Broadcast radio positioning system near-far effect elimination method, storage medium and terminal equipment

Abstract

The invention discloses a near-far effect elimination method for a broadcast type radio positioning system, a storage medium and terminal equipment. The near-far effect elimination method comprises the following steps: acquiring and preprocessing multi-source data; constructing a dynamic graph data structure; a Transform-GNN (Global Navigation Number) joint model is constructed; training and optimizing the model; and performing model evaluation and adaptive optimization. According to the method, global signal correlation analysis is realized through a self-attention mechanism of Transform, based on an alternately stacked GNN-Transform architecture, a residual connection and gating mechanism is introduced to dynamically fuse local-global features, and in combination with staged training and a dynamic weight adjustment strategy, the positioning precision and the interference suppression capability are optimized. The method supports multi-modal data fusion, significantly improves the anti-multipath performance and dynamic adaptability in a complex scene, solves the defects of high calculation complexity, strong hardware dependence and insufficient environment robustness of a traditional method, is suitable for near-far effect suppression under the condition that a satellite is unavailable, improves the system positioning precision, and has a wide application prospect. And the system can be better applied to high-precision positioning scenes such as indoor positioning and unmanned aerial vehicle navigation.

Description

Technical Field

[0001] The present invention relates to the field of wireless positioning technology, and in particular to a method for eliminating the near-far effect of a broadcast radio positioning system, a storage medium and a terminal device. Background Art

[0002] In the broadcast radio positioning system, since the signal propagation loss is proportional to the square of the distance, the signal strength of the ground base station closer to the radio positioning receiver terminal is much greater than the signal of the ground base station farther away. This causes the signal of the ground base station at a short distance to drown out the signal of the ground base station at a long distance. This phenomenon is called the near-far effect of the broadcast radio positioning system. The near-far effect is a key technical problem that needs to be solved in radio positioning. It affects the signal reception of the navigation system, causing the system positioning accuracy to be seriously reduced, or even unable to work normally, affecting the accuracy and reliability of the entire navigation system.

[0003] The GNSS system broadcasts ranging signals through satellites in medium and high orbits, and satellite navigation receivers receive the signals to complete real-time positioning and speed measurement. However, due to the high orbit of the satellite navigation system, the signal power is extremely weak when it reaches the ground, and it is easy to be blocked or maliciously interfered with by radar, communications, criminals, etc., causing the satellite navigation receiver to work abnormally. In the field of high security of national economy and people's livelihood, or in military applications, it is necessary to consider an emergency means of navigation and positioning to meet the navigation and positioning needs when the satellite signal is interfered with or the satellite system is paralyzed. For this reason, domestic and foreign research institutions have successively carried out research on related technologies of broadcast radio positioning systems.

[0004] The broadcast radio positioning system consists of two parts: a radio ground base station and a radio positioning receiver. Multiple radio ground base stations are deployed on the ground. The strength of the signal transmitted by a single station can effectively cover a hundred kilometers. The radio positioning receiver terminal simultaneously receives the ranging signals of multiple radio ground base stations to complete its own real-time positioning and speed measurement. However, due to the strong transmission signal of the ground base station, when the radio positioning receiver terminal is close to the ground base station, it will affect the receiver terminal's capture of the distant ground base station signal, causing the receiver terminal to reduce its positioning accuracy or even fail to locate.

[0005] Currently, the main solution to the near-far effect in broadcast radio positioning systems is based on radio ground base stations. By improving the signal spectrum and signal strength at the transmitting end of the ground base station, the purpose of eliminating the near-far effect is achieved. Common methods include out-of-band emission, frequency offset, frequency hopping, time hopping, and power adjustment. The first three methods are difficult to implement and require major modifications to the radio frequency part or correlator of the radio positioning receiver terminal. Therefore, they are not widely used at present. The time hopping method has achieved relatively good results in a small area. However, due to time slot division, the overall power of the transmitted time hopping signal decreases compared to the continuous signal, affecting the anti-interference performance of the system. The power adjustment method can eliminate the influence of the near-far effect to a certain extent, but it requires signal strength analysis according to the application scenario to achieve power adjustment. The system has low automation and is not suitable for flexible deployment of the system.

[0006] The invention patent "A Method for Suppressing the Near-Far Effect of Pseudosatellites Based on Multi-Constraint Beamforming" proposes an improved method for signal reception and processing of a radio positioning receiver terminal. The signals received by the antenna array are successively sampled, band-pass filtered, amplitude-phase error corrected, covariance matrix and its inverse matrix calculated, MUSIC direction finding, pseudosatellite and navigation satellite signal power estimation, constraint response vector calculation corresponding to the direction of the pseudosatellite signal, multi-constraint beamforming weight calculation, digital beamforming, and navigation solution to finally obtain local position information. The invention has stronger versatility and can work in an interference environment, is convenient and flexible, and does not require modification of the transmitting end pseudosatellite or the radio frequency of the receiving end. Summary of the Invention

[0007] The technical solution of the present invention to solve the above technical problems is to provide a method for eliminating the near-far effect in a broadcast radio positioning system, including the following steps:

[0008] S1: Multi-source data acquisition and preprocessing: Collect radio ground base station data, radio positioning receiver data, interference source data, and reflection source data; perform signal alignment, feature standardization, environmental perception fusion, and outlier repair processing on the data;

[0009] S2: Construct a dynamic graph data structure: Define ground base stations, receivers, interference sources, and fictional multipath reflection nodes as graph nodes, construct ground station-receiver connection edges, interference source-receiver connection edges, and spatio-temporal trajectory connection edges based on signal propagation relationships, and design a dynamic update mechanism to achieve real-time update of node positions and edge weights;

[0010] S3: Construct a Transformer-GNN joint model: Achieve local-global joint modeling through alternately stacked GNN layers and Transformer layers, where:

[0011] The GNN layer defines message functions based on node types to extract local topological features;

[0012] The Transformer layer introduces edge attribute encoding and captures global signal correlations through a sparse attention mechanism;

[0013] Residual connections and gating mechanisms are used to dynamically fuse GNN and Transformer features;

[0014] S4: Model training and optimization: A phased training strategy is adopted. First, the GNN layer is pre-trained, and then the Transformer-GNN model is jointly trained; the localization loss and interference suppression loss are balanced through dynamic weight adjustment (AdaLoss);

[0015] S5: Model evaluation and adaptive optimization: The localization accuracy is evaluated based on the root mean square error (RMSE) and mean absolute error (MAE). According to the evaluation results, the model structure parameters are adjusted, and data augmentation is used to improve the generalization ability.

[0016] Furthermore, in the step S2:

[0017] The node types include:

[0018] Radio ground base station nodes, whose feature vectors include three-dimensional coordinates, transmit power, carrier frequency, and pseudo-random code number;

[0019] Radio positioning receiver nodes, whose feature vectors include real-time position, speed, received signal strength vector, and carrier-to-noise ratio;

[0020] Interference source nodes, whose feature vectors include interference position, power, bandwidth, and type coding;

[0021] Fictitious multipath reflection nodes, whose feature vectors include reflector coordinates, incident angle, reflection coefficient, and path delay;

[0022] The edge weight calculation is carried out in the following way:

[0023] Ground station-receiver edge: Based on the free space path loss formula and the obstacle attenuation model;

[0024] Interference-receiver edge: Based on the integral of spectral overlap degree and spatial distance correlation;

[0025] Space-time trajectory edge: Based on the time decay factor or prediction confidence score.

[0026] Furthermore, the dynamic update mechanism in the step S2 includes:

[0027] The adjustable range of the update frequency is 50ms to 5000ms, with a default of 3000ms;

[0028] The reconstruction mechanism is triggered when a new interference source is detected or the signal is lost. The reconstruction conditions include:

[0029] The mutation of spectral energy exceeds 3 times the background noise threshold;

[0030] The signal is not detected for N consecutive epochs, where N is dynamically adjustable and the default value is 5.

[0031] Furthermore, the construction of the Transformer-GNN joint model in step S3 further includes:

[0032] The GNN layer adopts a graph convolutional network (GCN) or a graph attention network (GAT), and reduces the computational complexity through neighbor sampling or TopK aggregation;

[0033] The Transformer layer adopts a sparse attention mechanism and encodes the edge attributes as position embeddings;

[0034] The GNN layer and the Transformer layer are alternately stacked 3 to 5 times, and the order is the GNN→Transformer loop structure.

[0035] Furthermore, the implementation method of the residual connection and the gating mechanism is:

[0036] The residual connection directly adds the GNN output feature and the Transformer output feature;

[0037] The gating mechanism dynamically controls the feature fusion weight based on GRU, and selects the retained features through the update gate and the reset gate.

[0038] Furthermore, the dynamic weight adjustment (AdaLoss) in step S4 satisfies:

[0039] The joint loss function is , where is the positioning loss, is the interference suppression loss, and the weight λ is dynamically adjusted during the training process;

[0040] The training data is divided into a training set, a validation set, and a test set according to 6:3:1, and GPU acceleration parallel training is adopted.

[0041] Furthermore, the adaptive optimization in step S5 includes:

[0042] Adjust the number of GNN layers, the number of Transformer heads, and the hidden layer dimension according to the RMSE and MAE evaluation results;

[0043] Enhance the training set through data rotation and scaling to improve the adaptability of the model to the dynamic environment.

[0044] To solve the above technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for eliminating the near-far effect as described above.

[0045] To solve the above technical problems, the present invention also provides a terminal device for a broadcast radio positioning system, including a memory, a processor, and a computer program stored on the memory, wherein when the processor executes the program, the method for eliminating the near-far effect as described above is implemented.

[0046] Compared with the prior art, the present invention has the following technical effects:

[0047] 1. Creatively construct a Transformer-GNN dual-modal dynamic fusion network to break through the technical boundaries of traditional single models. Capture kilometer-level signal correlations through the global attention mechanism of Transformer, and combine the graph structure learning of GNN to analyze centimeter-level signal propagation characteristics, realizing cross-scale feature fusion in the spatial dimension.

[0048] 2. Initiate a signal equalization algorithm based on attention weights to eliminate intensity deviation through feature space mapping.

[0049] 3. Enhance robustness in a multi-physical-field coupling environment, construct an electromagnetic-space joint representation learning framework, and realize multipath propagation modeling through spatio-temporal graph convolution.

[0050] 4. Develop a parameter dynamic transfer learning mechanism.

[0051] 5. Innovatively introduce a meta-learning optimizer to achieve co-evolution of model parameters and environmental features. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the structures shown in these drawings without creative efforts.

[0053] Figure 1 It is a flowchart of the steps of the method for eliminating the near-far effect of the broadcast radio positioning system according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The present invention provides a method for eliminating the near-far effect of a broadcast radio positioning system, a storage medium, and a terminal device, aiming to design a method for eliminating the near-far effect of a broadcast radio positioning system based on a Transformer-GNN model.

[0055] The method for eliminating the near - far effect of the broadcast radio positioning system proposed by the present invention will be described in the following specific embodiments:

[0056] Embodiment 1: A method for eliminating the near - far effect of a broadcast radio positioning system, as Figure 1 shown, includes the following steps:

[0057] S1: Multi - source data acquisition and pre - processing: Acquire radio ground base station data, radio positioning receiver data, interference source data, and reflection source data; perform signal alignment, feature standardization, environmental perception fusion, and outlier repair processing on the data;

[0058] S2: Construct a dynamic graph data structure: Define ground base stations, receivers, interference sources, and fictional multipath reflection nodes as graph nodes, construct ground station - receiver connection edges, interference source - receiver connection edges, and spatio - temporal trajectory connection edges based on signal propagation relationships, and design a dynamic update mechanism to achieve real - time update of node positions and edge weights;

[0059] S3: Construct a Transformer - GNN joint model: Achieve local - global joint modeling through alternately stacked GNN layers and Transformer layers, where:

[0060] The GNN layer defines a message function based on node types and extracts local topological features;

[0061] The Transformer layer introduces edge attribute encoding and captures global signal associations through a sparse attention mechanism;

[0062] Adopt residual connections and gating mechanisms to dynamically fuse GNN and Transformer features;

[0063] S4: Model training and optimization: Adopt a phased training strategy, first pre - train the GNN layer, and then jointly train the Transformer - GNN model; balance the positioning loss and interference suppression loss through dynamic weight adjustment (AdaLoss);

[0064] S5: Model evaluation and adaptive optimization: Evaluate the positioning accuracy based on the root mean square error (RMSE) and mean absolute error (MAE), adjust the model structure parameters according to the evaluation results, and enhance the generalization ability through data augmentation.

[0065] Further, in the step S2:

[0066] The node types include:

[0067] Radio ground base station nodes, the feature vector includes three - dimensional coordinates, transmission power, carrier frequency, and pseudo - random code number;

[0068] A radio positioning receiver node, the feature vector includes real-time position, velocity, received signal strength vector, and carrier-to-noise ratio;

[0069] An interference source node, the feature vector includes interference position, power, bandwidth, and type coding;

[0070] A fictitious multipath reflection node, the feature vector includes reflecting surface coordinates, incident angle, reflection coefficient, and path delay;

[0071] The edge weight calculation is carried out in the following way:

[0072] Ground station - receiver edge: Based on the free space path loss formula and the obstacle attenuation model;

[0073] Interference - receiver edge: Based on the integral of spectral overlap degree and spatial distance correlation;

[0074] Space - time trajectory edge: Based on the time decay factor or the prediction confidence score.

[0075] Furthermore, the dynamic update mechanism in step S2 includes:

[0076] The adjustable range of the update frequency is 50ms to 5000ms, and the default value is 3000ms;

[0077] When a new interference source is detected or a signal is lost, the reconstruction mechanism is triggered, and the reconstruction conditions include:

[0078] The spectral energy mutation exceeds 3 times the background noise threshold;

[0079] The signal is not detected for N consecutive epochs, where N is dynamically adjustable and the default value is 5.

[0080] Furthermore, the construction of the Transformer - GNN joint model in step S3 further includes:

[0081] The GNN layer adopts a graph convolutional network (GCN) or a graph attention network (GAT), and reduces the computational complexity through neighbor sampling or TopK aggregation;

[0082] The Transformer layer adopts a sparse attention mechanism and encodes the edge attributes into position embeddings;

[0083] The GNN layer and the Transformer layer are alternately stacked 3 to 5 times, and the order is a GNN→Transformer loop structure.

[0084] Furthermore, the implementation method of the residual connection and the gating mechanism is:

[0085] The residual connection directly adds the GNN output feature and the Transformer output feature;

[0086] The gating mechanism dynamically controls the feature fusion weights based on GRU, and selects and retains features through the update gate and the reset gate.

[0087] Furthermore, the dynamic weight adjustment (AdaLoss) in step S4 satisfies:

[0088] The combined loss function is , where is the localization loss, is the interference suppression loss, and the weight λ is dynamically adjusted during the training process;

[0089] The training data is divided into a training set, a validation set, and a test set in a ratio of 6:3:1, and GPU acceleration is used for parallel training.

[0090] Furthermore, the adaptive optimization in step S5 includes:

[0091] Adjust the number of GNN layers, the number of Transformer heads, and the hidden layer dimension according to the RMSE and MAE evaluation results;

[0092] Enhance the training set through data rotation and scaling to improve the adaptability of the model to the dynamic environment.

[0093] To solve the above technical problems, the present invention also proposes a computer-readable storage medium storing a computer program, and when the program is executed by a processor, it implements the near-far effect elimination method as described above.

[0094] Embodiment 2: A method for eliminating the near-far effect in a broadcast radio positioning system. Based on the complementarity of the Transformer model and the GNN model in information processing mechanisms and modeling capabilities, a Transformer-GNN near-far effect elimination model is constructed based on global-local joint modeling and end-to-end learning capabilities to improve the overall positioning performance of the system.

[0095] The Transformer-GNN model combines the Transformer model with the Graph Neural Network (GNN). The combination is not simply a stacking of modules, but is based on the mechanisms of both, achieving the design effect through collaborative optimization. For the near-far effect problem in radio positioning systems, not only the local problem that signals from nearby ground base stations may overwhelm those from distant ground base stations needs to be solved, but also systematic issues such as the capture and tracking of all ground base station signals by the radio positioning receiver terminal, the effective geometric distribution of ground base stations, and the positioning accuracy of the radio positioning receiver terminal need to be considered. Through research and analysis, GNN mainly focuses on graph-structured data. By propagating and aggregating information between nodes and their neighbor nodes, it can deeply explore the local topological structure of the graph and the local connection relationships between nodes. The core advantage of the Transformer model is its self-attention mechanism. Therefore, this model has the ability to capture global information, can establish long-range dependencies in sequence data, and can analyze the global associations between elements. Therefore, the Transformer-GNN model formed by the collaborative combination of the two has both a global perspective and can control details, effectively solving the near-far problem.

[0096] The implementation idea of the Transformer-GNN model is as follows: First, use GNN to preliminarily process the graph-structured data of the broadcast radio positioning system, extract local signal features, and accurately model local topological relationships (such as the distance attenuation between users and beacons, the coupling strength between users and interference) through the message passing mechanism. Then, input the processed features into the Transformer. The Transformer further analyzes and integrates the signal features from a global perspective through the self-attention mechanism, discovers cross-regional signal associations (such as the collaborative positioning mode of distant beacons), makes up for the deficiencies in GNN modeling, and improves the global optimization ability of the model. GNN and Transformer share node embeddings and achieve end-to-end optimization through joint training. The feature representations of the two enhance each other, forming hierarchical information extraction. Through model training and optimization, adjust the processing weights for near and far signals, thereby effectively eliminating the influence of the near-far effect and improving the positioning accuracy of the radio positioning system receiver for receiving devices that transmit signals from ground stations at different near and far positions.

[0097] Specifically, it includes the following steps:

[0098] Step S1, multi-source data collection and preprocessing:

[0099] The collected multi-source data includes: radio ground base station data: coordinates, transmission power, signal spectrum characteristics (pseudo-random code, carrier frequency, etc.); radio positioning receivers: real-time position, received signal spectrum (signal strength, carrier-to-noise ratio), motion state (speed, acceleration), etc.; interference source data: spectrum scanning results (interference center frequency, bandwidth, power), dynamic position, etc.; reflection source data: such as the reflection surface position coordinates of buildings, the ground, or water surfaces, incident angles, reflection coefficients, and other data.

[0100] The data preprocessing work includes signal alignment of radio ground base station data: achieving multi-channel signal synchronization and performing signal timestamp alignment; feature standardization processing of radio positioning receivers: normalizing (Z-Score) features such as position information, signal power, and distance, and mapping them to the [0, 1] interval to improve the stability of model training; environmental perception fusion of interference source data: converting other auxiliary data or interference source data collected into obstacle masks for path loss correction; missing value and outlier detection and repair of reflection source data: identifying missing values and outliers through statistical analysis, filling in the missing values in the data using interpolation methods, and repairing outliers through machine learning algorithms.

[0101] Step S2, constructing a dynamic graph data structure:

[0102] The construction of the dynamic graph data structure mainly includes node (Node) definition, edge construction, and a dynamic graph update mechanism. ① Node definition: Each radio ground base station and radio positioning receiver is used as a node of the graph, and the node features are the data collected and preprocessed in step 1; ② Edges are constructed according to the signal propagation relationship, and a total of three types of edge construction are included. The first type is the connection edge between the radio ground station and the receiver. If a radio positioning receiver receives the signal of a certain radio ground base station, an edge can be constructed between the two; the second type is the connection edge between the interference source and the radio receiver. If a radio positioning receiver receives the signal of a certain interference source, an edge can be constructed between the two; the third type is the spatio-temporal trajectory connection edge, including historical connection edges and connection edges formed by spatio-temporal trajectory deduction based on the movement trajectory of the radio positioning receiver. ③ Dynamic update mechanism: The position of the moving node and the edge weight are updated with an adjustable spatio-temporal update mechanism, from 50ms to 5000ms, with each 10ms as an adjustable step, and the default update frequency is 3000ms. The spatio-temporal association storage time limit is from 5 frames to 20 frames, with each 5 frames as an adjustable step, and the default is 10 frames. When the system detects a new interference source or signal loss, the reconstruction mechanism is started.

[0103] The detailed description of constructing the dynamic graph data structure is as follows:

[0104] 1. Node (Node) parameter definition:

[0105] (1) Node type;

[0106] Radio ground base station node (GroundNode);

[0107] Physical meaning: Fixed or mobile radio ground base station emission source.

[0108] Eigenvector: ;

[0109] ① : Three-dimensional coordinates of the radio ground base station, unit: m;

[0110] ② : Transmit power, unit: dBm;

[0111] ③ : Carrier power, unit: MHz;

[0112] ④ PRN: Pseudo-random code number.

[0113] (2) Radio positioning receiver node (ReceiverNode);

[0114] Physical meaning: Radio positioning receiver to be positioned.

[0115] Eigenvector: ;

[0116] ① : Real-time position of the radio positioning receiver, unit: m;

[0117] ② , : Three-dimensional velocity, unit: m / s;

[0118] ③ : Received signal strength vector, dimension = number of radio ground base stations;

[0119] ④ : Carrier-to-noise ratio (unit: dB-Hz).

[0120] (3) Interference node (InterferenceNode);

[0121] Physical meaning: Active or passive interference source.

[0122] Eigenvector: ;

[0123] ① : Interference source position, unit: m;

[0124] ② : Interference power, unit: dBm;

[0125] ③ : Interference bandwidth, unit: MHz;

[0126] ④ : Interference source type code (e.g., 0: Jamming interference, 1: Deception interference; 2: New type of interference) (unit: dB-Hz).

[0127] (4) Fictitious multipath reflection node;

[0128] Physical meaning: The reflection surface nodes that constitute the multipath effect.

[0129] Eigenvector: ;

[0130] ① : Coordinates of the center point of the reflection surface, unit: m;

[0131] ② : Incident angle of the reflection source, unit: °;

[0132] ③ : Reflection coefficient;

[0133] ④ : Delay of the reflection path relative to the direct path, unit: ns.

[0134] 2. Definition of edge parameters;

[0135] (1) Ground-Receiver Link Edges

[0136] Weight calculation: Build a model based on path loss and obstacle blocking attenuation, +

[0137] ① : Free space path loss value, calculated by the following formula.

[0138] , Unit: km, Unit: MHz.

[0139] ② : Obstacle attenuation, unit dB, estimated through an empirical model.

[0140] (2) Interference-Receiver Link Edges;

[0141] Weight calculation: Build a model based on spectral overlap and spatial correlation, ;

[0142] ① : The spectral overlap energy integral of the interference signal within the radio navigation signal frequency band:

[0143] ;

[0144] : The center frequency of the radio navigation signal;

[0145] : The bandwidth of the radio navigation signal;

[0146] : The power spectral density function of the interference signal;

[0147] ② : The weight coefficient, which balances the relationship between the spectrum and the spatial position, ∈ [0, 1];

[0148] ③ : The real-time distance between the interference source l and the radio receiver i, unit: m.

[0149] (2) Temporal and spatial deduction edges;

[0150] ① Receiver-R history edges;

[0151] Connection method: Connect the current node of the radio receiver with the nodes at the past (T) moments.

[0152] Weight calculation: Calculate the weight based on the time decay factor, , is the decay coefficient, with a value range of 0.1 to 0.5, and t is the time.

[0153] ② Deduction edges;

[0154] Generation method: Predict the expected position of the radio receiver through algorithms such as LSMT, and generate fictional connection edges.

[0155] Weight calculation: Score the confidence of the prediction:

[0156] · ;

[0157] Among them, : The root mean square error between the predicted position and the actual position; Adjust in real time according to the real-time environmental changes and the maximum tolerable positioning error.

[0158] 3. Dynamic update mechanism:

[0159] (1)Adjustable spatio-temporal update mechanism;

[0160] The adjustable spatio-temporal update mechanism is used to update the positions of mobile nodes and edge weights. The update frequency is 50ms - 5000ms, with an adjustable step of 10ms each. The default update frequency is 3000ms. The spatio-temporal correlation storage time limit is 5 - 20 frames, with an adjustable step of 5 frames each, and the default is 10 frames. When the system detects a new interference source or signal loss, the reconstruction mechanism is started.

[0161] Update period: Typical value 3000ms (suitable for scenarios with a speed not greater than 3 Mach);

[0162] Calculation formula: .

[0163] (2)Reconstruction mechanism;

[0164] The reconstruction mechanism is enabled in the following two cases. ① New interference detection: Detect whether there is a sudden change in the spectral energy. Usually, three times the background noise level is taken as the detection threshold. ② Detect whether the received radio ground base station signal is lost. If the signal is not detected for N consecutive epochs, it is determined that the signal is lost. N can be adjusted dynamically, and the typical value is 5.

[0165] Step 3: Construct the Transformer-GNN joint model:

[0166] The Transformer-GNN model is an optimized model after the combination of Transformer and GNN. Therefore, the construction of the Transformer-GNN joint model includes the design of the GNN layer, the design of the Transformer layer, and the combined design of the Transformer and GNN models.

[0167] (1)GNN layer design:

[0168] Select the graph convolutional neural network (GCN) as the basic architecture. On this basis, determine the number of layers and the number of neurons in each layer, usually set to 2 - 4 layers. The optimal number of neurons is determined through simulation experiments, and the measurement indicators are model complexity and performance. The role of the GNN layer is to extract and update local features of graph-structured data and propagate neighbor node information. In the information aggregation stage, node types are distinguished to accurately capture local physical constraints, such as the geometric distance between the radio ground base station and the radio positioning receiver, the spectral influence range of the interference source, etc. Local physical constraints are realized through type-related message functions.

[0169] Message function definition: ;

[0170] Among them, : The feature vector of the source node i at layer l;

[0171] : The feature vector of the target node j at layer l, is a d-dimensional real vector space;

[0172] : Edge feature, is an e-dimensional real vector space;

[0173] : The generated message vector, which is consistent with the node dimension.

[0174] (2) Transformer layer design:

[0175] Construct a Transformer model that includes a multi-head attention mechanism and a feed-forward neural network. Set the number of heads of the attention mechanism, which can be set from 2 to 10 heads; set the dimension of the hidden layer, which can be set from 64 to 1024, and determine the optimal hyperparameters through cross-validation. Use the output of the GNN layer as the input of the Transformer layer, and use the Transformer layer to capture global information and adjust the signal feature weights. Inject edge attributes (such as path loss) as positional encoding in the Transformer. The edge feature encoding layer is implemented as follows:

[0176] Input: Original edge attributes , including path loss, interference source intensity, spectrum overlap degree, etc.

[0177] Encoding process:

[0178] : The weight matrix input to the hidden layer;

[0179] : The bias of the hidden layer;

[0180] : The projection of the hidden layer attention space;

[0181] : The output bias.

[0182] (3) The combination design of the Transformer and GNN models realizes the elimination of the near-far effect:

[0183] The combined design of the Transformer and GNN models completes the elimination of the near-far effect. The combination of the Transformer and GNN is not simply a module stacking, but based on the complementarity of the two in information processing mechanisms and modeling capabilities, and the purpose of eliminating the near-far effect is achieved through collaborative optimization. The combination of the two faces great design difficulties and implementation challenges, and it is necessary to solve system problems such as inconsistent feature representation distributions between the GNN and the Transformer, increased computational complexity, and difficulty in determining the weights of training stability. It is necessary to achieve the design goal through careful design and engineering optimization. The specific steps for implementing the Transformer-GNN fusion construction in this patent are as follows: ① Information input: At the beginning of the model, the node features are embedded; ② GNN layer design: Each layer of the GNN processes the information of local neighbors. GAT or GCN or other variants can be used, and at the same time, the integration of edge features is considered. Since the computational overhead is large after the combination of the GNN and the Transformer, and the multi-hop aggregation complexity of the GNN is O(K|E|), neighbor sampling or TopK aggregation is used in the GNN to reduce the amount of computation; ③ Transformer layer design: Processes the global information of all nodes. It is necessary to customize the position encoding to adapt to the graph structure. The complexity of the self-attention mechanism of the Transformer is 0(N2). Sparse attention or local window attention is used for the Transformer to reduce the complexity; ④ Residual connection and fusion normalization: The feature representation distributions of the GNN and the Transformer are inconsistent, and direct stacking will lead to performance degradation. The present invention introduces a residual connection and a gating mechanism to dynamically fuse features, and uses a joint loss function (localization loss + interference suppression loss) to achieve feature fusion normalization. In the implementation of the above several methods, a large number of simulations and tests are required to determine the weight coefficient control in the dynamic fusion process, and to design a reasonable residual connection, gating mechanism and joint loss function, so as to significantly improve the performance, stability and flexibility of the model. ⑤ Alternating stacked loop structure: The GNN layer and the Transformer layer are alternately stacked, usually alternating 3 to 5 times, so that the model can fully fuse local and global information to form a complete Transformer-GNN combined model. The design of the alternating order is GNN layer → Transformer layer → GNN layer → Transformer layer, and this cycle is repeated 3 to 5 times; ⑥ Information output: Aggregate the features after alternating processing, perform signal separation and interference suppression, and achieve the elimination of the near-far effect.

[0184] Residual connection: In the residual connection, the input feature is directly added to the output feature through a skip connection to avoid losing key information during the feature transfer process. In the Transformer-GNN model, the residual connection is used to fuse the features of the GNN and the Transformer.

[0185] GNN output features: ;

[0186] Transformer output features: Transformer( );

[0187] Residual connection: + .

[0188] Gating mechanism: mainly implemented based on GRU, controls feature fusion through dynamic weights, and selectively retains or ignores certain features. The key parameters of GRU are the update gate, reset gate, candidate state, and output parameters, which dynamically control feature fusion through the update gate, reset gate, and candidate state, and achieve efficient alignment and fusion of features in the Transformer-GNN joint model.

[0189] Joint loss function: The joint loss function guides the feature alignment of GNN and Transformer through multi-task learning, and optimizes the localization task and interference suppression task simultaneously.

[0190] The formula of the joint loss function is: ;

[0191] Where: is the localization loss.

[0192] is the interference suppression loss.

[0193] λ is the weight of the localization loss, and its value range is [0,1].

[0194] Step S4: Training the model and optimization:

[0195] In the training of the Transformer-GNN model, there is a possibility of conflict between the optimization objectives of GNN and Transformer, resulting in unstable training process. This patent solves the stability problem in the model training process through staged training and dynamic weight adjustment (AdaLoss). The model training is based on a parallel processing architecture and combined with GPU acceleration to improve the model training efficiency. ① Staged training. First, train GNN alone to achieve better performance on graph structure tasks. Then, based on the pre-trained GNN, jointly train Transformer and GNN to make it adapt to global dependency tasks. ② Dynamic weight adjustment (AdaLoss), dynamically adjust the weights according to the task loss to balance the localization and interference suppression tasks. The localization loss weight λ and the interference suppression loss weight 1−λ are dynamically adjusted during the training process. ③ Divide the preprocessed graph data into training set, validation set and test set according to the ratio of 6:3:1; ④ Iteratively train the model in the training set, calculate the loss each iteration and update the parameters through backpropagation; ⑤ Evaluate the model performance on the validation set, adjust the hyperparameters according to the validation results to prevent overfitting; ⑥ Verify whether the results are available on the test set.

[0196] Perform pre-training in the laboratory calibration environment and then adapt to the specific scenario through online learning.

[0197] Step S5: Model evaluation and adaptive optimization:

[0198] Evaluate whether the model trained in step 4 plays the role of eliminating the near-far effect. During the evaluation process, based on the radio ground base station signals with near-far effect interference scenarios, evaluate the positioning accuracy through the radio positioning receiver terminal.

[0199] The model evaluation includes two aspects: evaluation standard indicators and system positioning performance evaluation:

[0200] a. Evaluation standard indicators:

[0201] The evaluation standard indicators include two aspects: model accuracy (MSE), signal recognition rate (F1) and bit error rate (BER).

[0202] MSE = , is the true amplitude of the i-th signal, is the prediction result of the model for the i-th sample.

[0203] F1 = , and the F1 score (F1-Score) evaluates the model recognition accuracy, which is an important indicator to measure the model recognition accuracy.

[0204] Among them, is the accuracy rate, is the recall rate.

[0205] BER = , the BER is a core metric for measuring the performance of digital communication systems, representing the ratio of the number of error bits to the total number of bits during transmission.

[0206] Among them, is the number of bits that are inconsistent between the demodulated data at the receiving end and the original data at the transmitting end, is the total number of bits transmitted by the transmitting end.

[0207] To verify the accuracy (signal reconstruction accuracy) and recognition rate (target signal detection ability) of the Transformer-GNN model in signal separation and near-far effect cancellation tasks, the following experiments are designed:

[0208] Signal modulation signals: QPSK, 16-QAM, OFDM (different modulation complexities).

[0209] Noise environment: Additive white Gaussian noise (SNR range: 0 - 20 dB).

[0210] Near-far effect simulation: The power of the near signal is 10 - 30 dB higher than that of the far signal.

[0211] Number of signal sources: 10, with 1 being the near signal.

[0212] Number of experiments: 100 times.

[0213] Experimental method: Signal recognition and near-far effect cancellation suppression are carried out through traditional power control methods, Transformer methods, GNN methods, and Transformer-GNN methods.

[0214] Based on 100 experiments, the following experimental results are obtained:

[0215] Serial number Model MSE F1 BER (1E-3) 1 Traditional power control 0.151 0.73 15.3 2 Transformer 0.049 0.84 8.2 3 GNN 0.062 0.83 9.2 4 Transformer-GNN 0.03 0.91 5.1

[0216] Accuracy advantage:

[0217] The MSE of Transformer-GNN (0.03) is about 35% lower than that of pure Transformer, proving that the joint model can reconstruct signals more accurately.

[0218] Recognition rate improvement:

[0219] At low SNR (0 - 20 dB), the F1-score still reaches 0.91, significantly better than traditional methods.

[0220] Robustness verification:

[0221] In the extreme case where the maximum difference between the near and far powers is 30 dB, the BER remains below 5.1e-3.

[0222] b. System positioning performance evaluation:

[0223] Compare the changes in the positioning errors of radio positioning receivers at different distances before and after model processing to evaluate the effect of eliminating the near-far effect.

[0224] Based on the positioning accuracy of the radio positioning terminal as the model performance evaluation criterion, the evaluation indicators are the root mean square error (RMSE) and the mean absolute error (MAE) of the system positioning.

[0225] ;

[0226] ;

[0227] Among them, is the true position, is the predicted position.

[0228] After simulation, by suppressing the near-far effect through the Transformer-GNN model, the root mean square error (RMSE) of the system is reduced to within 5 meters, and the mean absolute error (MAE) is reduced to within 3 meters.

[0229] Optimization strategy: According to the above evaluation results, adjust the structural parameters such as the number of layers and neurons of the GNN and Transformer layers. Through a large amount of learning and testing, the system can have the ability of dynamic adaptive adjustment. By performing transformations such as rotation and scaling on the training data, the data volume is expanded, and the generalization ability of the model is improved.

[0230] Key points of the present invention:

[0231] The implementation idea of realizing anti-interference in the airspace based on the Transformer-GNN deep learning method is the key technical point that the present invention first protects;

[0232] Combining Transformer and GNN to realize a new Transformer-GNN model, this method is applicable to solving the problem of eliminating the near-far effect in the radio positioning system, which is the key technical point that the present invention needs to protect;

[0233] The construction, training and optimization methods of the Transformer-GNN model are the key technical points that the present invention needs to protect.

[0234] Advantages: Innovative fusion architecture: This patent innovatively combines two cutting-edge technologies, Transformer and GNN, to break the limitations of traditional single technology. Transformer is good at capturing long-distance dependencies, and GNN is efficient in processing graph structure data. The fusion of the two enables the model to obtain the correlation between global signals and finely analyze the propagation characteristics of local signals in radio positioning. This cross-domain fusion architecture is leading in the field of navigation, but the two are not simply superimposed. A lot of simulation and experiments are required in the process of joint model construction and model training to achieve the dynamic fusion goal of complementary advantages between the two.

[0235] Leading near-far effect processing capability: Conventional methods are often limited by signal strength differences when processing near-far effects, making it difficult to take into account both near and far targets. This invention uses the Transformer self-attention mechanism, does not rely on signal strength, treats near and far signal features equally, and mines the complex relationship between far-distance and near-distance signals through global modeling. Combined with GNN's in-depth understanding of local signal propagation characteristics, it accurately eliminates near-far effects and far-far signal equalization processing far exceeds traditional solutions.

[0236] Real-time high-precision positioning guarantee: Based on the powerful ability to eliminate the near-far effect, this patented technology greatly improves the positioning accuracy of the radio positioning system, especially for long-distance receiving devices. Theoretically, it is expected that the root mean square error (RMSE) can be reduced to less than 5 meters and the mean absolute error (MAE) can be reduced to less than 3 meters, far exceeding the positioning level of similar technologies, and providing reliable support for application scenarios with extremely high requirements for positioning accuracy, such as autonomous driving and precise flight of drones. The Transformer-GNN model has a complex structure, but by optimizing the algorithm and combining it with GPU acceleration, it can meet the strict real-time requirements of the radio navigation system and provide users with accurate location information in a timely manner.

[0237] Strong adaptability to complex environments: In complex electromagnetic environments, problems such as signal interference and multipath effects seriously affect navigation performance, and radio signals are susceptible to fluctuations due to environmental interference. This technology makes the model adaptable to complex data through normalization, missing value and outlier processing in data preprocessing, and data enhancement in model training. At the same time, its unique global and local information processing capabilities can effectively deal with the influence of interference signals and multipath signals, ensure stable and accurate elimination of far and near effects in complex environments, and maintain high-precision positioning, which is difficult for traditional navigation technology to achieve. In this environment, the traditional navigation algorithm has a large fluctuation range of positioning error. When the signal fluctuates, the positioning error can instantly increase to 20~30 meters or even larger, and the error recovery time is long. The Transformer-GNN model significantly reduces the fluctuation range of positioning error through continuous learning and analysis of global and local information of the signal. Even when the signal fluctuates severely, the positioning error can be quickly recovered, showing stronger stability.

[0238] Flexible model optimizability: The structural parameters of the Transformer-GNN model of this patent, such as the number of layers of GNN and Transformer layers, the number of neurons, the number of heads of Transformer, and the hidden layer dimension, can all be flexibly adjusted, and a dynamic adjustment mechanism can be formed through training. By means of cross-validation, experimental optimization, etc., the optimal model configuration can be customized according to different application scenarios, hardware conditions, and data characteristics, enabling the technology to exhibit the best performance under different requirements, demonstrating high flexibility and optimizability.

[0239] Embodiment 3: A terminal device of a broadcast radio positioning system, including a memory, a processor, and a computer program stored on the memory. When the processor executes the program, it performs the near-far effect elimination method as described in Embodiment 1.

[0240] As mentioned above, only the specific preferred embodiments of the present invention are described, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for eliminating the near-far effect of a broadcast radio positioning system, characterized in that: The following steps are involved: Step S1: multi-source data collection and preprocessing: collecting radio ground base station data, radio positioning receiver data, interference source data and reflection source data; Perform signal alignment, feature standardization, environmental perception fusion and outlier repair on the data; Step S2: Construct a dynamic graph data structure: define ground base stations, receivers, interference sources, and fictitious multipath reflection nodes as graph nodes, construct ground station-receiver connection edges, interference source-receiver connection edges, and spatiotemporal trajectory connection edges based on signal propagation relationships, and design a dynamic update mechanism to achieve real-time update of node positions and edge weights; Step S3: Constructing a Transformer-GNN joint model: Local-global joint modeling is achieved by alternately stacking GNN layers and Transformer layers, where: The GNN layer defines message functions based on node types and extracts local topological features; The Transformer layer introduces edge attribute encoding and captures global signal associations through a sparse attention mechanism; Use residual connection and gating mechanism to dynamically fuse GNN and Transformer features; Step S4: Model training and optimization: adopt a phased training strategy, first pre-train the GNN layer, and then jointly train the Transformer-GNN model; balance the positioning loss and interference suppression loss through dynamic weight adjustment (AdaLoss); Step S5: Model evaluation and adaptive optimization: Evaluate the positioning accuracy based on the root mean square error (RMSE) and mean absolute error (MAE), adjust the model structure parameters according to the evaluation results, and use data enhancement to improve the generalization ability.

2. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 1, characterized in that: In step S2: Node types include: For a radio ground base station node, the feature vector includes three-dimensional coordinates, transmission power, carrier frequency and pseudo-random code number; Radio positioning receiver node, the feature vector includes real-time position, velocity, received signal strength vector and carrier-to-noise ratio; Interference source node, the feature vector contains the interference location, power, bandwidth and type encoding; The fictitious multipath reflection node, the feature vector includes the reflection surface coordinates, incident angle, reflection coefficient and path delay; The edge weights are calculated as follows: Ground station-receiver side: based on the free space path loss formula and obstacle attenuation model; Interference-receiver side: based on the correlation between the integral of spectrum overlap and spatial distance; Space-time trajectory edges: based on time decay factors or prediction confidence scores.

3. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 1, characterized in that: The dynamic update mechanism of step S2 includes: The update frequency can be adjusted from 50ms to 5000ms, with the default setting being 3000ms; The reconstruction mechanism is triggered when a new interference source is detected or the signal is lost. The reconstruction conditions include: The spectrum energy mutation exceeds the threshold of 3 times the background noise; No signal is detected for N consecutive epochs, where N is dynamically adjustable and the default value is 5.

4. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 1, characterized in that: The construction of the Transformer-GNN joint model in step S3 further includes: The GNN layer uses a graph convolutional network (GCN) or a graph attention network (GAT) to reduce computational complexity through neighbor sampling or TopK aggregation; The Transformer layer uses a sparse attention mechanism to encode edge attributes as position embeddings; Alternately stack GNN layers and Transformer layers 3 to 5 times, in the order of GNN→Transformer loop structure.

5. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 4, characterized in that: The residual connection and gating mechanism are implemented as follows: The residual connection directly adds the GNN output features to the Transformer output features; The gating mechanism dynamically controls the feature fusion weights based on GRU, and selects retained features through update gates and reset gates.

6. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 1, characterized in that: The dynamic weight adjustment (AdaLoss) of step S4 satisfies: The joint loss function is ,in, For the positioning loss, To suppress the loss of interference, the weight λ is dynamically adjusted during the training process; The training data is divided into training set, validation set and test set in the ratio of 6:3:1, and GPU is used to accelerate parallel training.

7. The method for eliminating the near-far effect of a broadcast radio positioning system according to claim 1, characterized in that: The adaptive optimization of step S5 includes: Adjust the number of GNN layers, number of Transformer heads, and hidden layer dimensions based on RMSE and MAE evaluation results; The training set is enhanced by data rotation and scaling to improve the model's adaptability to dynamic environments.

8. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the near-far effect elimination method described in any one of claims 1 to 7 is implemented.

9. A broadcast radio positioning system terminal device, comprising a memory, a processor and a computer program stored in the memory, characterized in that: When the processor executes the program, the near-far effect elimination method described in any one of claims 1-7 is implemented.

Images