Unmanned aerial vehicle cluster multi-sensor fusion positioning method based on improved CPInformer

Abstract

This invention proposes a multi-sensor fusion localization method for UAV swarms based on an improved CPInformer. It utilizes information geometry theory to unify GNSS positioning data, INS positioning data, and cooperative ranging and direction-finding data into a three-dimensional Gaussian probability distribution, obtaining an information probability model. An improved CPInformer neural network is designed, using the expected position vectors of multiple navigation sources within a historical time window as input. This input is then processed through a data expansion layer for spatiotemporal feature embedding. The environmental encoder uses the error covariance matrix to construct an uncertainty mask matrix to adaptively modulate the probsparse self-attention score. The localization decoder fuses long- and short-time features through probsparse self-attention and cross-attention, and outputs the current three-dimensional positioning coordinates via a fully connected feedforward neural network. This invention achieves a deep integration of information geometry theory and neural network attention mechanisms, offering advantages such as high positioning accuracy, strong real-time performance, and support for random node access and free aggregation / dispersion. It is suitable for collaborative localization tasks in UAV swarms under complex environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) navigation and cooperative positioning technology, specifically to a multi-sensor fusion cooperative positioning method suitable for UAV swarm environments, and particularly to a multi-sensor fusion positioning method for UAV swarms based on an improved CPInformer. Background Technology

[0002] With the rapid development of smart cities and the low-altitude economy, UAV swarms have become a core application platform, and high-precision positioning of UAVs is the foundation for swarm collaborative operations. However, single navigation sources (such as GNSS and INS) have significant drawbacks: GNSS is susceptible to multipath interference in complex environments such as urban canyons, resulting in a significant decline in signal quality; INS suffers from long-term cumulative errors, and its positioning error diverges severely when operating alone. Therefore, multi-source UAV swarm integrated navigation systems capable of aggregating, coordinating, fusing, and optimizing information from multiple sensors and navigation sources have become a research focus for many scholars.

[0003] Currently, GNSS / INS fusion positioning systems based on Kalman filtering are the mainstream solution, offering good real-time performance and accuracy. However, their scalability is poor, making it difficult to achieve plug-and-play compatibility with multiple navigation sources. In recent years, neural networks have been introduced into the field of fusion positioning due to their powerful nonlinear modeling capabilities. They exhibit strong noise resistance and robustness, enabling end-to-end fusion of multi-source data. However, in cluster collaborative scenarios, their computational complexity and communication overhead are high, making it difficult to meet real-time requirements. Furthermore, existing methods do not fully consider the heterogeneity of navigation sources in terms of sampling rate, error characteristics, and data format, leading to difficulties in unified modeling of multi-source data.

[0004] In the field of information-geometric fusion positioning, existing research, such as "Information-Geometric Fusion Positioning Method for UAV Swarms Based on Low-Earth Orbit Satellites" (Publication No. CN116931031B), applies information geometry to UAV swarm positioning. Regarding the application of Informer networks to positioning, "Autonomous Satellite Positioning Accuracy Prediction Method and Device Based on Informer" (Publication No. CN116819581B) discloses the use of Informer for satellite positioning accuracy prediction. However, existing technologies have not yet solved the deep integration problem of information-geometric probability models and neural network attention mechanisms, nor have they proposed a lightweight neural network architecture suitable for real-time collaborative positioning of UAV swarms. In swarm collaborative positioning, how to achieve unrestricted access and free aggregation / dispersion between nodes while ensuring high accuracy and low latency remains a current technological bottleneck.

[0005] Therefore, there is an urgent need for a drone swarm positioning method that has high positioning accuracy, strong real-time performance, and can meet the requirements of unrestricted access and free aggregation and dispersal under high-speed drone maneuvering. This method is of great practical significance for building large-scale high-precision navigation and positioning systems in scenarios such as low-altitude economy and drone swarm collaborative operations. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a multi-sensor fusion localization method for UAV swarms based on an improved CPInformer, comprising the following steps:

[0007] Step 1: Construct a multi-source heterogeneous cooperative positioning system model for UAV swarms, acquire GNSS positioning data, INS positioning data, and cooperative ranging and direction finding data of each UAV, and use information geometry theory to model the GNSS positioning data, INS positioning data, and cooperative ranging and direction finding data into three-dimensional Gaussian probability distributions respectively, to obtain a unified format information probability model, which includes a position expectation vector and an error covariance matrix;

[0008] Step 2: Design an improved CPInformer neural network, where CPInformer stands for Cooperative Positioning Informer; the improved CPInformer neural network is composed of a data expansion layer, an environment encoder, and a localization decoder cascaded in sequence;

[0009] The overall input of the improved CPInformer neural network is: the position expectation vector of the multi-source navigation source information probability model of all times within the historical time window of length L generated in step 1, which constitutes the historical position expectation vector sequence, where L is the total number of time steps contained in the historical time window, and the value range is 50 to 200.

[0010] The overall output of the improved CPInformer neural network is: the three-dimensional positioning coordinates of the target UAV at the current moment;

[0011] The data expansion layer takes all L historical time position expectation vectors as input, performs numerical embedding, position encoding embedding and timestamp encoding embedding on the position expectation vectors of each historical time respectively, and adds the three embedding results element by element to output a comprehensive feature tensor that integrates spatiotemporal information; the error covariance matrix in step 1 is not input to the data expansion layer, but is sent to the environment encoder as an independent input branch;

[0012] The environmental encoder takes the comprehensive feature tensor as input, adopts a three - level cascaded feature extraction architecture, extracts the global deep features within the long - history window of all L historical moments, and outputs a global positioning feature matrix. The environmental encoder also receives the error covariance matrix as an independent input, constructs an uncertainty mask matrix using the error covariance matrix, and adaptively modulates the Probsparse self - attention scores inside the environmental encoder.

[0013] The positioning decoder takes the short - term historical position expectation vector sequence and the global positioning feature matrix as dual inputs. Among them, the short - term historical position expectation vector sequence is obtained by intercepting the subsequence of the most recent U moments from the position expectation vectors of L historical moments. The value range of U is 10 - 30, and U < L. The positioning decoder extracts the short - term motion features of the short - term historical position expectation vector sequence through Probsparse self - attention, then deeply fuses the short - term motion features with the global positioning feature matrix through Probsparse cross - attention, and outputs the three - dimensional positioning coordinates of the current moment through a fully - connected feed - forward neural network regression.

[0014] Step 3: In a distributed architecture, each UAV node independently runs the complete processes of Step 1 and Step 2. Each UAV node broadcasts its own positioning probability model parameters and ranging and direction - finding raw observations through an inter - vehicle link, without a central node, supporting the ad - hoc access and free aggregation of nodes.

[0015] A further preferred solution is that the information - geometric modeling method for collaborative ranging and direction - finding data in Step 1 is as follows:

[0016] Assume that the target node at receives ranging information and direction - finding information from the collaborative node , where is the azimuth angle, is the pitch angle; the positioning result of the collaborative node itself follows a Gaussian distribution , where represents the Gaussian probability distribution function that the positioning result of the collaborative node at obeys, represents the position expectation vector of the positioning result of the collaborative node at , represents the error covariance matrix of the positioning result of the collaborative node at ; then the target node based on the collaborative node The expected vector of the ranging and orientation positioning results is:

[0017]

[0018] The corresponding error covariance matrix is ​​obtained through error propagation:

[0019]

[0020] in These are the standard deviations of the ranging error, azimuth measurement error, and elevation angle measurement error, respectively. These are the Jacobian matrices of the target node position with respect to the range, azimuth, and elevation angles, respectively.

[0021] When the target node simultaneously receives ranging and direction-finding information from m cooperating nodes, the final cooperative localization probability distribution is obtained by fusing the information:

[0022]

[0023]

[0024] in The desired vector for the fused collaborative localization. This is the fused error covariance matrix.

[0025] In a further preferred embodiment, in step 2, the environment encoder uses a probsparse self-attention mechanism to calculate the input data and obtain the original self-attention score. ;

[0026] When calculating the Probsparse self-attention, the environment encoder also utilizes the error covariance matrix corresponding to each navigation source in the information probability model. Construct their respective uncertainty mask matrices The mask matrix is ​​used to characterize the degree to which the uncertainty of the localization results of each navigation source suppresses attention allocation, and it is constructed as follows:

[0027]

[0028] in It is an adjustable scaling parameter. The trace of the error covariance matrix is ​​used to quantify the total positioning uncertainty.

[0029] The uncertainty mask matrix With the original self-attention score Element-wise multiplication, i.e., the Hadamard product, yields the modulated adaptive attention score. :

[0030]

[0031] Using the adaptive attention score The original self-attention score is replaced in the subsequent weighted summation calculation, thereby enabling adaptive adjustment of attention weights based on the positioning uncertainty of each navigation source.

[0032] In a further preferred embodiment, in step 2, the input to the data extension layer is a historical location time-series data sequence. The sequence is based on the current time. It consists of the navigation source positioning results from multiple historical moments, specifically defined as:

[0033]

[0034] in The expected three-dimensional position vector obtained from GNSS calculation. This is the expected 3D position vector obtained from the INS algorithm. to They are respectively The expected three-dimensional position vector of the target node is obtained by the cooperative nodes through ranging and orientation measurement.

[0035] The data expansion layer performs three types of embedding encoding on the data vectors in the input sequence. The specific processing steps are as follows:

[0036]

[0037] in For the numerical embedding process, For the position encoding embedding process, The timestamp encoding embedding process involves adding the three embedding vectors element-wise to obtain a comprehensive feature tensor that integrates numerical information, temporal location information, and time period information. .

[0038] In a further preferred embodiment, the numerical embedding process employs a fully connected neural network to embed the expected three-dimensional position vector of each input. A linear mapping to a high-dimensional feature space yields the corresponding feature vector. , Embed the dimension of the model's features;

[0039] The position encoding embedding process uses a sine-cosine position encoding function to add absolute position information in temporal order to the position vectors at different times in the sequence. The calculation formula is as follows:

[0040]

[0041] in For the time step index in the sequence, For dimension indexing;

[0042] The timestamp encoding embedding process converts the sampling time corresponding to each data point into a periodic time feature vector, and maps it to... Embedding vectors of the same dimension.

[0043] In a further preferred embodiment, in step 2, the environment encoder adopts a three-level cascaded feature extraction architecture, wherein the first two levels are composed of a Probsparse self-attention module and a convolutional distillation module connected in series, and the third level is composed of a Probsparse self-attention module.

[0044] The input to the environment encoder is the comprehensive feature tensor output by the data expansion layer. Its dimensions are ,in This represents the total number of time steps included in the historical time window, ranging from 50 to 200 time steps. For the total number of navigation sources, , The number of cooperating nodes, where 2 represents two independent navigation sources: GNSS and INS; Embed the dimension of the model's features;

[0045] The processing flow at each level is as follows:

[0046] First-level processing: The comprehensive feature tensor... The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the first Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the first intermediate feature tensor. Its time dimension remains unchanged. ;Will The input is the first convolutional distillation module. This module first extracts local neighborhood features in the time dimension through one-dimensional convolution, then downsamples the time window length using max pooling, and outputs the first-level output feature tensor. Its time dimension is shortened to , dimension ;

[0047] Second-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the second Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the second intermediate feature tensor. Its time dimension remains unchanged. ;Will The second convolutional distillation module is input and processed in the same way as the first-level convolutional distillation module, outputting the second-level output feature tensor. Its time dimension is further shortened to , dimension ;

[0048] Third-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the third Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the third-level output feature tensor. Its time dimension remains unchanged. , dimension ;

[0049] The second-level output feature tensor and the third-level output feature tensor Perform upsampling interpolation on the time dimension to restore its time dimension to its original state. The upsampled feature tensor is obtained. and All dimensions ;

[0050] After aligning the time dimension , and By concatenating and fusing features along their respective dimensions, a multi-scale fused feature tensor is obtained, with the following dimensions: ;

[0051] The multi-scale fused feature tensor is then mapped through a fully connected linear layer to restore it to a unified feature dimension. This yields a fused feature tensor with a unified feature dimension, whose dimension is... Then, the fused feature tensor of the unified feature dimension is zero-padded in the time dimension from... Restore to The final output of the environment encoder is obtained: the global localization feature matrix. Its dimensions are .

[0052] In a further preferred solution, in step 2, the positioning decoder is composed of a single-layer Probsparse self-attention module, a single-layer Probsparse cross-attention module, and a fully-connected feed-forward neural network FNN cascaded in sequence;

[0053] The positioning decoder has two independent input sources:

[0054] First input: a sequence of short-term historical position expectation vectors , whose time window length is U, the value range is 10 to 30 time steps, and U < L. The sequence of short-term historical position expectation vectors is obtained by intercepting the subsequence of the nearest U moments from the position expectation vectors of L historical moments;

[0055] Second input: the global positioning feature matrix output by the environment encoder , whose dimension is ;

[0056] The processing process of the positioning decoder for the above two inputs is executed sequentially in the following stages:

[0057] Stage 1: Short-term feature self-extraction:

[0058] After the sequence of short-term historical position expectation vectors is processed by the embedding encoding of the data expansion layer, a short-term comprehensive feature tensor is obtained;

[0059] Pass through three independent linear projection layers respectively to be mapped into a query vector matrix , a key vector matrix , and a value vector matrix . Input the single-layer Probsparse self-attention module, adopt the sparse calculation method for sparse screening of query vectors, perform sparse self-attention calculation, and output a short-term motion feature tensor , which is used to characterize the immediate motion pattern of the drone in the recent short time window;

[0060] Stage 2: Query vector generation:

[0061] Map the short-term motion feature tensor through a linear projection layer into a cross-attention query vector matrix ;

[0062] Stage 3: Key-value pair generation:

[0063] Map the global positioning feature matrix through two independent linear projection layers into a cross-attention key vector matrix respectively and cross-attention value vector matrix ;in The projection dimensions of the cross-attention query vector and the cross-attention key vector; The projection dimension of the cross-attention value vector;

[0064] Phase Four: Cross-fusion of Long-Term and Short-Term Features

[0065] Will , , The single-layer Probsparse cross-attention module is used as a common input, and a sparse computation method is employed to process the cross-attention query vector matrix. The row vectors are sparsified and filtered to obtain the cross-attention score matrix;

[0066] The cross-attention score matrix and the cross-attention value vector matrix are used together. Weighted summation, outputting the long-term and short-term fusion feature tensor ;

[0067] Phase 5: Coordinate Regression Output:

[0068] The long-short time fusion feature tensor Flattened into a one-dimensional feature vector, it is input into the fully connected feedforward neural network (FNN);

[0069] The fully connected feedforward neural network consists of multiple stacked fully connected layers, with the last layer having an output dimension of 3. After linear activation, it outputs the current three-dimensional positioning coordinates.

[0070]

[0071] in This indicates a flattening operation.

[0072] A further optimized approach, employing a sparse computation method for sparse filtering of the query vector, involves the following process:

[0073] In any self-attention module, given a query vector matrix and key vector matrix ,in The length of the query vector sequence. Let be the sequence length of the key vector. The dimensions of the query vector and key vector;

[0074] Using sparsity index to measure the first query vectors The degree of concentration of the point integral distribution of all key vectors is calculated using the following formula:

[0075]

[0076] in Key vector matrix The OK;

[0077] For each query vector According to the sparsity index Sort the values ​​from largest to smallest, and select only the top ones. One query vector participates in the complete dot product operation, and the attention scores corresponding to the remaining query vectors are filled in with a uniform distribution.

[0078] Beneficial effects:

[0079] The UAV swarm multi-sensor fusion localization method based on the improved CPInformer proposed in this invention has the following beneficial effects:

[0080] 1. This invention establishes an information probability model for navigation sources, transforming GNSS, INS, and cooperative ranging and direction finding data from three heterogeneous navigation sources into a unified information probability function form, thus overcoming the limitations of traditional methods in processing data of different formats and frequencies.

[0081] 2. This invention embeds the error covariance matrix information from the information probability model into the mask calculation of the Probsparse attention mechanism, enabling the network to adaptively adjust information weights based on the uncertainties of each navigation source. This achieves a deep integration of information geometry theory and neural network architecture, rather than a simple loosely coupled concatenation. This mechanism allows the positioning system to dynamically optimize the fusion strategy when faced with navigation source data of varying quality, significantly improving the system's robustness and accuracy.

[0082] 3. This invention embeds and encodes the position expectation vector in the information probability model into the data extension layer, while independently inputting the error covariance matrix into the mask construction unit of the environment encoder. The two information streams are fused within the network through an attention modulation mechanism. This architectural design allows uncertain information to directly participate in the feature extraction process in the form of multiplicative gating, rather than simply being concatenated as additional features, thus possessing stronger physical interpretability.

[0083] 4. This invention reduces computational complexity from low to high by using the Probsparse attention mechanism for sparse computation. Reduce to This significantly reduces the computational overhead of neural networks, while maintaining high positioning accuracy and meeting the stringent requirements of real-time computation for UAV swarms.

[0084] 5. This invention adopts a distributed architecture, with each UAV as an independent solution unit, eliminating the need for a central node. It supports the random access and free aggregation and dispersal of nodes, effectively improving the robustness and scalability of the cluster positioning system, and meeting the navigation and positioning needs of large-scale UAV clusters in complex environments.

[0085] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0086] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0087] Figure 1 This is a diagram illustrating the architecture of a multi-machine, multi-source fusion collaborative positioning system in this embodiment.

[0088] Figure 2 This is a diagram of the CPInformer fusion module architecture.

[0089] Figure 3 This is a diagram comparing the output mechanisms of CPInformer and the traditional Informer.

[0090] Figure 4 A comparison of the training convergence speeds of CPInformer, AT-LSTM, and CNN-GRU in a typical noisy scenario with segmented uniform motion.

[0091] Figure 5 Comparison of localization accuracy between CPInformer, AT-LSTM, and CNN-GRU in a typical noisy scene with segmented uniform motion;

[0092] Figure 6 A comparison of the training convergence speed of CPInformer, AT-LSTM, and CNN-GRU in a high-noise scene of segmented circular motion;

[0093] Figure 7 A comparison of the localization accuracy of CPInformer, AT-LSTM, and CNN-GRU in a high-noise scene of segmented circular motion;

[0094] Figure 8 The images show a comparison of the trajectories in the field experiment. (a) is a real-time 3D positioning image of the four UAVs, (b) is a top-down view of the trajectories of the four UAVs, and (c) is a top-down image taken synchronously in the field. Detailed Implementation

[0095] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. This embodiment employs a multi-sensor fusion localization method for UAV swarms based on an improved CPInformer proposed in this invention, and verifies it in both simulation and real-world scenarios. The specific steps include:

[0096] Step 1: Construct a multi-source heterogeneous cooperative positioning system model for UAV clusters, acquire GNSS positioning data, INS positioning data, and cooperative ranging and direction finding data of each UAV, and use information geometry theory to model the GNSS positioning data, INS positioning data, and cooperative ranging and direction finding data into three-dimensional Gaussian probability distributions respectively, to obtain a unified information probability model. The information probability model includes a position expectation vector and an error covariance matrix.

[0097] like Figure 1 As shown, the UAV swarm adopts a distributed architecture, comprising 10 UAVs. Each UAV is equipped with a GNSS receiver, an inertial measurement unit (IMU), and a ranging and direction-finding module capable of communicating with neighboring UAVs. While acquiring their own navigation data, each node exchanges ranging and direction-finding information in real time via inter-UAV links to achieve collaborative positioning. To achieve unified fusion of heterogeneous navigation source data, this embodiment utilizes information geometry theory to model the positioning results from GNSS, INS, and collaborative ranging and direction-finding as three-dimensional Gaussian probability distributions.

[0098] The positioning error of GNSS can be modeled as follows:

[0099]

[0100] GDOP, or Geometric Precision Factor, is the range vector amplification factor between the receiver and the space satellite caused by GNSS ranging errors. It is the standard deviation of pseudorange error. The covariance matrix of satellite navigation and positioning results approximately follows a Gaussian distribution.

[0101]

[0102] in It is GNSS Location results at any given time for The actual location at any given moment.

[0103] For INS positioning, assuming after inertial navigation calibration... At time t, the error model of the inertial navigation source is:

[0104]

[0105] in express The three-dimensional position error vector of inertial navigation at any given time. Let be the mean vector of this error. Let be the covariance matrix of this error. This is the sampling period for inertial navigation data.

[0106] Let the attitude angle of the target to be located be... ,in These are the roll angle, pitch angle, and yaw angle, respectively. The direction cosine matrix determined by the attitude angles is: This is used to convert vectors in the machine system (b-frame) to the navigation system (n-frame). Let the output (specific force vector) of the accelerometer in the machine system be... The static zero bias of the accelerometer and gyroscope are respectively and Accelerometer triaxial noise standard deviation and the standard deviation of gyroscope three-axis noise for:

[0107] ,

[0108] Then after inertial navigation calibration At time t, the error model of the inertial navigation source is: ,in and The mean error of the previous time step and error covariance This is obtained recursively, which is a common method in this field. Zero bias of gyroscope and accelerometer zero bias This is determined by the accumulation over time, reflecting the physical characteristic that inertial navigation errors diverge over time; Determined by the noise standard deviation of the accelerometer and gyroscope, as well as the time duration, it reflects the magnitude of the cumulative increase in error uncertainty over time.

[0109] For cooperative ranging and direction finding data, assuming the target node... exist Receive data from the cooperating node at all times Distance information and direction finding information ,in It is the azimuth angle. Pitch angle; cooperative node Its own localization results follow a Gaussian distribution. ,in Represents a collaborative node exist The Gaussian probability distribution function followed by the positioning result at time t. Represents a collaborative node exist The expected position vector of the positioning result at any given time. Represents a collaborative node exist The error covariance matrix of the positioning results at time t;

[0110] Then the target node Based on collaborative nodes The expected vector of the ranging and orientation positioning results is:

[0111]

[0112] The corresponding error covariance matrix is ​​obtained through error propagation:

[0113]

[0114] in These are the standard deviations of the ranging error, azimuth measurement error, and elevation angle measurement error, respectively. These are the Jacobian matrices of the target node position with respect to the range, azimuth, and elevation angles, respectively.

[0115] When the target node simultaneously receives ranging and direction-finding information from m cooperating nodes, the final cooperative localization probability distribution is obtained by fusing the information:

[0116]

[0117]

[0118] in The desired vector for the fused collaborative localization. This is the fused error covariance matrix.

[0119] Through the above processing, the navigation source data, which originally had different formats and sampling frequencies, were unified into an information probability model, laying a foundation for consistent format for subsequent fusion.

[0120] Step 2: Design and improve the CPInformer neural network to achieve multi-source information fusion.

[0121] like Figure 2As shown, CPInformer stands for Cooperative Positioning Informer. Its core innovation lies in the customized improvement of the Informer network for UAV swarm cooperative positioning scenarios. The network consists of three core modules: a data extension layer, an environment encoder, and a positioning decoder. The overall input of the improved CPInformer neural network is the position expectation vector of the multi-source navigation source information probability model of all times within the historical time window of length L generated in step 1; the overall output of the improved CPInformer neural network is the three-dimensional positioning coordinates of the target UAV at the current time.

[0122] The data expansion layer takes all L historical time-time position expectation vectors as input, performs numerical embedding, position encoding embedding, and timestamp encoding embedding on the position expectation vectors of each historical time, and adds the three embedding results element by element to output a comprehensive feature tensor that integrates spatiotemporal information; the error covariance matrix in step 1 is not input to the data expansion layer, that is, it does not participate in the embedding encoding of the data expansion layer, but is sent to the environment encoder as an independent input branch to modulate the attention score.

[0123] The input to the data extension layer is a historical location time series data sequence. The sequence is based on the current time. It consists of the navigation source positioning results from multiple historical moments, specifically defined as:

[0124]

[0125] in The expected three-dimensional position vector obtained from GNSS calculation. This is the expected 3D position vector obtained from the INS algorithm. to They are respectively The expected three-dimensional position vector of the target node is obtained by the cooperative nodes through ranging and direction finding; all the above components are 3×1 column vectors;

[0126] The data expansion layer performs three types of embedding encoding on the data vectors in the input sequence. The specific processing steps are as follows:

[0127]

[0128] in For the numerical embedding process, For the position encoding embedding process, The timestamp encoding embedding process involves adding the three embedding vectors element-wise to obtain a comprehensive feature tensor that integrates numerical information, temporal location information, and time period information. .

[0129] The numerical embedding process employs a fully connected neural network to embed the expected three-dimensional position vector of each input. A linear mapping to a high-dimensional feature space yields the corresponding feature vector. , The feature embedding dimension of the model is 512 in this embodiment.

[0130] The position encoding embedding process uses a sine-cosine position encoding function to add absolute position information in temporal order to the position vectors at different times in the sequence. The calculation formula is as follows:

[0131]

[0132] in For the time step index in the sequence, For dimension indexing;

[0133] The timestamp encoding embedding process converts the sampling time corresponding to each data point into a periodic time feature vector, including periodic codes such as minutes, hours, and days of the week, and maps them to... Embedding vectors of the same dimension.

[0134] The final obtained comprehensive feature tensor Its dimensions are .

[0135] The environmental encoder extracts global deep features from the extended positioning data during the UAV's long historical movement process. For example... Figure 2 As shown, the environment encoder takes the comprehensive feature tensor as input, adopts a three-level cascaded feature extraction architecture, extracts global deep features within a long historical window of all L historical moments, and outputs a global localization feature matrix; the environment encoder also receives the error covariance matrix as independent input, uses the error covariance matrix to construct an uncertainty mask matrix, and adaptively modulates the Probsparse self-attention score inside the environment encoder.

[0136] In the three-level cascaded feature extraction architecture of the environment encoder, the first two levels are composed of a Probsparse self-attention module and a convolutional distillation module connected in series, and the third level is composed of a Probsparse self-attention module.

[0137] The input to the environment encoder is the comprehensive feature tensor output by the data expansion layer. Its dimensions are ; For the total number of navigation sources, , The number of cooperating nodes is taken as [number] in this embodiment. 2 represents two independent navigation sources: GNSS and INS; Embed the dimension of the model's features; The total number of time steps included in the historical time window, ranging from 50 to 200 time steps, is preferably 128 time steps in this embodiment.

[0138] The processing flow at each level is as follows:

[0139] First-level processing: The comprehensive feature tensor... The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the first Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the first intermediate feature tensor. Its time dimension remains unchanged. ;Will The input is the first convolutional distillation module. This module first extracts local neighborhood features in the time dimension through one-dimensional convolution, then downsamples the time window length using max pooling, and outputs the first-level output feature tensor. Its time dimension is shortened to , dimension ;

[0140] Second-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the second Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the second intermediate feature tensor. Its time dimension remains unchanged. ;Will The second convolutional distillation module is input and processed in the same way as the first-level convolutional distillation module, outputting the second-level output feature tensor. Its time dimension is further shortened to , dimension ;

[0141] Third-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the third Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the third-level output feature tensor. Its time dimension remains unchanged. , dimension ;

[0142] The second-level output feature tensor and the third-level output feature tensor Perform upsampling interpolation on the time dimension to restore its time dimension to its original state. The upsampled feature tensor is obtained. and All dimensions .

[0143] In each of the Probsparse self-attention modules, the following sparsity calculation method is used to perform sparsity filtering of the query vector:

[0144] Suppose that in a certain Probsparse self-attention module, given a query vector matrix... and key vector matrix ,in The length of the query vector sequence. Let be the sequence length of the key vector. The dimensions of the query vector and key vector;

[0145] Using sparsity index to measure the first query vectors The degree of concentration of the point integral distribution of all key vectors is calculated using the following formula:

[0146]

[0147] in Key vector matrix The OK;

[0148] For each query vector According to the sparsity index Sort the values ​​from largest to smallest, and select only the top ones. One query vector participates in the complete dot product operation, while the attention scores corresponding to the remaining query vectors are filled in with a uniform distribution. Through the above sparsification calculation, the computational complexity of the environment encoder is reduced from that of the standard Transformer. Reduce to In addition, to avoid degradation in deep networks, residual connections and layer normalization are set after the attention module in each Probsparse layer.

[0149] One of the key innovations of this invention lies in the deep integration of information geometric probability model parameters and the Probsparse attention mechanism. Specifically, when calculating Probsparse self-attention, the environment encoder utilizes the error covariance matrix corresponding to each navigation source in the information probability model. Construct their respective uncertainty mask matrices :

[0150]

[0151] in For adjustable scaling parameters, this embodiment preferably uses... , The trace of the error covariance matrix is ​​used to quantify the total uncertainty in positioning.

[0152] The uncertainty mask matrix Compared with the original self-attention score Element-wise multiplication yields the modulated adaptive attention score. :

[0153]

[0154] Using the adaptive attention score The original self-attention score is replaced in subsequent calculations. This represents element-wise multiplication (Hadamard product). Through this mechanism, when fusing multi-source positioning information, the network can adaptively adjust the attention weights based on the uncertainty of the probability distribution of each navigation source: the larger the trace of the covariance matrix (i.e., the higher the positioning uncertainty), the stronger the suppression of the corresponding attention score. This allows the positioning system to dynamically reduce the influence weight of unreliable navigation sources when facing situations such as degraded GNSS signal quality and increased INS cumulative error, thereby significantly improving the system's robustness and positioning accuracy. This mechanism directly embeds the navigation source uncertainty described by information geometry theory into the attention calculation process of the neural network, achieving a deep synergy between information geometry theory and deep learning architecture.

[0155] After obtaining the three-level output feature tensor, the first-level output feature tensor... Second-level output feature tensor and the third-level output feature tensor Perform multi-scale fusion. Because... and Time dimension and Inconsistent, therefore, the second-level output feature tensor is first... and the third-level output feature tensor Perform upsampling interpolation on the time dimension to restore its time dimension to its original state. The upsampled feature tensor is obtained. and All dimensions In this embodiment, the upsampling interpolation process adopts the linear interpolation method, that is, interpolation points are uniformly inserted between adjacent time steps, and the feature vector of the interpolation point is the arithmetic mean of the feature vectors of two adjacent time steps.

[0156] After aligning the time dimension , and By concatenating and fusing features along their respective dimensions, a multi-scale fused feature tensor is obtained, with the following dimensions: ;

[0157] The multi-scale fused feature tensor is then mapped through a fully connected linear layer to restore it to a unified feature dimension. This yields a fused feature tensor with a unified feature dimension, whose dimension is... Then, the fused feature tensor of the unified feature dimension is zero-padded in the time dimension from... Restore to The final output of the environment encoder is obtained: the global localization feature matrix. Its dimensions are .

[0158] In this embodiment, the first level: (128×11×512) is obtained through the first Probsparse self-attention module. (128×11×512), then passed through the first convolution distillation module (Conv1D kernel size 3, stride 1; MaxPooling pooling window 2, stride 2) to obtain (64×11×512).

[0159] Level 2: (64×11×512) is obtained through the second Probsparse self-attention module. (64×11×512), then obtained through the second convolution distillation module. (32×11×512).

[0160] Level 3: (32×11×512) is obtained through the third Probsparse self-attention module. (32×11×512).

[0161] Multi-scale fusion processing: (64×11×512) (32×11×512) and are fused. First, and are linearly interpolated and upsampled in the time dimension to restore to 64 time steps, obtaining and (both are 64×11×512). They are concatenated in the feature dimension to obtain a tensor of 64×11×1536, which is restored to 64×11×512 through a fully connected linear mapping layer (1536→512), and then restored to 128×11×512 through zero padding to obtain the global positioning feature matrix .

[0162] The positioning decoder separates the input sequence and the output sequence, obtains the instant motion feature through self-attention using short-term historical data, and then deeply fuses this short-term feature with the global feature through Probsparse cross-attention. Finally, the three-dimensional coordinates of the current moment are regressed by a fully connected network.

[0163] The positioning decoder is composed of a single-layer Probsparse self-attention module, a single-layer Probsparse cross-attention module, and a fully connected feed-forward neural network FNN cascaded in sequence.

[0164] The positioning decoder has two independent input sources:

[0165] First input: The short-term historical position expectation vector sequence , whose time window length is U, and the value range is 10 to 30 time steps, and U < L. The short-term historical position expectation vector sequence is obtained by intercepting the subsequence of the nearest U moments from the position expectation vectors of L historical moments;

[0166] Second input: The global positioning feature matrix output by the environment encoder, whose dimension is ;

[0167] The processing process of the positioning decoder for the above two inputs is executed in the following stages in sequence:

[0168] Stage 1: Short-term feature self-extraction:

[0169] The short-term historical position expectation vector sequence [[ID=*43*]]after being processed by the embedding encoding of the data expansion layer, obtains the short-term comprehensive feature tensor ;

[0170] Pass through three independent linear projection layers respectively to map them into the query vector matrix , the key vector matrix and the value vector matrix The input is a single-layer Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, executes sparsity self-attention calculation, and outputs a short-time motion feature tensor. This is used to characterize the instantaneous motion pattern of a drone within the most recent short time window;

[0171] Phase Two: Query Vector Generation

[0172] The short-time motion feature tensor Mapped to a cross-attention query vector matrix through a linear projection layer. ;

[0173] Phase 3: Key-value pair generation:

[0174] The global localization feature matrix The cross-attention key vector matrix is ​​mapped through two independent linear projection layers. and cross-attention value vector matrix ;in The projection dimensions of the cross-attention query vector and the cross-attention key vector; is the projection dimension of the cross-attention value vector; both are preset hyperparameters. In this embodiment, , .

[0175] Phase Four: Cross-fusion of Long-Term and Short-Term Features

[0176] Will , , The single-layer Probsparse cross-attention module is input together.

[0177] by Each row vector is used as a query vector sequence, with As a sequence of key vectors, a sparsity calculation method is used to perform sparsity filtering of the query vector. Specifically, for the first... query vectors (Right now The (row), its sparsity index is:

[0178]

[0179] in for The Okay. Yes. Calculate the sparsity index of each query vector and sort them in descending order, selecting only the top ones. One query vector participates in the complete dot product operation, and the attention scores corresponding to the remaining query vectors are filled in with a uniform distribution to obtain the cross-attention score matrix.

[0180] The cross-attention score matrix and the cross-attention value vector matrix are used together. Weighted summation, outputting the long-term and short-term fusion feature tensor It should be noted that the length of the output sequence of the cross-attention module is determined by the query vector. sequence length Determined by, rather than by, value vectors sequence length This decision ensures that the fused features remain temporally aligned with short-term historical location data.

[0181] Phase 5: Coordinate Regression Output:

[0182] The long-short time fusion feature tensor Flattened into a one-dimensional feature vector, it is input into the fully connected feedforward neural network (FNN);

[0183] The fully connected feedforward neural network (FNN) consists of multiple stacked fully connected layers: the first layer has an input dimension of... The output dimension is 512, and the activation function is ReLU; the second layer dimension is... The activation function is ReLU; the third dimension is... No activation function (linear output). The last layer outputs the current 3D positioning coordinates:

[0184]

[0185] in This indicates a flattening operation.

[0186] In this embodiment, the specific processing procedure of the positioning decoder is as follows:

[0187] Phase 1: Obtained through embedding encoding (16×11×512), obtained through three independent linear projection layers , , The result is obtained through Probsparse self-attention. (16×11×512).

[0188] Phase Two: Obtained by linear projection (16×11×64).

[0189] Phase Three: Obtained through two independent linear projections (128×11×64) and (128×11×512).

[0190] Phase Four: , , Input the Probsparse cross-attention module, to Each row vector is a query (sequence length 16), with... Using the key (sequence length 128), calculate the sparsity index for each query vector, filter them to obtain the attention score matrix, and then... Weighted summation, output (16×11×512).

[0191] Phase Five: Flattened into a one-dimensional vector of length 16×11×512=90112, it is then processed by a three-layer FNN (90112→512→256→3) to output three-dimensional coordinates. .

[0192] In step 3, the working mechanism and advantages of the distributed cooperative positioning architecture are reflected in its independence from a central node. Each UAV operates as an independent solution unit, running the complete process of steps 1 and 2. Each UAV node broadcasts its own positioning probability model parameters and raw ranging and direction-finding observations through inter-UAV links, resulting in low communication overhead. The cluster size can be dynamically scaled, new nodes can join at any time, and the departure of nodes does not affect system operation, achieving on-demand access and free aggregation and dispersal, significantly improving the system's robustness and scalability.

[0193] The simulation experiments and field tests are given below:

[0194] In this embodiment, a simulation scenario is set up in a space of 1000m×1000m×50m. Ten drones move along similar but non-overlapping trajectories. Each drone receives data from its own GNSS, INS, and the data from the other nine cooperating drones and performs CPInformer fusion positioning calculation.

[0195] The first group of experiments involved segmented uniform linear motion with a GNSS error standard deviation of 1m (normal noise); the second group involved uniform circular motion with a GNSS error standard deviation of 10m (high noise). Both experiments were conducted with the UAV speed set to approximately 10m / s, the cooperative ranging error standard deviation set to approximately 0.1m, the total simulation duration to 72s, and a positioning output time interval of 0.1s, totaling 720 positioning output time steps.

[0196] Table 1 Simulation Parameter Settings

[0197]

[0198] To verify the effectiveness of the method of the present invention, one of the drones was selected as the target and compared with two fusion localization algorithms based on AT-LSTM and CNN-GRU.

[0199] (1) Comparison of training convergence speed:

[0200] For neural network-based localization algorithms, a rapid decrease in the loss function during the initial training phase indicates that the model can quickly learn the basic data distribution, shortening the debugging cycle. This facilitates timely verification of the localization algorithm's feasibility, reduces computational resource consumption, and enables rapid extraction of localization features in new environments, thereby improving environmental adaptability. The training set includes 5040 time steps to allow all three methods to fully extract temporal features from the historical localization process.

[0201] pass Figure 4 and Figure 6 It is easy to see that in normal noise scenarios, the loss function of CPInformer tends to stabilize within about 20 training epochs, while AT-LSTM and CNN-GRU require about 35 and 40 epochs, respectively. Furthermore, in high noise scenarios, the convergence advantage of the method in this invention is even more pronounced. This indicates that the lightweight design of CPInformer enables it to learn data distribution characteristics more quickly.

[0202] (2) Comparison of positioning accuracy:

[0203] Since all three time-series algorithms extract features from historical positioning data, the quality of short historical data will be significantly reflected in the positioning error. Therefore, their positioning errors have a potentially alternating high and low temporal characteristics. And through... Figure 5 and Figure 7 It can be pointed out that in the comparison of the mean squared error of positioning over 720 time steps, the average error of CPInformer in normal noise scenarios is approximately 0.35m, AT-LSTM is approximately 0.52m, and CNN-GRU is approximately 0.48m; in high noise scenarios, the average error of the method of this invention is approximately 1.2m, while the comparative methods reach 1.8m and 1.6m respectively. Local error analysis shows that CPInformer can still maintain stable output during periods of abrupt changes in GNSS signals, demonstrating stronger anti-interference capability.

[0204] To better verify the simulation results, a field experiment was conducted in a 50m×50m×20m cubic space at the football field of the Chang'an Campus of Northwestern Polytechnical University. The experiment deployed four drones and four ground nodes. One drone remained stationary on the ground (simulating landing), two drones hovered, and one drone flew as the target drone in the experimental area.

[0205] The experiment employed the method of this invention to receive UAV positioning data in real time and display historical trajectory images. During the experiment, the target UAV flew along a predetermined route, and the ground station recorded the real-time positioning trajectory output by CPInformer, comparing it with the true value obtained through high-precision differential GNSS.

[0206] Figure 8 The results show that the positioning trajectory output by the method of this invention closely matches the actual trajectory, with an average three-dimensional position error of approximately 0.28m and a maximum error not exceeding 0.45m. The positioning output frequency reaches 10Hz, meeting real-time requirements. Furthermore, during the target aircraft's flight, the cooperating nodes stably provide ranging and direction-finding information, and the system operates smoothly without any positioning interruptions or divergences caused by dynamic changes in nodes. This field experiment verifies the reliability of the simulation results and demonstrates the effectiveness and reliability of the method of this invention in actual UAV swarm scenarios.

[0207] Through the above simulation and field experiments, the UAV swarm multi-sensor fusion collaborative localization method based on the improved CPInformer proposed in this invention has shown significant advantages in terms of positioning accuracy, convergence speed, noise resistance and real-time performance, and is suitable for UAV swarm collaborative localization tasks in complex environments.

[0208] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A multi-sensor fusion localization method for UAV swarms based on an improved CPInformer, characterized in that, It includes the following steps: Step 1: Construct a multi-source heterogeneous collaborative positioning system model for the UAV swarm, obtain the GNSS positioning data, INS positioning data and collaborative ranging and direction-finding data of each UAV, and use the information geometry theory to model the GNSS positioning data, INS positioning data and collaborative ranging and direction-finding data as three-dimensional Gaussian probability distributions respectively, to obtain an information probability model in a unified format. The information probability model includes a position expectation vector and an error covariance matrix; Step 2: Design an improved CPInformer neural network, which is composed of a data expansion layer, an environment encoder and a positioning decoder cascaded in sequence. The input of the improved CPInformer neural network is the position expectation vectors of the multi-source navigation source information probability models at all moments within the historical time window of length L generated in Step 1, forming a historical position expectation vector sequence, and the output is the three-dimensional positioning coordinates of the target UAV at the current moment; Step 3: Under the distributed architecture, each UAV node independently runs the complete processes of Step 1 and Step 2. Each UAV node broadcasts its own positioning probability model parameters and ranging and direction-finding raw observations through an inter-aircraft link, without a central node, to achieve multi-sensor fusion positioning of the UAV swarm.

2. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 1, characterized in that, In Step 2: The data expansion layer takes the position expectation vectors of all L historical moments as inputs, performs numerical embedding, position encoding embedding and timestamp encoding embedding on the position expectation vectors of each historical moment respectively, and adds the three embedding results element by element to output a comprehensive feature tensor that fuses spatio-temporal information. The error covariance matrix in Step 1 is not input into the data expansion layer, but is sent into the environment encoder as an independent input branch; The environment encoder takes the comprehensive feature tensor as input, adopts a three-level cascaded feature extraction architecture to extract the global deep features within the long historical window of all L historical moments, and outputs a global positioning feature matrix. The environment encoder also receives the error covariance matrix as an independent input, and uses the error covariance matrix to construct an uncertainty mask matrix to adaptively modulate the Probsparse self-attention scores inside the environment encoder; The positioning decoder takes the short-term historical position expectation vector sequence and the global positioning feature matrix as double inputs. Among them, the short-term historical position expectation vector sequence is obtained by intercepting a subsequence of the nearest U moments from the position expectation vectors of L historical moments, where U < L. The positioning decoder extracts the short-term motion features of the short-term historical position expectation vector sequence through Probsparse self-attention, and then deeply fuses the short-term motion features with the global positioning feature matrix through Probsparse cross-attention, and outputs the three-dimensional positioning coordinates at the current moment through a fully connected feedforward neural network regression.

3. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 1, characterized in that, The information geometry modeling method for the collaborative ranging and direction-finding data in Step 1 is: Assuming the target node exist Receive data from the cooperating node at all times Distance information and direction finding information ,in It is the azimuth angle. Pitch angle; cooperative node Its own localization results follow a Gaussian distribution. ,in Represents a collaborative node exist The Gaussian probability distribution function followed by the positioning result at time t. Represents a collaborative node exist The expected position vector of the positioning result at any given time. Represents a collaborative node exist The error covariance matrix of the positioning results at time t; Then the target node Based on collaborative nodes The expected vector of the ranging and orientation positioning results is: The corresponding error covariance matrix is obtained through error propagation: in These are the standard deviations of the ranging error, azimuth measurement error, and elevation angle measurement error, respectively. These are the Jacobian matrices of the target node position with respect to the range, azimuth, and elevation angles, respectively. When the target node simultaneously receives ranging and direction-finding information from m cooperating nodes, the final cooperative localization probability distribution is obtained by fusing the information: in The desired vector for the fused collaborative localization. This is the fused error covariance matrix.

4. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 2, characterized in that, In step 2, the environment encoder uses the Probsparse self-attention mechanism to calculate the input data and obtain the original self-attention score. ; The environmental encoder also utilizes the error covariance matrix corresponding to each navigation source in the information probability model. Construct their respective uncertainty mask matrices : in It is an adjustable scaling parameter. The trace of the error covariance matrix; The uncertainty mask matrix With the original self-attention score Element-wise multiplication yields the modulated adaptive attention score. : Using the adaptive attention score Replace the original self-attention score in subsequent calculations.

5. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 1, characterized in that, In step 2, the input to the data extension layer is a historical location time series data sequence. The sequence is based on the current time. It consists of the navigation source positioning results from multiple historical moments, specifically defined as: in The expected three-dimensional position vector obtained from GNSS calculation. This is the expected 3D position vector obtained from the INS algorithm. to They are respectively The expected three-dimensional position vector of the target node is obtained by the cooperative nodes through ranging and orientation measurement. The data expansion layer performs three types of embedding encoding on the data vectors in the input sequence. The specific processing steps are as follows: in For the numerical embedding process, For the position encoding embedding process, The process of embedding timestamp encoding; By summing the three embedding vectors element-wise, a comprehensive feature tensor integrating numerical information, temporal location information, and time period information is obtained. .

6. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 5, characterized in that, The numerical embedding process uses a fully connected neural network to linearly map each input three-dimensional position expectation vector to a high-dimensional feature space to obtain the corresponding feature vector; the position encoding embedding process uses a sine-cosine position encoding function to add the absolute position information in the temporal order to the position vectors at different times in the sequence; the timestamp encoding embedding process converts the sampling time corresponding to each data point into a periodic time feature vector.

7. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 2, characterized in that, In step 2, the environment encoder adopts a three-level cascaded feature extraction architecture, wherein the first two levels are composed of a Probsparse self-attention module and a convolutional distillation module connected in series, and the third level is composed of a Probsparse self-attention module. The input to the environment encoder is the comprehensive feature tensor output by the data expansion layer. Its dimensions are ,in This represents the total number of time steps included in the historical time window, ranging from 50 to 200 time steps. For the total number of navigation sources, , The number of cooperating nodes, where 2 represents two independent navigation sources: GNSS and INS; Embed the dimension of the model's features; The processing flow at each level is as follows: First-level processing: The comprehensive feature tensor... The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the first Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the first intermediate feature tensor. Its time dimension remains unchanged. ;Will The input is the first convolutional distillation module. This module first extracts local neighborhood features in the time dimension through one-dimensional convolution, then downsamples the time window length using max pooling, and outputs the first-level output feature tensor. Its time dimension is shortened to , dimension ; Second-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the second Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the second intermediate feature tensor. Its time dimension remains unchanged. ;Will The second convolutional distillation module is input and processed in the same way as the first-level convolutional distillation module, outputting the second-level output feature tensor. Its time dimension is further shortened to , dimension ; Third-level processing: The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is the third Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, performs sparsity self-attention calculation, and outputs the third-level output feature tensor. Its time dimension remains unchanged. , dimension ; The second-level output feature tensor and the third-level output feature tensor Perform upsampling interpolation on the time dimension to restore its time dimension to its original state. The upsampled feature tensor is obtained. and All dimensions ; After aligning the time dimension , and By concatenating and fusing features along their respective dimensions, a multi-scale fused feature tensor is obtained, with the following dimensions: ; The multi-scale fused feature tensor is then mapped through a fully connected linear layer to restore it to a unified feature dimension. This yields a fused feature tensor with a unified feature dimension, whose dimension is... Then, the fused feature tensor of the unified feature dimension is zero-padded in the time dimension from... Restore to The final output of the environment encoder is obtained: the global localization feature matrix. Its dimensions are .

8. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to claim 2, characterized in that, In step 2, the localization decoder is composed of a single-layer probsparse self-attention module, a single-layer probsparse cross-attention module, and a fully connected feedforward neural network (FNN) cascaded in sequence. The positioning decoder uses a short-time historical location expectation vector sequence. and the global localization feature matrix As a dual input; The localization decoder processes the two inputs in the following stages: Phase 1: Short-term feature extraction The short-time historical location expected vector sequence After embedding and encoding processing by the data expansion layer, the short-time comprehensive feature tensor is obtained. ; Will The query vector matrix is ​​mapped through three independent linear projection layers. Key vector matrix Sum value vector matrix The input is a single-layer Probsparse self-attention module, which uses a sparsity calculation method to perform sparsity filtering of the query vector, executes sparsity self-attention calculation, and outputs a short-time motion feature tensor. ; Phase Two: Query Vector Generation The short-time motion feature tensor Mapped to a cross-attention query vector matrix through a linear projection layer. ; Phase 3: Key-value pair generation: The global localization feature matrix The cross-attention key vector matrix is ​​mapped through two independent linear projection layers. and cross-attention value vector matrix ;in The projection dimensions of the cross-attention query vector and the cross-attention key vector; The projection dimension of the cross-attention value vector; Phase Four: Cross-fusion of Long-Term and Short-Term Features Will , , The single-layer Probsparse cross-attention module is used as a common input, and a sparse computation method is employed to process the cross-attention query vector matrix. The row vectors are sparsified and filtered to obtain the cross-attention score matrix; The cross-attention score matrix and the cross-attention value vector matrix are used together. Weighted summation, outputting the long-term and short-term fusion feature tensor ; Phase 5: Coordinate Regression Output: The long-short time fusion feature tensor Flattened into a one-dimensional feature vector, it is input into the fully connected feedforward neural network (FNN); The fully connected feedforward neural network consists of multiple stacked fully connected layers, with the last layer having an output dimension of 3. After linear activation, it outputs the current three-dimensional positioning coordinates. in This indicates a flattening operation.

9. The UAV swarm multi-sensor fusion localization method based on improved CPInformer according to any one of claims 7 or 8, characterized in that, The specific process of using sparse computation methods to perform sparse filtering of query vectors is as follows: In any self-attention module, given a query vector matrix and key vector matrix ,in The length of the query vector sequence. Let be the sequence length of the key vector. The dimensions of the query vector and key vector; Using sparsity index to measure the first query vectors The degree of concentration of the point integral distribution of all key vectors is calculated using the following formula: in Key vector matrix The OK; For each query vector According to the sparsity index Sort the values ​​from largest to smallest, and select only the top ones. One query vector participates in the complete dot product operation, and the attention scores corresponding to the remaining query vectors are filled in with a uniform distribution.

Images