional fingerprint and RSSI positioning which only use signal strength, the model of this invention combines deep learning with graph structures, making it more adaptable to the attenuation patterns of signals from multiple base stations. Whether in line-of-sight (LOS) or non-line-of-sight (NLOS) environments, the system can learn the statistical features of multipath propagation, achieving robust positioning in complex scenes.

[0088] This application also provides a positioning method, which specifically includes the following steps:

[0089] The signal strength indicators and location information of multiple distributed base stations in the target cell are obtained; a graph neural network model is used to process the signal strength indicators and location information of the multiple distributed base stations to obtain the predicted location information of terminal devices in the target cell. The graph neural network model is obtained through... Figure 1 The model is obtained by training using the method shown.

[0090] To achieve high-precision positioning of terminal devices within a target cell, signal strength metrics (such as RSRP values) and location information must first be collected from multiple distributed base stations deployed within the cell. Signal strength metrics represent the wireless communication quality between the terminal device and each RU (Radio Router), including information such as the distance from the terminal device to the base station, path loss, and potential multipath effects. Simultaneously, the location information of the base stations, i.e., their coordinates in two-dimensional or three-dimensional space, provides a static spatial context for the positioning model, helping the model understand the geometry and layout between base stations.

[0091] Then, a graph neural network model is used for deep processing of signal strength indicators and location information. The graph neural network model has already been used... Figure 1 The model training method shown is pre-trained to learn and optimize the ability to predict the location of terminal devices from signal strength indicators and location information. The graph neural network model makes full use of node (base station) features and edge (distance between nodes) attributes. By introducing cross-modal attention mechanisms and graph attention layers, it can effectively extract the correlation features between signal strength and spatial location, as well as the spatial relationship features between base stations, thereby improving the accuracy and robustness of location prediction.

[0092] Specifically, the processing flow of a graph neural network model involves extracting deep features from node features and edge attributes, dynamically fusing signal strength indicators and location information through a cross-modal attention mechanism, and then aggregating the feature information of nodes and their neighboring nodes using network structures such as graph attention layers. The ultimate goal of this series of operations is to generate the predicted location information of the terminal device at the output layer, such as two-dimensional coordinates (x, y).

[0093] It is worth noting that the training process of the graph neural network model is an offline process, employing a large amount of historical data and supervised learning algorithms to adjust the model parameters, enabling the model to accurately predict the location of the terminal device given signal strength and base station location. However, in the location prediction stage, i.e., the process of estimating the actual location of the terminal device, the model acts as an online application, directly processing real-time signal strength data and base station location information to produce immediate location results. This process ensures the real-time performance and accuracy of the positioning method, enabling it to provide reliable positioning services for terminal devices in complex and ever-changing wireless communication environments.

[0094] Figure 5 This is a structural diagram of a model training device according to an embodiment of this application, such as... Figure 5 As shown, the device includes:

[0095] The acquisition module 51 is used to acquire the signal strength indicators and location information of multiple distributed base stations in the target cell.

[0096] Graph construction module 52 is used to map distributed base stations as nodes in a graph structure, construct signal strength indicators and location information as feature representations of nodes, and determine the distance between distributed base stations corresponding to any two nodes as the edge attribute of the edge connecting any two nodes, wherein the signal strength indicators and location information are features of different modes.

[0097] The fusion module 53 is used to fuse the features of nodes in different modalities in the graph structure based on the cross-modal attention mechanism of the graph neural network model, so as to obtain the fused feature representation of the node.

[0098] The aggregation module 54 is used to aggregate the fused feature representation of a node with the fused feature representation of its neighboring nodes based on the edge attributes, so as to update the feature representation of the node.

[0099] The prediction module 55 is used to determine the predicted location information of terminal devices in the target cell based on the updated node characteristics.

[0100] Training module 56 is used to train a graph neural network model using supervised learning, and obtains a trained graph neural network model when a preset stopping condition is met.

[0101] Optionally, a cross-modal attention mechanism based on a graph neural network model fuses the features of different modalities of nodes in the graph structure to obtain the fused feature representation of the nodes. Specifically, this includes the following steps: A first linear transformation layer is used to linearly transform the signal strength index of the nodes to obtain a signal strength feature vector; a second linear transformation layer is used to linearly transform the position information of the nodes to obtain a position coordinate feature vector; using the signal strength feature vector as the query vector and the position coordinate feature vector as the key and value vectors, the dot product of the query vector and the key vector is calculated to obtain an attention score; the attention score is normalized to obtain cross-modal attention weights, and the value vectors are weighted and summed using the cross-modal attention weights to obtain cross-modal fused features; the cross-modal fused features are residually connected to the signal strength feature vector, and the residual connection result is processed using a nonlinear activation function to obtain the fused feature representation of the nodes.

[0102] Optionally, based on edge attributes, the fused feature representation of a node is aggregated with the fused feature representations of its neighboring nodes to update the node's feature representation. This involves the following steps: concatenating the fused feature representation of the node with the fused feature representations of its neighboring nodes to obtain a concatenated feature representation; calculating the attention coefficients between the node and its neighboring nodes using an attention calculation function; normalizing the attention coefficients to obtain normalized attention weights; weighting the fused feature representations of neighboring nodes using the normalized attention weights to obtain the neighbor aggregated features; and combining the neighbor aggregated features with the fused feature representation of the node, and processing the combination result using a non-linear activation function to obtain the updated feature representation of the node.

[0103] Optionally, based on the updated node features, the predicted location information of the terminal device in the target cell is determined, specifically including the following steps: The updated feature representation of the node is transformed and compressed using a first fully connected layer to obtain the node feature representation; the node feature representation is weighted using a second fully connected layer to obtain the initial contribution weight of the node to the positioning terminal device, wherein the output dimension of the second fully connected layer is a single-dimensional scalar; the initial contribution weight is normalized to obtain the contribution weight; the location information of the node is weighted and summed with the corresponding contribution weight to obtain the preliminary predicted location of the terminal device; the preliminary predicted location is transformed and mapped using a third fully connected layer to obtain the predicted location information of the terminal device, wherein the third fully connected layer has a two-dimensional output and includes a nonlinear activation function to map the preliminary location to the target coordinate space.

[0104] Optionally, distributed base stations are mapped to nodes in a graph structure. Signal strength indicators and location information are constructed as feature representations of nodes, and the distance between distributed base stations corresponding to any two nodes is determined as an edge attribute of the edge connecting the two nodes. Specifically, this includes the following steps: creating node identifiers corresponding to distributed base stations and establishing a mapping relationship between distributed base stations and the graph structure based on the node identifiers; constructing a signal strength feature vector based on the preprocessed reference signal received power value and constructing a location coordinate vector based on the two-dimensional coordinates of the distributed base station; concatenating the signal strength feature vector and the location coordinate vector to obtain the feature representation of the node; based on the mapping relationship, identifying two distributed base stations corresponding to two nodes connected by an undirected edge, calculating the Euclidean distance between the two distributed base stations, and determining the Euclidean distance as an edge attribute of the undirected edge.

[0105] Optionally, a supervised learning approach is used to train the graph neural network model. Under a preset stopping condition, the trained graph neural network model is obtained. Specifically, this includes the following steps: calculating the mean squared error loss between the predicted location information and the corresponding real location information; calculating the gradient of the trainable parameters of the graph neural network model based on the mean squared error loss, and updating the trainable parameters of the graph neural network model using an optimizer based on the gradient; evaluating the model performance in each training epoch on a validation dataset and recording the validation loss value; dynamically adjusting the learner's learning rate parameter according to a learning rate decay strategy; determining that the model performance of the graph neural network model will no longer improve when the validation loss value has not decreased in multiple consecutive training epochs; terminating the training process when the model performance of the graph neural network model will no longer improve, thus obtaining the trained graph neural network model.

[0106] Optionally, the learning rate parameter of the optimizer is dynamically adjusted according to the learning rate decay strategy, specifically including the following steps: initializing the initial learning rate parameter of the optimizer and obtaining the learning rate decay trigger condition; after each training epoch is completed, obtaining the current training epoch number and the corresponding validation loss value; based on the current training epoch number and validation loss value, determining whether the learning rate decay trigger condition is met; if the learning rate decay trigger condition is met, calculating the decay of the current learning rate parameter according to the preset learning rate decay strategy to obtain the updated learning rate parameter.

[0107] It should be noted that the above Figure 5 The modules in can be program modules (e.g., a set of program instructions that implements a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0108] It should be noted that, Figure 5 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0109] Figure 6 A hardware block diagram of a computer terminal for implementing a model training method is shown. Figure 6As shown, the computer terminal 60 may include one or more processors 602 (shown as 602a, 602b, ..., 602n in the figure) 602 (processor 602 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 604 for storing data, and a transmission module 606 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 6 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, the computer terminal 60 may also include... Figure 6 The more or fewer components shown, or having the same Figure 6 The different configurations shown.

[0110] It should be noted that the aforementioned one or more processors 602 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 60. As described in the embodiments of this application, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0111] The memory 604 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the model training method in this embodiment. The processor 602 executes various functional applications and data processing by running the software programs and modules stored in the memory 604, thereby realizing the above-mentioned model training method. The memory 604 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 604 may further include memory remotely located relative to the processor 602, and these remote memories can be connected to the computer terminal 60 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0112] The transmission module 606 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 60. In one example, the transmission module 606 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 606 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0113] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 60.

[0114] It should be noted here that, in some optional embodiments, the above... Figure 6 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 6 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0115] It should be noted that, Figure 6 The computer terminal shown is used to execute Figure 1 The model training method shown above is also applicable to this electronic device, and will not be repeated here.

[0116] This application also provides a non-volatile storage medium, which includes a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above model training method.

[0117] The non-volatile storage medium performs the following functions: Acquisition step: Acquire signal strength indicators and location information of multiple distributed base stations in the target cell; Graph construction step: Map the distributed base stations as nodes in a graph structure, construct signal strength indicators and location information as feature representations of the nodes, and determine the distance between any two distributed base stations as the edge attribute of the edge connecting any two nodes, where signal strength indicators and location information are features of different modalities; Cross-modal fusion step: Based on the cross-modal attention mechanism of the graph neural network model, fuse the features of different modalities of the nodes in the graph structure to obtain the fused feature representation of the nodes; Feature aggregation step: According to the edge attribute, aggregate the fused feature representation of the node with the fused feature representation of the node's neighboring nodes to update the node's feature representation; Location prediction step: Determine the predicted location information of the terminal device in the target cell based on the updated node features; Model training step: Train the graph neural network model using supervised learning, and obtain the trained graph neural network model when the preset stopping condition is met.

[0118] This application also provides an electronic device, including a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described model training method during runtime.

[0119] The processor runs a program that performs the following functions: Acquisition Step: Acquire signal strength indicators and location information of multiple distributed base stations in the target cell; Graph Construction Step: Map the distributed base stations to nodes in a graph structure, construct signal strength indicators and location information as feature representations of the nodes, and determine the distance between any two distributed base stations as the edge attribute of the edge connecting any two nodes, where signal strength indicators and location information are features of different modalities; Cross-Modal Fusion Step: Fusion of features of different modalities of nodes in the graph structure based on the cross-modal attention mechanism of the graph neural network model to obtain the fused feature representation of the nodes; Feature Aggregation Step: Aggregate the fused feature representation of the nodes with the fused feature representation of the nodes' neighboring nodes according to the edge attributes to update the feature representation of the nodes; Location Prediction Step: Determine the predicted location information of terminal devices in the target cell based on the updated node features; Model Training Step: Train the graph neural network model using supervised learning, and obtain the trained graph neural network model when the preset stopping conditions are met.

[0120] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0121] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0122] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0127] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A model training method, characterized in that, include: Acquisition steps: Obtain signal strength indicators and location information of multiple distributed base stations in the target cell; Graph construction steps: Map the distributed base stations as nodes in a graph structure, construct the signal strength index and the location information as feature representations of the nodes, and determine the distance between any two distributed base stations corresponding to any two nodes as the edge attribute of the edge connecting the two nodes, wherein the signal strength index and the location information are features of different modes; Cross-modal fusion step: The cross-modal attention mechanism based on the graph neural network model is used to fuse the features of the nodes in the graph structure of different modalities to obtain the fused feature representation of the nodes; Feature aggregation step: Based on the edge attributes, the fused feature representation of the node is aggregated with the fused feature representation of the node's neighboring nodes to update the node's feature representation; Location prediction step: Based on the updated node characteristics, determine the predicted location information of the terminal devices in the target cell; Model training steps: The graph neural network model is trained using supervised learning. When the preset stopping condition is met, the trained graph neural network model is obtained.

2. The method according to claim 1, characterized in that, A cross-modal attention mechanism based on a graph neural network model fuses the features of nodes from different modalities in the graph structure to obtain the fused feature representation of the nodes, including: The signal strength index of the node is linearly transformed using the first linear transformation layer to obtain the signal strength feature vector, and the position information of the node is linearly transformed using the second linear transformation layer to obtain the position coordinate feature vector. Using the signal strength feature vector as the query vector and the location coordinate feature vector as the key vector and value vector, the attention score is obtained by calculating the dot product of the query vector and the key vector; The attention scores are normalized to obtain cross-modal attention weights, and the value vectors are weighted and summed using the cross-modal attention weights to obtain cross-modal fusion features. The cross-modal fusion features are residually concatenated with the signal intensity feature vector, and the residual concatenation result is processed by a nonlinear activation function to obtain the fused feature representation of the node.

3. The method according to claim 1, characterized in that, Based on the edge attributes, the fused feature representation of the node is aggregated with the fused feature representations of the node's neighboring nodes to update the node's feature representation, including: The fused feature representation of the node is concatenated with the fused feature representation of the node's neighboring nodes to obtain the concatenated feature representation. The attention coefficient between the node and its neighboring nodes is obtained by using the attention calculation function to calculate the concatenated feature representation and the edge attributes of the edges connecting the node and its neighboring nodes. The attention coefficients are normalized to obtain normalized attention weights; The fused feature representations of the neighbor nodes are weighted and summed using the normalized attention weights to obtain the neighbor aggregation features; The neighbor aggregation features and the fused feature representation of the node are combined, and the combination result is processed by a nonlinear activation function to obtain the updated feature representation of the node.

4. The method according to claim 1, characterized in that, Based on the updated node characteristics, the predicted location information of terminal devices in the target cell is determined, including: The updated feature representation of the node is transformed and compressed using the first fully connected layer to obtain the node feature representation. The importance weights of the node feature representations are calculated using the second fully connected layer to obtain the initial contribution weights of the nodes to locating the terminal device. The output dimension of the second fully connected layer is a single-dimensional scalar. The initial contribution weights are normalized to obtain the final contribution weights. The location information of the node is weighted and summed with the corresponding contribution weight to obtain the preliminary predicted location of the terminal device; The preliminary predicted position is nonlinearly transformed and mapped using a third fully connected layer to obtain the predicted position information of the terminal device. The third fully connected layer has a two-dimensional output and includes a nonlinear activation function, which is used to map the preliminary position to the target coordinate space.

5. The method according to claim 1, characterized in that, The distributed base stations are mapped as nodes in a graph structure. The signal strength index and the location information are constructed as feature representations of the nodes. The distance between any two distributed base stations corresponding to any two nodes is determined as the edge attribute of the edge connecting the two nodes, including: Create node identifiers corresponding to the distributed base stations, and establish a mapping relationship between the distributed base stations and the graph structure based on the node identifiers; Based on the preprocessed reference signal received power value, a signal strength feature vector is constructed, and based on the two-dimensional coordinates of the distributed base station, a location coordinate vector is constructed. The signal strength feature vector and the location coordinate vector are concatenated to obtain the feature representation of the node; Based on the mapping relationship, two distributed base stations corresponding to two nodes connected by an undirected edge are identified, and the Euclidean distance between the two distributed base stations is calculated. The Euclidean distance is then determined as the edge attribute of the undirected edge.

6. The method according to claim 1, characterized in that, The graph neural network model is trained using supervised learning. Under preset stopping conditions, a fully trained graph neural network model is obtained, including: Calculate the mean square error loss value between the predicted location information and the corresponding true location information; The gradient of the trainable parameters of the graph neural network model is calculated based on the mean squared error loss value, and the trainable parameters of the graph neural network model are updated by the optimizer according to the gradient. Evaluate the model performance for each training epoch on the validation dataset and record the validation loss values; The learner's learner rate parameter is dynamically adjusted according to the learner rate decay strategy. If the validation loss value does not decrease in multiple consecutive training rounds, it is determined that the performance of the graph neural network model will no longer improve. If it is determined that the performance of the graph neural network model can no longer be improved, the training process is terminated, and the trained graph neural network model is obtained.

7. The method according to claim 6, characterized in that, The learner's learner rate parameter is dynamically adjusted according to the learner rate decay strategy, including: Initialize the initial learning rate parameters of the optimizer and obtain the learning rate decay trigger condition; After each training round is completed, obtain the current training round number and the corresponding validation loss value; Based on the current number of training rounds and the validation loss value, determine whether the learning rate decay trigger condition is met; If the learning rate decay trigger condition is met, the current learning rate parameter is decayed according to the preset learning rate decay strategy to obtain the updated learning rate parameter.

8. A positioning method, characterized in that, include: Obtain signal strength indicators and location information of multiple distributed base stations in the target cell; The signal strength indicators and location information of the multiple distributed base stations are processed using a graph neural network model to obtain the predicted location information of the terminal devices in the target cell. The graph neural network model is obtained by training using the model training method described in any one of claims 1 to 7.

9. A model training device, characterized in that, include: The acquisition module is used to acquire signal strength indicators and location information of multiple distributed base stations in the target cell; The graph construction module is used to map the distributed base stations as nodes in a graph structure, construct the signal strength index and the location information as feature representations of the nodes, and determine the distance between any two distributed base stations corresponding to any two nodes as the edge attribute of the edge connecting the two nodes, wherein the signal strength index and the location information are features of different modes; The fusion module is used to fuse the features of the nodes in the graph structure from different modes based on the cross-modal attention mechanism of the graph neural network model, so as to obtain the fused feature representation of the nodes. An aggregation module is used to aggregate the fused feature representation of the node with the fused feature representation of the node's neighboring nodes based on the edge attributes, so as to update the feature representation of the node. The prediction module is used to determine the predicted location information of terminal devices in the target cell based on the updated node characteristics. The training module is used to train the graph neural network model using supervised learning, and obtains the trained graph neural network model when a preset stopping condition is met.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, the device containing the non-volatile storage medium is controlled to perform the model training method according to any one of claims 1 to 7 and the localization method according to claim 8.

11. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the model training method of any one of claims 1 to 7 and the localization method of claim 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the model training method according to any one of claims 1 to 7 and the localization method according to claim 8.

Images