Unmanned aerial vehicle anti-interference passive positioning method and system based on GRU and particle filter fusion

Abstract

The invention relates to an unmanned aerial vehicle anti-interference passive positioning method and system based on GRU and particle filtering fusion, belongs to the field of navigation and positioning, and solves the problems that an unmanned aerial vehicle positioning signal is unstable and is easy to drop in a complex environment in the existing method. Comprising the following steps: acquiring the time difference of arrival of radio communication signals of at least one unmanned aerial vehicle by using a plurality of TDOA passive positioning nodes, and generating a historical state sequence of each unmanned aerial vehicle based on the time difference of arrival; wherein the state value in each historical state sequence is a TDOA observation value and comprises three-dimensional coordinate data and three-dimensional speed data; inputting the historical state sequence of each unmanned aerial vehicle into a plurality of pre-trained GRU neural networks, and generating proposed distribution of the state of each unmanned aerial vehicle at the next moment based on the GRU state sequences output by the plurality of networks; and obtaining a state prediction result of each unmanned aerial vehicle based on a proposal distribution fusion particle filtering algorithm of the state of each unmanned aerial vehicle at the next moment. And continuous, smooth and high-precision positioning of the unmanned aerial vehicle under the condition of discontinuous signals is realized.

Description

Technical Field

[0001] This invention relates to the field of navigation and positioning technology, and in particular to a method and system for anti-interference passive positioning of unmanned aerial vehicles based on the fusion of GRU and particle filter. Background Technology

[0002] Precise outdoor positioning of drones typically relies on GNSS (Global Navigation Satellite System, such as GPS and BeiDou). However, when facing unauthorized drone intrusion, GNSS positioning is often unavailable. In such cases, TDOA (Time Difference of Arrival) radio positioning technology based on ground base stations becomes an important and efficient passive positioning method for unauthorized drones. However, TDOA technology has many limitations and is susceptible to multipath effects caused by urban environments, non-line-of-sight propagation errors, and strong background noise interference, leading to problems such as positioning result jumps, increased positioning errors, reduced real-time performance, and prolonged signal loss.

[0003] There are two main existing solutions: The first is the traditional filtering method, such as Kalman filtering and extended Kalman filtering. Although these can smooth the UAV trajectory, their positioning prediction performance heavily relies on a pre-set motion model. UAVs are highly maneuverable devices with complex motion patterns, making it difficult to accurately describe the motion model, resulting in large position prediction errors. Furthermore, the error accumulation is severe when TDOA radio detection signals are lost for extended periods. Another solution is the particle filtering algorithm, which can handle nonlinear problems, but uses a simple random diffusion model in the prediction step, resulting in low particle sampling efficiency, requiring a large number of particles, and thus low computational efficiency. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a passive positioning method and system for UAVs based on the fusion of GRU and particle filtering, in order to solve the technical problems of unstable and easily dropped UAV positioning signals in complex environments in existing methods, so as to achieve continuous, smooth and high-precision positioning of UAVs under conditions of intermittent signals.

[0005] This invention provides a passive localization method for unmanned aerial vehicles (UAVs) based on the fusion of GRU and particle filter, comprising the following steps: Step S1: Collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate a historical state sequence for each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data; Step S2: Input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposed distribution of the next state of each UAV based on the GRU state sequence output by multiple networks. Step S3: Based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next moment, obtain the state prediction results of each UAV.

[0006] Furthermore, the proposed distribution of the next time step state includes the predicted mean of the next time step state. and forecast uncertainty ; Each network outputs a GRU state sequence including the three-dimensional coordinate data of the current state and the predicted value of the corresponding three-dimensional velocity data; The mean and covariance of the predicted values ​​of 3D coordinate data and 3D velocity data output from multiple pre-trained GRU neural networks are obtained by calculating the mean and covariance. and As shown below: ; in, These are the predicted mean values ​​for three-dimensional coordinates and three-dimensional velocity, respectively. represents the covariance of the three-dimensional coordinates and the three-dimensional velocity, respectively, indicating uncertainty; This represents a diagonal matrix.

[0007] Further, step S3 includes: Initialize the particle set, with each particle represented as ;in, This represents the particle state, corresponding to three-dimensional coordinate data and three-dimensional velocity data; This represents the weight corresponding to each particle; Particle sampling is performed based on the proposed distribution. ;in, The particle state at the next moment; Update the weight of each particle based on the TDOA observation value received at the current time; if no TDOA observation value is received, keep the particle weight unchanged. Perform weight normalization and calculate the effective particle count; determine whether to resample based on the effective particle count. If resampling is performed, the weights of each particle are reset; the predicted next time step is output. Final state estimation results for the drone.

[0008] Furthermore, the effective number of particles is calculated as follows: ; in, The effective number of particles; for Time of the first The weight of each particle; like When this happens, resampling is triggered to prevent particle degradation; among which, For the preset threshold value, This represents the total number of particles.

[0009] Furthermore, predicting the next moment. The final state estimation results of the UAV are shown below: ; in, For the next moment The final predicted state of the drone includes the drone's three-dimensional coordinate data and three-dimensional velocity data; This provides the three-dimensional coordinates and three-dimensional velocity data for each particle in the particle filter.

[0010] Furthermore, based on the current moment The received TDOA observations are used to update the weights of each particle, as shown below: ; in, For the current moment Received TDOA observations, for Time of the first The state of each particle; The likelihood function is used to measure the degree of matching between the particle state and the TDOA observation. for Time of the first The weight of each particle.

[0011] Further, step S1 includes: The TDOA passive positioning node records the timestamps of radio communication signals emitted by at least one UAV arriving at each TDOA passive positioning node in real time. The arrival time difference is obtained based on the timestamps, and the corresponding longitude, latitude, and altitude coordinate values ​​are calculated based on the arrival time difference. The longitude, latitude, and altitude coordinate values ​​are stored in the data queue. Calculate the spatial distance S between the newly obtained longitude and latitude and the latest longitude and latitude in the historical data queue. If S is less than or equal to a preset threshold, store the newly obtained longitude, latitude, and altitude values ​​into the corresponding historical data queue; otherwise, create a new data queue for storage. Delete the data queue that has not been updated with new 3D coordinate data for m consecutive seconds; determine the data queue that has been stored with 3D coordinate data n times consecutively as the data queue of the real UAV. In each data queue, the longitude, latitude, and altitude values ​​of each data queue element are converted into three-dimensional coordinate data in the world coordinate system. Based on adjacent 3D coordinate data and corresponding timestamps, the corresponding 3D velocity data is calculated. ; The three-dimensional coordinate data and the corresponding three-dimensional velocity data constitute the TDOA observations in the historical state sequence; the historical state sequences of multiple UAVs are obtained, with each UAV corresponding to one historical state sequence.

[0012] Furthermore, the spatial distance S between the newly obtained longitude and latitude and the latest longitude and latitude in the historical data queue is calculated as follows: ; in, , ; This indicates the newly acquired longitude and latitude. This indicates the longitude and latitude of the latest data in the historical data queue; The radius is the Earth's radius.

[0013] Furthermore, the pre-trained GRU neural network is obtained through the following process; Obtain a training dataset, which includes a UAV real state sequence composed of high-precision RTK positioning data and a UAV TDOA state sequence composed of TDOA observation data corresponding to the real state sequence. The GRU neural network is trained under supervision by using real state sequences as training labels and the corresponding UAV TDOA state sequences as input. During training, the mean squared error loss function is used; when the model's loss function value on the validation set is continuously... When the performance no longer decreases within a training cycle, training is terminated, and a pre-trained GRU neural network is obtained.

[0014] The present invention also discloses an anti-interference passive positioning system for unmanned aerial vehicles based on the fusion of GRU and particle filter, the system comprising a TDOA multi-target state sequence generation module M1, a GRU multi-target proposal distribution generation module M2, and a particle filter multi-target state estimation module M3; The TDOA multi-target state sequence generation module M1 is used to collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate historical state sequences of each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data. The GRU multi-objective proposal distribution generation module M2 is used to input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposal distribution of the next state of each UAV based on the GRU state sequence output by the multiple networks. The particle filter multi-target state estimation module M3 is used to obtain the state prediction results of each UAV based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next time step.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. The proposal distribution generated by the GRU of this invention can accurately guide particles to high-probability regions, greatly improving sampling efficiency. It can achieve the accuracy of thousands of particles in traditional particle filtering with a small number of particles (such as 100), significantly reducing the computational load, improving computational efficiency, facilitating real-time deployment, and making it more suitable for real-time processing on embedded platforms with limited computing resources. It overcomes the shortcomings of traditional particle filtering, which uses a random diffusion model, has extremely low particle sampling efficiency, requires a large number of particles to ensure accuracy, has high computational complexity, and is difficult to run in real time. 2. This invention utilizes the powerful sequence learning capability of the GRU neural network to directly learn the motion patterns of the UAV from historical state sequences. The generated proposal distribution can more accurately predict its future state, achieving high positioning accuracy and effectively tracking complex maneuvers. It significantly improves positioning accuracy and trajectory smoothness in complex scenarios. This overcomes the shortcomings of traditional filtering methods (such as Kalman filtering), which rely on pre-set motion models and cannot accurately describe the complex maneuvers of the UAV, leading to large prediction biases. 3. This invention, through the deep fusion of GRU prediction and particle filtering framework, can continue to provide a high-probability state estimate by relying on the predictive capability of GRU when the TDOA observation signal is lost. When the signal is recovered, it can immediately and seamlessly fuse new observations for correction, exhibiting strong anti-interference and fault tolerance, ensuring positioning continuity, and achieving continuous and stable positioning under discontinuous signal conditions. It overcomes the shortcomings of existing TDOA signals being easily lost due to interference, and the accumulation of errors in traditional methods when the signal is lost for a long time, leading to positioning interruptions.

[0016] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of a UAV anti-interference passive positioning method based on GRU and particle filter fusion in an embodiment of the present invention; Figure 2 This is a schematic diagram of the TDOA passive positioning node collecting radio communication signals from a UAV in an embodiment of the present invention; Figure 3 This is a schematic diagram of a UAV anti-interference passive positioning system module based on the fusion of GRU and particle filter in an embodiment of the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] To address the aforementioned technical challenges, a fusion-assisted positioning scheme is needed that can accurately predict complex motions, efficiently fuse discontinuous observation data, and achieve high computational efficiency. This invention addresses the problem of unstable and large-error positioning signals from unmanned aerial vehicles (UAVs) by utilizing GRU deep learning and particle filtering for fusion improvement, thereby achieving high-precision, fault-tolerant positioning for UAVs.

[0020] Example 1: The core effect of this invention lies in using multiple pre-trained GRU (Gated Recurrent Unit) networks to generate a proposal distribution, which is then integrated with a particle filtering algorithm. The proposal distribution serves as a guide for the direction of particle propagation in the particle filtering process, thereby organically combining the sequence prediction capability of GRU deep learning with the framework of particle filtering. This is used to solve the problem of efficient and accurate prediction and positioning of UAVs, effectively locking onto UAV targets and distinguishing between different UAV targets.

[0021] A specific embodiment of the present invention discloses a passive localization method for unmanned aerial vehicles (UAVs) based on the fusion of GRU and particle filtering, such as... Figure 1 As shown, it includes the following steps: Step S1: Collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate a historical state sequence for each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data; Step S2: Input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposed distribution of the next state of each UAV based on the GRU state sequence output by multiple networks. Step S3: Based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next moment, obtain the state prediction results of each UAV.

[0022] Step S1, specifically.

[0023] Step S1 includes: The TDOA passive positioning node records the timestamps of radio communication signals emitted by at least one UAV arriving at each TDOA passive positioning node in real time. The arrival time difference is obtained based on the timestamps, and the corresponding longitude, latitude, and altitude coordinate values ​​are calculated based on the arrival time difference. The longitude, latitude, and altitude coordinate values ​​are stored in the data queue. Calculate the spatial distance S between the newly obtained longitude and latitude and the latest longitude and latitude in the historical data queue. If S is less than or equal to a preset threshold, store the newly obtained longitude, latitude, and altitude values ​​into the corresponding historical data queue; otherwise, create a new data queue for storage. Delete the data queue that has not been updated with new 3D coordinate data for m consecutive seconds; determine the data queue that has been stored with 3D coordinate data n times consecutively as the data queue of the real UAV. In each data queue, the longitude, latitude, and altitude values ​​of each data queue element are converted into three-dimensional coordinate data in the world coordinate system. Based on adjacent 3D coordinate data and corresponding timestamps, the corresponding 3D velocity data is calculated. ; The three-dimensional coordinate data and the corresponding three-dimensional velocity data constitute the TDOA observations in the historical state sequence; the historical state sequences of multiple UAVs are obtained, with each UAV corresponding to one historical state sequence.

[0024] The TDOA (Time Difference of Arrival Passive Positioning Node) is a ground terminal monitoring station that integrates radio signal reception, high-precision timestamp marking, and data uploading functions. It does not need to actively transmit signals to the UAV ("passive" characteristic) but receives radio signals emitted by the UAV itself (such as remote control, telemetry, and image transmission signals, with mainstream frequency bands including 2.4GHz and 5.8GHz).

[0025] like Figure 2 As shown, this is a list of four deployed TDOA nodes. Figure 2 The terminal monitoring system (the location indicated by the four arrows in the middle).

[0026] Based on the deployment of four TDOA passive positioning nodes and radio detection drone nodes, the arrival time difference of the received radio communication signals emitted by the drones to multiple receiving stations at different locations is utilized. The TDOA passive positioning node does not measure the absolute time of the radio communication signal from the UAV transmitter to a single TDOA passive positioning node, but rather the relative time difference between the radio communication signal arriving at two or more TDOA passive positioning nodes, thereby calculating the UAV's position.

[0027] Based on the received radio communication signals transmitted by the UAVs, the four TDOA passive positioning nodes output the positioning results of all UAVs within a 5KM radius around the TDOA passive positioning nodes every 0.5 seconds. The positioning results do not include UAV identity information and are discrete three-dimensional position data. Therefore, it is necessary to perform distance-based identity matching between the new positioning results and the historical positioning results. If the calculated distance between a new result and the latest value of an old result is less than or equal to a preset distance threshold, it is regarded as the same UAV target, and the corresponding UAV's new positioning result data is stored in the historical data queue.

[0028] If the location data of a new drone does not match the historical drone data, it is considered a new target, and a new drone data queue is created to indicate the appearance of new drone data. For example, m is 10 seconds. If there are no new data updates for drones in a drone list for more than 10 consecutive seconds, it means that the drone may have flown out of the detection range or the signal has been completely interrupted, indicating that the drone signal has disappeared. The data queue is then deleted to avoid invalid data consuming resources. A distance calculation function based on the Haversine semi-versus formula is used to calculate the spatial distance S between the newly obtained longitude and latitude and the latest longitude and latitude in the historical data queue, in order to distinguish different UAV targets.

[0029] The spatial distance S between the newly acquired longitude and latitude and the latest longitude and latitude in the historical data queue is calculated as follows: ; in, , ; This indicates the newly acquired longitude and latitude. This indicates the longitude and latitude of the latest data in the historical data queue; The radius is the Earth's radius.

[0030] For example, n is 10 times; if 10 consecutive drone location data appear in a certain data queue, then it will be considered that a new drone has appeared.

[0031] During the location data update process, the drone speed data is calculated simultaneously. Historical latitude, longitude, altitude, and timestamps are used to calculate the speed information of the drone target at a certain location. At this point, the drone status information for each timestamp includes six-dimensional variables, and the six-dimensional data includes three-dimensional coordinate data and three-dimensional velocity data, i.e. The status information of all timestamps of each drone constitutes a continuous historical state sequence for each drone.

[0032] Step S1 is to convert the unordered and unidentified radio communication signals from multiple TDOA passive positioning nodes into a single target UAV state sequence with continuous time sequence that can be used by the GRU network.

[0033] Step S2 includes steps S21-S22.

[0034] Step S21: For the UAV state sequence of multiple single targets By using multiple trained GRU neural networks, a proposed distribution of the next state (location) of each drone is obtained.

[0035] The proposed distribution of the next time step state includes the predicted mean of the next time step state. and prediction uncertainty Σ t ; Each network outputs a GRU state sequence including the three-dimensional coordinate data of the current state and the predicted value of the corresponding three-dimensional velocity data; The mean and covariance of the predicted values ​​of 3D coordinate data and 3D velocity data output from multiple pre-trained GRU neural networks are obtained by calculating the mean and covariance. and As shown below: ; in, These are the predicted mean values ​​for three-dimensional coordinates and three-dimensional velocity, respectively. represents the covariance of the three-dimensional coordinates and the three-dimensional velocity, respectively, indicating uncertainty; This represents a diagonal matrix.

[0036] Based on the historical state sequence of each drone, a proposed distribution of the next state of each drone is obtained; this is done independently.

[0037] For example, there may be 3 or 5 pre-trained GRU neural networks. The specific choice should be based on a balance between scene complexity and computational resources. (1) If real-time performance and lightweight design are emphasized: select 3 GRU neural networks. Calculate the mean and covariance of the outputs of the 3 networks to reduce the prediction bias of a single model to a certain extent, while controlling the computational cost and meeting the real-time requirements of UAV positioning; (2) To achieve higher robustness and accuracy, select 5 GRU neural networks. More parallel outputs of models can further reduce prediction variance. In complex scenarios such as multipath effects and non-line-of-sight propagation, the reliability of the proposed distribution can be significantly improved by estimating the mean and covariance through multi-model fusion, thereby optimizing the localization results of subsequent particle filtering.

[0038] The output covariance matrix (usually diagonal) represents the uncertainty of the GRU in this prediction.

[0039] Using multiple pre-trained GRU neural networks avoids the bias risk of a single model and improves accuracy through a reasonable number of models.

[0040] The innovation of this step lies in using multiple offline pre-trained GRU neural networks to generate the proposal distribution for the next time step. Instead of the traditional GRU generating a prediction for the next time step, it prepares for step S3 to determine the diffusion direction of the particle in the next time step.

[0041] Step S22: Obtain the pre-trained GRU neural network.

[0042] The pre-trained GRU neural network is obtained through the following process; Obtain a training dataset, which includes a UAV real state sequence composed of high-precision RTK (Real-time Kinematic) positioning data and a UAV TDOA state sequence composed of TDOA observation data corresponding to the real state sequence. The GRU neural network is trained under supervision by using real state sequences as training labels and the corresponding UAV TDOA state sequences as input. During training, the mean squared error loss function is used; when the model's loss function value on the validation set is continuously... When the performance no longer decreases within a training cycle, training is terminated, and a pre-trained GRU neural network is obtained.

[0043] For example, the number of consecutive training cycles The value is 10, but it can be changed according to specific needs in practical applications.

[0044] The training samples consist of high-precision RTK positions (sample labels) of drones and corresponding TDOA observation data (sample data). The training samples cover the flight trajectories of drones within the TDOA base station's detection coverage area, including special scenarios such as sharp turns and hovering. The training samples are divided into training, validation, and test sets according to a preset ratio.

[0045] For example, the preset ratio is 7:2:1; the training set is used for training, and the validation set is used for validating the trained model.

[0046] The training set is used to iteratively update the parameters of the GRU neural network, enabling the model to learn the mapping relationship from TDOA observation data to the actual state of the UAV (RTK position and velocity). The validation set is used to evaluate the model's performance during training, determine whether the model is overfitting or underfitting, and determine when to terminate training (such as terminating training when the validation set loss no longer decreases for a preset number of consecutive periods). The test set is used to simulate real-world scenarios after model training to perform a final, unbiased performance evaluation of the model and test its generalization ability.

[0047] The GRU neural network training uses mean squared error as the loss function, which is defined as the squared mean of the Euclidean distance between the state vector predicted by the GRU neural network and the reference RTK state vector, to ensure that the GRU neural network model can accurately predict the position and velocity dynamics of the UAV at the same time.

[0048] The training is terminated when the loss of the GRU neural network on the validation set no longer decreases significantly within a preset number of consecutive epochs.

[0049] The purpose of step S2 is to process the historical state sequence of each UAV through multiple pre-trained GRU neural networks, and output the proposed distribution of the state of each UAV at the next moment (including the predicted mean and prediction uncertainty), so as to provide accurate sampling basis for subsequent particle filtering.

[0050] Step S3, specifically.

[0051] To address the issues of unstable TDOA detection signals and the need to improve the accuracy of predicted positioning, a proposed distributed fusion particle filter algorithm based on GRU is proposed to correct UAV positioning data and perform accurate UAV position prediction.

[0052] Step S3 includes: Initialize the particle set, with each particle represented as ;in, This represents the particle state, corresponding to three-dimensional coordinate data and three-dimensional velocity data; This represents the weight corresponding to each particle; Particle sampling is performed based on the proposed distribution. ;in, The particle state at the next moment; Update the weight of each particle based on the TDOA observation value received at the current time; if no TDOA observation value is received, keep the particle weight unchanged. Perform weight normalization and calculate the effective particle count; determine whether to resample based on the effective particle count. If resampling is performed, the weights of each particle are reset; the predicted next time step is output. Final state estimation results for the drone.

[0053] The first step is to initialize the particle set. Based on the latest real-time location data of a confirmed drone from step S1, N particles are randomly generated within the initial possible location area of ​​the drone; for example, N is 100 particles for subsequent operations. Each particle is represented as... , represents a state. Represents particles status ; The initial possible location area is determined based on the maximum location range that appears in the state sequence of a single target UAV.

[0054] This represents the weight corresponding to each particle. The initial weights are equal and uniformly distributed, as shown below. ; in, To initialize the number of particles.

[0055] Particle state The initial distribution is as shown in formula (3): ; in, The initial mean of the particle state is represented by a six-dimensional vector, whose values ​​are equal to the average values ​​of the initial three-dimensional coordinates and three-dimensional velocities of the particle, respectively. The variance, representing the uncertainty of the particle's state, is determined based on the maximum deviation observed in historical data. Formula (4) means It follows a uniform distribution during initialization; its physical meaning is to represent the distribution of particles during initialization.

[0056] The purpose of the first step is to generate N initial UAV state particles in the initial state, laying the foundation for subsequent particle filtering and GRU model fusion prediction.

[0057] The second step involves sampling based on the proposed distribution of the UAV's next state obtained from the GRU. This step is the particle prediction step, where, for each particle, the traditional method is no longer used. Calculation method. Among them, The next position of a particle is calculated based on its previous position using a prior distribution; random noise is used; this calculation method is the position prediction function of traditional particle filtering. In this invention, particle prediction is performed using the proposal distribution of the UAV's next state generated by GRU, and samples are directly taken from the intelligent proposal distribution generated by GRU, as shown below: ; in, This represents the particle state at the next moment.

[0058] In this invention, the time sequence number starts from 0, and the current time is... The next moment is .

[0059] The beneficial effect of this step is that the propagation of the particle swarm is no longer a blind random walk, but is guided by the proposal distribution generated by GRU to the most likely state region, which greatly improves the efficiency and quality of particle sampling and can achieve higher accuracy with fewer particles.

[0060] The innovation of this step lies in the fact that traditional particle prediction does not utilize the latest observational information, but is based on prior distributions. Particles are sampled from historical process models; if the process model is inaccurate, particles will spread to incorrect regions. This step utilizes the proposed distribution for the next time step generated by GRU. This determines the direction of particle diffusion in the next moment, achieving intelligent prediction.

[0061] For existing drone scenarios, this step can reduce the uncertainty of drone position prediction and improve positioning accuracy.

[0062] The third step is to update the observations. If the actual TDOA observation data is received at the current time... Then calculate the weight of each particle.

[0063] Based on the current time The received TDOA observations are used to update the weights of each particle, as shown below: ; in, For the current moment Received TDOA observations, for Time of the first The state of each particle; The likelihood function is used to measure the degree of matching between the particle state and the TDOA observation. for Time of the first The weight of each particle.

[0064] If no observations are received, the particle weights remain unchanged, which ensures the continuity of position estimation even when observations are missing.

[0065] The third step involves updating the observed position of the drone using the equipment. In a drone scenario, this step can use the actual TDOA observations of the drone to assign weights to the predicted values ​​and then perform a fusion operation between the actual and predicted values.

[0066] The fourth step is weight normalization and resampling, which involves normalizing the updated weights: ; in, For particle serial numbers, This represents the total number of particles.

[0067] The effective number of particles is calculated as follows: ; in, The effective number of particles; for Time of the first The weight of each particle; like When this happens, resampling is triggered to prevent particle degradation; among which, For the preset threshold value, This represents the total number of particles.

[0068] For example, a predefined threshold value, Values In practical applications, modifications can be made according to specific needs.

[0069] After resampling, the weights are reset to: ; The fifth step is to output the drone at the specified time. The location estimate is as follows, and the covariance estimate is also output as a measure of the uncertainty of the estimated location.

[0070] Predicting the next moment The final state estimation results of the UAV are shown below: ; in, For the next moment The final predicted state of the drone includes the drone's three-dimensional coordinate data and three-dimensional velocity data; This provides the three-dimensional coordinates and three-dimensional velocity data for each particle in the particle filter.

[0071] The variables are six-dimensional, including the UAV's three-dimensional coordinates and three-dimensional velocity data.

[0072] Step S3 serves to use the proposal distribution generated by GRU as the core guide for particle filtering. Through a closed-loop mechanism of "intelligent sampling - observation update - dynamic resampling", the prediction prior of deep learning is optimally fused with TDOA observation data, and finally outputs high-precision and high-reliability UAV state estimation results.

[0073] Example 2: A specific embodiment of the present invention discloses a passive positioning system for unmanned aerial vehicles (UAVs) based on the fusion of GRU and particle filter, thereby realizing the passive positioning method for UAVs based on the fusion of GRU and particle filter in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1.

[0074] like Figure 3 As shown, an anti-interference passive positioning system for unmanned aerial vehicles based on the fusion of GRU and particle filter is disclosed. The system includes a TDOA multi-target state sequence generation module M1, a GRU multi-target proposal distribution generation module M2, and a particle filter multi-target state estimation module M3. The TDOA multi-target state sequence generation module M1 is used to collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate historical state sequences of each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data. The GRU multi-objective proposal distribution generation module M2 is used to input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposal distribution of the next state of each UAV based on the GRU state sequence output by the multiple networks. The particle filter multi-target state estimation module M3 is used to obtain the state prediction results of each UAV based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next time step.

[0075] Since the system in this embodiment and the method in Embodiment 1 are related and can be referenced from each other, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the above method embodiment, it also possesses the corresponding technical effects of the above method embodiment.

[0076] In summary, the UAV anti-interference passive positioning method and system based on GRU and particle filter fusion according to the embodiments of the present invention has the following beneficial effects: 1. The proposal distribution generated by the GRU of this invention can accurately guide particles to high-probability regions, greatly improving sampling efficiency. It can achieve the accuracy of thousands of particles in traditional particle filtering with a small number of particles (such as 100), significantly reducing the computational load, improving computational efficiency, facilitating real-time deployment, and making it more suitable for real-time processing on embedded platforms with limited computing resources. It overcomes the shortcomings of traditional particle filtering, which uses a random diffusion model, has extremely low particle sampling efficiency, requires a large number of particles to ensure accuracy, has high computational complexity, and is difficult to run in real time. 2. This invention utilizes the powerful sequence learning capability of the GRU neural network to directly learn the motion patterns of the UAV from historical state sequences. The generated proposal distribution can more accurately predict its future state, achieving high positioning accuracy and effectively tracking complex maneuvers. It significantly improves positioning accuracy and trajectory smoothness in complex scenarios. This overcomes the shortcomings of traditional filtering methods (such as Kalman filtering), which rely on pre-set motion models and cannot accurately describe the complex maneuvers of the UAV, leading to large prediction biases. 3. This invention, through the deep fusion of GRU prediction and particle filtering framework, can continue to provide a high-probability state estimate by relying on the predictive capability of GRU when the TDOA observation signal is lost. When the signal is recovered, it can immediately and seamlessly fuse new observations for correction, exhibiting strong anti-interference and fault tolerance, ensuring positioning continuity, and achieving continuous and stable positioning under discontinuous signal conditions. It overcomes the shortcomings of existing TDOA signals being easily lost due to interference, and the accumulation of errors in traditional methods when the signal is lost for a long time, leading to positioning interruptions.

[0077] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0078] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A passive localization method for unmanned aerial vehicles (UAVs) based on the fusion of GRU and particle filtering, characterized in that, Includes the following steps: Step S1: Collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate a historical state sequence for each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data; Step S2: Input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposed distribution of the next state of each UAV based on the GRU state sequence output by multiple networks. Step S3: Based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next moment, obtain the state prediction results of each UAV.

2. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 1, characterized in that, The proposed distribution of the next time step state includes the predicted mean of the next time step state. and forecast uncertainty ; Each network outputs a GRU state sequence including the three-dimensional coordinate data of the current state and the predicted value of the corresponding three-dimensional velocity data; The mean and covariance of the predicted values ​​of 3D coordinate data and 3D velocity data output from multiple pre-trained GRU neural networks are obtained by calculating the mean and covariance. and As shown below: ； in, These are the predicted mean values ​​for three-dimensional coordinates and three-dimensional velocity, respectively. represents the covariance of the three-dimensional coordinates and the three-dimensional velocity, respectively, indicating uncertainty; This represents a diagonal matrix.

3. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 2, characterized in that, Step S3 includes: Initialize the particle set, with each particle represented as ;in, This represents the particle state, corresponding to three-dimensional coordinate data and three-dimensional velocity data; This represents the weight corresponding to each particle; Particle sampling is performed based on the proposed distribution. ;in, The particle state at the next moment; Update the weight of each particle based on the TDOA observation value received at the current time; if no TDOA observation value is received, keep the particle weight unchanged. Perform weight normalization and calculate the effective particle count; determine whether to resample based on the effective particle count. If resampling is performed, the weights of each particle are reset; the predicted next time step is output. Final state estimation results for the drone.

4. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 3, characterized in that, The effective number of particles is calculated as follows: ； in, The effective number of particles; for Time of the first The weight of each particle; like When this happens, resampling is triggered to prevent particle degradation; among which, For the preset threshold value, This represents the total number of particles.

5. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 3, characterized in that, Predicting the next moment The final state estimation results of the UAV are shown below: ； in, For the next moment The final predicted state of the drone includes the drone's three-dimensional coordinate data and three-dimensional velocity data; This provides the three-dimensional coordinates and three-dimensional velocity data for each particle in the particle filter.

6. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 5, characterized in that, Based on the current time The received TDOA observations are used to update the weights of each particle, as shown below: ； in, For the current moment Received TDOA observations, for Time of the first The state of each particle; The likelihood function is used to measure the degree of matching between the particle state and the TDOA observation. for Time of the first The weight of each particle.

7. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 1, characterized in that, Step S1 includes: The TDOA passive positioning node records the timestamps of radio communication signals emitted by at least one UAV arriving at each TDOA passive positioning node in real time. The arrival time difference is obtained based on the timestamps, and the corresponding longitude, latitude, and altitude coordinate values ​​are calculated based on the arrival time difference. The longitude, latitude, and altitude coordinate values ​​are stored in the data queue. Calculate the spatial distance S between the newly obtained longitude and latitude and the latest longitude and latitude in the historical data queue. If S is less than or equal to a preset threshold, store the newly obtained longitude, latitude, and altitude values ​​into the corresponding historical data queue; otherwise, create a new data queue for storage. Delete the data queue that has not been updated with new 3D coordinate data for m consecutive seconds; determine the data queue that has been stored with 3D coordinate data n times consecutively as the data queue of the real UAV. In each data queue, the longitude, latitude, and altitude values ​​of each data queue element are converted into three-dimensional coordinate data in the world coordinate system. Based on adjacent 3D coordinate data and corresponding timestamps, the corresponding 3D velocity data is calculated. ; The three-dimensional coordinate data and the corresponding three-dimensional velocity data constitute the TDOA observations in the historical state sequence; the historical state sequences of multiple UAVs are obtained, with each UAV corresponding to one historical state sequence.

8. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to claim 7, characterized in that, The spatial distance S between the newly acquired longitude and latitude and the latest longitude and latitude in the historical data queue is calculated as follows: ； in, , ; This indicates the newly acquired longitude and latitude. This indicates the longitude and latitude of the latest data in the historical data queue; The radius is the Earth's radius.

9. The UAV anti-interference passive positioning method based on GRU and particle filter fusion according to any one of claims 1-8, characterized in that, The pre-trained GRU neural network is obtained through the following process; Obtain a training dataset, which includes a UAV real state sequence composed of high-precision RTK positioning data and a UAV TDOA state sequence composed of TDOA observation data corresponding to the real state sequence. The GRU neural network is trained under supervision by using real state sequences as training labels and the corresponding UAV TDOA state sequences as input. During training, the mean squared error loss function is used; when the model's loss function value on the validation set is continuously... When the performance no longer decreases within a training cycle, training is terminated, and a pre-trained GRU neural network is obtained.

10. A passive positioning system for unmanned aerial vehicles (UAVs) based on the fusion of GRU and particle filtering, characterized in that, The system includes a TDOA multi-objective state sequence generation module M1, a GRU multi-objective proposal distribution generation module M2, and a particle filter multi-objective state estimation module M3; The TDOA multi-target state sequence generation module M1 is used to collect the time difference of arrival of radio communication signals of at least one UAV using multiple TDOA passive positioning nodes, and generate historical state sequences of each UAV based on the time difference of arrival; wherein, the state value in each historical state sequence is a TDOA observation value, including three-dimensional coordinate data and three-dimensional velocity data. The GRU multi-objective proposal distribution generation module M2 is used to input the historical state sequence of each UAV into multiple pre-trained GRU neural networks, and generate the proposal distribution of the next state of each UAV based on the GRU state sequence output by the multiple networks. The particle filter multi-target state estimation module M3 is used to obtain the state prediction results of each UAV based on the proposed distribution fusion particle filter algorithm of the state of each UAV at the next time step.

Images