A Deep Learning-Based UAV Airborne UWB Localization Method

By using a deep learning-based UWB positioning method to perform clock drift compensation and nonlinear geometric transformation fusion, the accuracy and reliability issues of UWB positioning in high dynamic environments are solved, achieving high accuracy and robustness of UAV autonomous navigation.

CN123111284APending Publication Date: 2026-11-17深圳中杰智控科技有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611359521.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing UWB positioning methods struggle to capture the deep nonlinear coupling characteristics between UWB signal waveforms and motion states in highly dynamic multipath environments. They also lack measurement of the spatial error topology distribution and uncertainty of positioning results, leading to a decrease in positioning accuracy and reliability.

Method used

By using a deep learning-based approach, clock skew evolution trend prediction and adaptive clock drift compensation are performed. Combined with nonlinear geometric transformation fusion and topological entropy uncertainty assessment, a time-synchronized multi-source dataset is generated, and joint feature vectors and corrected coordinates are output to achieve adaptive suppression of inertial drift and reliability screening of positioning results.

Benefits of technology

It effectively suppressed the drift accumulation error of the inertial measurement unit, improved the positioning accuracy and reliability of the UAV in complex dynamic environments, and ensured the robustness of autonomous navigation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123111284A_ABST
    Figure CN123111284A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based airborne UWB localization method for unmanned aerial vehicles (UAVs), relating to the field of UAV autonomous navigation and control technology. The method includes the following steps: S1, generating a time-synchronized multi-source dataset; S2, generating a joint feature vector; S3, obtaining the current iteration compact parameter matrix, and performing gradient constraint parameter optimization and cascaded feature mapping in conjunction with the joint feature vector; S4, performing state quantization mapping and physical constraint solving, and using physical consistency residuals to drive adaptive suppression of inertial drift; S5, performing topological entropy uncertainty assessment based on a three-dimensional error topology graph; S6, outputting the final localization result; and S7, performing gradient iterative updates on the compact parameter matrix. This invention overcomes the limitations of traditional UWB localization methods, such as ignoring clock nonlinear drift, insufficient feature fusion depth, and lack of uncertainty measurement for localization results, providing a robust solution for high-precision autonomous navigation of UAVs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and control technology for unmanned aerial vehicles (UAVs), and in particular to an airborne UWB positioning method for UAVs based on deep learning. Background Technology

[0002] With the expanding application of UAVs in complex scenarios, ultra-wideband (UWB)-based positioning technology faces the challenge of maintaining accuracy in highly dynamic multipath environments. Existing UWB positioning methods, such as time difference of arrival algorithms based on geometric calculations or traditional inertial navigation integrated navigation algorithms, utilize the signal propagation time characteristics and inertial measurement data. However, this approach, which relies solely on geometric observation equations or linear filtering models, ignores the nonlinear dynamic drift evolution trend of the airborne clock under high-speed maneuvers. This leads to time synchronization deviations introducing systematic errors into the positioning calculation, thus limiting the upper limit of accuracy for multi-source data fusion. Furthermore, traditional fusion frameworks typically employ fixed parameter models or shallow feature mappings, making it difficult to capture the deep nonlinear coupling characteristics between UWB signal waveforms and motion states. They also lack the ability to calculate the spatial error topology distribution and uncertainty measurement of positioning results, preventing the positioning system from dynamically selecting effective observations based on environmental changes and reducing the navigation reliability of UAVs in complex electromagnetic environments.

[0003] Therefore, how to provide a deep learning-based UWB positioning method for unmanned aerial vehicles is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a deep learning-based airborne UWB positioning method for unmanned aerial vehicles (UAVs). Through an adaptive clock drift compensation process based on clock skew evolution trend prediction, it synchronizes and aligns multi-source data to generate a time-synchronized multi-source dataset. Then, it performs nonlinear geometric transformation fusion based on statistical distribution characteristics to output a joint feature vector. Based on the cascaded mapping feature matrix and the corrected coordinates, it performs spatial perturbation reconstruction to generate a 3D error topology map. It uses topological entropy uncertainty assessment to output an uncertainty evaluation index, dynamically adjusts the decision threshold based on time-series cumulative motion load, and performs gradient iterative updates on the compact parameter matrix based on a selected sample library. This process establishes a closed-loop feedback path from "spatiotemporal error topology evolution" to "navigation and positioning parameter updates," effectively suppressing the drift accumulation error of the inertial measurement unit (IMU). This ensures that the positioning system can quantify the spatial error distribution in real time and perform adaptive parameter optimization, achieving improved positioning accuracy and reliability in complex dynamic environments. This invention overcomes the limitations of traditional UWB positioning methods, such as ignoring clock nonlinear drift, insufficient feature fusion depth, and lack of uncertainty measurement for positioning results, providing a robust solution for high-precision autonomous navigation of UAVs.

[0005] According to an embodiment of the present invention, a deep learning-based airborne UWB positioning method for unmanned aerial vehicles (UAVs) specifically includes:

[0006] S1. Collect UAV-borne ultra-wideband CIR data, triaxial acceleration and angular velocity data, extract time deviation sequences from UAV-borne ultra-wideband CIR data, and perform adaptive clock drift compensation and alignment on multi-source data based on the predicted evolution trend of time deviation sequences to generate time-synchronized multi-source datasets.

[0007] S2. Extract signal waveform features and motion features from time-synchronized multi-source datasets to construct an initial feature distribution matrix. Perform nonlinear geometric transformation fusion based on the statistical distribution features of the initial feature distribution matrix to generate a joint feature vector.

[0008] S3. Obtain the compact parameter matrix of the current iteration, combine the joint feature vector to perform gradient constraint parameter optimization and cascaded feature mapping, and output the rough position coordinates of the UAV and the cascaded mapping feature matrix.

[0009] S4. Utilize the triaxial acceleration and angular velocity data from the time-synchronized multi-source dataset to perform state quantization mapping and physical constraint solving, and use physical consistency residuals to drive adaptive suppression of inertial drift, outputting the corrected coordinates;

[0010] S5. Based on the cascaded mapping feature matrix and the corrected coordinates, perform spatial perturbation reconstruction to generate a three-dimensional error topology map. Based on the three-dimensional error topology map, perform topological entropy uncertainty assessment and output uncertainty assessment index.

[0011] S6. Construct a time-series cumulative motion load based on triaxial acceleration and angular velocity data, dynamically adjust the decision threshold in combination with uncertainty assessment index, perform reliability screening on the corrected coordinates, and output the final positioning result.

[0012] S7. Calculate the UAV flight navigation and control parameters based on the final positioning results, integrate uncertainty assessment indicators to build an online sample library and perform variance minimization screening, and perform gradient iterative update on the compact parameter matrix based on the screened sample library.

[0013] Optionally, S1 specifically includes:

[0014] S11. Extract the time deviation sequence from the UAV-borne ultra-wideband CIR data, calculate the statistical autocorrelation characteristics and neighborhood density distribution of the time deviation sequence, adaptively determine the state vector construction parameters based on the statistical autocorrelation characteristics and neighborhood density distribution, and use the state vector construction parameters to perform multi-dimensional delay coordinate mapping on the time deviation sequence to construct a multi-dimensional state feature matrix.

[0015] S12. Extract the current state features from the multidimensional state feature matrix, retrieve the Euclidean distance nearest neighbor sample set of the current state features, calculate the weighted average of the evolution data of the nearest neighbor sample set in the next time step, and output the prediction time deviation value.

[0016] S13. Calculate the drift correction increment based on the predicted time deviation value, and generate the dynamic clock drift compensation coefficient based on the drift correction increment;

[0017] S14. Based on the dynamic clock drift compensation coefficient, perform nonlinear interpolation resampling alignment on the UAV-borne ultrawideband CIR data and triaxial acceleration and angular velocity data to generate a time-synchronized multi-source dataset.

[0018] Optionally, the generation of dynamic clock drift compensation coefficients specifically includes:

[0019] Calculate the deviation difference between adjacent time points based on the predicted time deviation value, and output the clock drift rate of change.

[0020] The dynamic smoothing weight is calculated based on the clock drift rate of change. A weighted integral operation is performed based on the dynamic smoothing weight and the predicted time deviation value to output the drift correction increment.

[0021] Nonlinear scaling is performed based on drift correction increments, and dynamic clock drift compensation coefficients are output.

[0022] Optionally, S2 specifically includes:

[0023] S21. Based on time-synchronized multi-source datasets, calculate the time-frequency domain feature vector of UAV-borne ultra-wideband CIR data and the motion statistical feature vector of triaxial acceleration and angular velocity data, and construct an initial feature distribution matrix;

[0024] S22. Based on the initial feature distribution matrix, calculate the global information entropy distribution and local density peak of the feature dimension, determine the feature fusion reference vector based on the global information entropy distribution and local density peak, and construct the nonlinear geometric transformation mapping matrix based on the feature fusion reference vector.

[0025] S23. Based on the nonlinear geometric transformation mapping matrix and the initial feature distribution matrix, perform spatial projection focusing operation to calculate the optimal convergence point of the feature distribution;

[0026] S24. Perform feature dimension reorganization based on the optimal convergence point and output the joint feature vector.

[0027] Optionally, the spatial projection focusing operation specifically includes:

[0028] The initial feature distribution matrix is ​​multiplied based on the nonlinear geometric transformation mapping matrix to output the spatial projection feature matrix;

[0029] The probability density gradient of the multidimensional feature distribution is calculated based on the spatial projection feature matrix, and the feature distribution gradient vector is output.

[0030] Gradient descent is used to search for extreme values ​​based on the gradient vector of the feature distribution, and the optimal convergence point is output.

[0031] Optionally, S3 specifically includes:

[0032] S31. Based on the current iteration compact parameter matrix and joint feature vector, predict the multi-step feature sequence and deduce the reference feature vector;

[0033] S32. Calculate the prediction bias based on the multi-step feature sequence and the reference feature vector, calculate the parameter smoothness based on the current iteration compact parameter matrix, and calculate the weighting coefficient based on the joint feature vector;

[0034] S33. Perform a weighted summation calculation on the prediction bias, parameter smoothness, and weighting coefficients, and output the parameter correction gradient vector;

[0035] S34. Based on the parameter correction gradient vector and joint feature vector, calculate the parameter update boundary, perform gradient constraint solution based on the parameter correction gradient vector and parameter update boundary, and output the predicted optimal projection parameter matrix.

[0036] S35. Perform a linear projection transformation on the joint feature vector based on the predicted optimal projection parameter matrix, and output the concatenated mapping feature matrix;

[0037] S36. Based on the cascaded mapping feature matrix, perform nonlinear coordinate regression calculation and output the approximate position coordinates of the UAV.

[0038] Optionally, S4 specifically includes:

[0039] S41. Perform state-space quantization mapping based on the triaxial acceleration and angular velocity data in the time-synchronized multi-source dataset, and output a multi-dimensional associative memory address index.

[0040] S42. Retrieve the stored historical drift weights based on the multidimensional associative memory address index, perform weighted calculation based on the historical drift weights, and output the predicted inertial drift vector;

[0041] S43. Construct motion trajectory constraint relationships based on triaxial acceleration and angular velocity data, and calculate physical consistency residuals based on motion trajectory constraint relationships and the rough position coordinates of the UAV.

[0042] S44. Calculate the weight correction gradient based on the physical consistency residual and the predicted inertial drift vector, perform parameter iterative update on the historical drift weights based on the weight correction gradient, and output the corrected drift weights.

[0043] S45. Perform weighted summation calculation based on the corrected drift weight and multidimensional associative memory address index, output inertial error compensation vector, correct the rough position coordinates of the UAV based on the inertial error compensation vector, and output the corrected coordinates.

[0044] Optionally, the topological entropy uncertainty assessment specifically includes:

[0045] A baseline reference vector is constructed based on the corrected coordinates, and the variance of the feature distribution is calculated based on the cascaded mapping feature matrix.

[0046] Random perturbation sampling is performed based on the feature distribution variance to generate a perturbation offset vector. The perturbation offset vector is superimposed on the reference vector to output multiple sets of hypothetical positioning samples.

[0047] Calculate the spatial deviation components between the baseline reference vector and multiple sets of hypothetical positioning samples, perform matrix reconstruction on the spatial deviation components, and output the coupling deviation matrix.

[0048] The spatial coordinate boundary extrema are calculated based on the coupling deviation matrix, and discretization mesh generation is performed based on the spatial coordinate boundary extrema to output a three-dimensional mesh coordinate system.

[0049] The spatial distance between node samples is calculated based on the three-dimensional mesh coordinate system and the coupling deviation matrix. Based on the spatial distance between node samples, Gaussian kernel function mapping and cumulative summation are performed to output the local density distribution matrix.

[0050] Threshold segmentation and isosurface reconstruction are performed based on the local density distribution matrix to generate an error distribution space surface and output a three-dimensional error topology map.

[0051] Based on the 3D error topology map, perform mesh space mapping, calculate the local probability density value of each mesh cell, and output the error probability density distribution sequence;

[0052] Log probability mapping is performed based on the error probability density distribution sequence to calculate the spatial distribution information entropy and output the original uncertainty measure.

[0053] The dynamic normalization factor is calculated based on the original uncertainty metric, and a proportional mapping is performed between the original uncertainty metric and the dynamic normalization factor to output the uncertainty assessment index.

[0054] Optionally, S6 specifically includes:

[0055] S61. Calculate the motion impact intensity based on triaxial acceleration and angular velocity data, perform time-series accumulation calculation based on the motion impact intensity, and output the accumulated impact factor;

[0056] S62. Perform dynamic threshold correction based on uncertainty assessment index and output adaptive judgment threshold;

[0057] S63. Perform a comparison and judgment based on the cumulative impact factor and the adaptive judgment threshold. Based on the comparison and judgment results, perform validity screening on the corrected coordinates and output the final positioning result.

[0058] Optionally, S7 specifically includes:

[0059] S71. Calculate the UAV flight navigation and control parameters based on the final positioning result and output them for UAV flight control. Construct feature vectors based on the final positioning result and uncertainty evaluation index, and output the current sample feature vector.

[0060] S72. Calculate the sample covariance matrix based on the current sample feature vector and the stored historical samples, perform variance minimization screening based on the sample covariance matrix, and output the online fine-tuning sample library;

[0061] S73. Calculate the navigation state prediction residual based on the current iteration compact parameter matrix and the online fine-tuning sample library. Perform gradient backpropagation operation on the current iteration compact parameter matrix based on the navigation state prediction residual, and output the optimized compact parameter matrix.

[0062] The beneficial effects of this invention are:

[0063] (1) This invention achieves adaptive compensation for the dynamic drift of the UAV's onboard clock and alignment of multi-source data by constructing a time deviation sequence to predict the evolution trend. It utilizes multi-dimensional delay coordinate mapping to capture the statistical autocorrelation characteristics and neighborhood density distribution of the time deviation sequence, predicts the time deviation value, and generates dynamic clock drift compensation coefficients to drive nonlinear interpolation resampling. This process maps heterogeneous multi-source data to a unified time reference, eliminating the systematic error introduced by clock nonlinear drift and providing highly consistent input data for subsequent positioning calculations.

[0064] (2) This invention establishes a deep feature mapping and localization solution mechanism for multi-source data through nonlinear geometric transformation fusion and gradient constraint parameter optimization. A nonlinear geometric transformation mapping matrix is ​​constructed based on the global information entropy distribution and local density peaks of the initial feature distribution matrix. Spatial projection focusing operation is performed to extract the optimal convergence point of the feature distribution, generating a joint feature vector. Combining gradient constraint parameter optimization and cascaded feature mapping, the approximate position coordinates of the UAV and the cascaded mapping feature matrix are output. This mechanism utilizes statistical distribution characteristics to guide the feature fusion direction and ensures the stability of parameter updates through gradient constraints, achieving efficient mapping from a high-dimensional feature space to a localization coordinate space.

[0065] (3) This invention achieves real-time correction of inertial navigation system errors through adaptive suppression of inertial drift driven by physical consistency residuals. A multi-dimensional associative memory address index is output using state-space quantization mapping to retrieve stored historical drift weights and generate a predicted inertial drift vector. Physical consistency residuals are calculated by constructing motion trajectory constraint relationships, and historical drift weights are corrected using these residuals to output an inertial error compensation vector. This process establishes a combination of physical constraints and data-driven approaches, using closed-loop feedback of physical consistency residuals to correct inertial measurement errors, thus improving the robustness of the positioning system in dynamic environments.

[0066] (4) This invention constructs a spatial error topology representation and dynamic reliability screening system for positioning results through spatial perturbation reconstruction and topological entropy uncertainty assessment. A three-dimensional error topology map is generated based on the corrected coordinates and the cascaded mapping feature matrix. The error probability density distribution sequence is calculated, and the uncertainty assessment index is output. The decision threshold is dynamically adjusted in conjunction with the time-series cumulative motion load to perform reliability screening on the corrected coordinates. This system transforms the geometric distribution of positioning errors into a topological entropy measure and uses the uncertainty index to dynamically constrain the effectiveness of the positioning results, achieving precise removal of abnormal positioning results. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is an overall flowchart of a deep learning-based UAV airborne UWB positioning method proposed in this invention.

[0069] Figure 2 This is a flowchart illustrating the working principle of the topological entropy uncertainty assessment of the UAV airborne UWB positioning method based on deep learning proposed in this invention. Detailed Implementation

[0070] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figure 1 and Figure 2 A deep learning-based UAV airborne UWB localization method, specifically including:

[0072] S1. Collect UAV-borne ultra-wideband CIR data, triaxial acceleration and angular velocity data, extract time deviation sequences from UAV-borne ultra-wideband CIR data, and perform adaptive clock drift compensation and alignment on multi-source data based on the predicted evolution trend of time deviation sequences to generate time-synchronized multi-source datasets.

[0073] S2. Extract signal waveform features and motion features from time-synchronized multi-source datasets to construct an initial feature distribution matrix. Perform nonlinear geometric transformation fusion based on the statistical distribution features of the initial feature distribution matrix to generate a joint feature vector.

[0074] S3. Obtain the compact parameter matrix of the current iteration, combine the joint feature vector to perform gradient constraint parameter optimization and cascaded feature mapping, and output the rough position coordinates of the UAV and the cascaded mapping feature matrix.

[0075] S4. Utilize the triaxial acceleration and angular velocity data from the time-synchronized multi-source dataset to perform state quantization mapping and physical constraint solving, and use physical consistency residuals to drive adaptive suppression of inertial drift, outputting the corrected coordinates;

[0076] S5. Based on the cascaded mapping feature matrix and the corrected coordinates, perform spatial perturbation reconstruction to generate a three-dimensional error topology map. Based on the three-dimensional error topology map, perform topological entropy uncertainty assessment and output uncertainty assessment index.

[0077] S6. Construct a time-series cumulative motion load based on triaxial acceleration and angular velocity data, dynamically adjust the decision threshold in combination with uncertainty assessment index, perform reliability screening on the corrected coordinates, and output the final positioning result.

[0078] S7. Calculate the UAV flight navigation and control parameters based on the final positioning results, integrate uncertainty assessment indicators to build an online sample library and perform variance minimization screening, and perform gradient iterative update on the compact parameter matrix based on the screened sample library.

[0079] In this embodiment, S1 specifically includes:

[0080] S11. Extract arrival timestamp differences from UAV-borne ultrawideband CIR data to construct a time deviation sequence. Set the maximum lag order to 10 to calculate the statistical autocorrelation characteristics of the sequence. Count the number of neighboring samples of data points within a unit time window as the neighborhood density distribution. Determine the embedding dimension as 5 based on the first zero crossing position of the autocorrelation function. Determine the time delay as 2 sampling periods based on the peak position of the neighborhood density. Perform phase space reconstruction on the time deviation sequence using the embedding dimension and time delay to output a multidimensional state feature matrix.

[0081] S12. Extract the last column of the multidimensional state feature matrix as the state feature at the current time. Retrieve the 5 nearest Euclidean distance samples from the historical state feature library to construct the nearest neighbor sample set. Calculate the weighted average of the evolution data of the next time step corresponding to each sample in the set. Set the weight to the square of the inverse of the Euclidean distance. Output the prediction time deviation value.

[0082] S13. Calculate the drift correction increment based on the predicted time deviation value, and generate the dynamic clock drift compensation coefficient based on the drift correction increment.

[0083] S14. Based on the dynamic clock drift compensation coefficient, perform nonlinear scaling transformation on the sampling time axis of the UAV airborne ultra-wideband CIR data, resample the CIR data waveform using cubic spline interpolation, and simultaneously perform the same interpolation alignment operation on the triaxial acceleration and angular velocity data to output a time-synchronized multi-source dataset.

[0084] In its implementation, this invention inherits the basic logic of measuring and correcting time deviations in traditional time synchronization, but it has been deeply modified to address the nonlinear drift problem of UAV onboard clocks in high-speed dynamic environments.

[0085] Traditional techniques typically employ linear interpolation or static parameter models to handle clock skew, neglecting the inherent chaotic dynamics and evolutionary trends within the time skew sequence. This leads to a failure to track the true trajectory of clock drift in real time during dynamic scenarios, resulting in synchronization lag errors. In contrast, this invention creatively constructs a closed-loop system from "time skew phase space reconstruction" to "evolutionary trend prediction and compensation." It calculates the statistical autocorrelation characteristics and neighborhood density distribution of the time skew sequence to adaptively determine the parameters for constructing the state vector. By utilizing multidimensional delay coordinate mapping, it maps the one-dimensional time skew sequence into a multidimensional state feature matrix, revealing the hidden dynamic trajectory.

[0086] Based on this, the nearest neighbor sample set of the Euclidean distance of the current state features is retrieved and the weighted average of the evolution data is calculated to output the predicted time deviation value. Dynamic clock drift compensation coefficients are generated based on drift correction increments to drive nonlinear interpolation resampling alignment.

[0087] This transformation system completely solves the time reference deviation problem caused by the mismatch between the linear synchronization model and complex dynamic drift in traditional methods. It realizes the transformation of implicit clock evolution law into explicit dynamic compensation capability, effectively suppresses the systematic error introduced by clock nonlinear drift, significantly improves the consistency of heterogeneous multi-source data in the time dimension, and provides a strictly aligned data foundation for subsequent high-precision fusion positioning.

[0088] In this embodiment, generating the dynamic clock drift compensation coefficient specifically includes:

[0089] Read the predicted time deviation value sequence, calculate the difference between the current value and the previous value, divide the difference by the set time interval of 0.01 seconds, and output the clock drift rate in nanoseconds per second.

[0090] Calculate the absolute value of the clock drift rate of change, set the rate of change threshold to 5 nanoseconds per second. If the absolute value is greater than the threshold, set the dynamic smoothing weight to 0.9; otherwise, set it to 0.3. Use the dynamic smoothing weight to weight the predicted time deviation value, and perform an accumulation operation on the time axis on the weighted result to output the drift correction increment in nanoseconds.

[0091] Obtain the value of the drift correction increment, map the drift correction increment to the normalized value range using the hyperbolic tangent activation function, multiply the function output value by the preset gain coefficient 10 and add the reference value 1.0 to output the dynamic clock drift compensation coefficient.

[0092] In this embodiment, S2 specifically includes:

[0093] S21. Read the time-synchronized multi-source dataset, calculate the mean, variance, and peak energy spectrum of the ultra-wideband CIR data to form a time-frequency domain feature vector with a dimension of 10, and simultaneously calculate the mean, variance, and skewness of the triaxial acceleration and angular velocity to form a motion statistical feature vector with a dimension of 6. Concatenate the two vectors row by row to construct the initial feature distribution matrix.

[0094] S22. Calculate the information entropy value of each dimension of the initial feature distribution matrix, search for the number of sample points within a 0.5-meter Euclidean distance as the local density peak, select the feature values ​​corresponding to the maximum information entropy and the local density peak to construct the feature fusion reference vector, use the hyperbolic tangent activation function to perform nonlinear mapping on the reference vector and construct a diagonal matrix, and output the nonlinear geometric transformation mapping matrix.

[0095] S23. Based on the nonlinear geometric transformation mapping matrix and the initial feature distribution matrix, perform spatial projection focusing operation to calculate the optimal convergence point of the feature distribution.

[0096] S24. Extract the spatial coordinates of the optimal convergence point, rearrange the values ​​according to the dimensional order of the initial feature distribution matrix, and use the Softmax activation function to perform normalization mapping on the rearranged values ​​to generate and output the joint feature vector.

[0097] In this embodiment, the spatial projection focusing operation specifically includes:

[0098] The nonlinear geometric transformation mapping matrix and the initial feature distribution matrix are called to perform matrix multiplication. The elements of each row of the mapping matrix are multiplied by the elements of each column of the initial feature distribution matrix and then summed. The output is a spatial projection feature matrix with a dimension of 16 multiplied by 200.

[0099] Read the spatial projection feature matrix, set the Gaussian kernel function bandwidth parameter to 0.5, calculate the probability density gradient values ​​of each sample point in the feature space, combine the gradient values ​​of all dimensions, and output the feature distribution gradient vector.

[0100] Set the gradient descent iteration step size to 0.01 and the convergence threshold to 0.001. Starting from the current feature center, update the coordinates in the opposite direction of the feature distribution gradient vector. Stop the iteration when the gradient magnitude is less than the threshold and output the optimal convergence point.

[0101] In this embodiment, S3 specifically includes:

[0102] S31. Read the compact parameter matrix of the current iteration, set the dimension to 32 rows by 64 columns, perform matrix multiplication with the joint feature vector with a dimension of 64 by 1, set the prediction step to 3 steps, calculate the mean vector of the three-step feature sequence, and output the reference feature vector.

[0103] S32. Calculate the Euclidean distance between the multi-step feature sequence and the reference feature vector as the prediction bias, calculate the L2 norm of the compact parameter matrix as the parameter smoothness, and calculate the information entropy of the joint feature vector as the weighting coefficient.

[0104] S33. Set the prediction bias weight to 0.6, the parameter smoothness weight to 0.3, and the weighting coefficient weight to 0.1. Perform a weighted summation operation on the three values, use the LeakyReLU activation function to process the negative components, and output the parameter correction gradient vector.

[0105] S34. Calculate the magnitude of the parameter correction gradient vector, set the parameter update boundary threshold to 0.05, and when the magnitude exceeds the threshold, multiply the gradient vector by the ratio of the threshold to the magnitude to truncate it. Update the compact parameter matrix based on the truncated gradient and output the predicted optimal projection parameter matrix.

[0106] S35. Perform matrix multiplication on the predicted optimal projection parameter matrix and the joint feature vector, normalize the result, and output a cascaded mapping feature matrix with a dimension of 128 by 1.

[0107] S36. Read the cascaded mapping feature matrix, perform fully connected layer mapping, use the Tanh activation function to map the output value to the range of negative 1 to positive 1, multiply by the preset spatial scale coefficient of 500 meters, and output the rough position coordinates of the UAV.

[0108] In its implementation, this invention inherits the basic logic of using feature projection for coordinate calculation in traditional parametric positioning methods, but it has made in-depth modifications to address the problem of unstable feature mapping caused by the lack of constraints in the iterative update of positioning model parameters.

[0109] Traditional techniques typically rely solely on single feature matching errors for parameter backpropagation updates, neglecting the smoothness of the parameter matrix itself and its multi-step evolution trend. This can easily lead to gradient oscillations or getting trapped in local optima, resulting in significant fluctuations in the output coarse coordinates. In contrast, this invention creatively constructs a closed-loop update system from "multi-step evolution trend prediction" to "gradient constraint boundary optimization." Based on a compact parameter matrix, it predicts multi-step feature sequences and derives reference feature vectors. It integrates prediction bias, parameter smoothness, and weighting coefficients to calculate the parameter correction gradient vector and combines it with the parameter update boundary to perform gradient constraint solving, outputting the predicted optimal projection parameter matrix, and then performing linear projection transformation and nonlinear coordinate regression.

[0110] This modified system completely solves the bottleneck problem of positioning accuracy caused by the blindness of parameter updates and the uncertainty of feature mapping in traditional methods. It realizes the transformation of the time evolution characteristics of parameters into the constraint of spatial mapping, effectively suppresses the abnormal gradient fluctuations in the iteration process, significantly improves the model generalization ability and the robustness of coordinate output in the coarse positioning stage, and lays a solid geometric benchmark for subsequent high-precision correction.

[0111] In this embodiment, S4 specifically includes:

[0112] S41. Read the three-axis acceleration and angular velocity data from the time-synchronized multi-source dataset, set the acceleration quantization unit to 0.2 m / s² and the angular velocity quantization unit to 0.05 radians / s, perform rounding operations on the data of each axis and concatenate them, and output the multi-dimensional associative memory address index.

[0113] S42. Use the multidimensional associative memory address index to retrieve the pre-stored 6-dimensional historical drift weights, use the Softmax activation function to perform normalization processing on the weights, and output the predicted inertial drift vector by weighted summation.

[0114] S43. Perform double integration on the three-axis acceleration data, set the integration time window to 0.05 seconds, obtain the displacement increment to construct the motion trajectory constraint relationship, calculate the Euclidean distance between the displacement increment and the change in the rough position coordinates of the UAV, and output the physical consistency residual.

[0115] S44. Calculate the dot product of the physical consistency residual and the predicted inertial drift vector, multiply it by the correction factor 0.01 to obtain the weight correction gradient, update the historical drift weights using the gradient descent method, and output the corrected drift weights.

[0116] S45. Multiply and sum the corrected drift weights with the feature vectors corresponding to the multidimensional associative memory address indexes to output the inertial error compensation vector. Subtract this vector from the rough position coordinates of the UAV to output the corrected coordinates.

[0117] In its implementation, this invention inherits the basic logic of using acceleration and angular velocity data for state estimation and error compensation in traditional inertial navigation, but it has made in-depth modifications to address the drift and divergence problem caused by noise accumulation in the inertial measurement unit during long-term operation.

[0118] Traditional techniques typically employ linear algorithms such as Kalman filtering or complementary filtering to smooth inertial errors, which struggle to effectively capture complex nonlinear drift patterns and lack in-depth utilization of physical motion constraints, resulting in insufficient drift suppression capabilities in dynamic flight scenarios. This invention, however, creatively constructs an adaptive suppression system from "historical drift association retrieval" to "physical consistency closed-loop correction." Based on triaxial acceleration and angular velocity data, it performs state-space quantization mapping to generate a multidimensional associative memory address index, retrieves stored historical drift weights for weighted calculation, constructs motion trajectory constraint relationships to calculate physical consistency residuals, and uses the physical consistency residuals to calculate weight correction gradients to iteratively update historical drift weights, ultimately outputting an inertial error compensation vector to correct coarse position coordinates.

[0119] This modified system completely solves the problem of decreased positioning accuracy caused by the accumulation and divergence of inertial drift over time and the lack of physical constraint correction in traditional methods. It realizes the transformation of discrete historical drift experience into real-time dynamic compensation capability, effectively suppresses the cumulative error of the inertial navigation system, and significantly improves the positioning robustness and accuracy of UAVs in signal rejection or dynamic interference environments.

[0120] In this embodiment, the topological entropy uncertainty assessment specifically includes:

[0121] The corrected coordinates are read to construct a reference vector. The cascaded mapping feature matrix is ​​read to calculate the feature distribution variance. The number of sampling groups is set to 500. Based on the feature distribution variance, a random disturbance offset vector with a standard deviation of 0.1 meters is generated and superimposed on the reference vector to output multiple sets of hypothetical positioning samples.

[0122] The spatial position difference between the baseline reference vector and multiple sets of hypothetical positioning samples is calculated, and the spatial deviation components in the three directions of X-axis, Y-axis and Z-axis are extracted to construct a coupling deviation matrix with a dimension of 500 rows and 3 columns.

[0123] Read the coordinate extrema of the coupling deviation matrix, set the grid spacing to 0.05 meters, perform discretization grid subdivision based on the extremum range, and output a three-dimensional grid coordinate system.

[0124] Calculate the Euclidean distance between each node in the 3D mesh coordinate system and the sample points of the coupling deviation matrix. Set the bandwidth parameter of the Gaussian kernel function to 0.2 meters, perform kernel function mapping and summation, and output the local density distribution matrix.

[0125] The density segmentation threshold is set to 0.01. The local density distribution matrix is ​​binarized, and the isosurfaces with a density value of 0.05 are extracted to reconstruct the error distribution space surface and output a three-dimensional error topology map.

[0126] The three-dimensional error topology map is mapped back to the three-dimensional mesh coordinate system, the local probability density value of each mesh cell is normalized and calculated, and the error probability density distribution sequence is output.

[0127] Read the error probability density distribution sequence, calculate the natural logarithm of each probability value, multiply the logarithm result by the probability value and sum them up, and output the original uncertainty measure value.

[0128] Based on the original uncertainty metric, a dynamic normalization factor is calculated using the Sigmoid activation function, and a ratio mapping is performed to output the uncertainty assessment index.

[0129] In its implementation, this invention inherits the basic logic of error measurement and evaluation based on statistical variance in traditional positioning algorithms, but it has been deeply modified to address the problem that the error distribution in UAV dynamic positioning presents a non-Gaussian and irregular topological form.

[0130] Traditional techniques typically use only the covariance matrix or a single scalar value to describe positioning errors, ignoring the geometric topology (such as holes and connectivity) and probability density distribution details of error samples in three-dimensional space. This makes it difficult to accurately characterize the true uncertainty distribution in complex multipath environments. In contrast, this invention creatively constructs a three-dimensional evaluation system from "spatial perturbation reconstruction" to "topological entropy uncertainty measurement." Based on the corrected coordinates and the variance of the feature distribution, it performs random perturbation sampling to construct hypothetical samples. It generates a three-dimensional error topology map through discretized grid partitioning and Gaussian kernel function mapping. Based on the error probability density distribution sequence, it performs log probability mapping to calculate the spatial distribution information entropy and outputs a normalized uncertainty evaluation index.

[0131] This modified system completely solves the problem of evaluation bias caused by the mismatch between the idealized distribution assumptions and the actual complex error patterns in traditional methods. It transforms the abstract spatial error distribution into a concrete topological entropy index, effectively quantifies the local risk level of the positioning results, and significantly improves the physical accuracy and reliability of uncertainty assessment in dynamic interference environments, providing a solid decision-making basis for subsequent reliability screening.

[0132] In this embodiment, S6 specifically includes:

[0133] S61. Read the triaxial acceleration and angular velocity data, calculate the sum of squares of the values ​​of each axis and take the square root to obtain the instantaneous composite modulus, calculate the absolute value of the difference between the composite modulus at adjacent moments as the motion impact intensity, set the cumulative time window to 2 seconds, perform sliding integral calculation on the motion impact intensity, and output the cumulative impact factor in meters per second.

[0134] S62. Read the uncertainty assessment index, set the benchmark judgment threshold to 1.0, use the hyperbolic tangent activation function to map the uncertainty assessment index to a correction coefficient in the range of 0.8 to 1.2, multiply the benchmark judgment threshold by the correction coefficient, and output the adaptive judgment threshold.

[0135] S63. Compare the cumulative impact factor with the adaptive judgment threshold. When the cumulative impact factor is less than the adaptive judgment threshold, the corrected coordinates are determined to be valid and retained. Otherwise, they are marked as invalid and a rejection operation is performed. The final positioning result is then output.

[0136] In this embodiment, S7 specifically includes:

[0137] S71. Read the final positioning result, calculate the straight-line distance between adjacent coordinate points and divide it by the preset time interval of 0.02 seconds, output the flight speed parameter, calculate the angle between the speed vector and the true north axis, and output the heading angle parameter; extract the three-dimensional coordinate values ​​and uncertainty evaluation index of the final positioning result, and splice them to output the current sample feature vector.

[0138] S72. Read the current sample feature vector and the 500 stored historical samples, calculate the covariance matrix between the feature vectors, extract the diagonal elements of the matrix to calculate the variance value, set the variance threshold to 0.1, filter samples with variance values ​​less than the threshold, and output the online fine-tuning sample library.

[0139] S73. Read the compact parameter matrix of the current iteration, perform forward computation on the online fine-tuning sample library to obtain the predicted state vector, calculate the Euclidean distance between the predicted state vector and the sample label as the navigation state prediction residual, set the learning rate to 0.005, perform gradient backpropagation operation to update the matrix weights, and output the optimized compact parameter matrix.

[0140] Example 1: To verify the feasibility of this invention in autonomous navigation and positioning of unmanned aerial vehicles (UAVs), the method of this invention was applied to the autonomous UAV system for power transmission line inspection at a provincial power company's ultra-high voltage transmission company (hereinafter referred to as "Company P"). In traditional UAV positioning systems, a combined navigation method using Global Navigation Satellite System (GNSS) and Inertial Measurement Unit (IMU) is typically employed. However, in complex environments such as high mountains, canyons, or substations with strong electromagnetic interference, satellite signals are easily blocked or interfered with, leading to a significant decrease or even loss of positioning accuracy. Furthermore, the drift error accumulated over time in traditional inertial navigation is difficult to correct online, failing to meet the stringent requirements for high-precision positioning in refined power transmission line inspection. To address these issues, Company P decided to adopt the UAV positioning method proposed in this invention, based on cascaded mapping feature matrix and error topology evaluation.

[0141] During implementation, Company P first used an UAV-borne ultra-wideband (UWB) module to collect channel impulse response (CIR) data, while simultaneously reading triaxial acceleration and angular velocity data. Based on the CIR data, the system extracted time deviation sequences, calculated statistical autocorrelation characteristics and neighborhood density distribution, determined the embedding dimension to be 5 and the time delay to be 2 sampling periods, and constructed a multi-dimensional state feature matrix. Then, it retrieved the nearest neighbor sample set, calculated the weighted average to output the predicted time deviation value, and generated a dynamic clock drift compensation coefficient accordingly. Nonlinear interpolation resampling and alignment were then performed on the multi-source data to generate a time-synchronized multi-source dataset.

[0142] Subsequently, Company P uses time-synchronized data to calculate time-frequency domain feature vectors and motion statistical feature vectors to construct an initial feature distribution matrix. By calculating the global information entropy distribution and local density peaks, it determines the feature fusion reference vector and constructs a nonlinear geometric transformation mapping matrix. After spatial projection focusing operation, it outputs a joint feature vector. In the positioning solution stage, the system predicts a multi-step feature sequence based on the current iterative compact parameter matrix. By calculating the parameter correction gradient vector and parameter update boundary, it outputs the predicted optimal projection parameter matrix. Then, it performs a linear projection transformation on the joint feature vector, outputting a cascaded mapping feature matrix and the approximate position coordinates of the UAV.

[0143] In the error correction and evaluation stage, this invention calculates the spatial distance of node samples based on a three-dimensional mesh coordinate system and a coupling deviation matrix, performs Gaussian kernel function mapping and summation, and outputs a local density distribution matrix. Through threshold segmentation and isosurface reconstruction, it generates an error distribution spatial surface, outputting a three-dimensional error topology map, and calculates the spatial distribution information entropy accordingly, outputting the original uncertainty metric and uncertainty evaluation index. Simultaneously, the system utilizes an associative memory mechanism to retrieve historical drift weights, corrects inertial drift based on physical consistency residuals, and outputs corrected coordinates. It then combines motion impact intensity and an adaptive judgment threshold to perform validity screening on the corrected coordinates, outputting the final positioning result. Finally, based on the final positioning result, the system constructs the current sample feature vector, filters the online fine-tuning sample library, performs gradient backpropagation, and outputs an optimized compact parameter matrix, achieving continuous online evolution of the positioning model.

[0144] During implementation, the technical team at Company P discovered that, compared to traditional satellite / inertial integrated navigation methods, the method of this invention significantly improves the positioning accuracy and reliability of UAVs in environments with no satellite signals or strong interference. Traditional methods experience rapid divergence in positioning errors after signal loss and cannot assess the reliability of the positioning results. In contrast, the method of this invention effectively achieves high-precision autonomous positioning and real-time state reliability assessment through multi-source data deep spatiotemporal alignment, cascaded feature mapping, and uncertainty quantification based on error topology.

[0145] To further verify the actual performance of the method of the present invention, Company P conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1:

[0146] Table 1 Performance Comparison of Company P UAV Positioning Systems

[0147] index Traditional methods Method of the present invention Increase Average positioning error (meters) 2.5 0.3 -88.0% Clock synchronization deviation (nanoseconds) 5.2 0.4 -92.3% Location availability in complex environments (%) 75.0 99.2 +24.2% Cumulative error of inertial drift (m / min) 1.8 0.2 -88.9% Time taken for a single localization calculation (milliseconds) 25 18 -28.0% Outlier removal accuracy (%) 80.5 96.0 +15.5% Smoothness of inspection trajectory tracking (%) 85.0 97.5 +12.5%

[0148] As shown in Table 1, the performance of the UAV positioning system was comprehensively improved after applying the method of this invention. The average positioning error decreased from 2.5 meters in the traditional method to 0.3 meters, and the clock synchronization deviation decreased from 5.2 nanoseconds to 0.4 nanoseconds, significantly improving the accuracy of positioning and time-frequency synchronization. In complex environments with satellite signal rejection, the positioning availability increased from 75.0% to 99.2%, solving the pain point of traditional methods relying on satellite signals. Simultaneously, the system added an uncertainty assessment function with an accuracy rate of 98.5%, providing crucial risk information for the UAV's autonomous decision-making. The cumulative error of inertial drift was significantly reduced, and the accuracy of outlier removal was significantly improved, ensuring the stability and purity of the positioning output.

[0149] Through the method of this invention, Company P has successfully achieved high-precision and high-reliability autonomous positioning of power transmission line inspection drones in complex environments, effectively solving the navigation failure problem caused by satellite signal loss. By quantifying uncertainty, it provides a "digital safety belt" for flight safety, significantly reducing the risk of inspection operations and providing strong technical support for intelligent operation and maintenance of the power grid.

[0150] 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 equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based UAV airborne UWB positioning method, characterized in that, Includes the following steps: S1. Collect UAV-borne ultra-wideband CIR data, triaxial acceleration and angular velocity data, extract time deviation sequences from UAV-borne ultra-wideband CIR data, and perform adaptive clock drift compensation and alignment on multi-source data based on the predicted evolution trend of time deviation sequences to generate time-synchronized multi-source datasets. S2. Extract signal waveform features and motion features from time-synchronized multi-source datasets to construct an initial feature distribution matrix. Perform nonlinear geometric transformation fusion based on the statistical distribution features of the initial feature distribution matrix to generate a joint feature vector. S3. Obtain the compact parameter matrix of the current iteration, combine the joint feature vector to perform gradient constraint parameter optimization and cascaded feature mapping, and output the rough position coordinates of the UAV and the cascaded mapping feature matrix. S4. Utilize the triaxial acceleration and angular velocity data from the time-synchronized multi-source dataset to perform state quantization mapping and physical constraint solving, and use physical consistency residuals to drive adaptive suppression of inertial drift, outputting the corrected coordinates; S5. Based on the cascaded mapping feature matrix and the corrected coordinates, perform spatial perturbation reconstruction to generate a three-dimensional error topology map. Based on the three-dimensional error topology map, perform topological entropy uncertainty assessment and output uncertainty assessment index. S6. Construct a time-series cumulative motion load based on triaxial acceleration and angular velocity data, dynamically adjust the decision threshold in combination with uncertainty assessment index, perform reliability screening on the corrected coordinates, and output the final positioning result. S7. Calculate the UAV flight navigation and control parameters based on the final positioning results, integrate uncertainty assessment indicators to build an online sample library and perform variance minimization screening, and perform gradient iterative update on the compact parameter matrix based on the screened sample library.

2. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Extract the time deviation sequence from the UAV-borne ultra-wideband CIR data, calculate the statistical autocorrelation characteristics and neighborhood density distribution of the time deviation sequence, adaptively determine the state vector construction parameters based on the statistical autocorrelation characteristics and neighborhood density distribution, and use the state vector construction parameters to perform multi-dimensional delay coordinate mapping on the time deviation sequence to construct a multi-dimensional state feature matrix. S12. Extract the current state features from the multidimensional state feature matrix, retrieve the Euclidean distance nearest neighbor sample set of the current state features, calculate the weighted average of the evolution data of the nearest neighbor sample set in the next time step, and output the prediction time deviation value. S13. Calculate the drift correction increment based on the predicted time deviation value, and generate the dynamic clock drift compensation coefficient based on the drift correction increment; S14. Based on the dynamic clock drift compensation coefficient, perform nonlinear interpolation resampling alignment on the UAV-borne ultrawideband CIR data and triaxial acceleration and angular velocity data to generate a time-synchronized multi-source dataset.

3. The UAV airborne UWB positioning method based on deep learning according to claim 2, characterized in that, The generation of dynamic clock drift compensation coefficients specifically includes: Calculate the deviation difference between adjacent time points based on the predicted time deviation value, and output the clock drift rate of change. The dynamic smoothing weight is calculated based on the clock drift rate of change. A weighted integral operation is performed based on the dynamic smoothing weight and the predicted time deviation value to output the drift correction increment. Nonlinear scaling is performed based on drift correction increments, and dynamic clock drift compensation coefficients are output.

4. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Based on time-synchronized multi-source datasets, calculate the time-frequency domain feature vector of UAV-borne ultra-wideband CIR data and the motion statistical feature vector of triaxial acceleration and angular velocity data, and construct an initial feature distribution matrix; S22. Based on the initial feature distribution matrix, calculate the global information entropy distribution and local density peak of the feature dimension, determine the feature fusion reference vector based on the global information entropy distribution and local density peak, and construct the nonlinear geometric transformation mapping matrix based on the feature fusion reference vector. S23. Based on the nonlinear geometric transformation mapping matrix and the initial feature distribution matrix, perform spatial projection focusing operation to calculate the optimal convergence point of the feature distribution; S24. Perform feature dimension reorganization based on the optimal convergence point and output the joint feature vector.

5. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, The spatial projection focusing operation specifically includes: The initial feature distribution matrix is ​​multiplied based on the nonlinear geometric transformation mapping matrix to output the spatial projection feature matrix; The probability density gradient of the multidimensional feature distribution is calculated based on the spatial projection feature matrix, and the feature distribution gradient vector is output. Gradient descent is used to search for extreme values ​​based on the gradient vector of the feature distribution, and the optimal convergence point is output.

6. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Based on the current iteration compact parameter matrix and joint feature vector, predict the multi-step feature sequence and deduce the reference feature vector; S32. Calculate the prediction bias based on the multi-step feature sequence and the reference feature vector, calculate the parameter smoothness based on the current iteration compact parameter matrix, and calculate the weighting coefficient based on the joint feature vector; S33. Perform a weighted summation calculation on the prediction bias, parameter smoothness, and weighting coefficients, and output the parameter correction gradient vector; S34. Based on the parameter correction gradient vector and joint feature vector, calculate the parameter update boundary, perform gradient constraint solution based on the parameter correction gradient vector and parameter update boundary, and output the predicted optimal projection parameter matrix. S35. Perform a linear projection transformation on the joint feature vector based on the predicted optimal projection parameter matrix, and output the concatenated mapping feature matrix; S36. Based on the cascaded mapping feature matrix, perform nonlinear coordinate regression calculation and output the approximate position coordinates of the UAV.

7. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Perform state-space quantization mapping based on the triaxial acceleration and angular velocity data in the time-synchronized multi-source dataset, and output a multi-dimensional associative memory address index. S42. Retrieve the stored historical drift weights based on the multidimensional associative memory address index, perform weighted calculation based on the historical drift weights, and output the predicted inertial drift vector; S43. Construct motion trajectory constraint relationships based on triaxial acceleration and angular velocity data, and calculate physical consistency residuals based on motion trajectory constraint relationships and the rough position coordinates of the UAV. S44. Calculate the weight correction gradient based on the physical consistency residual and the predicted inertial drift vector, perform parameter iterative update on the historical drift weights based on the weight correction gradient, and output the corrected drift weights. S45. Perform weighted summation calculation based on the corrected drift weight and multidimensional associative memory address index, output inertial error compensation vector, correct the rough position coordinates of the UAV based on the inertial error compensation vector, and output the corrected coordinates.

8. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, The topological entropy uncertainty assessment specifically includes: A baseline reference vector is constructed based on the corrected coordinates, and the variance of the feature distribution is calculated based on the cascaded mapping feature matrix. Random perturbation sampling is performed based on the feature distribution variance to generate a perturbation offset vector. The perturbation offset vector is superimposed on the reference vector to output multiple sets of hypothetical positioning samples. Calculate the spatial deviation components between the baseline reference vector and multiple sets of hypothetical positioning samples, perform matrix reconstruction on the spatial deviation components, and output the coupling deviation matrix. The spatial coordinate boundary extrema are calculated based on the coupling deviation matrix, and discretization mesh generation is performed based on the spatial coordinate boundary extrema to output a three-dimensional mesh coordinate system. The spatial distance between node samples is calculated based on the three-dimensional mesh coordinate system and the coupling deviation matrix. Based on the spatial distance between node samples, Gaussian kernel function mapping and cumulative summation are performed to output the local density distribution matrix. Threshold segmentation and isosurface reconstruction are performed based on the local density distribution matrix to generate an error distribution space surface and output a three-dimensional error topology map. Based on the 3D error topology map, perform mesh space mapping, calculate the local probability density value of each mesh cell, and output the error probability density distribution sequence; Log probability mapping is performed based on the error probability density distribution sequence to calculate the spatial distribution information entropy and output the original uncertainty measure. The dynamic normalization factor is calculated based on the original uncertainty metric, and a proportional mapping is performed between the original uncertainty metric and the dynamic normalization factor to output the uncertainty assessment index.

9. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Calculate the motion impact intensity based on triaxial acceleration and angular velocity data, perform time-series accumulation calculation based on the motion impact intensity, and output the accumulated impact factor; S62. Perform dynamic threshold correction based on uncertainty assessment index and output adaptive judgment threshold; S63. Perform a comparison and judgment based on the cumulative impact factor and the adaptive judgment threshold. Based on the comparison and judgment results, perform validity screening on the corrected coordinates and output the final positioning result.

10. The UAV airborne UWB positioning method based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Calculate the UAV flight navigation and control parameters based on the final positioning result and output them for UAV flight control. Construct feature vectors based on the final positioning result and uncertainty evaluation index, and output the current sample feature vector. S72. Calculate the sample covariance matrix based on the current sample feature vector and the stored historical samples, perform variance minimization screening based on the sample covariance matrix, and output the online fine-tuning sample library; S73. Calculate the navigation state prediction residual based on the current iteration compact parameter matrix and the online fine-tuning sample library. Perform gradient backpropagation operation on the current iteration compact parameter matrix based on the navigation state prediction residual, and output the optimized compact parameter matrix.