Wind profile radar spectral moment estimation method based on GAT-Transform hybrid model
By using a GAT-Transformer hybrid model, radar echo data is preprocessed and features are extracted to construct a heterogeneous map and estimate the wind profile. This solves the problem of insufficient accuracy in wind profile spectral moment estimation in complex atmospheric environments and achieves high-precision wind profile spectral moment estimation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-04-14
AI Technical Summary
Under complex atmospheric conditions, especially in low signal-to-noise ratio environments, wind profiler radar spectral moment estimation is difficult to effectively filter out pure atmospheric components. Clutter masks the signal, resulting in insufficient accuracy of wind profiler spectral moment estimation.
A method based on the GAT-Transformer hybrid model is adopted to preprocess radar echo data, construct heterogeneous maps and perform feature extraction and encoding, combine in-layer and out-of-layer aggregation modules and Transformer models to estimate wind profiles, and use relative beam parallel training and wind vector synthesis to achieve high-precision wind profile spectral moment estimation.
High-precision wind profile spectral moment estimation was achieved in complex atmospheric environments, improving the robustness and accuracy of the estimation. Experimental results show that the correlation coefficients with radiosonde data reached 0.949 and 0.934, which are significantly better than existing methods.
Smart Images
Figure CN121856970A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind profiler radar signal processing technology, and in particular to a method for estimating the spectral moment of wind profiler radar based on the GAT-Transformer hybrid model. Background Technology
[0002] Boundary layer wind profiler radar, as a key tool for atmospheric remote sensing, receives echo signals scattered by Bragg scattering through atmospheric turbulence and continuously and at high resolution detects horizontal wind speeds and vertical wind profile profiles within the boundary layer. It has important application value in atmospheric dynamics, numerical weather prediction, boundary layer inversion, and low-altitude navigation.
[0003] After processing radar echo data, a range-Doppler spectrum with two dimensions—range and Doppler—can be obtained. Each range gate in the range dimension contains a Doppler power spectrum. Ideally, a clean atmospheric signal follows a Gaussian distribution in its power spectrum. Its zeroth, first, and second moments correspond to the power value, average Doppler velocity, and spectral width of the atmospheric signal, respectively. By correctly estimating the spectral moments of the atmospheric signal in each range gate within the range-Doppler spectrum and concatenating the estimation results, a continuous wind profile can be obtained. However, under complex atmospheric conditions, especially in low signal-to-noise ratio environments, the power spectrum is easily contaminated, exhibiting a cluttered, multi-peaked shape. The atmospheric signal is masked by clutter, making it difficult to effectively filter out clean atmospheric components, which greatly challenges wind profile spectral moment estimation.
[0004] In recent years, researchers have proposed a series of signal processing algorithms to solve the problem of wind profile spectral moment estimation. Early methods included the Vaisala Single Peak Selection Program (SPP), which identifies the highest signal peak in the power spectrum as the atmospheric signal to determine the profile. While effective under ideal conditions, it is prone to failure under low signal-to-noise ratios or clutter interference. The Fuzzy Logic Algorithm (FL) suppresses clutter by using statistical power spectrum features (such as variance and gradient), effectively separating atmospheric and clutter signals. The Adaptive Moment Estimation Method (AME) improves spectral moment recognition by setting dynamic thresholds. The Multi-Parameter Cost Function Method (MPCF) introduces wind shear constraints and multi-parameter optimization to achieve an adaptive estimation process. However, most of these methods are based on correlations with local wind profiles, which can easily lead to erroneous wind profile predictions. Summary of the Invention
[0005] Purpose of the invention: The purpose of this invention is to provide a wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model. By constructing an overall heterogeneous map of the range Doppler spectrum and integrating the overall characteristics of the range Doppler, high-precision spectral moment estimation can be achieved, accurately capturing the trend of the wind profile.
[0006] Technical solution: A method for estimating the spectral moment of wind profile radar based on the GAT-Transformer hybrid model, comprising the following steps: S1, preprocesses all power spectra in the range Doppler spectrum generated by the radar echo; S2, construct a heterogeneous map for the preprocessed power spectrum; input the range Doppler spectrum of the relative beam into a GAT-Transformer model respectively to obtain the meridional wind speed profile estimation results and zonal wind speed profile estimation results of a single range Doppler spectrum; then synthesize the meridional wind speed profile estimation results and zonal wind speed profile estimation results into a wind vector and output the final estimated complete wind profile. The GAT-Transformer model comprises a graph attention network and a Transformer model. The graph attention network extracts depth features from heterogeneous graphs, and an intra-layer aggregation module is used to integrate the features of each range gate and merge them into a depth feature matrix. Then, the extracted depth features are height-position encoded to obtain a position encoding matrix. The position encoding matrix is fused with the depth feature matrix and input into the pre-trained Transformer model for spectral moment estimation, outputting the longitude / zonal wind speed profile estimation results of a single range-Doppler spectrum.
[0007] Furthermore, the detailed steps for preprocessing all power spectra are as follows: S11, perform a three-point moving average processing on the entire spectrum; S12, perform noise level estimation and noise filtering; S13, search for ground clutter in the power spectrum and filter out ground clutter using clutter cancellation method.
[0008] Furthermore, in step S2, a heterogeneous graph is constructed. ,in, , Corresponding to nodes and edges in a heterogeneous graph, the construction process includes the following steps: S21, for each distance gate, select the top 7 signal peaks with the highest power values as nodes. If the number of signal peaks is less than 7, select all peaks; or if the current peak value is less than 7... When the selection stops, stop immediately. This represents the noise level. S22, for each signal peak, the features include: peak signal-to-noise ratio, spectral width, average meridional velocity, skewness, kurtosis, peak power, distance gate height, and full-spectrum signal-to-noise ratio; the input matrix is constructed from the features of all nodes. , For feature dimension, This represents the number of nodes in the current power spectrum. It is the set of real numbers; S23, construct two different types of edges: intra-layer edges connecting nodes within the same distance gate, connecting only adjacent points in terms of velocity, are bidirectional edges; inter-layer edges connecting nodes at adjacent distance gates, where nodes are not connected if the difference between them exceeds the wind shear threshold, are unidirectional edges from low to high. The expression for the wind shear threshold is as follows: , in, For wind shear threshold, The distance between adjacent gates is measured in km.
[0009] Furthermore, the process of integrating the features of each distance gate using an intra-layer aggregation module and merging them into a depth feature matrix includes the following steps: S31, Deep features are extracted using a graph attention network. The training process of the graph attention network is as follows: , in, and They represent different nodes. Indicates the first Layer Time Node Hidden features, number of layers When it is 0, For the feature input matrix ; Represents the set of all adjacent nodes of a node. The weight parameter matrix, Attention coefficient; For the Sigmoid function, This is the merging process; S32, perform another inter-layer aggregation process between each distance gate to obtain the final output features of each distance gate: , in, For distance gate Each node in, Here is the weight matrix for each node. The output features of the distance gate; S33, obtain the output matrix of the graph attention network. ,in For the distance to the gate, To output the number of features, It is the set of real numbers.
[0010] Furthermore, the process of obtaining the longitude / zonal wind speed profile estimation results of a single distance-Doppler spectrum includes the following steps: S41, Calculate the frequency factor: , in, For frequency factors, Indexed by feature dimension; The location information is subjected to cross-sine and cosine coding to obtain the coding matrix. : , By fusing with deep features, a feature matrix is obtained. : ; S42, the feature matrix Inputting into a pre-trained Transformer model, the output is a single-frame distance-Doppler spectrum estimation of the longitude / zonal wind speed profile.
[0011] Furthermore, for the GAT-Transformer model, during a single training process, two sets of distance-Doppler spectra at the same moment and in opposite directions are input into a GAT-Transformer model for parallel training. The meridional wind profiles extracted from two opposing beams are combined into a vertical profile. If the wind profile is east-west, the zonal vertical wind profile is obtained. : , in, This is the eastward wind profile. This is the westerly wind profile. The beam pitch angle; If the wind profile runs north-south, the meridional vertical wind profile is obtained: , in, This is the northward wind profile. Southward wind profile; Generate zonal and meridional wind profiles: , , Wherein, U is the zonal vertical wind profile and V is the meridional vertical wind profile; Beamforming is placed in the layer before the output layer, and a machine learning backpropagation process is introduced for training to output the final estimated complete wind profile.
[0012] Compared with the prior art, the significant advantages of this invention are as follows: 1. This invention constructs a holistic heterogeneous map of the range-Doppler spectrum, integrating the overall features of range-Doppler. It utilizes the GAT model to extract deep features from the heterogeneous map to accelerate training, and combines an intra-layer aggregation module to enhance the representational ability of features within each range gate for potential atmospheric nodes. 2. This invention encodes the depth features at height and uses the Transformer model to effectively achieve high-precision wind profile sequence estimation capability; 3. This invention introduces parallel training of relative beam power spectrum and wind vector synthesis formula, providing physical guidance for the originally "black box" training network of machine learning, and enhancing the machine learning method's ability to accurately learn profiles. Experimental results show that the estimation method of this invention can accurately capture the wind profile trend and achieve high-precision spectral moment estimation. Compared with radiosonde data, the correlation coefficients of zonal wind and meridional wind are 0.949 and 0.934, respectively, which can be used for wind profile radar wind field detection. Attached Figure Description
[0013] Figure 1 A flowchart of the wind profiler radar spectral moment estimation method provided in an embodiment of the present invention; Figure 2 The diagram shows the radial wind profile estimation results of the GAT-Transformer model under different beams, where (a) is the east beam, (b) is the west beam, (c) is the north beam, and (d) is the south beam. Figure 3 A comparative diagram showing the performance of wind profile spectral moment estimation using the FL, AME, MPCF, VDA, and GAD-Transformer methods; Figure 4 This diagram illustrates the performance comparison of wind speed estimation for the entire test set using the FL, AME, MPCF, VDA, and GAD-Transformer methods. Detailed Implementation
[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0015] like Figure 1 As shown, a wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model includes the following steps: Step 1: Perform preprocessing on all power spectra in the range Doppler spectrum generated by the radar echo; First, a three-point moving average is performed. The specific operation process is as follows: calculate the average power value of a certain frequency point and the power values of the two adjacent frequency points on the left and right, and replace the power value at that position.
[0016] Next, noise level estimation and noise filtering are performed. The noise level estimation method uses the segmented level method, which divides the full power spectrum into 8 segments, calculates the average power value of each segment, and selects the minimum average power value of the segment without atmospheric signal peaks as the noise level. The noise filtering process is as follows: (1) in, This is the current frequency point position. This is the power spectrum value after noise filtering. This is the original power spectrum value. This represents the noise level.
[0017] Finally, ground clutter is filtered out using the clutter cancellation method. The specific operation process is as follows: compare the power values at the symmetrical frequency points on both sides of the zero frequency, replace the larger value with the absolute value of the power values at the symmetrical frequency points on both sides of the zero frequency, and set the smaller value to zero.
[0018] Step 2: Construct an isomer map from the preprocessed power spectrum; Heterogeneous diagram is ,in, and These correspond to nodes and edges in the heterogeneous graph, respectively. The specific construction process is as follows: Step 21, perform heterogeneous graph node processing For each distance gate, select the top 7 signal peaks with the highest power values as nodes. If the number of signal peaks is less than 7, select all peaks; or if the current peak value is less than 7... When that happens, stop selecting immediately.
[0019] Step 22: Select eight features for each heterogeneous graph node. For each signal peak, the specific features include: peak signal-to-noise ratio (SNR), spectral width, average meridional velocity, skewness, kurtosis, peak power, range gate height, and full-spectrum SNR. The first six features are local features unique to each peak, representing only the characteristics of that peak; the last two features are shared features of the current range gate, representing the spatiotemporal physical information of the current node. Organize the features of all nodes into the input matrix. middle, For feature dimension, This represents the number of nodes in the current power spectrum. It is the set of real numbers.
[0020] Step 23, complete the heterogeneous graph edges The graph is constructed using two different types of edges: intra-layer edges connecting nodes within the same distance gate (since only one node in each distance gate is selected, to reduce graph complexity, only velocity-adjacent points are connected), and inter-layer edges connecting nodes between adjacent distance gates (where no connection is made if the difference between nodes exceeds the wind shear threshold), and unidirectional edges from low to high shear. The expression for the wind shear threshold is as follows: (2) in, For wind shear threshold, The distance between adjacent gates is the gate spacing.
[0021] Step 3 involves extracting depth features from heterogeneous graphs using a Graph Attention Network (GAT), and introducing an intra-layer aggregation module to integrate the features of each distance gate and merge them into a depth feature matrix. The specific steps include: Step 31: Extract deep features using GAT; The GAT processing procedure is as follows: (3) in, and They represent different nodes. Indicates the first Layer Time Node Hidden features, number of layers When it is 0, That is, the feature input matrix ; Represents the set of all adjacent nodes of a node. The weight parameter matrix, This represents the attention coefficient. For the Sigmoid function, This is the merging process.
[0022] The GAT structure consists of two identical sub-modules, each containing two convolutional layers, each with a dimension of 64. The first convolutional layer contains four attention heads, and the second convolutional layer contains one attention head. A normalization layer is applied at the end of each sub-module.
[0023] Step 32: Set the intra-layer aggregation module to the layer before the last output layer of GAT; Based on the hard physical mechanism that only one node is selected in each distance gate, an inter-layer aggregation is performed between each distance gate to finally obtain the output features of each distance gate: (4) in, , Distance gates The individual nodes and their corresponding weight matrices in G; m It is the set of all nodes in the same distance gate; For distance gate The output characteristics.
[0024] Step 33: Obtain the depth features output by GAT; merge the output features of each distance gate to obtain the depth feature matrix output by GAT. ,in For the distance to the gate, This is the number of output features.
[0025] Step 4 involves encoding the height and location of the extracted depth features and training them using a Transformer model; the specific steps include: Step 41: Encode the height location of the extracted depth features; First, calculate the frequency factor: (5) in, For frequency factors, Indexed by feature dimensions.
[0026] The location information is subjected to cross-sine and cosine coding to obtain the location coding matrix. : (6) The position encoding matrix T is fused with the depth feature matrix Y to obtain the feature matrix. : (7) Feature matrix As input to the Transformer encoder. Step 42, convert the feature matrix The input is fed into a pre-trained Transformer model, which outputs a single-frame distance-Doppler spectrum estimate of the meridional wind speed profile. The Transformer model includes four attention heads, and all layers are set to 64 dimensions.
[0027] Step 5 involves parallel training of the range Doppler spectrum of the relative beams, incorporating a wind vector synthesis formula during the training process to output the final estimated complete wind profile. The specific steps include: Step 51: Perform parallel training on the relative dual-beam data; during one training process, train a GAT-Transformer model for each of the two sets of distance Doppler spectra at the same time and in opposite directions (east and west, north and south) and perform parallel training.
[0028] Step 52: Combine the meridional wind profiles extracted from the two relative beams into a vertical profile. If the wind profiles are east-west, obtain the zonal vertical wind profile. : (8) in, This is the eastward wind profile. This is the westerly wind profile. The beam pitch angle.
[0029] If the wind profile runs north-south, the meridional vertical wind profile is obtained: (9) in, This is the northward wind profile. This is the southerly wind profile.
[0030] Based on the vertical beam and beams in each direction, zonal and meridional wind profiles are generated. The expression for the beam synthesis formula is as follows: (10) (11) Where U is the zonal vertical wind profile and V is the meridional vertical wind profile.
[0031] Step 53: Beamforming is set in the layer preceding the output layer, and a machine learning backpropagation process is introduced for training. This beamforming enhances the model's physical guidance and profile learning ability, ultimately generating high-precision zonal and meridional wind profiles.
[0032] The dataset used in this invention was obtained by the Atmospheric Radiation Measurement (ARM) facility in Barrow, Alaska, USA. The detection equipment was the NSAC1 central facility located on the north slope of Alaska. The data consists of range-Doppler spectra in the east, south, west, and north directions. Each spectrum contains 40 range gates with a range resolution of 62.55 m, and each range gate has a 64-length Doppler power spectrum. The velocity resolution is 0.28 m / s. The label data consists of radiosonde data collected from the same site. Data was collected from January 1, 2008 to January 31, 2010. After quality control screening, 1964 power spectra, including 78,560 signal peak nodes, were selected for training. During training, the training-to-test ratio was 8:2, and the learning rate was 0.001.
[0033] Figure 2The figures show the radial wind profile estimation results of the GAT-Transformer model under four beam patterns. The data was collected at 5:33 AM local time on February 15, 2008. As can be seen from the figures, above the 15-range gate, the power spectrum at each range gate contains significant clutter and noise. Furthermore, between range gates 13 and 15, anomalous propagation may occur due to atmospheric waveguide trapping, causing the signal to radiate into the surrounding ground environment. The power spectrum then only exhibits strong zero-frequency clutter, while the actual atmospheric component is masked. The proposed GAT-Transformer model-based estimation profile, fitted with the training results, yields more accurate wind speed estimates, demonstrating better profile prediction capabilities.
[0034] Figure 3 This paper compares the performance of five methods for estimating wind profiler spectral moments. The left side shows the signal-to-noise ratio at different altitudes, while the middle and right sides show the comparison results for zonal (U) and meridional (V) winds, respectively. It can be seen that the wind profiler spectral moment estimation method proposed in this invention exhibits good consistency with radiosonde (RS) data at different altitudes and signal-to-noise ratios, outperforming the other four existing methods.
[0035] Figure 4 To compare the performance of five methods for wind speed estimation across the entire test set, from... Figure 4 As can be seen, the algorithm proposed in this invention achieves excellent estimation accuracy. The correlation coefficients of zonal wind and meridional wind with the radiosonde data are 0.949 and 0.934, respectively, which are much higher than the other four algorithms.
[0036] A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the wind profiler radar spectral moment estimation method of the present invention. The processor contains a kernel that retrieves the corresponding program unit from memory.
[0037] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of the wind profiler radar spectral moment estimation method of the present invention.
[0038] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.
[0039] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the wind profiler radar spectral moment estimation method of the present invention.
[0040] In summary, this invention solves the problems of poor generalization and insufficient estimation accuracy of wind profiler radar spectral moment estimation methods in complex atmospheric environments, effectively improving the robustness and estimation accuracy of wind profiler spectral moment estimation methods. There are many methods and approaches to implement this technical solution; the above description is only a preferred embodiment of this invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for estimating the spectral moment of wind profile radar based on a GAT-Transformer hybrid model, characterized in that, Includes the following steps: S1, preprocesses all power spectra in the range Doppler spectrum generated by the radar echo; S2, construct a heterogeneous map for the preprocessed power spectrum; input the range Doppler spectrum of the relative beam into a GAT-Transformer model respectively to obtain the meridional wind speed profile estimation results and zonal wind speed profile estimation results of a single range Doppler spectrum; then synthesize the meridional wind speed profile estimation results and zonal wind speed profile estimation results into a wind vector and output the final estimated complete wind profile. The GAT-Transformer model comprises a graph attention network and a Transformer model. The graph attention network extracts depth features from heterogeneous graphs, and an intra-layer aggregation module is used to integrate the features of each range gate and merge them into a depth feature matrix. Then, the extracted depth features are height-position encoded to obtain a position encoding matrix. The position encoding matrix is fused with the depth feature matrix and input into the pre-trained Transformer model for spectral moment estimation, outputting the longitude / zonal wind speed profile estimation results of a single range-Doppler spectrum.
2. The wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model according to claim 1, characterized in that, The detailed steps for preprocessing all power spectra are as follows: S11, perform a three-point moving average processing on the entire spectrum; S12, perform noise level estimation and noise filtering; S13, search for ground clutter in the power spectrum and filter out ground clutter using clutter cancellation method.
3. The wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model according to claim 1, characterized in that, In step S2, a heterogeneous graph is constructed. ,in, , Corresponding to nodes and edges in a heterogeneous graph, the construction process includes the following steps: S21, for each distance gate, select the top 7 signal peaks with the highest power values as nodes. If the number of signal peaks is less than 7, select all peaks; or if the current peak value is less than 7... When the selection stops, stop immediately. This represents the noise level. S22, for each signal peak, the features include: peak signal-to-noise ratio, spectral width, average meridional velocity, skewness, kurtosis, peak power, distance gate height, and full-spectrum signal-to-noise ratio; the input matrix is constructed from the features of all nodes. , For feature dimension, This represents the number of nodes in the current power spectrum. It is the set of real numbers; S23, construct two different types of edges: intra-layer edges connecting nodes within the same distance gate, connecting only adjacent points in terms of velocity, are bidirectional edges; inter-layer edges connecting nodes at adjacent distance gates, where nodes are not connected if the difference between them exceeds the wind shear threshold, are unidirectional edges from low to high. The expression for the wind shear threshold is as follows: in, For wind shear threshold, The distance between adjacent gates is measured in km.
4. The wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model according to claim 1, characterized in that, The process of integrating the features of each distance gate using an intra-layer aggregation module and merging them into a depth feature matrix includes the following steps: S31, Deep features are extracted using a graph attention network. The training process of the graph attention network is as follows: in, and They represent different nodes. Indicates the first Layer Time Node Hidden features, number of layers When it is 0, For the feature input matrix ; Represents the set of all adjacent nodes of a node. The weight parameter matrix, Attention coefficient; For the Sigmoid function, This is the merging process; S32, perform another inter-layer aggregation process between each distance gate to obtain the final output features of each distance gate: in, For distance gate Each node in, Here is the weight matrix for each node. The output features of the distance gate; S33, obtain the output matrix of the graph attention network. ,in For the distance to the gate, To output the number of features, It is the set of real numbers.
5. The wind profile radar spectral moment estimation method based on the GAT-Transformer hybrid model according to claim 1, characterized in that, The process of obtaining the longitude / zonal wind speed profile estimation results of a single distance-Doppler spectrum includes the following steps: S41, Calculate the frequency factor: in, For frequency factors, Indexed by feature dimension; The location information is subjected to cross-sine and cosine coding to obtain the coding matrix. : By fusing with deep features, a feature matrix is obtained. : ； S42, the feature matrix Inputting into a pre-trained Transformer model, the output is a single-frame distance-Doppler spectrum estimation of the longitude / zonal wind speed profile.
6. The wind profiler radar spectral moment estimation method based on the GAT-Transformer hybrid model according to claim 1, characterized in that, For the GAT-Transformer model, during a single training process, two sets of distance-Doppler spectra at the same moment and in opposite directions are input into a GAT-Transformer model for parallel training. The meridional wind profiles extracted from two opposing beams are combined into a vertical profile. If the wind profile is east-west, the zonal vertical wind profile is obtained. : in, This is the eastward wind profile. This is the westerly wind profile. The beam pitch angle; If the wind profile runs north-south, the meridional vertical wind profile is obtained: in, This is the northward wind profile. Southward wind profile; Generate zonal and meridional wind profiles: Wherein, U is the zonal vertical wind profile and V is the meridional vertical wind profile; Beamforming is placed in the layer before the output layer, and a machine learning backpropagation process is introduced for training to output the final estimated complete wind profile.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 6.
9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.