An unmanned aerial vehicle trajectory prediction method based on frequency perception and complementary damaged view enhancement
By employing frequency sensing and complementary damaged view enhancement, the low-frequency and high-frequency features of UAV trajectories are decoupled, and missing information is adaptively repaired. This solves the problems of scale imbalance and signal occlusion in UAV trajectory prediction, achieving high-precision trajectory prediction results.
Patent Information
- Application Number
- CN202611534779.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-03
AI Technical Summary
Existing UAV trajectory prediction technologies suffer from poor prediction stability when dealing with scale imbalances between the altitude and horizontal axes, mixed high and low frequency features, signal masking, and communication packet loss. In particular, they struggle to achieve high-precision trajectory prediction in complex airspace environments.
A frequency-aware and complementary damaged view enhancement method is adopted. Low-frequency and high-frequency features are decoupled by trainable grouped one-dimensional convolutional wavelet decomposition units to construct a complementary damaged view. The missing information is adaptively repaired by a gated residual enhancement module, and finally prediction is performed through a GRU temporal recursive network.
It effectively improves the stability and accuracy of UAV trajectory prediction, enhances the robustness and generalization ability of the model in complex environments, and enables strong engineering applicability and prediction robustness across all scenarios on multi-rotor UAVs.
Smart Images

Figure CN123030912A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent transportation, aircraft navigation, and airspace big data analysis, and in particular to a method for predicting the trajectory of unmanned aerial vehicles (UAVs) based on frequency perception and complementary damaged view enhancement. Background Technology
[0002] With the explosive growth of the low-altitude economy and drone logistics in smart cities, the density of drone traffic in urban airspace is showing a continuous upward trend. High-precision drone trajectory prediction is a key technology for ensuring airspace traffic safety, avoiding aerial conflicts, and proactive route replanning. However, existing drone trajectory prediction technologies face three major problems in practical applications:
[0003] 1. Imbalance between the altitude and horizontal axes: In flight logs obtained from UAV flights, the horizontal axis (East / North) typically spans several kilometers, while the vertical altitude (Up) often remains within a certain altitude range, with minimal variance in altitude variation compared to horizontal motion. When optimizing absolute physical meters, the gradient of the neural network is dominated by the horizontal axis, leading to a significant static prediction bias in the altitude axis.
[0004] 2. Mixed High and Low Frequency Characteristics: The motion of UAVs exhibits both the smoothness of long flight paths determined by navigation points (low frequency) and the instantaneous jitter caused by sudden gusts of wind, obstacle avoidance maneuvers, or sensor noise (high frequency). Existing methods typically use end-to-end temporal networks for direct fitting, which cannot adaptively separate and model the frequency components with different physical characteristics.
[0005] 3. Signal masking and communication packet loss: Urban canyons, electromagnetic shielding from tall buildings, or temporary communication interruptions can cause continuous interruptions in sensor data at low frequencies or discrete noise at high frequencies. Traditional models exhibit extremely poor prediction stability when faced with such incomplete information.
[0006] This invention proposes a frequency domain sensing and complementary damaged view prediction framework, which solves the problem of high non-convergence by local coordinate translation and reset, and significantly improves the generalization accuracy of the model in extreme environments by utilizing learnable wavelet transform and transposed convolution masking mechanism. Summary of the Invention
[0007] The purpose of this invention is to provide a UAV trajectory prediction method based on frequency perception and complementary damaged view enhancement, so as to improve the problems existing in the prior art.
[0008] This invention is implemented as follows: a UAV trajectory prediction method based on frequency sensing and complementary damaged view enhancement, specifically including the following steps:
[0009] S1. Obtain the historical flight state time series data of the UAV under test. After preprocessing and coordinate transformation, perform relative coordinate centering with the end of the observation window as the reference to construct the relative state feature tensor.
[0010] S2. Through trainable grouped one-dimensional convolutional wavelet decomposition units, the relative state feature tensor is decomposed in the time dimension into a multi-scale frequency domain decomposition, decoupled into a low-frequency component representing the flight path trend and a high-frequency component representing instantaneous disturbance.
[0011] S3. Construct a complementary damage view, generate a continuous masking matrix corresponding to the low-frequency component and a discrete random masking matrix corresponding to the high-frequency component, and perform masking on one of the two frequency domain components by randomly selecting a branch to obtain an asymmetric masking damage feature pair, in which one frequency domain feature is damaged and the other frequency domain feature is intact.
[0012] S4. Construct a gated residual enhancement module, using the complete features in the damaged feature pair as the reference source, adaptively repair the missing information of the damaged features and perform feature enhancement, and output the optimized low-frequency features and high-frequency features.
[0013] S5. The optimized low-frequency and high-frequency features are input into the GRU time-series recursive network for extrapolation to obtain the predicted frequency domain components. These components are then reconstructed into a time-domain relative displacement sequence via inverse discrete wavelet and finally restored to an absolute spatial trajectory.
[0014] More preferably, step S1 specifically includes:
[0015] S101. Acquire ten-dimensional raw time-series data containing latitude, longitude, elevation, vacuum velocity, vertical velocity, ground speed, track angle, heading angle, airborne thrust, and physical drag; perform one-dimensional linear interpolation on the raw time-series data to resample to an equal-interval time grid with a step size of 1 second; correct outliers, missing values, and duplicate records with positive and negative infinity; and remove trajectory sequences with a length of less than 50 frames after resampling.
[0016] S102. The three-dimensional geographic coordinates of the UAV are converted into geocentric rectangular coordinates using the standard geodetic transformation formula. Then, the first sampling point of the original time series data is used as the local coordinate projection origin, and the local northeastern ENU rectangular coordinates are mapped through station-centered coordinate rotation transformation.
[0017] S103. Using the last frame of the historical observation sequence as the local reference origin, perform origin centering alignment on all coordinates within the observation window to generate a relative feature time series.
[0018] S104. The ten-dimensional physical features of each sampling point are mapped to a 64-dimensional latent space through a fully connected linear projection layer with shared weights, and the corresponding temporal embedding vectors are fused to construct a relative state feature tensor with dimensions of batch × observation duration × feature dimension.
[0019] More preferably, in step S2, the grouped one-dimensional convolutional wavelet decomposition unit has 64 input channels, 128 output channels, a kernel length of 2, a stride of 2, 64 groups, and no bias term; the kernel weights are initialized to standard Haar wavelet decomposition coefficients, including independent low-pass and high-pass filter operators, which output low-frequency and high-frequency components respectively; the kernel weights are jointly trained end-to-end with the overall model through the Adam optimizer.
[0020] More preferably, in step S3, the low-frequency continuous masking matrix is generated as follows: the low-frequency component is calculated as a sliding window feature fluctuation score in the feature channel dimension and time axis dimension, wherein the fluctuation score is the sum of the absolute values of the first-order differences of the feature values of adjacent time steps within the window; the fluctuation score is denoised and binarized by a preset smoothness threshold to generate a binarized feature map, wherein the region with a fluctuation score less than the preset smoothness threshold is marked as 1, and the region with a fluctuation score greater than or equal to the preset smoothness threshold is marked as 0; the binarized feature map is temporally diffused using a one-dimensional transposed convolution kernel with all 1s, and after normalization by a sign function, it is inverted to obtain a low-frequency continuous masking matrix composed of 0s and 1s, wherein 0 represents the feature masking at the corresponding position, and 1 represents the feature retention at the corresponding position.
[0021] More preferably, in step S3, the high-frequency discrete random masking matrix follows a Bernoulli distribution with a parameter of 0.8. Each element in the matrix is sampled independently, and is assigned a value of 1 with an 80% probability to retain the original features, and is assigned a value of 0 with a 20% probability to clear the corresponding features. This is used to simulate the instantaneous signal loss scenario of airborne sensor interference, electromagnetic pulse interference, and data link packet loss.
[0022] More preferably, in step S3, the specific rules for the random selection of branches are as follows: a decision random number uniformly sampled in the [0,1] interval is introduced, and a predetermined branch probability of 0.5 is used for judgment; when the decision random number is less than 0.5, continuous occlusion is applied to the low-frequency component, while the high-frequency component remains intact; when the decision random number is greater than or equal to 0.5, discrete occlusion is applied to the high-frequency component, while the low-frequency component remains intact; the occlusion mechanism is only enabled during the model training stage, and no occlusion operation is performed during the inference and evaluation stage, where the complete low-frequency and high-frequency components are directly input.
[0023] More preferably, step S4 specifically includes: concatenating the damaged features of the current frequency domain with the corresponding complete reference features in the channel dimension to obtain a hybrid feature tensor with doubled channel dimensions; based on the hybrid feature tensor, using two independent learnable weight matrices and bias terms, calculating adaptive correlation selection weights through the Sigmoid activation function and candidate residual features through the tanh activation function; multiplying the adaptive correlation selection weights and candidate residual features element by element and superimposing them onto the original damaged features to obtain the reconstructed and enhanced frequency domain features.
[0024] More preferably, in step S5, the hidden layer feature channel dimension of the GRU temporal recursive network is 64, and the predicted low-frequency component and predicted high-frequency component are output with a length of 1 / 2 of the prediction duration through autoregressive recursion.
[0025] More preferably, the inverse discrete wavelet reconstruction is implemented using trainable one-dimensional transposed grouped convolution, with 128 input channels, 64 output channels, a kernel size of 2, and a stride of 2. After reconstructing the temporal relative displacement prediction sequence, the local reference origin coordinates at the end of the observation window are superimposed to restore the absolute spatial trajectory.
[0026] More preferably, the model training employs a multi-task joint loss function for end-to-end optimization, the expression of which is: ;in, and These are the future relative temporal domain position prediction sequence and the corresponding actual displacement label sequence, respectively; Indicates the absolute error of the L1 norm; and These are for predicting low-frequency components and predicting high-frequency components, respectively. To output the linear layer parameter matrix, , These are the frequency domain ground truth values obtained from the same wavelet decomposition of the true labels. The square of the L2 norm, To balance the hyperparameters.
[0027] Compared with existing technologies, this invention has the following advantages: By constructing a multi-scale frequency domain coupling and relative coordinate resetting framework, this invention achieves adaptive decoupling between the long-term flight path smoothing trend and the instantaneous vertical / horizontal fluctuation characteristics of UAVs. This effectively improves the prediction bias problems caused by static offset of the Up axis and gradient dominance due to inconsistent physical scales in traditional trajectory prediction algorithms, and enhances the consistency of representation of motion features in different dimensions. By utilizing complementary occlusion views and cross-frequency gated residual linkage mechanisms, the model's feature self-repair capability and robustness are strengthened in scenarios of satellite signal blockage and communication packet loss. This effectively filters high-frequency random noise from wind fields and sensor measurement noise interference, improving the stability of trajectory prediction under complex interference environments. At the same time, by combining frequency domain target optimization and kinematic feature perception strategies without graph structure constraints, this invention not only significantly improves the collaborative prediction accuracy of global flight intention and local dynamic dynamics in complex airspace wind field environments, but also eliminates the computational cost and reliance on prior maps for complex airspace topology mapping. It possesses strong engineering applicability and robust prediction across all scenarios for various types of civilian multi-rotor UAVs. Attached Figure Description
[0028] Figure 1 This is a flowchart of the method steps of the present invention.
[0029] Figure 2 This is a diagram of the frequency domain decomposition architecture of the present invention.
[0030] Figure 3 This is a schematic diagram of the mask construction of the complementary damaged view of the present invention.
[0031] Figure 4 This is a schematic diagram comparing the two-dimensional latitude and longitude trajectory prediction of the model of this invention.
[0032] Figure 5 This is a schematic diagram comparing the time-series height prediction of the model in this invention. Detailed Implementation
[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.
[0034] Example 1: Please refer to Figures 1-5 This embodiment provides a method for predicting UAV trajectories based on frequency sensing and complementary damaged view enhancement, including the following steps:
[0035] Step S1: Spatiotemporal Rectangular Projection and Relative Coordinate Transformation; Obtain the historical flight state sequence of the UAV under test, and project the high-dimensional latitude, longitude, altitude, and multi-dimensional dynamic characteristics into a spatiotemporal feature tensor in the local northeast-sky (ENU) coordinate system. ,in, C represents the historical observation time step, and C represents the feature dimension; the last frame of the observation window is used as the reference point. Perform local relative coordinate centering to construct a relative state feature tensor. Step S1 specifically includes:
[0036] S101: Acquire raw time-series data of the UAV's historical flight status. The raw time-series data includes ten physical features: latitude, longitude, altitude, vacuum speed, vertical speed, ground speed, track angle, heading angle, airborne thrust, and physical drag. Perform linear interpolation on the raw time-series data to eliminate intermittent recording breaks caused by sampling errors from different sensors and electromagnetic interference, so that the physical timestamps of all sampling points are aligned equally and the time sequence is complete.
[0037] Specifically, linear interpolation is performed independently for each of the ten physical features, using one-dimensional linear interpolation and allowing boundary extrapolation. The equally spaced time grid is constructed based on the time range of the UAV trajectory itself, i.e., an arithmetic progression grid with a step size of 1.0 second, rounded down from the earliest physical timestamp to the latest physical timestamp, and interpolated to this unified grid for each feature dimension. The rules for handling missing and outlier values are as follows: positive and negative infinity are replaced with missing values; duplicate records with the same timestamp are deduplicated, with the first record retained and sorted by time; missing values are first filled forward and then backward, with zeros added if still missing; trajectories with an original length or a length less than 50 frames after resampling are completely discarded. This length threshold is higher than the total length of 44 frames (30 historical observation frames and 14 future labels) required for subsequent model input, reserving redundant length for sliding window sampling and data pruning to avoid insufficient effective training samples due to excessively short sequences.
[0038] In this embodiment, the raw time-series data of the UAV's historical flight status is obtained from the publicly available UAV Delivery Simulation Dataset (ATS-UAV Delivery Dataset). This dataset is based on real urban ground food delivery orders and generated using the open-source air traffic simulator ATS, with a data acquisition frequency of 1.0Hz. During the preprocessing stage, the dataset is divided at the individual UAV trajectory file level: after sorting all individual trajectory files, the first 80% constitutes the training set, and the last 20% constitutes the test set. Any complete trajectory of a UAV belongs to only one set, achieving physical isolation between the training and test data. In this embodiment, no independent validation set is set; during training, the average displacement error on this 20% isolated test set is used to evaluate and save the optimal model weights.
[0039] S102: To eliminate floating-point overflow and instability caused by directly performing addition, subtraction, multiplication, and division calculations on large-scale spherical geographic coordinates in the neural network, the system utilizes standard geodetic transformation formulas to convert the UAV's three-dimensional geographic coordinates at time t for any given time t within the time series. Convert to geocentric rectangular coordinates Subsequently, the position of the first sampling point of the original time-series data of this flight state was used as the origin of the local coordinate projection. Using the station-centered coordinate rotation transformation matrix, the geocentric rectangular coordinates are transformed. After performing translation and rotation transformations, the local ENU rectangular coordinates at time t are obtained. ,in The coordinates are eastward. The coordinates are north-facing. The coordinates are celestial coordinates, with superscript. This indicates transpose.
[0040] S103: Addressing the significant physical scale imbalance and altitude invariance of the UAV's Up axis relative to the East-North plane, this embodiment uses the ENU rectangular coordinate system. Perform a local relative coordinate transformation and set the last frame of the current historical observation sequence (i.e., the Tth frame). obs Frame, in this embodiment T obs =30, meaning a total of 30 seconds at a sampling frequency of 1.0Hz. The ENU rectangular coordinates are based on the local reference origin. , This represents the ENU coordinates corresponding to the last frame of the observation window; it also represents the historical time series of the observation window (values). ) and the future time series to be predicted (values) All ENU Cartesian coordinates within the range are centered and aligned to the origin, generating relative feature time series. = ,in, This represents the ENU coordinates at time t relative to the local reference origin. The relative position vector, , Indicates the prediction time step (in this embodiment) =14, that is, predicting the next 14 seconds (14 frames) at a sampling frequency of 1.0Hz); the above transformation makes the prediction starting point zero in three-dimensional space, which greatly reduces the large-scale static bias of the height axis.
[0041] S104: Input the ten-dimensional physical features of each sampling point into the linear projection layer. The linear projection layer is a fully connected mapping network that shares weights across time steps. It independently maps the ten-dimensional features of each time step and projects them uniformly to... Latent space (in this embodiment, the latent space dimension) ).
[0042] At the same time, regarding history Each time step, set the shape to The learnable parameter tensor is used as the temporal embedding, initialized with random values, and trained together with the other network parameters, and broadcast in the batch dimension; it should be clear that no corresponding temporal embedding is generated for future predicted time series. Specifically, the 64-dimensional latent feature vector obtained by linear projection of each historical time t is fused element-wise with the temporal embedding vector corresponding to that time, to construct a relative state feature tensor containing complete spatiotemporal motion semantics. .
[0043] Step S2: Trainable wavelet multi-scale decoupling; such as Figure 2 As shown, trainable grouped one-dimensional convolutional wavelet decomposition units are used to decompose the relative state feature tensor in the time dimension. Multi-scale decomposition is performed to decouple it into low-frequency components that characterize the general trend of the flight path. High-frequency components characterizing instantaneous disturbances .
[0044] Specifically, to effectively separate the long-cycle macroscopic flight path trend of the UAV (low-frequency signal) from short-cycle maneuvers such as wind resistance interference and instantaneous obstacle avoidance (high-frequency signal), this embodiment constructs a set of trainable grouped one-dimensional convolutional layers. This convolutional layer has 64 input channels, 128 output channels, a kernel length of 2, a stride of 2, and 64 groups. It does not include a bias term, and the time dimension is downsampled from 30 frames to 15 frames after convolution. The 64 groups correspond to each input feature channel independently performing a one-dimensional wavelet filter operation, with no shared convolution parameters between groups. The 128 output channels are divided into two paths: the first 64 paths are the low-frequency components output by low-pass filtering, and the last 64 paths are the high-frequency components output by high-pass filtering. This grouped one-dimensional convolutional layer includes low-pass and high-pass filter operators, and the convolution kernel weight parameters of the low-pass and high-pass filter operators are... Initialized with standard Haar wavelet decomposition coefficients, the output low- and high-frequency decomposition results are represented by the following parameterized mapping operator: ; ;in, The input is the relative state feature tensor. , These are the low-frequency and high-frequency parameter decoupling mapping functions, respectively. , These are the low-frequency and high-frequency components obtained from decoupling, respectively. During backpropagation of the data, this convolutional layer is jointly optimized and updated end-to-end by the Adam optimizer (learning rate 0.001) and the remaining parameters of the network. The weights remain fixed during the inference phase, ensuring that the decomposed features are fully adapted to the startup mechanical response features of a specific type of UAV.
[0045] Step S3: Constructing a transposed convolutional complementary corrupted view; as follows Figure 3 As shown, a complementary damage view is constructed, and the low-frequency components are calculated. The sliding smoothness is scored along the time series, and a low-frequency continuous occlusion matrix is generated in the parallel computing core using a one-dimensional transposed convolution operator. At the same time, for high-frequency components Perform discrete random masking to generate high-frequency discrete masking matrices By generating low-frequency and high-frequency masking damage features through asymmetric feature destruction, a pair of masking damage features is achieved. .
[0046] S301, Low-frequency continuous occlusion matrix generation;
[0047] First, calculate the low-frequency components. The sliding window feature fluctuation score is calculated based on the feature channel dimension and the time axis dimension, and the score is calculated as follows: Where c represents the feature channel number, and C represents the total number of feature channels; i represents the temporal position number of the sliding window, and , is the length of the low-frequency sequence after wavelet downsampling, and w is the length of the sliding window; Low-frequency components The element value at time position j in channel c. Characteristic fluctuation score. The kinematic fluctuation intensity of local temporal features is quantified by accumulating the absolute values of the first-order differences of features at adjacent time steps within the window. In this embodiment, the number of feature channels C is set to 64, and the length of the low-frequency sequence after wavelet downsampling is... The value is 15, and the sliding window length w is 4.
[0048] Furthermore, based on the aforementioned characteristic fluctuation score The calculation results are used to set a smoothness scoring threshold. Denoising and binarization are performed to determine the state, and binarized feature maps are generated through filtering. In this embodiment, the preset smoothness scoring threshold... Values The binarized feature map The elements in for: ;
[0049] Subsequently, a one-dimensional convolution is used to process the binarized feature map. Perform temporal backward diffusion to generate a low-frequency continuous masking matrix. : In the formula The weights of the one-dimensional transpose convolution kernel of length w are all 1s (all 1s means that the weights at all positions in the convolution kernel are 1). In this embodiment, w=4. This represents a one-dimensional transpose convolution operation with a stride of 1 and padding of 0. This represents the mathematical symbol for a step function (which outputs 1 when the input is greater than 0, 0 when the input is equal to 0, and -1 when the input is less than 0; because...). Since all are non-negative, the output of the transposed convolution is always non-negative. The values consisting of only 0 and 1 (0 represents occlusion and 1 represents preservation) are used to normalize the numerical nonlinear projection after transpose convolution to a binary matrix state of {0,1}.
[0050] The length of the output sequence of this one-dimensional transpose convolution operation is: The occlusion matrix is perfectly aligned with the original time series length of 15 for the low-frequency components, ensuring a one-to-one match between the dimensions of the occlusion matrix and the feature tensor. This continuous occlusion mechanism generates continuous occlusion regions for maneuver segments with drastic feature fluctuations, forcing the model to utilize the trend information of smooth segments to repair the missing features of the maneuver segments. Compared to random occlusion, this is more consistent with the continuous physical characteristics of the UAV flight trajectory.
[0051] S302, Generation of high-frequency discrete random masking matrix; targeting high-frequency fluctuation components. Generate a high-frequency discrete random masking matrix , It follows a Bernoulli distribution with a parameter of 0.8, i.e. Specifically, by performing Bernoulli independent sampling, Each feature element in the matrix is assigned a probability value independently; each element has an 80% probability of being assigned a value of 1 (indicating that the original high-frequency feature at that position will be preserved without loss in subsequent calculations), and a 20% probability of being assigned a value of 0 (indicating that the high-frequency feature at that position will be forcibly cleared and discarded).
[0052] This multi-point discrete random masking mechanism based on Bernoulli distribution effectively simulates the intermittent instantaneous loss of telemetry and control signals caused by high-frequency resonance of airborne sensors, sudden electromagnetic pulse interference, or short-term packet loss of data links when a real UAV is flying in complex airspace.
[0053] S303, Branch Selection and Generation of Damaged Features by Occlusion;
[0054] Introducing Bernoulli decision random variables , It follows a uniform distribution in the interval [0,1]. To randomly determine the specific frequency domain component branch to be corrupted in the current training iteration, this invention employs a custom branch hedging selector to achieve asymmetric feature corruption. The output of the feature corruption... The complete operational rules are represented by the following piecewise function: In the formula This represents the element-wise Hadamard product, where p is a preset branch selection probability; in this embodiment, p = 0.5. In this embodiment, the decision random number... In order to be in Branch decision random numbers uniformly sampled over an interval, when When the branch probability is less than the preset branch probability of 0.5, perform the following on the low-frequency trend component: Continuous masking, while high-frequency components remain intact; otherwise, perform masking on high-frequency components. Discrete random masking is applied, while low-frequency components remain intact.
[0055] It should be clarified that the aforementioned occlusion mechanism is only enabled during the model training phase, and occlusion disruption is not performed during the inference and evaluation phase; that is, it is directly disabled. and This enables lossless spatial location prediction.
[0056] Step S4: Cross-frequency gated residual adaptive enhancement; Construct a cross-frequency gated residual enhancement module, using the undamaged complementary component as the reference source, and utilize the gate mechanism to adaptively repair and enhance the masked damaged features. By analyzing the damaged features, cross-scale adaptive compensation and dynamic weighting are achieved to obtain an optimized frequency feature representation. and The undamaged complementary component here is the complete feature of another branch that complements the frequency domain of the currently damaged branch: when the low-frequency branch is damaged, the complete high-frequency branch feature is used as the reference source; when the high-frequency branch is damaged, the complete low-frequency branch feature is used as the reference source. Damaged features are repaired through cross-frequency domain information complementarity, which is the core meaning of "cross frequency".
[0057] The occlusion damage feature pairs obtained in step S3 under the current channel are defined as the damage feature branch input. Simultaneously, the undamaged complete control channel features obtained in step S3 corresponding to the current channel are defined as the corresponding complete reference feature branch input. Where B is the batch size and d is the hidden layer channel dimension.
[0058] After determining the input source, the concatenated channel-level hybrid feature tensor as follows: ∈ That is, after splicing, the batch dimension B and the time dimension The feature channel dimension remains unchanged, but increases from d to 2d.
[0059] Based on the hybrid feature tensor Adaptive relevance selection weights With nonlinear candidate residual characteristic matrix The calculation formula is as follows: ; ;in, and These are the corresponding Sigmoid function and hyperbolic tangent function, respectively; and It is a learnable weight matrix with an input feature dimension of 2d (i.e., 128 dimensions) and an output feature dimension of d (i.e., 64 dimensions). and This is the corresponding bias term.
[0060] By injecting adaptive relevance selection weights and candidate residual features into the corresponding damaged feature branches, the reconstructed and enhanced frequency feature representations are obtained. ,in This formula achieves dynamic adaptive fusion of multidimensional features in the channel latent space and local noise filtering, ultimately outputting an optimized low-frequency feature representation. High-frequency feature representation .
[0061] Step S5: Frequency domain recursive prediction and time domain inverse transform reconstruction; representing the optimized frequency features. and In the input sequence predictor, future predictions are recursively made. The predicted low-frequency and high-frequency components at each time step are reconstructed into predicted values of the future time-domain trajectory relative displacement using a trainable inverse discrete wavelet transform. And by superimposing the local origin, it is restored to the true absolute future three-dimensional spatial trajectory.
[0062] S501, Frequency Domain Temporal Recursive Prediction: First, the optimized frequency domain features are fed into a GRU temporal recursive network for extrapolation prediction. The sequence predictor contains two GRU temporal networks with identical structures and independent parameters, corresponding to the low-frequency branch and the high-frequency branch respectively. These two networks perform temporal extrapolation inference in parallel. The hidden layer feature channel dimension of this network is set to 64, and the output length through autoregression is... / 2 position prediction low frequency components With prediction of high-frequency components ;
[0063] S502, Inverse Discrete Wavelet Reconstruction and Absolute Coordinate Restoration; and The input is a trainable inverse discrete wavelet reconstruction unit. The convolutional kernel weights of this reconstruction unit are not shared with the convolutional kernel weights of the forward wavelet decomposition unit in step S2, and there are no orthogonal constraints; both participate independently in end-to-end training. This reconstruction unit is implemented using one-dimensional transposed grouped convolution, with 128 input channels, 64 output channels, a kernel size of 2, and a stride of 2. Through the mapping by the reconstruction unit, the predicted values are reconstructed and mapped into a future relative displacement prediction sequence. : In the formula This represents the inverse discrete wavelet reconstruction mapping function formed by one-dimensional transpose grouped convolution.
[0064] Then, the local reference origin coordinates saved in step S1 are used. By performing broadcast accumulation operations in the spatial dimension, the physical absolute ENU rectangular spatial trajectory can be obtained through lossless reconstruction. In the formula This represents the sequence number of the future time step to be predicted, and its value range is... ; The first output of the model represents the... A three-dimensional relative displacement vector for each future time step.
[0065] S503. Construction of Multi-Task Joint Loss Function: To achieve collaborative training of the prediction head (i.e., the GRU temporal recursive network in step S501) and the model, a multi-task joint loss optimization function is designed. : ;in, and These are the future relative temporal domain position prediction sequence and the corresponding actual displacement label sequence, respectively; Indicates the absolute error of the L1 norm. The square of the L2 norm; and These are the future predicted low-frequency component and the predicted high-frequency component output by the sequence predictor in step S501, respectively. To output the linear layer parameter matrix, The same projection matrix is shared by low-frequency and high-frequency branches. Its function is to project the frequency domain features in the high-dimensional hidden space to a 3-dimensional physical representation space so as to calculate the feature alignment loss. , The wavelet spectral distribution feature benchmark corresponding to the decoupled ground value trajectory is generated as follows: During the training phase, using the "grouped one-dimensional convolutional wavelet decomposition unit" described in S2, the future true displacement label sequence is processed. Multi-scale time-axis filtering decomposition is performed to directly extract the low-frequency and high-frequency energy distributions as the reference ground truth. The weights of the decomposition units are updated in real time with the network weights in step S2 to ensure consistency in the frequency domain feature space. To preset the balance hyperparameter (in this embodiment, we take...) This joint loss-constrained predictor performs end-to-end regression of the trajectory in the time domain, while also conforming to the true macro- and micro-energy distribution at the low-pass and high-pass boundaries in the frequency domain, eliminating the constant cumulative bias of the height axis and trajectory distortion.
[0066] Example 2: Based on Example 1, the prediction results of this invention are as follows: Figure 4 , Figure 5 As shown, Figure 4 The line graph shows the performance of the invention on the two-dimensional horizontal geographic trajectory on the ATS-UAV Delivery Dataset. It can be seen that the latitude and longitude prediction lines are highly fitted to the actual trajectory, and the maximum planar position deviation is only within 0.5 meters. The model can still track the horizontal heading trend of the UAV very well under complex wind fields. Figure 5This demonstrates a dynamic comparison between the predicted altitude and the actual altitude. During the cruising level flight phase, where the altitude change is minimal, the predicted curve and the actual altitude line remain closely aligned, with the maximum vertical deviation controlled within 2 centimeters. This effectively overcomes the technical problem that traditional algorithms are prone to constant drift or divergence on the Up axis.
[0067] To more objectively evaluate the model's global generalization and reconstruction performance under incomplete physical information, this experiment conducted comprehensive inference tests on a segregated 20% test dataset. Experimental results show that the model's average displacement error (ADE) is 2.9088m, the final displacement error (FDE) is 6.0491m, the mean square error (MSE) is 11.1380, the mean absolute error (MAE) is only 1.0818m, and the mean absolute percentage error (MAPE) is as low as 1.2724% under massive unknown delivery samples. Therefore, this invention can stably predict the future short-to-medium-term 3D trajectory of UAVs with meter-level accuracy, possessing excellent engineering application potential.
[0068] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.
Claims
1. A method for predicting the trajectory of a UAV based on frequency sensing and complementary damaged view enhancement, characterized in that, Includes the following steps: S1. Obtain the historical flight state time series data of the UAV under test. After preprocessing and coordinate transformation, perform relative coordinate centering with the end of the observation window as the reference to construct the relative state feature tensor. S2. Through trainable grouped one-dimensional convolutional wavelet decomposition units, the relative state feature tensor is decomposed in the time dimension into a multi-scale frequency domain decomposition, decoupled into a low-frequency component representing the flight path trend and a high-frequency component representing instantaneous disturbance. S3. Construct a complementary damage view, generate a continuous masking matrix corresponding to the low-frequency component and a discrete random masking matrix corresponding to the high-frequency component, and perform masking on one of the two frequency domain components by randomly selecting a branch to obtain an asymmetric masking damage feature pair, in which one frequency domain feature is damaged and the other frequency domain feature is intact. S4. Construct a gated residual enhancement module, using the complete features in the damaged feature pair as the reference source, adaptively repair the missing information of the damaged features and perform feature enhancement, and output the optimized low-frequency features and high-frequency features. S5. The optimized low-frequency and high-frequency features are input into the GRU time-series recursive network for extrapolation to obtain the predicted frequency domain components. These components are then reconstructed into a time-domain relative displacement sequence via inverse discrete wavelet and finally restored to an absolute spatial trajectory.
2. The method according to claim 1, characterized in that, Step S1 specifically includes: S101. Acquire ten-dimensional raw time-series data containing latitude, longitude, elevation, vacuum velocity, vertical velocity, ground speed, track angle, heading angle, airborne thrust, and physical drag; perform one-dimensional linear interpolation on the raw time-series data to resample to an equal-interval time grid with a step size of 1 second; correct outliers, missing values, and duplicate records with positive and negative infinity; and remove trajectory sequences with a length of less than 50 frames after resampling. S102. The three-dimensional geographic coordinates of the UAV are converted into geocentric rectangular coordinates using the standard geodetic transformation formula. Then, the first sampling point of the original time series data is used as the local coordinate projection origin, and the local northeastern ENU rectangular coordinates are mapped through station-centered coordinate rotation transformation. S103. Using the last frame of the historical observation sequence as the local reference origin, perform origin centering alignment on all coordinates within the observation window to generate a relative feature time series. S104. The ten-dimensional physical features of each sampling point are mapped to a 64-dimensional latent space through a fully connected linear projection layer with shared weights, and the corresponding temporal embedding vectors are fused to construct a relative state feature tensor with dimensions of batch × observation duration × feature dimension.
3. The method according to claim 1, characterized in that, In step S2, the grouped one-dimensional convolutional wavelet decomposition unit has 64 input channels, 128 output channels, a kernel length of 2, a stride of 2, 64 groups, and no bias term; the kernel weights are initialized to standard Haar wavelet decomposition coefficients, including independent low-pass and high-pass filter operators, which output low-frequency and high-frequency components respectively; the kernel weights are jointly trained end-to-end with the overall model through the Adam optimizer.
4. The method according to claim 1, characterized in that, In step S3, the low-frequency continuous masking matrix is generated by calculating the sliding window feature fluctuation score of the low-frequency component in the feature channel dimension and the time axis dimension. The fluctuation score is the sum of the absolute values of the first-order differences of the feature values of adjacent time steps within the window. The fluctuation score is denoised and binarized by a preset smoothness threshold to generate a binarized feature map, wherein the region where the fluctuation score is less than the preset smoothness threshold is marked as 1, and the region where the fluctuation score is greater than or equal to the preset smoothness threshold is marked as 0. A one-dimensional transposed convolution kernel with all 1s is used to perform temporal diffusion on the binarized feature map. After normalization by the sign function, the kernel is inverted to obtain a low-frequency continuous occlusion matrix composed of 0 and 1, where 0 represents the occlusion of the corresponding position feature and 1 represents the preservation of the corresponding position feature.
5. The method according to claim 1, characterized in that, In step S3, the high-frequency discrete random masking matrix follows a Bernoulli distribution with a parameter of 0.
8. Each element in the matrix is sampled independently, and is assigned a value of 1 with an 80% probability to retain the original features, and is assigned a value of 0 with a 20% probability to clear the corresponding features. This is used to simulate the instantaneous signal loss scenario of airborne sensor interference, electromagnetic pulse interference and data link packet loss.
6. The method according to claim 1, characterized in that, In step S3, the specific rules for the random selection of branches are as follows: a decision random number uniformly sampled in the [0,1] interval is introduced, and a decision is made in combination with a preset branch probability of 0.5; when the decision random number is less than 0.5, continuous masking is applied to the low-frequency component, while the high-frequency component remains intact; when the decision random number is greater than or equal to 0.5, discrete masking is applied to the high-frequency component, while the low-frequency component remains intact. The masking mechanism is only enabled during the model training phase. Masking is not performed during the inference and evaluation phase, and the complete low-frequency and high-frequency components are directly input.
7. The method according to claim 1, characterized in that, Step S4 specifically includes: concatenating the damaged features of the current frequency domain with the corresponding complete reference features in the channel dimension to obtain a hybrid feature tensor with doubled channel dimensions; based on the hybrid feature tensor, using two independent learnable weight matrices and bias terms, calculating adaptive correlation selection weights through the Sigmoid activation function and candidate residual features through the tanh activation function; multiplying the adaptive correlation selection weights and candidate residual features element-wise and superimposing them onto the original damaged features to obtain the reconstructed and enhanced frequency domain features.
8. The method according to claim 1, characterized in that, In step S5, the hidden layer feature channel dimension of the GRU temporal recursive network is 64, and the predicted low-frequency component and predicted high-frequency component are output with a length of 1 / 2 of the prediction duration through autoregressive recursion.
9. The method according to claim 1, characterized in that, The inverse discrete wavelet reconstruction is achieved using trainable one-dimensional transposed grouped convolution with 128 input channels, 64 output channels, a kernel size of 2, and a stride of 2. After reconstructing the temporal relative displacement prediction sequence, the coordinates of the local reference origin at the end of the observation window are superimposed to restore the absolute spatial trajectory.
