A Sensing and Localization Method for Time-Sensitive Targets in Highly Dynamic and Complex Electromagnetic Environments

Abstract

This invention discloses a perception and localization method for time-sensitive targets in highly dynamic and complex electromagnetic environments. The method constructs a hierarchical collaborative perception system: First, a multi-agent cluster utilizes a space-frequency anti-aliasing generation network to complete and invert sparsely acquired electromagnetic environment data, achieving coarse target localization through peak detection and Gaussian weighting. Second, a UAV approaches the target area, acquiring high-precision visual and electromagnetic data, which is then fused and decoded at the feature level through a dual-modal fusion network, outputting centimeter-level precise localization coordinates. Finally, the precise localization coordinates are used as real-time observations for adaptive Kalman filtering. A signal-to-noise ratio decision mechanism dynamically selects either a sparse or deep Gaussian process regression model, estimating and adjusting the measurement noise covariance online, thereby achieving multi-source information fusion and continuous, stable, and adaptive tracking of the target state.

Description

Technical Field

[0001] This invention relates to electronic reconnaissance technology, specifically to a method for sensing and locating time-sensitive targets in highly dynamic and complex electromagnetic environments. Background Technology

[0002] While existing technologies for accurately sensing and locating time-sensitive targets in highly dynamic and complex electromagnetic environments have made progress, they all suffer from significant limitations. In practical applications, unmanned swarms face complex and ever-changing environments, primarily characterized by high dynamism and strong interference. On one hand, high-value targets such as modern radar and communication equipment possess agile radiation parameter modulation capabilities, resulting in extremely short reliable signal interception windows, making them highly dynamic electromagnetically time-sensitive targets. These targets appear and disappear randomly, and location opportunities are strictly limited to brief time windows. On the other hand, in adversarial environments, strong electromagnetic interference becomes a significant factor hindering the execution of unmanned swarm missions. Strong electromagnetic interference not only degrades the signal-to-interference-plus-noise ratio of intercepted signals, drastically compressing the effective communication distance between individuals, leading to a degradation or limitation of swarm communication capabilities, and causing communication link interruptions between some individuals, thus fragmenting the unmanned swarm system and preventing it from completing autonomous collaborative tasks; it also significantly weakens the payload's ability to sense and locate target radiation sources, increasing the uncertainty of environmental situational awareness. This highly dynamic and strong electromagnetic adversarial environment presents unprecedented and severe challenges to the intelligent perception of unmanned swarm systems.

[0003] Currently, radiation source localization technologies are mainly divided into two categories: active localization and passive localization. While active localization systems can achieve high positioning accuracy, their active signal transmission makes them highly susceptible to revealing their location, severely limiting their application in scenarios with extremely high concealment requirements. Passive localization, on the other hand, can achieve precise positioning simply by receiving radiation source signals through a receiver, and it offers high security. Current passive localization methods primarily rely on single sensors or simple multi-sensor combinations. However, single sensors experience a sharp deterioration in signal-to-noise ratio under complex electromagnetic interference, resulting in significant tracking errors for high-speed moving targets and an inability to effectively distinguish between multiple radiation source targets. Furthermore, relying solely on analyzing the intensity and frequency of electromagnetic signals for target localization makes it difficult to achieve high-precision positioning of radiation source targets in complex electromagnetic environments. Therefore, it is necessary to utilize multi-source, multi-modal data to improve positioning performance.

[0004] In the field of data fusion, traditional methods struggle to meet the demands of time-sensitive target perception and localization in highly dynamic and complex electromagnetic environments. Multimodal data (such as electromagnetic spectrum, radar, infrared, and photoelectric data) are complementary in practical applications, but traditional data fusion methods lack effective means to process them. Data collected by different sensors differ in resolution, spatiotemporal characteristics, and other aspects. Traditional methods struggle to perform accurate spatiotemporal alignment and feature fusion on this data, failing to fully leverage the advantages of multimodal data and hindering improvements in positioning accuracy.

[0005] In target state estimation, traditional methods typically employ fixed models and parameters, making it difficult to adapt to the highly dynamic characteristics of time-sensitive targets and the changing complex electromagnetic environments. The motion state of a time-sensitive target (such as velocity and acceleration) may change rapidly, and traditional models cannot adjust in time to accurately track these changes, leading to inaccurate target state estimation. Furthermore, traditional methods often lack the ability to quantify the uncertainty of the estimation results, failing to provide comprehensive and reliable information support for subsequent decision-making.

[0006] In summary, existing technologies have many shortcomings in sensing and locating time-sensitive targets in highly dynamic and complex electromagnetic environments, making it difficult to meet the needs of high-precision, high-reliability positioning and accurate estimation of target status in practical applications. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a solution for the difficulty of achieving high-precision and high-reliability perception, positioning and dynamic state estimation of time-sensitive targets in highly dynamic and complex electromagnetic environments by constructing an integrated technology system of multi-agent collaborative hierarchical perception and adaptive dynamic tracking.

[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is a sensing and localization method for time-sensitive targets in highly dynamic and complex electromagnetic environments, comprising the following steps:

[0009] A multi-agent cluster collects electromagnetic environment data according to a preset discrete grid, inputs it into a space-frequency anti-aliasing generation network, completes the electromagnetic environment data of the target area, and obtains the electromagnetic field distribution tensor of the target area.

[0010] Peak detection is performed on the electromagnetic field distribution tensor, and a Gaussian weighted average is applied to the area around the peak coordinates to output the coarse positioning coordinates of the target.

[0011] The high-precision visual data and electromagnetic data collected by multiple agents in the coarse positioning area determined by the coarse positioning coordinates are preprocessed to obtain visual tensors and electromagnetic data tensors. The visual tensors and electromagnetic data tensors are then input into a dual-modal fusion network to output the fine positioning coordinates of each agent for the detected target.

[0012] Using the precise positioning coordinates output by the multi-agent system as real-time observations, combined with historical observation data, the system dynamically adjusts parameters through adaptive filtering and fuses multi-source information to output a continuous and stable dynamic state of the target, thereby achieving continuous tracking of highly dynamic targets. During the update process, the measurement noise covariance matrix of the adaptive filter is adaptively adjusted by choosing between a sparse Gaussian process regression model and a deep Gaussian process regression model based on the real-time signal-to-noise ratio decision.

[0013] In the initial sensing stage, this invention performs electromagnetic environment inversion based on the incomplete electromagnetic field strength of the target radiation source intercepted within the observation window, and builds a space-frequency anti-aliasing generation network. Next, it performs coarse positioning of the radiation source target based on the electromagnetic field distribution generated by the space-frequency anti-aliasing generation network. Then, as information accumulates during the approach to the target, multi-modal data is fused to achieve fine positioning of the target. Finally, a Gaussian process regression module is introduced into the Kalman filter target state estimation and the model parameters are adaptively updated to form a closed-loop collaborative sensing system for adaptive target tracking, completing the sensing and positioning of time-sensitive targets in highly dynamic and complex electromagnetic environments.

[0014] The beneficial effects of this invention are:

[0015] (1) Effectively realize the perception, localization and state estimation of high-value time-sensitive targets that are fast-moving and randomly appearing.

[0016] (2) Effectively realize the complete supplementation of electromagnetic environment data in the target area and the collaborative aggregation of observation data of multiple agents, enhance the anti-interference capability at the data level, and optimize the underlying data support capability of the perception and positioning system.

[0017] (3) Effectively achieve hierarchical and progressive precise positioning and dynamic adaptive continuous tracking of target state, enhance the system's adaptability to the high dynamic characteristics of the target, and optimize the continuity and stability of perception and positioning in complex scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a space-frequency anti-aliasing generator network;

[0019] Figure 2 This is a structural diagram of a dual-modal fusion network;

[0020] Figure 3 This is a flowchart of the sensing and positioning method in an embodiment. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0022] Example

[0023] like Figure 1 As shown, the space-frequency anti-aliasing generation network includes a fusion encoder and a space-frequency anti-aliasing generator.

[0024] Perception-based localization comprises two stages: coarse target localization and fine target localization. The coarse target localization stage utilizes a space-frequency anti-aliasing generation network, while the fine target localization stage employs a dual-modal fusion network. The following is a detailed introduction to the space-frequency anti-aliasing generation network and the dual-modal fusion network:

[0025] The fusion encoder is responsible for extracting and fusing geo-electromagnetic heterogeneous features, comprising two parallel two-layer convolutional unit branches, a feature fusion layer, and an activation layer. The inputs to the two convolutional unit branches are the geographic tensor G and the electromagnetic environment tensor E, respectively. The two branches extract the spatial features of the two types of input tensor data and output them to the feature fusion layer. The feature fusion layer performs a linear transformation to achieve heterogeneous feature fusion and outputs the result to the ReLU activation layer. The ReLU activation layer enhances the nonlinear expression of the fused heterogeneous features, ultimately generating a low-dimensional latent variable z carrying the core style features and outputting it to the space-frequency anti-aliasing generation network. In this embodiment, parallel two-layer 3*3 convolutional units are used with a stride of 1 and padding of 1 to ensure that the spatial resolution remains unchanged.

[0026] The space-frequency anti-aliasing generator is based on the third-generation style generative adversarial network StyleGAN-v3 architecture. The core of StyleGAN-v3 is to construct a generator that remains stable under any continuous geometric transformation. The latent variable z is first input into the style mapping layer in StyleGAN-v3. The style mapping layer transforms the latent variable z into a style vector containing scaling and offset factors through three fully connected layers (3×FC). Style vectors First, the first spatial frequency style convolutional block, Block-1, is input. After the spatial frequency style convolutional block integrates spatial feature convolution (SF-Conv), adaptive instance normalization (AdaIN), and ReLU activation, it is input to the bilinear interpolation upsampling unit to improve feature resolution. The upsampled features are then input to the second spatial frequency style convolutional block, Block-2. The output of spatial frequency style convolutional block 2 is then upsampled by bilinear interpolation and sequentially input to the third spatial frequency style convolutional block, Block-3, and the fourth spatial frequency style convolutional block, Block-4. After another bilinear interpolation upsampling, it is finally input to the fifth spatial frequency style convolutional block, Block-5. After Block-5 completes feature reconstruction, it outputs the electromagnetic field distribution tensor that can cover the target region, generated by the spatial-frequency anti-aliasing generator, through the output layer.

[0027] Specifically, the training process of the space-frequency anti-aliasing generator network is as follows:

[0028] 1) Download electromagnetic environment data and corresponding real electromagnetic field distribution tags E for the region of interest from the space-spectrum database. true The electromagnetic environment data includes the original geographic tensor G and the original electromagnetic environment tensor E. The electromagnetic environment data is normalized to eliminate the impact of dimensional differences on network training.

[0029] In this embodiment, the normalized electromagnetic environment data is arranged in a discrete grid array. This discrete grid array is a three-dimensional regular spatial sampling array capable of covering the target detection area. The grid density is dynamically adapted according to the area range and positioning accuracy requirements. The normalized electromagnetic environment data includes geographic coordinates and the electromagnetic spectrum; wherein, the geographic coordinates are represented as... , Indicates the index of the raster acquisition point. They represent the first The original longitude, latitude, and altitude of each grid acquisition point, with the superscript T indicating transpose; electromagnetic spectrum. Frequency point In the Original electromagnetic field strength values ​​at each grid acquisition point, frequency index , The total number of frequency points; simultaneously, the actual electromagnetic field distribution tensor at the corresponding grid acquisition point is extracted as the label E for network training. true ;

[0030] 2) Initialize the learnable parameters of the fusion encoder and the space-frequency anti-aliasing generator, set the optimizer to Adam, and define the loss function as the mean square error between the reconstructed electromagnetic field distribution tensor output by the network and the true global electromagnetic field distribution label;

[0031] 3) Input the global geographic tensor and global electromagnetic tensor of the electromagnetic environment data into the fusion encoder. After feature extraction, fusion and activation, a low-dimensional latent variable z is generated. Input the latent variable z into the space-frequency anti-aliasing generator. After style mapping, generation decoding and output layer mapping, the reconstructed electromagnetic field distribution tensor is obtained.

[0032] 4) Calculate the loss value of the reconstructed tensor and label based on the loss function, and traverse each layer of the network through the backpropagation algorithm to synchronously update all learnable parameters of the fusion encoder and the space-frequency anti-aliasing generator;

[0033] In this embodiment, the loss function of the space-frequency anti-aliasing generator network is calculated based on the reconstructed tensor and labels output by the network. for:

[0034] ;

[0035] in, The reconstructed tensor output by the network is located in space. Frequency The element value at that position; For height indexing, For width index; To label the spatial location of the true global electromagnetic field distribution Frequency The element value at that position; The total number of elements in the tensor. This represents the maximum height. This is the maximum width. This represents the maximum frequency value.

[0036] 5) Repeat steps 4)-5) until the loss value converges, stop training and save the optimal network parameters to obtain the trained space-frequency anti-aliasing generator network.

[0037] like Figure 2 As shown, the dual-modal fusion network includes a feature extraction module, a multimodal fusion module, a latent space mapping module, and a consistency decoding module.

[0038] The feature extraction module includes a visual feature extraction branch and an electromagnetic feature extraction branch. In the visual feature extraction branch, the input layer receives a visual tensor, which is then processed by a pre-trained ResNet-50 for hierarchical feature extraction. The feature dimension is then compressed through a 1×1 convolutional layer, and finally output as a high-dimensional visual feature vector through a global average pooling layer. In the electromagnetic feature extraction branch, the input layer receives an electromagnetic data tensor. After the data dimension is integrated by a channel concatenation unit, the spatial field strength features are extracted sequentially through three cascaded parallel convolutional units. Each parallel convolutional unit consists of a 3×3 convolutional Conv and ReLU activation. The output is then mapped through a fully connected layer to produce a high-dimensional electromagnetic feature vector. The visual feature vector and the electromagnetic feature vector undergo a dot product operation to complete the initial feature association before being input into the multimodal fusion module.

[0039] In the multimodal fusion module, the features input after initial association are aligned by a linear transformation layer. Then, the attention calculation layer calculates the association weights of electromagnetic and visual features to enhance effective information. Finally, the high-dimensional fusion feature tensor is output to the latent space mapping module through the fusion output layer.

[0040] In the latent space mapping module, the fused feature tensor is first converted into vector form by a flattening layer, and then the latent space feature is mapped by two cascaded fully connected layer units. The fully connected layer unit consists of a fully connected layer and ReLU activation. The latent space mapped feature input is consistent with the decoding module.

[0041] In the consistency decoding module, the features mapped in the latent space are restored to their dimensionality by a fully connected decoding layer, and then the coordinate data distribution is restored by an inverse normalization layer, finally outputting the precise positioning coordinates of the target.

[0042] Specifically, the training process of the dual-modal fusion network is as follows:

[0043] 1) Download visual-electromagnetic pairing data and corresponding real-world location coordinate labels from the multimodal target localization database, and perform preprocessing on the data;

[0044] In this embodiment, the high-precision multimodal data includes visual data and electromagnetic data; wherein, the visual data is a high-precision rasterized image, the data format is a three-dimensional tensor, and the three channels correspond to RGB grayscale values ​​respectively, carrying target contour and texture information; the electromagnetic data is... ;

[0045] The preprocessing process for high-precision multimodal data is as follows:

[0046] Coordinate system alignment: A rigid transformation is used to unify the coordinates of electromagnetic data based on the geographic coordinate system and visual data based on the visual sensor coordinate system. The transformation formula is as follows:

[0047] ;

[0048] in, This represents the electromagnetic data tensor adapted to the visual sensor coordinate system after coordinate alignment. for Rotation matrix, It is a 3D translation vector;

[0049] Finally, the electromagnetic data tensor after coordinate system alignment is... Normalization processing, the normalized element values ​​are all Within the range.

[0050] 2) Initialize the learnable parameters of each module of the network, set the optimizer to Adam, and define the loss function as the Euclidean distance loss between the network output positioning coordinates and the real label;

[0051] 3) The preprocessed visual tensor and electromagnetic data tensor are input into the feature extraction module, respectively. After preliminary association, multimodal fusion, latent space mapping and consistency decoding, the predicted positioning coordinates are obtained. This includes longitude, latitude, and altitude;

[0052] 4) Calculate the loss between the predicted value and the true label based on the loss function, and iterate through each layer of the network using the backpropagation algorithm to synchronously update all learnable parameters; in this embodiment, the loss function of the dual-modal fusion network... : ;in, The positioning coordinates for high-precision multimodal data prediction of the model are in Dimensional components, These correspond to longitude, latitude, and altitude, respectively. The components of the target's true geographic coordinates in dimension d correspond one-to-one with the predicted coordinate dimensions;

[0053] 5) Repeat steps 3)-4) until the loss value converges, stop training and save the optimal network parameters to obtain the trained bimodal fusion network.

[0054] like Figure 3 As shown, the perception and localization process using a space-frequency anti-aliasing generation network and a dual-modal fusion network includes the following steps:

[0055] 1) A multi-agent cluster collects electromagnetic environment data according to a preset discrete grid, and inputs it into a trained space-frequency anti-aliasing generation network. The space-frequency anti-aliasing generation network completes the electromagnetic environment data of the target area to obtain complete electromagnetic field distribution data covering the target area.

[0056] Specifically, the process of electromagnetic data completion using the space-frequency anti-aliasing generator and the fusion encoder in the space-frequency anti-aliasing generator is as follows:

[0057] (1.1) The collected discrete geographic dataset Electromagnetic data Spatial interpolation using Kriging is performed to fill grid gaps, yielding geographic and electromagnetic environment tensors. Normalization is then applied to the interpolated tensors to ensure the data distribution is compatible with the training set. Discrete geographic dataset. Composed of discrete geographic coordinates, electromagnetic data Composed of discrete electromagnetic spectrum;

[0058] (1.2) The normalized tensor is input into the trained fusion encoder in parallel. After parallel convolution feature extraction, linear transformation fusion and ReLU activation, a low-dimensional latent variable z carrying the real-time environment features is generated.

[0059] (1.3) Input the latent variable z into the trained space-frequency anti-aliasing generator, convert it into a style vector w that adapts to the real-time environment through the style mapping layer, and then generate complete electromagnetic field distribution data between grid points.

[0060] (1.4) By mapping the generator output layer, the complete electromagnetic field distribution tensor covering the target region is obtained. .

[0061] 2) Perform peak detection on the electromagnetic field distribution tensor, and output coarse target location coordinates by taking a Gaussian weighted average around the peak coordinates;

[0062] (2.1) Peak detection: for the complete electromagnetic field distribution tensor The average field strength is calculated across the entire frequency domain, and the three-dimensional coordinates corresponding to the maximum average field strength are selected as the peak field strength coordinates. The specific calculation formula is as follows:

[0063] ;

[0064] Where, is the total number of frequency points. Electromagnetic field distribution tensor Mid-3D coordinates At frequency The electromagnetic field strength at that location;

[0065] (2.2) Gaussian weighted average: based on the peak field strength coordinates Centered on a target, a 3D neighborhood grid is selected within a preset range; the field strength weighted value of each grid within this neighborhood is calculated, and the weighted centroid is used as the coarse positioning coordinate of the target. :

[0066] ;

[0067] in, These represent the offsets of the neighboring raster relative to the peak coordinates in three dimensions. The three-dimensional Gaussian weighting function for the neighborhood grid. This represents the electromagnetic field strength value of the corresponding grid in the neighborhood.

[0068] 3) High-precision multimodal data is collected in the coarse localization area when the multi-agent approaches the target. After preprocessing, the high-precision multimodal data is input into the trained dual-modal fusion network. The dual-modal fusion network outputs the fine localization coordinates of each UAV to the detected target.

[0069] Specifically, the process of the dual-modal fusion network outputting precise localization coordinates is as follows:

[0070] (3.1) Perform preprocessing on the collected visual and electromagnetic data, including coordinate system alignment and normalization, to ensure that the data distribution is from the same source as the training set to obtain the visual tensor and electromagnetic data tensor.

[0071] (3.2) The preprocessed visual tensor and electromagnetic data tensor are respectively input into the trained feature extraction module: the visual tensor is input into the input layer of the visual feature extraction branch, and after feature extraction by the pre-trained ResNet-50 layer, dimensionality compression by the 1×1 convolutional layer, and processing by the global average pooling layer, a high-dimensional visual feature vector is output; the electromagnetic data tensor is input into the input layer of the electromagnetic feature extraction branch, and after dimensionality integration by the channel splicing unit, field strength feature extraction by three cascaded parallel convolutional units, and mapping by the fully connected layer, a high-dimensional electromagnetic feature vector is output. The two types of feature vectors are initially associated by the dot product operation.

[0072] (3.3) Input the pre-associated features into the trained multimodal fusion module and latent space mapping module: first, align the dimensions through a linear transformation layer, strengthen the effective information through an attention calculation layer, and generate a high-dimensional fusion feature tensor through a fusion output layer; then, convert it into vector form through a flattening layer, and complete the latent space feature mapping through two cascaded fully connected layer units;

[0073] (3.4) Input the latent space mapped features into the trained consistency decoding module: restore the feature dimensions through the fully connected decoding layer, restore the coordinate data distribution through the inverse normalization layer, and finally output the agent. At the present moment Target precise positioning coordinates , , This represents the total number of agents in a multi-agent cluster.

[0074] 4) The precise positioning coordinates output by the multi-agent system are used as the real-time position reference. Combined with the historical observation data accumulated by the multi-agent system, parameters are dynamically adjusted through adaptive filtering and multi-source information is fused to output the continuous and stable dynamic state of the target, thereby achieving continuous tracking of highly dynamic targets.

[0075] In this embodiment, the adaptive filtering process is as follows:

[0076] (4.1) State prediction;

[0077] (4.1.1) Construct a Kalman filter prediction model for each agent based on the historical observation data accumulated by multiple agents:

[0078] ;

[0079] in, Indicates the first An intelligent agent in The target state vector at any given time includes position, velocity, and acceleration; For the first An intelligent agent in The state transition matrix at time t; For the first An intelligent agent in Gaussian modeling error at time t, Indexing for intelligent agents;

[0080] (4.1.2) Based on the target state vector of the previous time step Update the prior target state vector Based on the state covariance matrix of the previous time step Update the prior state covariance matrix The state covariance matrix is ​​the same as the state estimation error covariance matrix. For the first The adjustable parameters of each agent; the initial values ​​of the target state vector and state covariance matrix are obtained by taking into account the moment when the system first obtains the target's precise localization coordinates, combined with conservative assumptions about the target's initial dynamics;

[0081] (4.2) Adaptive selection of measurement model;

[0082] (4.2.1) Calculate the signal-to-noise ratio of each agent:

[0083] ;

[0084] ;

[0085] ;

[0086] in, For the first An intelligent agent in Signal-to-noise ratio at any given moment. Indicates the first An intelligent agent in Measure signal power at all times. Indicates the first An intelligent agent in Real-time ambient noise power The sliding time; Indicates the first An intelligent agent in Precise positioning coordinates at all times In Axial components, Axis represents a three-dimensional coordinate system Any axis in the middle, Indicates the first An intelligent agent in Prior target state vector at any time The position in the middle Axial components; The preset sliding window length;

[0087] (4.2.2) will With threshold In comparison, if Then, a sparse Gaussian process regression model is used to output the first... The prediction mean square error of each agent Sparse Gaussian process regression models are suitable for scenarios with low measurement noise and stable signals;

[0088] like The deep Gaussian process regression model is used to recursively calculate the first... The prediction mean square error of each agent ;

[0089] (4.2.3) Based on Construct the state covariance matrix , diag represents a diagonal matrix;

[0090] (4.3) Multi-source data fusion;

[0091] (4.3.1) Calculate the performance of each agent in... Optimal estimate of the target state variable at time t ;

[0092] Among them, the observation matrix , For intelligent agents At any moment The posterior state vector, satisfying:

[0093] ;

[0094] ;

[0095] ;

[0096] in, For intelligent agents At any moment Local Kalman gain, intelligent agent At any moment Prior observation estimates; posterior state vector For the target state vector At the present moment The optimal estimate;

[0097] (4.3.2) Fusion of prediction results from each agent;

[0098] Prediction results of all agents using the inverse variance weighting method Perform fusion to output the global prior state mean vector. :

[0099] ;

[0100] in, Let be the weighting coefficients, and satisfy: , It also serves as an agent index, where N is the total number of agents;

[0101] (4.3.3) Kalman filter parameter fusion;

[0102] The global Kalman gain is calculated based on the local Kalman gains of each agent using the same weighting strategy. ;

[0103] (4.4) Status update and dynamic output;

[0104] The final state of the target is updated based on the prediction results of the fusion of each intelligent agent. :

[0105] ;

[0106] This outputs the target's continuous and stable dynamic state. This includes three-dimensional position, velocity, and acceleration, enabling continuous tracking of highly dynamic targets; among which, the global posterior state vector... Global prior measurement prediction value .

[0107] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A perception positioning method for time-sensitive targets in a high dynamic complex electromagnetic environment, characterized in that, The method comprises the following steps: The multi-agent cluster collects electromagnetic environment data according to a preset discrete lattice array, inputs the data into an air-frequency anti-aliasing generation network, and completes the electromagnetic environment data of a target area to obtain an electromagnetic field distribution tensor of the target area; Peak value detection is performed on the electromagnetic field distribution tensor, and Gaussian weighted average is performed around the peak value coordinates to output target coarse positioning coordinates; High-precision visual data and electromagnetic data collected by the multi-agent in a coarse positioning area determined by the coarse positioning coordinates are preprocessed to obtain a visual tensor and an electromagnetic data tensor, and the visual tensor and the electromagnetic data tensor are input into a dual-modal fusion network to output the fine positioning coordinates of each agent to the detection target; The fine positioning coordinates output by the multi-agent are used as real-time observation, historical observation data is combined, adaptive filtering is used to dynamically adjust parameters and fuse multi-source information, and a continuous and stable dynamic state of the target is output to realize continuous tracking of the high-dynamic target; in the updating process of the adaptive filtering, the measurement noise covariance matrix is adaptively adjusted in the sparse Gaussian process regression model and the deep Gaussian process regression model according to the real-time signal-to-noise ratio.

2. The method of claim 1, wherein, The air-frequency anti-aliasing generation network comprises a fusion encoder and an air-frequency anti-aliasing generator, and the training steps of the air-frequency anti-aliasing generation network comprise: Electromagnetic environment data and corresponding real electromagnetic field distribution tensor labels are obtained, and the data is normalized; The geographic tensor and the electromagnetic tensor in the electromagnetic environment data are input into the fusion encoder, and low-dimensional latent variables z are generated after feature extraction, fusion and activation; The latent variables z are input into the air-frequency anti-aliasing generator, are converted into style vectors w through a style mapping layer, and are sequentially decoded through multiple air-frequency style convolution blocks to obtain reconstructed electromagnetic field distribution tensors; the air-frequency anti-aliasing generator is based on the StyleGAN-v3 architecture of the third generation of style generation adversarial networks; the style vectors w are obtained by converting the latent variables z through three fully connected layers in the style mapping layer; The mean square error between the reconstructed electromagnetic field distribution tensors and the real electromagnetic field distribution tensor labels is used as a loss function, network parameters are updated through back propagation until convergence, and the trained air-frequency anti-aliasing generation network is obtained.

3. The method of claim 2, wherein, The air-frequency style convolution block sequentially performs spatial feature convolution SF-Conv, adaptive instance normalization AdaIN, ReLU activation and bilinear interpolation upsampling operation.

4. The method of claim 1, wherein, The specific method for performing peak value detection on the electromagnetic field distribution tensor and performing Gaussian weighted average around the peak value coordinates to output target coarse positioning coordinates is as follows: Peak detection: on electromagnetic field distribution tensor Full frequency domain field strength average calculation is performed, and the three-dimensional coordinates corresponding to the maximum value of the field strength average are selected, that is, the field strength peak coordinates : ； where N is the total number of frequency points, denotes the electromagnetic field distribution tensor in three-dimensional coordinates at a frequency point ; frequency point index , is the total number of frequency points; Gaussian weighted average: with field strength peak coordinate Select three-dimensional neighborhood grid in preset range with the centroid as the center; calculate the field strength weighted value of each grid in the neighborhood, and take the weighted centroid as the coarse positioning coordinate of the target : ； wherein, the sub-tables indicate the amount of offset in three dimensions of the neighborhood grid relative to the peak coordinate, is a three-dimensional Gaussian weighting function for the neighborhood grid, represents the electromagnetic field strength value of the corresponding grid within the neighborhood.

5. The method of claim 1, wherein, The dual-modal fusion network comprises a feature extraction module, a multi-modal fusion module, a latent space mapping module and a consistency decoding module, and the training steps comprise: Visual-electromagnetic paired data that has been subjected to coordinate system alignment and normalization preprocessing and corresponding real fine positioning coordinate labels are obtained; The preprocessed visual data is input into a visual feature extraction branch in the feature extraction module to output a visual feature vector; and the preprocessed electromagnetic data is input into an electromagnetic feature extraction branch in the feature extraction module to output an electromagnetic feature vector; After the visual feature vector is preliminarily associated with the electromagnetic feature vector, the multi-modal fusion module is input for feature fusion, and a high-dimensional fusion feature tensor is output; The high-dimensional fusion feature tensor is input into the latent space mapping module to obtain a latent space feature vector; The latent space feature vector is input into the consistency decoding module to obtain a predicted fine positioning coordinate; The Euclidean distance between the predicted fine positioning coordinate and the real fine positioning coordinate label is taken as a loss function, and the parameters of the dual-modal fusion network are updated through back propagation until convergence.

6. The method of claim 1, wherein, The preprocessing includes aligning the coordinate system of the electromagnetic data with the sensor coordinate system of the visual data through rigid transformation, and normalizing the aligned electromagnetic data to obtain the electromagnetic data tensor; the visual tensor is obtained from the aligned visual data.

7. The method of claim 1, wherein, The adaptive adjustment state covariance matrix calculation mode in the sparse Gaussian process regression model and the deep Gaussian process regression model according to the real-time signal-to-noise ratio decision specifically includes: calculating the signal-to-noise ratio of each agent at time t , is an index for the agent; Will With threshold In comparison, if Then, the sparse Gaussian process regression model is used to calculate the first... The prediction mean square error of each agent ;like The deep Gaussian process regression model is used to calculate the first... The prediction mean square error of each agent ; Further based on Constructing the state covariance matrix of each agent at time t diag denotes a diagonal matrix.

8. The method of claim 1, wherein, The signal-to-noise ratio of each intelligent agent at time t The calculation method is as follows: ； ； ； in, For the first An intelligent agent in Signal-to-noise ratio at any given moment. Indicates the first An intelligent agent in Measure signal power at all times. Indicates the first An intelligent agent in Real-time ambient noise power The sliding time; Indicates the first An intelligent agent in Precise positioning coordinates at all times In Axial components, Axis represents a three-dimensional coordinate system Any axis in the middle, Indicates the first An intelligent agent in Prior target state vector at any time The position in the middle Axial components; Set the preset sliding window length.

9. The method of claim 8, wherein, The specific mode for continuously tracking the high-dynamic target is that the multi-agent output fine positioning coordinate is taken as real-time observation, combined with historical observation data, and through adaptive filtering, the parameters are dynamically adjusted, multi-source information is fused, and the target continuous stable dynamic state is output. target state vector of the last time instant updating the prior target state vector ; state covariance matrix obtained according to the adaptive adjustment of the last time instant updating the prior state covariance matrix , adjustable parameter of the jth agent the state covariance matrix at time t and the prior state covariance matrix the local Kalman gain of each agent at time t : , is the observation matrix, and the superscript T denotes the transpose According to the target fine positioning coordinates Computing the agent At time The posterior state vector ; For the agent At time The prior observation estimate value, ; The optimal estimation value of the target state quantity of each agent at the current time ;​ calculating global Kalman gain based on local Kalman gain of each agent ; is a weighting coefficient, and satisfies: , is also an agent index, and N is the total number of agents The inverse variance weighting method is used to fuse the prediction results of all agents and output the global prior state mean vector : ； updating the target final state based on the fused prediction results of the intelligent agents : ； Thus output target continuous stable dynamic state , including three-dimensional position, speed, acceleration, to achieve the continuous tracking of high dynamic target; wherein the global posterior state vector , global prior measurement prediction value .

Images