Learning enhanced kalman filtering method and related device for forest emergency positioning

Abstract

The application provides a learning enhanced Kalman filtering method for forest emergency positioning and a related device, time sequence measurement data of an unmanned aerial vehicle and a ground target node is collected; the time sequence measurement data is input into an error compensation network to output ranging error compensation and an observation noise covariance matrix; ranging compensation is performed on the time sequence ranging data, the ranging data compensated by error and state error of the ground target node are fed back to a state prediction network to output a state transition matrix and a process noise covariance matrix; state prediction is performed based on the state transition model and the process noise covariance matrix, Kalman filtering fusion is performed on the predicted state, the compensated ranging data and the observation noise covariance matrix, three-dimensional position estimation of the ground target node is determined, measurement correction and prior prediction accuracy are jointly improved, time-varying noise and nonlinear motion modeling are performed on the basis of retaining the optimality of traditional filtering, and positioning accuracy in a complex forest environment is improved.

Description

Technical Field

[0001] This invention relates to the field of emergency positioning technology, and in particular to a learning-enhanced Kalman filter method and related apparatus for emergency positioning in forest areas. Background Technology

[0002] Emergency positioning technology in complex environments such as mountains and forests is crucial for disaster relief and search and rescue; however, existing positioning methods still have significant limitations. The performance of Global Navigation Satellite Systems (GNSS) deteriorates sharply in scenarios with limited satellite signals, such as dense forests, prompting positioning technologies based on wireless communication networks to become an important supplementary solution. Kalman filtering (KF) is widely used for state estimation, but faces severe challenges in practical applications. Traditional methods heavily rely on accurate state transition and observation models, while the dynamic motion of UAVs and targets in forest environments makes linear modeling extremely difficult. Furthermore, traditional methods make overly idealistic assumptions about noise characteristics, requiring process noise and observation noise to follow a zero-mean Gaussian distribution with known covariance, which is seriously inconsistent with the actual forest environment. Although Kalman filtering (KF) can achieve the optimal solution for state estimation in linear Gaussian systems, its effectiveness depends on a series of relatively strict assumptions. Specifically, KF requires: 1) a state transition model... and observation model It must be accurate and linear; 2) Process noise covariance and observation noise covariance The distribution must be known and strictly follow a zero-mean Gaussian distribution. However, in emergency scenarios such as mountainous and forested areas, these assumptions are often difficult to meet. Due to the influence of complex terrain, it is difficult to design an accurate distribution. or accurately depict Meanwhile, non-line-of-sight (NLoS) errors caused by forest obstruction often lead to... It exhibits time-varying and non-zero mean characteristics, which severely affects the convergence and predictive performance of KF. Summary of the Invention

[0003] This invention provides a learning-enhanced Kalman filter method and related device for emergency positioning in forest areas, which solves the shortcomings of traditional positioning technology in emergency scenarios in forest areas, such as strong model dependence, idealized noise assumptions, and insufficient error compensation capabilities, which make it difficult to meet the actual needs of positioning accuracy and stability.

[0004] This invention provides a learning-enhanced Kalman filter method for emergency positioning in forest areas, comprising: Collect time-series observation data of UAVs and ground target nodes; The time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV are input into the state feedback TimesNet network, which outputs the ranging error compensation and the observation noise covariance matrix. The state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV are input into the state feedback LSTM network, and the state transition matrix and process noise covariance matrix are output. Based on the state transition model and the process noise covariance matrix, the prior state of the ground target node is predicted to obtain the predicted state. The ranging values ​​in the time-series observation data are compensated based on the ranging error compensation amount to obtain the compensated ranging data; The predicted state, the compensated ranging data, and the observation noise covariance matrix are fused using an extended Kalman filter to determine the current three-dimensional position and velocity of the ground target node.

[0005] The learning-enhanced Kalman filter method for emergency positioning in forest areas provided by the present invention further includes preprocessing of the time-series observation data, specifically including: Convert geographic coordinates to a local Euclidean coordinate system; A sliding window method is used to extract observation sequences from multiple consecutive time points; Data augmentation is performed on the observation sequence, including random translation, random rotation, and random flipping of the coordinates; The coordinates and distance measurements are normalized to obtain the temporal feature sequence used as network input.

[0006] According to the learning-enhanced Kalman filtering method for emergency positioning in forest areas provided by the present invention, the step of inputting the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV into a state feedback TimesNet network, and outputting the ranging error compensation and observation noise covariance matrix, includes: The time-series observation data, historical state error sequence, and pitch angle estimation sequence between ground target and UAV are combined to form the input sequence of the state feedback TimesNet network. Perform a Fast Fourier Transform on each input sequence to extract a preset number of dominant frequency components and their corresponding amplitudes; Based on the dominant frequency component, the one-dimensional time-domain sequence is reconstructed into a two-dimensional tensor, and spatiotemporal features are extracted and aggregated through two-dimensional convolution kernels. The aggregated features are input into two branches to predict the non-line-of-sight (NLoS) ranging error compensation and the observation noise covariance matrix, respectively.

[0007] According to the learning-enhanced Kalman filter method for emergency positioning in forest areas provided by the present invention, the step of inputting the state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV into a state feedback LSTM network, and outputting a state transition model and a process noise covariance matrix, includes: The state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV are used as the inputs to the state feedback LSTM network. The state feedback LSTM network extracts temporal features through at least two fully connected input layers and at least one LSTM layer, and outputs a state transition matrix and a process noise covariance matrix for constructing a state transition model through at least one fully connected output layer.

[0008] According to the learning-enhanced Kalman filtering method for emergency positioning in forest areas provided by the present invention, the prior state prediction of ground target nodes based on the state transition model and the process noise covariance matrix includes: Obtain the posterior state estimate from the previous time step; The posterior state estimate is calculated using the state transition model to obtain the prior state prediction at the current time. Based on the process noise covariance matrix and the posterior prediction error covariance of the previous time step, the prior prediction error covariance of the current time step is updated.

[0009] According to the learning-enhanced Kalman filtering method for emergency positioning in forest areas provided by the present invention, the step of fusing the predicted state, the compensated ranging data, and the observation noise covariance matrix using extended Kalman filtering includes: Calculate the partial derivatives of the ranging function with respect to the target position and velocity state to obtain the Jacobian matrix as the observation matrix; Calculate the Kalman gain based on the observation matrix, the prior prediction error covariance, and the observation noise covariance matrix. By combining the Kalman gain with the compensated ranging data, a state update is performed on the prior state prediction value to obtain the posterior state estimate value. Update the posterior prediction error covariance; The position component is extracted from the posterior state estimate as the three-dimensional position estimate of the target node, and the velocity component is extracted as the velocity estimate of the target node.

[0010] According to the learning-enhanced Kalman filter method for emergency positioning in forest areas provided by this invention, the network loss function of the state feedback TimesNet network is a weighted sum of posterior position error loss, ranging error compensation loss, and observation noise variance estimation loss. The network loss function of the state feedback LSTM network is a weighted sum of position loss and velocity prediction loss, where the velocity prediction loss is the mean square error between the predicted velocity and the actual velocity.

[0011] The present invention also provides an emergency positioning device for forest areas, comprising: The acquisition module is used to collect time-series observation data from the UAV and ground target nodes; The observation correction module is used to receive the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV, and output the ranging error compensation and observation noise covariance matrix through the built-in state feedback TimesNet network. The process prediction module is used to receive the state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV. Through the built-in state feedback LSTM network, it outputs the state transition model and the process noise covariance matrix. The filtering estimation module is used to predict the prior state based on the state transition model and the process noise covariance matrix, and to compensate the ranging value based on the ranging error compensation amount. Finally, the predicted state, the compensated ranging data and the observation noise covariance matrix are fused by extended Kalman filtering to determine the three-dimensional position and velocity of the ground target node at the current moment.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the learning-enhanced Kalman filter method for emergency positioning in forest areas as described in any of the preceding claims.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the learning-enhanced Kalman filter method for emergency positioning in forest areas as described above.

[0014] The learning-enhanced Kalman filter method and related device for emergency positioning in forest areas provided by this invention achieves a joint improvement in measurement correction and prior prediction accuracy by fusing a state feedback TimesNet network and a state feedback LSTM network. At the same time, by adopting a state error feedback mechanism and an adaptive Kalman filter framework, it achieves high-precision modeling of time-varying noise and nonlinear motion while retaining the optimality of traditional filtering theory, thus significantly improving the positioning accuracy and stability in complex forest environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the learning-enhanced Kalman filter method for emergency positioning in forest areas provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the dataset construction process provided in an embodiment of the present invention; Figure 3 This is a comparative diagram of the ranging correction error (CDF) of each algorithm provided in the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the positioning error CDF of each algorithm provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the LAKF-NRC algorithm framework provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the functional structure of the forest area emergency positioning device provided in an embodiment of the present invention; Figure 7 This is a functional structure diagram of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Figure 1 The flowchart of the learning-enhanced Kalman filter method for emergency positioning in forest areas provided in this embodiment of the invention is as follows: Figure 1 As shown, the learning-enhanced Kalman filter method for emergency positioning in forest areas provided in this embodiment of the invention includes: Step 101: Collect time-series ranging data between the UAV and ground target nodes; In this embodiment of the invention, a real-world dataset of dynamic positioning based on unmanned aerial vehicles (UAVs) was constructed. The data collection locations covered representative forest environments and were completed using ISAC networking equipment.

[0019] Step 102: Input the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV into the state feedback Timesnet network, and output the ranging error compensation and observation noise covariance matrix. Step 103: Perform ranging compensation on the time-series ranging data based on the ranging error compensation amount. Input the state error of the previous moment, the corrected ranging result of the current moment, and the estimated pitch and azimuth angles from the target to the UAV into the state prediction LSTM network, and output the state transition matrix and process noise covariance matrix. Step 104: Based on the state transition matrix and process noise covariance matrix, perform state prediction, and fuse the predicted state with the compensated ranging data and observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node.

[0020] Traditional Kalman filtering, widely used for state estimation, relies on accurate state transition and observation models. However, the dynamic motion of UAVs and targets in forest environments makes linear modeling extremely difficult. The effectiveness of Kalman filtering depends on a series of relatively strict assumptions, such as the state transition model. and observation model Must be accurate and linear; process noise covariance and observation noise covariance The distribution must be known and strictly follow a zero-mean Gaussian distribution. However, in emergency scenarios such as mountainous and forested areas, these assumptions are often difficult to meet. Due to the influence of complex terrain, it is difficult to design an accurate distribution. or accurately depict Meanwhile, non-line-of-sight (NLoS) errors caused by forest obstruction often lead to... It exhibits time-varying and non-zero mean characteristics, which severely affects the convergence and predictive performance of KF.

[0021] The learning-enhanced Kalman filter method for emergency positioning in forest areas provided in this invention collects time-series measurement data from UAVs and ground target nodes; inputs the features of the time-series measurement data into a state feedback network, outputting a ranging error compensation amount and an observation noise covariance matrix; performs ranging compensation on the time-series ranging data based on the ranging error compensation amount; inputs the error-compensated ranging data and the state error feedback features of the ground target node into a state prediction network, outputting a state transition matrix and a process noise covariance matrix; performs state prediction based on the state transition matrix and the process noise covariance matrix; and fuses the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node. By fusing the state feedback network and the state prediction network, the accuracy of measurement correction and prior prediction is jointly improved. Furthermore, by employing a state error feedback mechanism and an adaptive Kalman filter framework, high-precision modeling of time-varying noise and motion is achieved while retaining the optimality of traditional filtering theory, thus improving positioning accuracy in complex forest environments.

[0022] Based on any of the above embodiments, such as Figure 2 As shown, time-series ranging data between UAVs and ground target nodes were collected using handheld and airborne ISAC networking devices. These devices support the Two-Way Ranging (TWR) protocol, operate at a frequency of 1.4 GHz, have a transmit power of 35 dBm, and a bandwidth of 10 MHz. Under line-of-sight (LoS) conditions, ranging accuracy can reach the meter level, with a maximum ranging distance of 1.5 kilometers. Data collection was conducted in representative mountainous forest environments such as Arxan in Inner Mongolia and Dinghu Mountain in Zhaoqing, Guangdong. Multiple DJI M300 UAVs were equipped with airborne networking devices and recorded real-time flight trajectories through an airborne RTK system. Within the forest area, testers carried handheld networking devices and RTK modules to conduct walking or stationary tests. During the experiment, the UAVs flew at altitudes between 150 and 300 meters above the ground, and their flight trajectories included hovering and rectangular scanning. Ranging and Received Signal Strength Indicator (RSSI) data were collected between all nodes in the network at a frequency of 2.5 Hz. The experiment collected approximately 70,000 time-series measurement samples across various scenarios, including Loss of Space (LoS), sparse forest, and dense forest.

[0023] The learning-enhanced Kalman filter method for emergency positioning in forest areas also includes: preprocessing the time-series observation data, specifically including: Step 201: Convert the geographic coordinates to a local Euclidean coordinate system; Step 202: Use a sliding window to extract the observation sequence of N consecutive time points; Step 203: Normalize the coordinates and distance values ​​to obtain the distance measurement sequence.

[0024] This invention, in an embodiment supporting the collaborative localization of a single ground node by two UAVs, extracts representative sequences from the original dataset and performs several preprocessing steps to improve data usability. First, latitude, longitude, and altitude data are uniformly converted to standard Euclidean space. All sequences are upsampled and downsampled by half in the time dimension to optimize anchor point geometry and improve localization feasibility when the number of UAVs is limited. For sequences containing only a single UAV and a stationary ground node, staggered sampling and random selection of starting points simulate a dual-UAV localization scenario. For moving ground targets and multiple UAVs, two UAVs are selected as anchor points. The feature vector at each time step contains 14 dimensions: .

[0025] The first three dimensions are the three-dimensional coordinates of the ground target to be located, while the rest are the three-dimensional positions of the two UAVs, their respective distances from the target and RSSI values, as well as the timestamp of data acquisition.

[0026] In this embodiment of the invention, each sequence undergoes normalization and enhancement processing when constructing the training and test sets. First, the initial position of the ground target is translated to the origin of the coordinate system. To prevent the localization model from overfitting to a fixed initialization pattern, all coordinate dimensions in the sequence are randomly translated within the range of [-10m, 10m]. Subsequently, the entire sequence is randomly rotated in the xy-plane. Furthermore, the xy-axis is flipped with a 50% probability to utilize the spatial symmetry in the localization process and improve the model's generalization ability. After the above processing, a training set containing 267,000 time-series samples and an independent test set containing 26,700 time-series samples are finally constructed.

[0027] Based on any of the above embodiments, the state feedback network includes a first branch and a second branch. The step of inputting the time-series feature data into the state feedback network and outputting the ranging error compensation amount and the observation noise covariance matrix includes: Step 301: Concatenate the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV to form the input vector of the state prediction network; Step 302: Perform a Fast Fourier Transform on the input feature sequence to extract the top k dominant frequency components and their corresponding amplitudes; Step 303: Extract spatiotemporal features from the time-domain sequences corresponding to each dominant frequency component using a two-dimensional convolution kernel; Step 304: Predict the NLoS ranging error compensation amount through the first branch, and predict the observation noise covariance matrix through the second branch.

[0028] In this embodiment of the invention, the state feedback network is TimesNet, such as... Figure 3As shown, for each input feature sequence, the top k components with the largest amplitudes and their corresponding frequencies in the frequency domain are first extracted using a Fast Fourier Transform (FFT). Based on these frequencies, the original one-dimensional time series is resampled and converted into k two-dimensional tensors. Subsequently, two-dimensional convolutional kernels are used to extract features from these tensors, and the features are aggregated to learn the periodic features of the time series.

[0029] In forest environments, the distance the ranging signal travels through the canopy affects the ranging error, and this distance can be reflected by the drone's pitch angle relative to the user. Therefore, the proposed algorithm feeds back the user's position and accuracy output from the adaptive Kalman filter (AKF) at time t-1 to the network input at time t, thereby enabling the calculation of the drone's pitch angle relative to the user and aiding in the estimation of the ranging error. The network input and output at time t are defined as follows: enter: This contains multiple time series of length seq, with the last element corresponding to the state at time t. These sequences represent the measured RSSI values, the raw ranging results, and feedback sequences for pitch angle and position error, respectively. Specifically, the feedback value at time t is calculated based on the AKF output at time t-1. For example, the pitch angle at time t is based on historically estimated user positions. The position of the UAV at time t is calculated, and the position error at time t is... .

[0030] Output: , representing the correction term for the original ranging result, and the observation noise variance after correction at time t, respectively. .

[0031] To improve the accuracy of the correction term and variance estimate for each ranging result, the loss function of the state feedback network consists of a weighted sum of three parts: posterior position error loss, correction term loss, and corrected ranging observation noise variance estimate loss. The overall loss function is expressed as follows: , For dynamically located targets, posterior position loss Calculate the mean square error (MSE) between the predicted posterior location and the true location: , in, and Let represent the predicted position and the actual position of the i-th sample at time t, respectively. For a sequence of length T i For the i-th sample, the loss is accumulated from the s-th time to T. i-mTime. Since an initial sequence needs to be accumulated for frequency domain analysis, and the loss at some time t depends on the true value at time t+m, the loss calculation for each segment of time series data begins some time after the start and ends early at the end.

[0032] Correction term loss Calculate the corrected distance measurement The mean square error between the distance and the true distance is defined as: , Corrected range measurement observation noise variance estimation loss The mean absolute error (MAE) between the predicted observation noise variance and the statistical variance is calculated and expressed as: , in, The observation noise variance is obtained by calculating the sample variance of the residuals between the corrected measured values ​​and the true values ​​within a local time window.

[0033] Based on any of the above embodiments, the step of inputting the error-compensated ranging data and the state error feedback of the ground target node into the state prediction network, and outputting the state transition matrix and the process noise covariance matrix, includes: Step 401: Combine the state error of the previous moment, the corrected ranging result of the current moment, and the estimated pitch and azimuth angles from the target to the UAV to form the input vector of the state prediction network. Step 402: Input the concatenated features into the state prediction network. The state prediction network extracts temporal features through two fully connected input layers and one LSTM layer; it outputs the state transition matrix through the first branch of a fully connected output layer and the process noise covariance matrix through the second branch.

[0034] Based on any of the above embodiments, the state prediction based on the state transition matrix and the process noise covariance matrix includes: Step 501: Obtain the state estimate from the previous moment; Step 502: Predict the prior state using the state transition matrix; Step 503: Update the prior error covariance matrix based on the process noise covariance matrix; Step 504: Decompose the output of the prior state prediction into position components and velocity components, and apply physical constraints to the velocity components. Step 505: Output the constrained velocity component and the position component as the predicted state vector.

[0035] This invention, in its embodiment, uses an AKF module to simultaneously predict the three-dimensional position and velocity state of a target. The process is as follows: First, predict the prior state: in, The prior state representing the three-dimensional position and velocity at time t is defined as follows: Similarly, Let t be the posterior state at time t-1. For the state transition model, it is defined as: in, The output of the state feedback LSTM network reflects the change in three-dimensional velocity at time t.

[0036] Next, estimate the covariance of the prior prediction error: in, Let be the posterior prediction error covariance at time t-1. The process noise covariance at time t is defined as: in and All of these are the outputs of the state feedback LSTM network at time t.

[0037] Then, calculate the Kalman gain: in, The Jacobian matrix is ​​obtained from the partial derivatives of the ranging function with respect to the target position and velocity state. Based on the dataset settings, there are two calibrated ranging inputs at the same time. To observe the noise covariance, it is composed of the outputs of two independent state feedback TimesNet networks at time t.

[0038] Then, predict the posterior state: in, This represents the corrected ranging value output by the TimesNet network at time t.

[0039] Finally, update the covariance of the posterior prediction error: And then proceed to the next iteration.

[0040] Based on any of the above embodiments, the step of fusing the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node includes: Step 601: Obtain the observation matrix; Step 602: Calculate the Kalman gain based on the observation noise covariance matrix, the prior prediction error covariance, and the observation matrix. Step 603: Perform state update and error covariance matrix update on the predicted state vector; Step 604: Extract the position components from the updated state vector as the 3D position estimate of the target node.

[0041] To improve the accuracy of nonlinear prior state prediction, this embodiment of the invention predicts the changes in three-dimensional velocity and simultaneously estimates the corresponding process noise covariance. This achieves more accurate state priors. The network's input and output at time t are defined as follows: enter: This includes corrected ranging values ​​and observation noise variance estimates from the outputs of two state feedback TimesNet networks, as well as feedback information (including pitch, azimuth, and user position error estimates). The feedback values ​​at time t are calculated based on the AKF output at time t-1. For example, the pitch and azimuth angles at time t are derived from historically estimated user positions. The state error input at time t is calculated from the UAV's position at time t. .

[0042] Output: , represent the change in three-dimensional velocity, and the process noise variance estimate of position and velocity at time t, respectively.

[0043] In this embodiment of the invention, the network loss function of the state prediction network is a weighted sum of the position error loss and the velocity prediction loss, and the velocity prediction loss is the mean square error between the predicted velocity and the actual velocity.

[0044] To improve the accuracy of 3D velocity prediction and process noise variance estimation, the network loss function is designed as a weighted sum of posterior position error loss and velocity prediction loss. The overall loss function is as follows: , For 3D velocity prediction, velocity prediction loss The mean squared error between predicted and actual speeds is measured as follows: It's important to note that the updated velocity is not applied immediately, but rather used in the position prediction at the next time step. Because velocity changes are predicted simultaneously, the true value of the relevant process noise variance is difficult to obtain directly. and The prediction is mainly based on the main loss. Indirectly driven. Although this part does not have an independent supervised loss term, the physical consistency design of LAKF-NRC ensures that the prediction results are consistent with their physical meaning.

[0045] Based on any of the above embodiments, this invention proposes a learning-enhanced adaptive Kalman filter (LAKF-NRC) framework, as follows: Figure 4 As shown, the theoretical optimality of Kalman filtering (KF) is maintained when fusing prediction and observation information. In the observation correction stage, a TimeNet network based on state feedback is designed to extract and analyze frequency domain features, correct NLoS ranging errors caused by forest occlusion, and predict the corrected observation noise covariance. This also makes the error approach a zero-mean Gaussian distribution. In the prior prediction optimization stage, an LSTM network based on state feedback was designed to predict the state transition model. and process noise covariance This also makes the process noise approach a zero-mean Gaussian distribution.

[0046] The framework of this invention was evaluated through ablation experiments and compared with EKF, UKF, and KalmanNet (which directly predicts Kalman gain) in dynamic localization tasks. The results validated the effectiveness of each design and the overall performance advantage of the method. Simulation parameters were set as follows: for each training sequence sample, loss calculation started at time step s=60 and terminated at time step m=5 before the end of the sequence. The sequence length seq used for frequency domain analysis was 50. The weighting coefficients of the loss function were... The NLoS ranging compensation state feedback TimesNet and the prior prediction optimized state feedback LSTM are trained alternately per epoch.

[0047] Table 1 Comparison of ablation experiments and baseline methods

[0048] Table 1 presents the ablation experiments and baseline comparison results of the LAKF-NRC framework. Complete LAKF-NRC in The lowest value (27.382) was obtained, indicating that its dynamic positioning accuracy is the highest.

[0049] Replacing the state-feedback TimesNet with a state-feedback LSTM for NLoS ranging compensation, or replacing the state-feedback LSTM for prior prediction optimization with a state-feedback KalmanNet, both lead to a significant performance degradation. Compared to method 2), method 1) performs better in... It decreased by 65.38%, in The efficiency was reduced by 60.96%, demonstrating the effectiveness of TimesNet in frequency domain feature extraction and utilization. Compared to method 3, method 1) achieved a 60.96% improvement. It decreased by 70.36%, in The performance was reduced by 19.43%, further validating the necessity of maintaining the optimality of KF theory in the system. Furthermore, in the state-feedback LSTM used for prior prediction optimization, feedback from historical states is indispensable due to the need to predict velocity changes; while in the state-feedback TimesNet with NLoS ranging compensation, removing state feedback also leads to performance degradation. Compared to method 4), method 1) achieves a 19.43% reduction in performance. It decreased by 16.98%, in The figure decreased by 23.14%, highlighting the value of the feedback mechanism.

[0050] TimesNet does not introduce state feedback for observation noise covariance Predictions that require ranging correction will also lead to increased losses. Similarly, removing the state feedback LSTM from the process noise covariance... or state transition model The learning ability of the system will also continuously increase positioning and ranging errors, indicating the effectiveness of joint prediction. Compared with methods 5), 6), 7), 8), 9), and 10), method 1) is more effective in... The figures decreased by 52.32%, 75.23%, 79.06%, 54.41%, 69.43%, and 75.66% respectively. The figures decreased by 21.10%, 82.60%, 82.60%, 10.35%, 27.51%, and 17.25%, respectively.

[0051] Furthermore, EKF and UKF, which serve as baselines, do not employ any learning-based components, and KalmanNet does not incorporate the LAKF-NRC framework. All three methods underperform the proposed method in dynamic localization tasks, further validating the overall effectiveness of the proposed framework. Compared to methods 11), 12), and 13), method 1) performs better in... The figures decreased by 84.61%, 84.22%, and 82.13% respectively. The average decrease was 82.60%.

[0052] like Figure 4 and Figure 5As shown, this embodiment of the invention compares the dynamic ranging and localization performance of the trained model and the baseline method, and presents the cumulative distribution function (CDF) curves. The method based on the LAKF-NRC framework consistently outperforms other methods in all metrics. Regarding ranging error, method 1) is 31.11% and 7.63% lower than methods 2) and 3) which also use the LAKF-NRC framework at 90% CDF, respectively. The proposed state feedback TimesNet performs comparably to the LSTM of method 2) in terms of large error correction, but performs better in terms of small error correction. Compared with methods 11), 12), and 13) that do not use the LAKF-NRC framework, method 1) reduces the ranging error by 61.05% at 90% CDF, thanks to the NLoS error compensation mechanism.

[0053] For dynamic positioning error, Method 1) is 36.22% and 2.39% lower than Methods 2) and 3) respectively at 90% CDF, both of which use the LAKF-NRC framework. The proposed state feedback LSTM exhibits more stable performance in dynamic positioning. Compared with Method 1), the KalmanNet approach has an approximately 1.1% higher proportion of samples with positioning errors greater than 30 meters, which is also reflected in Table 1. The higher reason. Compared with the baseline methods 11), 12) and 13 that do not adopt the LAKF-NRC framework, method 1) reduced the localization error at 90% CDF by 60.41%, 60.09% and 53.17% respectively, further highlighting the effectiveness of the proposed LAKF-NRC framework.

[0054] This invention provides a learning-enhanced Kalman filter (KF) method for emergency positioning in forest areas, aiming to address the key limitations of traditional KF in practical positioning tasks, especially under NLoS conditions in mountainous forest areas. It preserves the theoretical optimality of KF when fusing prediction and observation, integrating two types of neural networks during the filtering process to synergistically improve the accuracy of observation correction and prior prediction, achieving a joint improvement in the accuracy of measurement correction and prior prediction. For observation correction, a state feedback-based TimesNet is designed to extract frequency domain features, correct ranging errors caused by NLoS, and predict the observation noise covariance. For prior prediction optimization, a state feedback-based LSTM network is developed to learn the state transition model and process noise covariance. Ultimately, this improves the overall filtering performance and achieves higher-precision dynamic positioning. Furthermore, a real UAV dynamic positioning dataset is constructed, and the proposed method is evaluated on this dataset. The results show that LAKF-NRC consistently outperforms EKF, UKF, and KalmanNet methods in dynamic localization tasks, demonstrating robust and high-precision performance. It provides a practical and generalizable solution for localization problems in complex scenarios and achieves an organic combination of learning models and classical estimation algorithms based on physical consistency.

[0055] The forest area emergency positioning device provided by the present invention is described below. The forest area emergency positioning device described below can be referred to in correspondence with the learning-enhanced Kalman filter method for forest area emergency positioning described above.

[0056] Figure 6 The functional structure diagram of the forest area emergency positioning device provided in the embodiments of the present invention is as follows: Figure 6 As shown, the forest area emergency positioning device provided in this embodiment of the invention includes: The acquisition module 601 is used to acquire time-series ranging data between the UAV and ground target nodes; The first input module 602 is used to input the time-series observation data, historical state error sequence and pitch angle estimation sequence between ground target and UAV into the state feedback Timesnet network, and output the ranging error compensation amount and observation noise covariance matrix. The second input module 603 is used to perform ranging compensation on the time-series ranging data based on the ranging error compensation amount. It inputs the state error of the previous moment, the corrected ranging result of the current moment, and the estimated pitch angle and azimuth angle from the target to the UAV into the state prediction LSTM network, and outputs the state transition matrix and process noise covariance matrix. The estimation module 604 is used to predict the state based on the state transition matrix and the process noise covariance matrix, and to fuse the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node.

[0057] The forest area emergency positioning device provided in this invention collects time-series ranging data between a UAV and a ground target node; inputs the time-series ranging data into a state feedback network, which outputs a ranging error compensation amount and an observation noise covariance matrix; performs ranging compensation on the time-series ranging data based on the ranging error compensation amount; feeds back the error-compensated ranging data and the state error of the ground target node into a state prediction network, which outputs a state transition matrix and a process noise covariance matrix; performs state prediction based on the state transition matrix and the process noise covariance matrix; and fuses the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node. By fusing the state feedback network and the state prediction network, the device achieves a joint improvement in measurement correction and prior prediction accuracy. Furthermore, by employing a state error feedback mechanism and an adaptive Kalman filtering framework, it achieves high-precision modeling of time-varying noise and nonlinear motion while retaining the optimality of traditional filtering theory, thereby improving positioning accuracy in complex forest environments.

[0058] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The memory 730 includes computer programs, an operating system, and acquired data. The processor 710 can call the logical instructions in the memory 730 to execute a learning-enhanced Kalman filter method for emergency positioning in forest areas. This method includes: acquiring time-series ranging data between a UAV and a ground target node; inputting the time-series ranging data into a state feedback network, outputting a ranging error compensation amount and an observation noise covariance matrix; performing ranging compensation on the time-series ranging data based on the ranging error compensation amount; feeding back the error-compensated ranging data and the state error of the ground target node into a state prediction network, outputting a state transition matrix and a process noise covariance matrix; performing state prediction based on the state transition matrix and the process noise covariance matrix; and fusing the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node.

[0059] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a learning-enhanced Kalman filter method for emergency positioning in forest areas provided by the methods described above. The method includes: collecting time-series ranging data between a UAV and a ground target node; inputting the time-series ranging data into a state feedback network and outputting a ranging error compensation amount and an observation noise covariance matrix; performing ranging compensation on the time-series ranging data based on the ranging error compensation amount; feeding back the error-compensated ranging data and the state error of the ground target node into a state prediction network and outputting a state transition matrix and a process noise covariance matrix; performing state prediction based on the state transition matrix and the process noise covariance matrix; and fusing the predicted state with the compensated ranging data and the observation noise covariance matrix using Kalman filtering to determine the three-dimensional position estimate of the ground target node.

[0061] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0062] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A learning-enhanced Kalman filter method for emergency positioning in forest areas, characterized in that, include: Collect time-series observation data of UAVs and ground target nodes; The time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV are input into an error compensation network composed of a state feedback TimesNet network, and the ranging error compensation and observation noise covariance matrix are output. The state error of the previous moment, the corrected ranging result of the current moment, and the estimated pitch and azimuth angles from the target to the UAV are input into the state prediction network composed of the state feedback LSTM network, and the output is the state transition model and the process noise covariance matrix. Based on the state transition model and the process noise covariance matrix, the prior state of the ground target node is predicted to obtain the predicted state. The ranging values ​​in the time-series observation data are compensated based on the ranging error compensation amount to obtain the compensated ranging data; The predicted state, the compensated ranging data, and the observation noise covariance matrix are fused using an extended Kalman filter to determine the current three-dimensional position and velocity of the ground target node.

2. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1, characterized in that, It also includes preprocessing the time-series observation data, specifically including: Convert geographic coordinates to a local Euclidean coordinate system; A sliding window method is used to extract observation sequences from multiple consecutive time points; Data augmentation is performed on the observation sequence, including random translation, random rotation, and random flipping of the coordinates; The coordinates and distance measurements are normalized to obtain the temporal feature sequence used as network input.

3. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1, characterized in that, The process of inputting the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV into the state feedback TimesNet network, and outputting the ranging error compensation and observation noise covariance matrix, includes: The time-series observation data, historical state error sequence, and pitch angle estimation sequence between ground target and UAV are combined to form the input sequence of the state feedback TimesNet network. Perform a Fast Fourier Transform on each input sequence to extract a preset number of dominant frequency components and their corresponding amplitudes; Based on the dominant frequency component, the one-dimensional time-domain sequence is reconstructed into a two-dimensional tensor, and spatiotemporal features are extracted and aggregated through two-dimensional convolution kernels. The aggregated features are input into two branches to predict the non-line-of-sight ranging error compensation and the observation noise covariance matrix, respectively.

4. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1, characterized in that, The process of inputting the state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV into the LSTM network, and outputting the state transition model and the process noise covariance matrix, includes: The state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV are used as the inputs to the state feedback LSTM network. The state feedback LSTM network extracts temporal features through two fully connected input layers and one LSTM layer, and outputs the three-dimensional velocity change and process noise covariance matrix for constructing the state transition model through a fully connected output layer.

5. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1 or 4, characterized in that, The prior state prediction of ground target nodes based on the state transition model and process noise covariance matrix includes: Obtain the posterior state estimate from the previous time step; The posterior state estimate is calculated using the state transition model to obtain the prior state prediction at the current time. Based on the process noise covariance matrix and the posterior prediction error covariance of the previous time step, the prior prediction error covariance of the current time step is updated.

6. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1 or 5, characterized in that, The step of fusing the predicted state, the compensated ranging data, and the observation noise covariance matrix using extended Kalman filtering includes: Calculate the partial derivatives of the ranging function with respect to the target position and velocity state to obtain the Jacobian matrix as the observation matrix; Calculate the Kalman gain based on the observation matrix, the prior prediction error covariance, and the observation noise covariance matrix. By combining the Kalman gain with the compensated ranging data, a state update is performed on the prior state prediction value to obtain the posterior state estimate value. Update the posterior prediction error covariance; The position component is extracted from the posterior state estimate as the three-dimensional position estimate of the target node, and the velocity component is extracted as the velocity estimate of the target node.

7. The learning-enhanced Kalman filter method for emergency positioning in forest areas according to claim 1, characterized in that, The network loss function of the state feedback TimesNet network is a weighted sum of the posterior position error loss, the ranging error compensation loss, and the observation noise variance estimation loss; the network loss function of the state feedback LSTM network is a weighted sum of the position loss and the velocity prediction loss, wherein the velocity prediction loss is the mean square error between the predicted velocity and the actual velocity.

8. An emergency positioning device for forest areas, characterized in that, include: The acquisition module is used to collect time-series observation data from the UAV and ground target nodes; The observation correction module is used to receive the time-series observation data, historical state error sequence, and pitch angle estimation sequence between the ground target and the UAV, and output the ranging error compensation and observation noise covariance matrix through the built-in state feedback TimesNet network. The state transition model prediction module is used to receive the state error of the ground target node at the previous moment, the corrected ranging result at the current moment, and the estimated pitch and azimuth angles from the target to the UAV. It outputs the state transition matrix and process noise covariance matrix through the LSTM network. The filtering estimation module is used to predict the prior state based on the state transition model and the process noise covariance matrix, and to compensate the ranging value based on the ranging error compensation amount. Finally, the predicted state, the compensated ranging data and the observation noise covariance matrix are fused by extended Kalman filtering to determine the three-dimensional position and velocity of the ground target node at the current moment.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the learning-enhanced Kalman filter method for emergency positioning in forest areas as described in any one of claims 1 to 7.

10. A non-transitory readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the learning-enhanced Kalman filter method for emergency positioning in forest areas as described in any one of claims 1 to 7.

Images