A wireless array single base station positioning method and system based on multi-dimensional electromagnetic field map

Abstract

This invention discloses a wireless array single-base station positioning method and system based on a multi-dimensional electromagnetic field map, belonging to the field of indoor / underground space positioning technology. It utilizes high-precision trajectory acquisition methods in conjunction with the array single base station to construct an electromagnetic field map containing multi-dimensional features in the time, frequency, and spatial domains. A vector orientation method is used to construct a local relative coordinate system for the base station, converting absolute coordinates to relative coordinates to decouple the model from the geographical environment. Geometric projection preprocessing is performed based on the current scene height difference to eliminate the influence of installation height, and domain labels are introduced into the input features to adapt to different moving targets. Then, a TCN-Transformer hybrid model with a fusion channel attention mechanism is constructed and trained. Finally, the coordinates are restored using an inverse transformation matrix in new scenes, supporting rapid fine-tuning with limited data. This invention achieves high-precision, high-efficiency, and strong generalization capability positioning under a single base station.

Description

Technical Field

[0001] This invention belongs to the field of indoor / underground space positioning technology, and specifically relates to a wireless array single base station positioning method and system based on multidimensional electromagnetic field diagrams, which is applicable to wireless positioning systems with array receiving capabilities such as ultra-wideband (UWB), pseudo-satellites, and 5G. Background Technology

[0002] Currently, with the rapid development of the Internet of Things, intelligent manufacturing, and autonomous driving technologies, high-precision positioning in indoor and underground spaces is playing an increasingly important role in warehousing and logistics, personnel management, and robot navigation. Since satellite signals cannot reach indoor spaces, positioning technologies based on radio signals (such as ultra-wideband (UWB), pseudosatellites, and 5G) have become the core support for solving the "last mile" positioning problem, possessing significant social and economic value.

[0003] Traditional wireless positioning systems primarily rely on geometric measurement principles, but face numerous challenges in practical applications: multi-base station joint calculation (such as TDOA) requires complex clock synchronization and high hardware deployment costs. Traditional single-base station geometric calculation (AoA + ranging) suffers severe signal phase and time-of-flight distortion in non-line-of-sight (NLOS) and multipath-rich environments, leading to a sharp degradation in positioning performance.

[0004] To overcome the limitations of geometric methods, fingerprint localization technology has been extensively studied. However, existing fingerprint localization methods suffer from the following bottlenecks: First, they are inefficient to construct; traditional static mesh acquisition methods are time-consuming and labor-intensive, making them difficult to implement in large-scale scenarios. Second, they have poor dynamic adaptability; static fingerprints cannot reflect the Doppler effect and dynamic occlusion features when the target is moving. Third, they have weak generalization ability; existing models are usually strongly coupled with the absolute coordinates and base station installation height of a specific scene, and the model fails once the environment changes or the base station location changes. Fourth, they have poor multi-target adaptability; the motion characteristics and signal occlusion patterns of pedestrians and vehicles are completely different, and traditional single models cannot simultaneously take them into account. Summary of the Invention

[0005] To address the technical problems mentioned in the background, this invention proposes a wireless array single-base station localization method and system based on a multi-dimensional electromagnetic field map. It achieves dynamic and automated construction of the electromagnetic field map using laser SLAM technology, decouples the coordinate system from the physical environment using vector orientation and geometric height projection techniques, introduces domain adaptive labeling into the TCN-Transformer network with a fused channel attention mechanism, and finally achieves high-precision, high-efficiency, and strong generalization localization under a single base station through an adaptive fine-tuning strategy in the new environment.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A generalized localization method for a single base station of a wireless array based on a multidimensional electromagnetic field map includes the following steps:

[0008] Step 1: Bind the positioning terminal to the trajectory acquisition device and move it continuously in the area to be tested to obtain the spatial trajectory mapping of the positioning terminal; use the array single base station to collect multi-dimensional radio frequency signal features covering the time domain, frequency domain and spatial domain, and construct an electromagnetic field map containing dynamic features by time alignment of the multi-dimensional radio frequency signal features and spatial trajectory mapping.

[0009] Step 2: Based on the acquired electromagnetic field map, a local coordinate system with the array single base station as the origin is constructed using the vector orientation method. The transformation parameters from the absolute coordinate system to the local coordinate system are calculated to convert the absolute spatial position in the electromagnetic field map constructed in Step 1 into a relative coordinate position; where the absolute coordinate system is the coordinate system of the trajectory acquisition device.

[0010] Step 3: Obtain the height difference between the base station and the positioning terminal in the current scene, and perform geometric dimensionality reduction projection on the ranging value; according to the motion characteristics of the positioning target, add explicit domain labels to the input features;

[0011] Step 4: Construct a hybrid neural network model that includes a temporal convolutional network with a channel attention mechanism and a Transformer encoder; train the model using the electromagnetic field map data processed in Step 2 and the explicit domain labels in Step 3 to learn the mapping relationship between multidimensional signal features and local relative coordinates.

[0012] Step 5: During the localization process in the new environment, measure the absolute coordinates and orientation reference point of the new base station, and calculate the inverse transformation matrix from the absolute coordinate system to the local coordinate system in the new scene. If the environmental differences cause the accuracy of the pre-trained model to decrease, collect electromagnetic field map data of the new environment, freeze the TCN layer parameters for extracting physical features at the front end of the model, unfreeze the Transformer and output layer at the back end, and use the new electromagnetic field map data to quickly fine-tune the pre-processed model. During real-time localization, the model outputs relative coordinates after fine-tuning the training, and calculates the absolute position through inverse transformation.

[0013] Furthermore, the specific composition of the multidimensional radio frequency signal characteristics in step 1 includes: time-domain characteristics: including ranging observations based on time-of-flight calculations, and channel impulse response characterizing the distribution of signal energy with time delay;

[0014] Frequency domain characteristics: including carrier phase observations and channel frequency response characterizing the fading characteristics of the signal at different frequency components;

[0015] Spatial characteristics: including the phase difference matrix between each antenna element in the array receiving device, which characterizes the spatial distribution information of the signal's angle of arrival.

[0016] Furthermore, the specific method of step 2 is as follows:

[0017] Obtain the absolute coordinates of the base station center Select any point along the positive X-axis in the base station coordinate system as the orientation reference point and obtain its absolute coordinates. Calculate the rotation angle:

[0018] ,

[0019] Rotation matrix from absolute coordinate system to local coordinate system Translation vector They are represented as follows:

[0020] ,

[0021] ,

[0022] For each training data point in the electromagnetic field diagram, its contained spatial location is determined from absolute coordinates. Convert to local relative coordinates of the base station This achieves decoupling between model training and the coordinate system;

[0023] in, .

[0024] Furthermore, step 3 is specifically implemented as follows: To address the uncertainty of the base station installation height and carrier height, a physical prior is introduced during the feature extraction stage; the base station height is... The height of the positioning terminal is The height difference is expressed as:

[0025] ,

[0026] For the original distance measurement value Perform geometric dimensionality reduction projection to obtain the projected planar distance value. for:

[0027] ,

[0028] For scenarios involving mixed human and vehicle localization, a one-dimensional explicit domain label scalar is added to the input feature vector. ; Obtain the feature vector as input for the next deep learning step :

[0029] ,

[0030] in, Each represents any time step The dimension-reduced distance characteristics, array phase, channel impulse response, and domain adaptive flags related to the time array base station.

[0031] Furthermore, step 4 is specifically implemented as follows: a deep neural network is constructed, with a stacked residual temporal convolutional network used for the feature extraction layer; each residual block embeds an SE-Block, which automatically suppresses antenna signals severely affected by multipath interference by learning the importance weights of each antenna channel; the temporal modeling layer uses a Transformer encoder, which captures long-distance temporal dependencies using a self-attention mechanism; the output layer regresses and predicts the two-dimensional local coordinates relative to the base station; the objective function is iteratively optimized based on the AdamW optimizer and the SmoothL1 loss function; a random masking strategy is introduced during training to simulate signal packet loss and improve robustness.

[0032] Feature extraction layer: Extracts local features using dilated causal convolution; for the input sequence convolution kernel and expansion rate The convolution operation is defined as:

[0033] ;

[0034] To suppress multipath interference, SE-Block is introduced to calculate the channel weight vector. Recalibrate the feature map:

[0035] ;

[0036] in, This represents the intermediate feature tensor output by the dilated convolution, which contains unweighted multi-channel temporal features. This represents the attention-weighted enhanced feature tensor;

[0037] Temporal modeling layer: Introduces positional encoding (PE) to preserve sequence order information, utilizes multi-head self-attention mechanism to calculate global temporal correlation, and outputs context vector;

[0038] Objective function: The optimization objective is constructed using a smooth L1 loss function.

[0039] ,

[0040] in When the error is small, it manifests as squared loss, and when the error is large, it manifests as absolute value loss, thereby enhancing the robustness of the system; This represents the total number of samples in the current training batch. This represents the local relative coordinates of the base station predicted by the model for the i-th sample; This represents the local relative coordinates of the real base station corresponding to the i-th sample.

[0041] Furthermore, the specific method for fine-tuning the training of some network layers in step 5 is as follows: collect electromagnetic field map data in the new scene, and repeat steps 2 and 3 to achieve the same preprocessing; freeze the weight parameters of the temporal convolutional network module and SE-Block used to extract the underlying physical signal features in the hybrid neural network model; unfreeze the weight parameters of the Transformer encoder module and the output layer; use the electromagnetic field map data of the new scene to iteratively update the parameters of the unfrozen layer at a learning rate lower than the initial training rate, so that the model can adapt to the multipath distribution characteristics of the new environment;

[0042] Freeze phase: Lock the front-end parameters so that their gradients do not participate in backpropagation, i.e. To retain the general physical layer signal extraction capability;

[0043] Update phase: Apply gradient descent to update the backend parameters:

[0044] ;

[0045] in, This represents the base parameters / front-end parameters; this part of the network is responsible for extracting the underlying physical features of the signal. For the backend parameters, this part of the network is responsible for mapping abstract signal features to specific geometric space coordinates. Since the multipath reflection structure changes in the new environment, it needs to be retrained. For gradient operators, To fine-tune the learning rate, set it to 1 / 10 of the initial training learning rate;

[0046] Finally, using the formula Restore the relative coordinates output by the model to the absolute coordinates in the new scene.

[0047] A wireless array single base station generalized positioning system is used to implement the above-mentioned wireless array single base station generalized positioning method based on multidimensional electromagnetic field map, including a signal acquisition module, a trajectory acquisition module, a data preprocessing module, a model training and fine-tuning module, and a coordinate post-processing module.

[0048] The signal acquisition module completes the acquisition of multi-dimensional radio frequency signal features covering the time, frequency, and spatial domains in step 1; the trajectory acquisition module completes the trajectory acquisition in step 1; the data preprocessing module is responsible for coordinate system decoupling and standardization in step 2 and step 3; the model training module completes the related tasks in step 4; the fine-tuning module is responsible for the fine-tuning training of the pre-trained model in step 5; and the coordinate post-processing module converts relative coordinates into absolute positions through inverse transformation calculation.

[0049] Due to the adoption of the above technical solution, the beneficial effects of this invention compared with the prior art are as follows:

[0050] 1. A multi-dimensional electromagnetic field map integrating the time, frequency, and spatial domains was constructed. This overcomes the limitations of traditional fingerprint positioning, which relies solely on the energy domain (RSSI). It fully utilizes ranging and CIR waveforms in the time domain, phase fading in the frequency domain, and array angle-of-arrival features in the spatial domain, significantly improving location accuracy. 2. Extremely high database construction efficiency and data completeness. Utilizing a dynamic database construction mode assisted by laser SLAM for "walk-and-collect" operation, efficiency is several times higher than traditional static grid acquisition, and it records the dynamic characteristics of the target in motion.

[0051] 2. The model achieves decoupling from the geographical environment, altitude, and diverse targets (strong generalization ability). Through "vector orientation" and "base station local coordinate system," the model no longer "memorizes" the map coordinates of specific rooms, but instead learns the physical mapping between signal characteristics and relative positions, eliminating the need for retraining due to base station fine-tuning. Geometric height projection technology eliminates the impact of base station installation height changes on the model; and by introducing explicit domain labeling, it solves the problem of a single model being unable to simultaneously adapt to conflicting characteristics of pedestrians (low-speed non-rigid bodies) and vehicles (high-speed rigid bodies).

[0052] 3. Strong anti-interference capability and low deployment cost. The hybrid network structure integrating channel attention mechanism can effectively suppress NLOS multipath interference. Combined with adaptive fine-tuning strategy, the system only needs to collect a very small amount of data to quickly adapt to new and complex metal environments. Attached Figure Description

[0053] Figure 1 This is a flowchart of a wireless array single base station positioning method based on a multidimensional electromagnetic field map in an embodiment of the present invention.

[0054] Figure 2 This is a geometric diagram illustrating the construction of the local coordinate system of the base station and the vector orientation method in an embodiment of the present invention.

[0055] Figure 3 This is a spatiotemporal fusion neural network model architecture that includes a channel attention mechanism in an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] This invention designs a wireless array single base station positioning method and system based on multidimensional electromagnetic field maps, such as... Figure 1 As shown, it includes the following steps:

[0058] Step 1: Construct a multi-dimensional electromagnetic field map. The positioning terminal is bound to a high-precision trajectory acquisition device (such as laser SLAM) and continuously moved within the area to be measured. Multi-dimensional radio frequency signal features covering the time, frequency, and spatial domains are collected using an array of single base stations. Specifically, this includes: ranging observations and channel impulse response (CIR) in the time domain, carrier phase observations and channel frequency response (CFR) in the frequency domain, and the array antenna phase difference matrix (representing AoA) in the spatial domain. The multi-dimensional signal features are aligned with the high-precision absolute trajectory acquired by the high-precision trajectory acquisition device through time alignment, constructing an electromagnetic field map containing dynamic features.

[0059] Step 2, Coordinate System Decoupling and Standardization. A local relative coordinate system with the array single base station as the origin is constructed using the vector orientation method. The transformation parameters from the absolute coordinate system (high-precision trajectory acquisition device coordinate system) to the local coordinate system (single base station coordinate system) are calculated, and the absolute spatial position in the electromagnetic field diagram constructed in Step 1 is converted into relative coordinates.

[0060] Step 3, Geometric and Domain Adaptation Feature Preprocessing. Obtain the height difference between the base station and the positioning terminal in the current scene, and perform geometric dimensionality reduction projection on the ranging values ​​to eliminate the influence of height. Based on the motion characteristics of the positioning target (e.g., person or vehicle), add explicit domain labels to the input features.

[0061] Step 4, Model Construction and Training. A hybrid neural network model is constructed, incorporating a Temporal Convolutional Network (TCN) with a Channel Attention (SE-Block) mechanism and a Transformer encoder. The model is trained using the electromagnetic field map data processed in Step 2 and the explicit domain labels from Step 3, learning the mapping relationship between multidimensional signal features and local relative coordinates.

[0062] Step 5: When the system is deployed in a completely new environment, measure the absolute coordinates and orientation reference point of the new base station, and automatically calculate the inverse transformation matrix for the new scene. If environmental differences cause a decrease in the accuracy of the pre-trained model, collect a small amount of electromagnetic field map data in space, freeze the TCN layer parameters for extracting physical features at the front end of the model, unfreeze the Transformer and output layers at the back end, and perform rapid fine-tuning training using the new electromagnetic field map data. During real-time positioning, the model outputs relative coordinates, and the absolute position is calculated through inverse transformation.

[0063] The specific method for step 2 is as follows:

[0064] Obtain the absolute coordinates of the base station center Choose any point along the positive X-axis in the base station coordinate system as the orientation reference point, such as... Figure 2 As shown, obtain its absolute coordinates. Calculate the rotation angle:

[0065] ,

[0066] Then the rotation matrix from the absolute coordinate system to the local coordinate system of the base station Translation vector They can be represented as:

[0067] ,

[0068] ,

[0069] For each training data point in the electromagnetic field diagram, its contained spatial location is determined from absolute coordinates. Convert to local relative coordinates of the base station This achieves decoupling of model training from the coordinate system.

[0070] ,

[0071] The specific method for step 3 is as follows:

[0072] To address the uncertainty of base station installation height and carrier height, a physical prior is introduced during the feature extraction stage. Let the base station height be... The height of the positioning terminal is The height difference can then be expressed as:

[0073] ,

[0074] For the original distance measurement value By performing geometric dimensionality reduction projection, the projected planar distance value can be obtained. for:

[0075] ,

[0076] For scenarios involving mixed human and vehicle localization, a one-dimensional explicit domain label scalar is added to the input feature vector. This design essentially provides a "context switch" for the neural network, enabling it to dynamically adjust its internal activation paths based on the target type.

[0077] Therefore, the feature vector for the next deep learning input can be obtained. :

[0078] ,

[0079] in, Each represents any time step The dimension-reduced distance characteristics (time domain), array phase (frequency domain and spatial domain), channel impulse response (time domain), and domain adaptive flag bits related to the time array base station.

[0080] The specific method for step 4 is as follows:

[0081] A spatiotemporal fusion neural network model is constructed, with the feature extraction layer employing stacked residual temporal convolutional networks (Residual TCNs). For example... Figure 3 As shown, each residual block embeds an SE-Block, which automatically suppresses antenna signals severely affected by multipath interference by learning the importance weights of each antenna channel. The timing modeling layer employs a Transformer encoder, utilizing a self-attention mechanism to capture long-distance timing dependencies. The output layer regresses and predicts the two-dimensional local coordinates relative to the base station. The objective function is iteratively optimized based on the AdamW optimizer and the SmoothL1 loss function. A random masking strategy is introduced during training to simulate signal packet loss and improve robustness.

[0082] Feature extraction layer (TCN module): Extracts local features using dilated causal convolution. For the input sequence convolution kernel and expansion rate The convolution operation is defined as:

[0083] ,

[0084] To suppress multipath interference, SE-Block is introduced to calculate the channel weight vector. Recalibrate the feature map:

[0085] ,

[0086] in, This represents the intermediate feature tensor output by dilated convolution, which contains unweighted multi-channel temporal features. This represents the attention-weighted enhanced feature tensor;

[0087] Temporal modeling layer (Transformer module): Introduces positional encoding (PE) to preserve sequence order information, utilizes multi-head self-attention mechanism to calculate global temporal correlation, and outputs context vector.

[0088] Objective function: The optimization objective is constructed using the Smooth L1 Loss function.

[0089] ,

[0090] in When the error is small, it manifests as squared loss, and when the error is large, it manifests as absolute value loss, thereby enhancing the robustness of the system. This represents the total number of samples in the current training batch (Batch Size). This represents the local relative coordinates of the base station predicted by the model for the i-th sample. This represents the true local relative coordinates (real value) of the base station corresponding to the i-th sample.

[0091] The specific method for step 5 is as follows:

[0092] Collect a small amount of electromagnetic field map data in the new scene and perform the same preprocessing as in steps 2 and 3; freeze the weight parameters of the Temporal Convolutional Network (TCN) module and SE-Block in the hybrid neural network model used to extract the underlying physical signal features; unfreeze the weight parameters of the Transformer encoder module and the output layer; use the electromagnetic field map data of the new scene to iteratively update the parameters of the unfrozen layer at a learning rate lower than the initial training rate, so that the model can adapt to the multipath distribution characteristics of the new environment.

[0093] Freezing phase: Locking the front-end parameters (including TCN and SE-Block) so that their gradients do not participate in backpropagation, i.e. To retain the general physical layer signal extraction capability.

[0094] Update phase: Gradient descent is applied only to the backend parameters (including Transformer and fully connected layers).

[0095] ,

[0096] in, This represents the base / front-end parameters. This part of the network is responsible for extracting the underlying physical characteristics of the signal (such as pulse waveform structure and phase transition patterns). These characteristics are determined by the signal transmission mechanism and are independent of the environment, so they should remain unchanged in new scenarios. These are the backend parameters; this part of the network is responsible for mapping abstract signal features to specific geometric coordinates. Because the multipath reflection structure has changed in the new environment, the mapping relationship from features to coordinates has changed, therefore retraining is required. For gradient operators, To fine-tune the learning rate, it is usually set to 1 / 10 of the initial training learning rate.

[0097] Finally, using the formula Restore the relative coordinates output by the model to the absolute coordinates in the new scene.

[0098] In summary, this invention proposes a wireless array single-base station localization method based on multidimensional electromagnetic field maps. By utilizing laser SLAM technology to dynamically and automatically construct electromagnetic field maps containing time-frequency-spatial multidimensional features, and by tightly coupling the TCN-Transformer network (which decouples vector-oriented coordinates, performs geometric height projection preprocessing, domain adaptive labeling, and integrates channel attention mechanisms) into a unified deep learning framework, this invention solves the problems faced by traditional fingerprint localization technologies, such as high database construction and maintenance costs, lack of dynamic features, and strong dependence of the model on absolute environmental coordinates and base station installation. This method achieves high-precision, high-efficiency, and strong generalization capabilities in complex indoor multipath scenarios.

[0099] This embodiment also provides a system for implementing a wireless array single base station positioning method based on a multidimensional electromagnetic field map, which includes a signal acquisition module, a trajectory acquisition module, a data preprocessing module, a model training and fine-tuning module, and a coordinate post-processing module.

[0100] The signal acquisition module completes the acquisition of multi-dimensional radio frequency signal features covering the time, frequency, and spatial domains in step 1; the trajectory acquisition module completes the acquisition of absolute trajectory in step 1; the data preprocessing module is responsible for the two tasks of coordinate system decoupling and standardization in step 2 and geometric and domain adaptation feature preprocessing in step 3; the model training module completes the related tasks in step 4; the fine-tuning module is responsible for the fine-tuning training of the pre-trained model in step 5; and the coordinate post-processing module converts relative coordinates into absolute positions through inverse transformation calculation.

[0101] For those skilled in the art, various modifications and variations can be made to this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of the claims of this invention.

Claims

1. A generalized positioning method for a single base station of a wireless array based on a multidimensional electromagnetic field map, characterized in that, Includes the following steps: Step 1: Bind the positioning terminal to the trajectory acquisition device and move it continuously in the area to be tested to obtain the spatial trajectory mapping of the positioning terminal; use the array single base station to collect multi-dimensional radio frequency signal features covering the time domain, frequency domain and spatial domain, and construct an electromagnetic field map containing dynamic features by time alignment of the multi-dimensional radio frequency signal features and spatial trajectory mapping. Step 2: Based on the acquired electromagnetic field map, a local coordinate system with the array single base station as the origin is constructed using the vector orientation method. The transformation parameters from the absolute coordinate system to the local coordinate system are calculated to convert the absolute spatial position in the electromagnetic field map constructed in Step 1 into a relative coordinate position; where the absolute coordinate system is the coordinate system of the trajectory acquisition device. Step 3: Obtain the height difference between the base station and the positioning terminal in the current scene, and perform geometric dimensionality reduction projection on the ranging value; according to the motion characteristics of the positioning target, add explicit domain labels to the input features; Step 4: Construct a hybrid neural network model that includes a temporal convolutional network with a channel attention mechanism and a Transformer encoder; train the model using the electromagnetic field map data processed in Step 2 and the explicit domain labels in Step 3 to learn the mapping relationship between multidimensional signal features and local relative coordinates. Step 5: During the localization process in the new environment, measure the absolute coordinates and orientation reference point of the new base station, and calculate the inverse transformation matrix from the absolute coordinate system to the local coordinate system in the new scene. If the environmental differences cause the accuracy of the pre-trained model to decrease, collect electromagnetic field map data of the new environment, freeze the TCN layer parameters for extracting physical features at the front end of the model, unfreeze the Transformer and output layer at the back end, and use the new electromagnetic field map data to quickly fine-tune the pre-processed model. During real-time localization, the model outputs relative coordinates after fine-tuning the training, and calculates the absolute position through inverse transformation.

2. The wireless array single base station generalized positioning method based on multidimensional electromagnetic field maps according to claim 1, characterized in that, The specific composition of the multidimensional radio frequency signal characteristics in step 1 includes: time domain characteristics: including ranging observations based on time-of-flight calculations, and channel impulse response characterizing the distribution of signal energy with time delay; Frequency domain characteristics: including carrier phase observations and channel frequency response characterizing the fading characteristics of the signal at different frequency components; Spatial characteristics: including the phase difference matrix between each antenna element in the array receiving device, which characterizes the spatial distribution information of the signal's angle of arrival.

3. The wireless array single base station generalized positioning method based on multidimensional electromagnetic field maps according to claim 2, characterized in that, The specific method for step 2 is as follows: Obtain the absolute coordinates of the base station center Select any point along the positive X-axis in the base station coordinate system as the orientation reference point and obtain its absolute coordinates. Calculate the rotation angle: ； Rotation matrix from absolute coordinate system to local coordinate system Translation vector They are represented as follows: ， ， For each training data point in the electromagnetic field diagram, its contained spatial location is determined from absolute coordinates. Convert to local relative coordinates of the base station This achieves decoupling between model training and the coordinate system; in, .

4. The wireless array single base station generalized positioning method based on multidimensional electromagnetic field maps according to claim 3, characterized in that, The specific method of step 3 is as follows: To address the uncertainty of the base station installation height and carrier height, a physical prior is introduced during the feature extraction stage; the base station height is... The height of the positioning terminal is The height difference is expressed as: ， For the original distance measurement value Perform geometric dimensionality reduction projection to obtain the projected planar distance value. for: ， For scenarios involving mixed human and vehicle localization, a one-dimensional explicit domain label scalar is added to the input feature vector. ; Obtain the feature vector as input for the next deep learning step : ， in, Each represents any time step The dimension-reduced distance characteristics, array phase, channel impulse response, and domain adaptive flags related to the time array base station.

5. A generalized positioning method for a single base station of a wireless array based on a multidimensional electromagnetic field map according to claim 4, characterized in that, The specific method of step 4 is as follows: a deep neural network is constructed, and the feature extraction layer adopts a stacked residual temporal convolutional network; each residual block embeds an SE-Block, which automatically suppresses antenna signals severely affected by multipath interference by learning the importance weights of each antenna channel; the temporal modeling layer adopts a Transformer encoder, which uses a self-attention mechanism to capture long-distance temporal dependencies; the output layer regresses and predicts the two-dimensional local coordinates relative to the base station; the objective function is iteratively optimized based on the AdamW optimizer and the SmoothL1 loss function; a random masking strategy is introduced during training to simulate signal packet loss and improve robustness; Feature extraction layer: Extracts local features using dilated causal convolution; for the input sequence convolution kernel and expansion rate The convolution operation is defined as: ； To suppress multipath interference, SE-Block is introduced to calculate the channel weight vector. Recalibrate the feature map: ； in, This represents the intermediate feature tensor output by the dilated convolution, which contains unweighted multi-channel temporal features. This represents the attention-weighted enhanced feature tensor; Temporal modeling layer: Introduces positional encoding (PE) to preserve sequence order information, utilizes multi-head self-attention mechanism to calculate global temporal correlation, and outputs context vector; Objective function: The optimization objective is constructed using a smooth L1 loss function. ， in When the error is small, it manifests as squared loss, and when the error is large, it manifests as absolute value loss, thereby enhancing the robustness of the system; This represents the total number of samples in the current training batch. This represents the local relative coordinates of the base station predicted by the model for the i-th sample; This represents the local relative coordinates of the real base station corresponding to the i-th sample.

6. The wireless array single base station generalized positioning method based on multidimensional electromagnetic field maps according to claim 5, characterized in that, The specific method for fine-tuning the network layers of the model in step 5 is as follows: collect electromagnetic field map data in the new scene, and repeat steps 2 and 3 to achieve the same preprocessing; freeze the weight parameters of the temporal convolutional network module and SE-Block used to extract the underlying physical signal features in the hybrid neural network model; unfreeze the weight parameters of the Transformer encoder module and the output layer; use the electromagnetic field map data of the new scene to iteratively update the parameters of the unfrozen layer at a learning rate lower than the initial training rate, so that the model can adapt to the multipath distribution characteristics of the new environment. Freeze phase: Lock the front-end parameters so that their gradients do not participate in backpropagation, i.e. To retain the general physical layer signal extraction capability; Update phase: Apply gradient descent to update the backend parameters: ； in, This represents the base parameters / front-end parameters; this part of the network is responsible for extracting the underlying physical features of the signal. For the backend parameters, this part of the network is responsible for mapping abstract signal features to specific geometric space coordinates. Since the multipath reflection structure changes in the new environment, it needs to be retrained. For gradient operators, To fine-tune the learning rate, set it to 1 / 10 of the initial training learning rate; Finally, using the formula Restore the relative coordinates output by the model to the absolute coordinates in the new scene.

7. A wireless array single base station generalized positioning system, used to implement the wireless array single base station generalized positioning method based on a multidimensional electromagnetic field map as described in any one of claims 1 to 6, characterized in that: It includes a signal acquisition module, a trajectory acquisition module, a data preprocessing module, a model training and fine-tuning module, and a coordinate post-processing module; The signal acquisition module completes the acquisition of multi-dimensional radio frequency signal features covering the time, frequency, and spatial domains in step 1; the trajectory acquisition module completes the trajectory acquisition in step 1; the data preprocessing module is responsible for coordinate system decoupling and standardization in step 2 and step 3; the model training module completes the related tasks in step 4; the fine-tuning module is responsible for the fine-tuning training of the pre-trained model in step 5; and the coordinate post-processing module converts relative coordinates into absolute positions through inverse transformation calculation.

Images