DOA estimation method based on 3D spherical conformal array and physical augmentation network
By designing a DOA estimation method using a three-dimensional spherical conformal array and a physical augmentation network, the problems of non-uniform array structure and insufficient compensation for mutual coupling effects in the swarm intelligence perception of micro UAVs are solved. This method achieves uniform perception and high-precision DOA estimation across the entire space, supports full polarization pattern recognition, and adapts to the dynamic flight attitude of UAVs.
Patent Information
- Application Number
- CN202611585035.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies for intelligent perception of micro UAVs suffer from problems such as uneven array structure design, limited compensation for mutual coupling effects, low DOA estimation accuracy, and insufficient polarization recognition capabilities, making them unable to adapt to the dynamic flight attitude and complex electromagnetic environment of UAVs.
A DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network is adopted. A 16-element small oblique truncated half-cube spherical array is designed. The physical manifold is reconstructed by decoupling the covariance matrix features and transforming the conjugate congruence matrix. Polarization state identification and DOA parameter decoupling estimation are achieved by combining a polarization router and a Softmax classifier.
It achieves 360° full-space uniform perception without blind spots, improves the mutual coupling compensation effect and DOA estimation accuracy, supports full polarization pattern recognition, adapts to complex electromagnetic environments, and improves estimation accuracy and robustness.
Smart Images

Figure CN123066682A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of antenna engineering and artificial intelligence technology, specifically to a DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network. Background Technology
[0002] With the rapid development of UAV technology, low-altitude airspace mobile swarm sensing networks have become a research hotspot in the fields of 6G communication and intelligent detection. Direction of arrival (DOA) estimation and polarization state identification are the core tasks of UAV electromagnetic sensing, directly determining the accuracy of radiation source localization and signal decoupling. To adapt to micro UAV platforms, the sensing system must simultaneously meet three core requirements: omnidirectional spatial coverage, miniaturized and compact layout, and robust estimation in complex environments. Existing technologies have formed a related research system around array structure design, mutual coupling effect compensation, joint DOA and polarization estimation algorithms, and machine learning.
[0003] The research results in related technical directions mainly fall into two categories. One is a compact array DOA estimation scheme based on an improved subspace algorithm that eliminates mutual coupling matrices through a subarray selection strategy. This scheme uses a fixed uniform linear array (ULA) for DOA estimation, achieving only one-dimensional DOA estimation and failing to achieve spherical omnidirectional perception. It also exhibits poor adaptability to dynamic attitudes such as pitch and roll during UAV flight, and its perception performance in non-orthogonal directions is severely degraded. The other approach reduces the impact of mutual coupling through pure data-driven methods.
[0004] While existing technologies have achieved certain research results in areas such as electromagnetic sensing of UAVs, DOA and polarization estimation, mutual coupling compensation, and spherical array design, none of them have performed end-to-end optimization for the specific application scenario of intelligent sensing of micro UAVs. All kinds of solutions have insurmountable technical defects, and these defects are compounded, as follows:
[0005] 1. Array structure design defects: Irregular spherical arrays have problems such as uneven spatial sampling and low isolation between elements, resulting in severe degradation of sensing performance in the spherical pole region; although planar and linear array structures are compact and easy to integrate, they can only achieve local spatial sensing, and the performance in non-orthogonal directions degrades, making them unsuitable for the dynamic flight attitude of UAVs in pitch and roll; although regular arrays based on three-dimensional spherical conformal arrays can achieve omnidirectional uniform sensing, the number of elements is large, resulting in a large antenna array volume.
[0006] 2. Limited effectiveness of mutual coupling effect compensation: Pure data-driven deep learning schemes ignore the physical laws of electromagnetic wave propagation, and the model relies only on sample fitting. The estimation accuracy decreases in low signal-to-noise ratio environments, and the anti-interference ability of mutual coupling, noise and other interferences is weak. Traditional subarray selection and matrix inversion methods require accurate solution of the mutual coupling matrix. In engineering, array errors will lead to a decrease in compensation effect and poor practicality. Summary of the Invention
[0007] The purpose of this invention is to propose a DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network to solve the problems mentioned in the background art.
[0008] To achieve the aforementioned objectives, the first technical solution adopted by this invention is: a DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network, comprising the following steps:
[0009] S1: Design a 16-element small rhombohedral half-cube spherical array based on spherical t-design, wherein the spherical array consists of 16 orthogonal bipolar dipole units;
[0010] S2: Based on the spherical array receiving electromagnetic radiation source signal, the electromagnetic radiation source signal is preprocessed, and the preprocessed signal features are decoupled into physical manifold features and statistical features. The covariance matrix features and statistical features are concatenated to form a 596-dimensional heterogeneous feature vector. ;in The unfolded 16×16 covariance matrix features carry array manifold information; These are complex mappings, amplitude, and phase statistical descriptors for signals, representing the physical properties of electromagnetic signals as statistical characteristics;
[0011] S3: Input the covariance matrix feature x into the matrix coupling layer, and restore the covariance matrix feature to a 16×16 two-dimensional complex covariance matrix according to the array physical topology. Perform conjugate congruential matrix transformation to realize physical manifold reconstruction and obtain the calibrated covariance matrix.
[0012] S4: The covariance matrix calibrated through the matrix coupling layer is re-expanded into one-dimensional features, along with statistical features. After fusion, the input is an adaptive affine calibration layer, on which calibration parameters are dynamically generated. Based on the calibration parameters, calibration with statistical distribution alignment is achieved, thereby reducing the impact of signal noise on the basis of S3 physical manifold reconstruction.
[0013] S5: Input the feature signal processed by S4 into the polarization router. The polarization distribution features of the features are perceived by the Softmax classifier. The feature signals of different linear polarization states of 0°, 45°, 90° and 135° are directed to the corresponding polarization expert subnets. Each polarization expert subnet is an independent neural network model built for a specific polarization state. It contains a shared feature encoding layer and a two-branch regression structure for pitch angle and azimuth angle estimation, respectively, thereby realizing the decoupled estimation of DOA parameters under polarization state conditions.
[0014] S6: Introducing comprehensive spatial angle error As evaluation metrics, the output includes estimated values of DOA pitch angle, azimuth angle, and polarization state identification results.
[0015] Furthermore, the orthogonal dual-polarized dipole unit uses a Rogers Ro4003 dielectric substrate as its base. Two pairs of sector dipoles are placed at 45° angles on the front side of the dielectric substrate. A metal cavity is set around the dielectric substrate, and a parasitic sector structure is provided on the back side. Metal pillars are provided at symmetrical positions of the coaxial cable. The 16 units are based on a small rhombic semi-cube layout structure, and 16 square face centers are selected as phase centers to achieve 360° uniform coverage of the entire space.
[0016] Furthermore, in step S1, the small rhombic truncated half-cube point structure contains 24 vertices, which are evenly distributed on the unit sphere. It contains 18 congruent squares and 8 congruent equilateral triangles. The phase center of the 16 array units is the face center of the remaining square after removing the face centers of the two orthogonal squares at the top and bottom.
[0017] Furthermore, the formula for the calibrated covariance matrix in step S3 is:
[0018] ;
[0019] in For learnable complex mutually coupled matrices, To observe the covariance matrix, This is the calibrated covariance matrix.
[0020] Furthermore, the original signal (electromagnetic radiation source signal) is preprocessed, including feature extraction and covariance matrix calculation, to obtain statistical features and covariance features.
[0021] Furthermore, the polarization states mentioned in step S5 include four full polarization modes: 0°, 45°, 90°, and 135°. The polarization router uses a Softmax classifier to accurately distribute signals of different polarization modes.
[0022] Furthermore, the statistical features include the real and imaginary parts of the complex mapping of the signal, the mean and variance of the signal amplitude, and the distribution characteristics of the signal phase.
[0023] Furthermore, in step S6, the spatial angle error is integrated. The calculation formula is as follows:
[0024] ;
[0025] In the formula, , For the actual pitch and azimuth angles of the signal, , The estimated signal elevation and azimuth angles are given.
[0026] Due to the application of the above-mentioned technical solution, the present invention has the following advantages compared with the prior art:
[0027] 1. Achieve 360° full-space uniform perception without blind spots: The 16-unit small oblique half-cube spherical array designed based on the spherical t-design theory breaks through the spatial coverage limitations of planar and linear arrays, solves the problem of uneven sampling of irregular spherical arrays, and achieves 360° full-space uniform perception without blind spots.
[0028] 2. Significantly improves mutual coupling compensation effect and estimation accuracy: By introducing conjugate congruential matrix transformation through the matrix coupling layer, the physical mechanism of electromagnetic field is explicitly integrated into the deep learning model, restoring the geometric orthogonality of the array manifold at the physical level. Compared with existing scalar uncalibrated schemes, this significantly improves the comprehensive spatial angle root mean square error. The accuracy was improved by 25.33% from 1.88° to 1.50°; compared to the adaptive statistical alignment scheme, The accuracy improved by 11.33% from 1.67° to 1.50°; compared to a purely data-driven bias scaling learning scheme, The accuracy was improved by 14% from 1.71° to 1.50°.
[0029] 3. Achieve 100% accurate recognition of all polarization modes: Through polarization routing and multi-task decoupling design, it supports the recognition of 0°, 45°, 90° and 135° all polarization modes, and the polarization classification accuracy reaches 100%. Compared with the existing solutions that only support dual polarization modes, the polarization recognition capability and accuracy are significantly improved, and it is suitable for multi-polarization signal perception in complex electromagnetic environments. Attached Figure Description
[0030] Figure 1 This shows a structural diagram of an antenna element provided in an embodiment of the present invention;
[0031] Figure 2 The diagram shows an antenna array structure provided in an embodiment of the present invention;
[0032] Figure 3 This diagram illustrates the simulation results of the array internal cells provided in an embodiment of the present invention.
[0033] Figure 4 This diagram illustrates the DOA estimation and spatial angle error of a 16-cell array provided in an embodiment of the present invention.
[0034] Figure 5 The diagram shows the structure of the expert network based on physical manifold matrix coupling provided in the embodiment of the present invention.
[0035] Figure 6 The cumulative distribution of DOA estimation error provided in the embodiment of the present invention is shown. Detailed Implementation
[0036] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.
[0037] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or system that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or systems.
[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0039] like Figures 1-6 As shown, this embodiment of the invention provides a DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network, including the following steps:
[0040] S1: Design a 16-element small rhombic half-cube spherical array based on spherical t-design. The spherical array consists of 16 orthogonal bipolar dipole units.
[0041] S2: Based on the spherical array receiving electromagnetic radiation source signal, the electromagnetic radiation source signal is preprocessed, and the preprocessed signal features are decoupled into physical manifold features and statistical features. The covariance matrix features and statistical features are concatenated to form a 596-dimensional heterogeneous feature vector. ;in The unfolded 16×16 covariance matrix features carry array manifold information; These are complex mappings, amplitude, and phase statistical descriptors for signals, representing the physical properties of electromagnetic signals as statistical characteristics;
[0042] S3: Input the covariance matrix feature x into the matrix coupling layer, and restore the covariance matrix feature to a 16×16 two-dimensional complex covariance matrix according to the array physical topology. Perform conjugate congruential matrix transformation to realize physical manifold reconstruction and obtain the calibrated covariance matrix.
[0043] S4: The covariance matrix calibrated through the matrix coupling layer is re-expanded into one-dimensional features, along with statistical features. After fusion, the input is an adaptive affine calibration layer, on which calibration parameters are dynamically generated. Based on the calibration parameters, calibration with statistical distribution alignment is achieved, thereby reducing the impact of signal noise on the basis of S3 physical manifold reconstruction.
[0044] S5: Input the feature signal processed by S4 into the polarization router. The polarization distribution features of the features are perceived by the Softmax classifier. The feature signals of different linear polarization states of 0°, 45°, 90° and 135° are directed to the corresponding polarization expert subnets. Each polarization expert subnet is an independent neural network model built for a specific polarization state. It contains a shared feature encoding layer and a two-branch regression structure for pitch angle and azimuth angle estimation, respectively, thereby realizing the decoupled estimation of DOA parameters under polarization state conditions.
[0045] Specifically, dimensional decoupling angle estimation: Configure an independent pitch angle for each polarization expert subnet ( ) and azimuth ( Experts return to the head; among them The expert network focuses on learning the variation pattern of the array's radial phase difference. The expert network focuses on learning the azimuth-oriented ring symmetry and combines it with a physical compensation layer to solve the phase jump problem at the poles of the spherical coordinate system; through parameter decoupling design, it eliminates... and Reduce parameter interference between estimates and improve the accuracy of angle estimation.
[0046] Polarization state identification: Through the classification output of the polarization expert subnet, the polarization state of the signal is accurately identified, and the polarization classification result is output.
[0047] S6: Introducing comprehensive spatial angle error As evaluation metrics, the output includes estimated values of DOA pitch angle, azimuth angle, and polarization state identification results.
[0048] It should be noted that DOA stands for Direction of Arrival, which refers to the spatial angle from which an electromagnetic signal propagates from the radiation source to the sensing array, including the elevation angle. (Angle with the horizontal plane) and azimuth (The angle between the direction and due north).
[0049] According to an embodiment of the present invention, the orthogonal dual-polarized dipole unit uses a Rogers Ro4003 dielectric substrate as the base. Two pairs of sector dipoles are placed at 45° angles on the front side of the dielectric substrate. A metal cavity is provided around the dielectric substrate, and a parasitic sector structure is provided on the back side. Metal pillars are provided at symmetrical positions of the coaxial cable. The 16 units are based on a small rhombic half-cube layout structure, and 16 square face centers are selected as phase centers to achieve 360° uniform coverage of the entire space.
[0050] According to an embodiment of the present invention, the small rhombic truncated half-cube point arrangement structure in step S1 contains 24 vertices, which are evenly distributed on the unit sphere, and includes 18 congruent squares and 8 congruent equilateral triangles. The phase center of the 16 array units is the face center of the remaining square after removing the face centers of the two orthogonal squares at the top and bottom.
[0051] It should be noted that the array meets the requirements of uniform cell spacing and engineering symmetry, which facilitates the design of power dividers and adapts to the miniaturization needs of micro UAVs.
[0052] Array performance indicators: Through structural optimization, the reflection coefficient of the array dual-polarization port is ≤-15dB, the polarization isolation of a single unit is ≥25dB, and the common polarization isolation of adjacent units is ≥21dB, achieving 360° uniform coverage across the entire space.
[0053] According to an embodiment of the present invention, the formula for the calibrated covariance matrix in step S3 is:
[0054] ;
[0055] in For learnable complex mutually coupled matrices, To observe the covariance matrix, This is the calibrated covariance matrix.
[0056] According to an embodiment of the present invention, the original signal is preprocessed, including feature extraction and covariance matrix calculation, to obtain statistical features and covariance features.
[0057] According to an embodiment of the present invention, the polarization state in step S5 includes four full polarization modes: 0°, 45°, 90°, and 135°. The polarization router performs accurate diversion of signals of different polarization modes through a Softmax classifier.
[0058] According to embodiments of the present invention, the statistical features include the real and imaginary parts of the complex mapping of the signal, the mean and variance of the signal amplitude, and the distribution characteristics of the signal phase.
[0059] According to an embodiment of the present invention, in step S6, the spatial angle error is integrated. The calculation formula is as follows:
[0060] ;
[0061] In the formula, , For the actual pitch and azimuth angles of the signal, , The estimated signal elevation and azimuth angles are given.
[0062] In one specific embodiment of the present invention: the fabrication of orthogonal bipolar dipole units includes:
[0063] 1. A Rogers Ro4003 dielectric substrate 1 with a thickness of 0.813 mm and a relative permittivity of 3.55 is selected as the substrate. Two pairs of sector dipoles 2 are etched at 45° on the front side of the substrate. The dipole arm length is 14 mm and the sector angle is 60°.
[0064] 2. Two 50-ohm coaxial cables 3 are selected. The inner conductor is welded to the same side wall of the sector dipole 2 to achieve excitation, and the outer conductor is welded to the annular pad with a diameter of 2mm on the back of the substrate to ensure the stability of signal transmission.
[0065] 3. A 1.5mm thick metal ground plane is set below the dielectric substrate 1, and a 16mm high metal cavity 4 is added around the ground plane to serve as a reflector for the dipole and improve the radiation gain.
[0066] 4. A pair of parasitic fan-shaped structures 5 are etched horizontally on the back side of the dielectric substrate 1, with an arm length of 9.5 mm and a fan angle of 50°, to broaden the antenna impedance bandwidth.
[0067] 5. A matrix groove 6 with a size of 14mm×16mm is symmetrically designed on the four side walls of the metal cavity 4 to reduce the antenna resonant frequency and optimize impedance matching;
[0068] 6. At a position symmetrical to the coaxial cable 3, two metal pillars 7 with a diameter of 2mm (consistent with the diameter of the outer conductor of the coaxial cable) are vertically welded. The top of the metal pillars passes through the dielectric substrate 1 and is welded to the sector dipole 2, and the bottom is connected to the metal cavity 4 to improve the unit isolation.
[0069] 7. After completing the above steps, an orthogonal dual-polarized dipole unit is obtained. The unit is tested and the dual-polarized port reflection coefficient is ≤-23 dB and the polarization isolation is ≥40 dB, which meets the design requirements.
[0070] Furthermore, the assembly of a 16-element spherically symmetrical array:
[0071] 1. Based on the spherical t-design theory (spherical t-design: a theory of uniform point distribution on a spherical surface, which ensures that polynomials of degree not exceeding t strictly satisfy the quadrature rule to achieve uniform point distribution on a unit sphere), a unit sphere with a radius of 48mm is constructed, and a small rhombic truncated half cube is used as the point distribution structure to determine its 18 square face centers as candidate phase center points;
[0072] 2. Remove the two orthogonal square face centers at the top and bottom, and select the remaining 16 face centers as the phase centers of the final array units to ensure uniform unit spacing;
[0073] 3. The 16 orthogonal bipolar dipole units were fixed at the 16 phase centers respectively, and the power divider and the RF front end were connected according to the engineering symmetry requirements to complete the assembly of the 16-unit small rhombic half-cube spherical array.
[0074] 4. Simulation tests were conducted on the assembled array, and the results showed that the reflection coefficient of the dual-polarized ports was ≤-15dB, the polarization isolation of a single unit was ≥25dB, and the common polarization isolation of adjacent units was ≥21dB, achieving 360° uniform coverage in the entire space, and it can be directly integrated into a micro UAV platform.
[0075] Specifically, Figure 3 Test results between antenna elements and their neighboring elements are presented. Except for the test port, all other ports are connected to a 50-ohm matched load. The simulation and test results of the reflection coefficient and isolation of the array elements are shown in the figure. The reflection coefficient of the dual-polarized ports is less than -15dB, and the polarization isolation of a single element exceeds 25dB, while the common polarization isolation between adjacent elements exceeds 21dB.
[0076] Furthermore, the training and testing of the expert system based on physical manifold reconstruction are as follows:
[0077] The experimental dataset construction includes:
[0078] Generates operating frequency of 2.4 GHz, polarization states of 0°, 45°, 90°, and 135°, and DOA angle coverage. : 0°~180° Electromagnetic signals ranging from 0° to 360° are received by an assembled 16-unit spherical array. After preprocessing the original signals, 596-dimensional heterogeneous feature vectors are extracted, and training, testing, and validation sets are constructed according to a ratio of 0.7:0.15:0.15. The training set contains 8,000 samples, while the testing and validation sets each contain 1,700 samples, covering the entire spatial angle and all polarization modes.
[0079] The construction and training of expert systems includes:
[0080] 1. Build an expert system based on physical manifold reconstruction. The core modules include a 16×16 complex matrix coupling layer, a 596-dimensional adaptive affine layer, a 4-class polarization router, 4 polarization expert subnets, and 8 angle expert regression heads.
[0081] 2. Set the complex cross-coupling matrix C of the matrix coupling layer as a learnable parameter, and initialize it with a random orthogonal matrix; set the weights and biases of the adaptive affine layer as learnable parameters; and use a Softmax classifier for the polarization router. and The expert networks are all 3-layer fully connected networks, and the activation function used is ReLU;
[0082] Figure 5 The paper presents a physical manifold matrix-coupled expert network architecture. First, the covariance features are matrix-coupled and then concatenated with statistical features using an adaptive affine transformation. Next, signal polarization is predicted, and data is routed to the corresponding expert model through polarization classification. Finally, the network achieves the desired result. and The prediction.
[0083] 3. Based on comprehensive spatial angular error Root mean square error ( The loss function is denoted as , and the Adam optimizer is used. The learning rate is set to 1e-4, the number of training epochs is 200, and the batch size is 64. The system is trained on the training set to obtain a converged training model.
[0084] The final polarization and DOA estimation results are shown in Table 1.
[0085] Table 1:
[0086] Model Name Polarization classification accuracy <![CDATA[RMSE θ ]]> <![CDATA[RMSE ρ ]]> Accuracy (ρ < 3.2∘) Physical manifold reconstruction 100% 0.92° 1.50° 96.89% Adaptive statistical alignment 100% 0.97° 1.67° 95.20% Scalar calibration 100% 1.01° 1.71° 94.35% Pure MLP model 100% 1.12° 1.88° 91.81%
[0087] In summary, by combining the high-symmetry spherical array design at the hardware level with the physical enhancement deep learning algorithm at the software level, 360° full-space uniform perception is achieved, effectively compensating for the mutual coupling effect of compact arrays, improving the accuracy and robustness of joint estimation of DOA and polarization state in complex electromagnetic environments, while taking into account array miniaturization, algorithm real-time performance and engineering feasibility, and adapting to the application requirements of UAV swarm intelligence perception.
[0088] This invention significantly improves the mutual coupling compensation effect and estimation accuracy:
[0089] By introducing a conjugate congruential matrix transformation through a matrix coupling layer, the physical mechanism of electromagnetic fields is explicitly integrated into the deep learning model, restoring the geometric orthogonality of the array manifold at the physical level. Compared with existing scalar uncalibrated schemes, this approach achieves a higher overall spatial root mean square error. The accuracy was improved by 25.33% from 1.88° to 1.50°; compared to the adaptive statistical alignment scheme, The accuracy improved by 11.33% from 1.67° to 1.50°; compared to a purely data-driven bias scaling learning scheme, The accuracy was improved by 14% from 1.71° to 1.50°.
[0090] This invention achieves 100% accurate recognition of all polarization modes: through polarization routing and multi-task decoupling design, it supports the recognition of 0°, 45°, 90° and 135° all polarization modes, and the polarization classification accuracy reaches 100%. Compared with the existing schemes that only support dual polarization modes, the polarization recognition capability and accuracy are qualitatively improved, and it is suitable for multi-polarization signal perception in complex electromagnetic environments.
Claims
1. A DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network, characterized in that, Includes the following steps: S1: Design a 16-element small rhombohedral half-cube spherical array based on spherical t-design, wherein the spherical array consists of 16 orthogonal bipolar dipole units; S2: Based on the spherical array receiving electromagnetic radiation source signal, the electromagnetic radiation source signal is preprocessed, and the preprocessed signal features are decoupled into physical manifold features and statistical features. The covariance matrix features and statistical features are concatenated to form a 596-dimensional heterogeneous feature vector. ;in The unfolded 16×16 covariance matrix features carry array manifold information; These are complex mappings, amplitude, and phase statistical descriptors for signals, representing the physical properties of electromagnetic signals as statistical characteristics; S3: Input the covariance matrix feature x into the matrix coupling layer, and restore the covariance matrix feature to a 16×16 two-dimensional complex covariance matrix according to the array physical topology. Perform conjugate congruential matrix transformation to realize physical manifold reconstruction and obtain the calibrated covariance matrix. S4: The covariance matrix calibrated through the matrix coupling layer is re-expanded into one-dimensional features, along with statistical features. After fusion, the input is placed into an adaptive affine calibration layer, where calibration parameters are dynamically generated. Based on these calibration parameters, calibration is achieved by aligning the statistical distribution. S5: Input the feature signal processed by S4 into the polarization router. The polarization distribution features of the features are perceived by the Softmax classifier. The feature signals of different linear polarization states of 0°, 45°, 90° and 135° are directed to the corresponding polarization expert subnets. Each polarization expert subnet is an independent neural network model built for a specific polarization state. It contains a shared feature encoding layer and a two-branch regression structure for pitch angle and azimuth angle estimation, respectively, thereby realizing the decoupled estimation of DOA parameters under polarization state conditions. S6: Introducing comprehensive spatial angle error As evaluation metrics, the output includes estimated values of DOA pitch angle, azimuth angle, and polarization state identification results.
2. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 1, characterized in that, The orthogonal dual-polarized dipole unit uses a Rogers Ro4003 dielectric substrate as its base. Two pairs of sector dipoles are placed at 45° angles on the front side of the dielectric substrate. A metal cavity is set around the dielectric substrate, and a parasitic sector structure is provided on the back side. Metal pillars are provided at symmetrical positions of the coaxial cable. The 16-unit small rhombic semi-cube spherical array is a small rhombic semi-cube point arrangement structure, selecting 16 square face centers as phase centers to achieve 360° full-space uniform coverage.
3. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 2, characterized in that, In step S1, the small rhomboid half-cube point layout structure contains 24 vertices, which are evenly distributed on the unit sphere. It contains 18 congruent squares and 8 congruent equilateral triangles. The phase center of the 16 array units is the face center of the remaining square after removing the face centers of the two orthogonal squares at the top and bottom.
4. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 3, characterized in that, The formula for the calibrated covariance matrix in step S3 is: ; in For learnable complex mutually coupled matrices, Represented as a matrix Conjugate congruent matrix To observe the covariance matrix, This is the calibrated covariance matrix.
5. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 1, characterized in that, The polarization states mentioned in step S5 include four full polarization modes: 0°, 45°, 90°, and 135°. The polarization router uses a Softmax classifier to accurately distribute signals of different polarization modes.
6. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 1, characterized in that, The statistical features include the real and imaginary parts of the complex mapping of the signal, the mean and variance of the signal amplitude, and the distribution characteristics of the signal phase.
7. The DOA estimation method based on a three-dimensional spherical conformal array and a physical augmentation network as described in claim 1, characterized in that, In step S6, the spatial angle error is integrated. The calculation formula is as follows: ; In the formula, , For the actual pitch and azimuth angles of the signal, , The estimated signal elevation and azimuth angles are given.
