Sparse array DOA (direction of arrival) estimation method based on adaptive decision level fusion

Abstract

The invention discloses a sparse array DOA (direction of arrival) estimation method based on adaptive decision level fusion. The method comprises the following steps: establishing a steering vector model for describing receiving signals and channel transmission characteristics of a sparse array; executing exploratory rotation measurement, and constructing an initial historical candidate angle set through incoherent decision-level fusion; entering a self-adaptive verification decision stage, performing clustering analysis on the initial historical candidate angle set to determine an optimal candidate angle, and calculating the global consistency cost to judge whether the angle estimation algorithm is converged or not; and repeating the steps, and carrying out iterative updating and outputting. Direction of arrival estimation is reconstructed into a sequential decision process, and a self-adaptive feedback mechanism is introduced to intelligently plan a measurement process, so that the defect of low efficiency of an existing fixed rotation strategy is overcome. According to the invention, the measurement overhead can be automatically adjusted according to the complexity of a scene, and redundant measurement in a fixed strategy is avoided.

Description

Technical Field

[0001] This application belongs to the field of array signal processing technology, specifically relating to a sparse array direction-of-arrival estimation method based on adaptive decision-level fusion. Background Technology

[0002] High-precision direction-of-arrival (AOA) estimation in array signal processing is a fundamental problem in cutting-edge applications such as radar detection and autonomous driving. To reduce hardware costs and achieve narrower beamwidths to improve angular resolution, sparse arrays with wide element spacing are often used in engineering applications. However, this design introduces spatial aliasing, leading to grating lobe ambiguity. This inherent ambiguity prevents traditional high-resolution algorithms such as Multiple Signal Classification (MUSIC) and Estimation of Signal Parameters by Rotation Invariant Techniques (ESPRIT) from accurately estimating the AOA. Therefore, effectively suppressing grating lobe ambiguity to achieve high-precision angle estimation remains a key engineering challenge in this field.

[0003] In related technologies, schemes for achieving direction of arrival (DOA) estimation through physical rotating arrays have emerged. The core principle of this type of technology is to eliminate grating lobe ambiguity by utilizing the consistency of the spectral peaks corresponding to the true angles at different viewing angles. However, existing physical rotating array technologies still have significant drawbacks in engineering applications, making it difficult to meet the needs of real-world scenarios. Some schemes employ fixed, non-adaptive rotation strategies, relying on inefficient qualitative comparisons, which can easily generate a large number of redundant measurements in high signal-to-noise ratio scenarios, resulting in a waste of measurement resources. Other schemes are only suitable for ideal scenarios with a single signal source and no multipath interference, failing to consider multipath propagation in real-world applications, thus limiting their applicability. Summary of the Invention

[0004] This application provides a sparse array direction-of-arrival estimation method based on adaptive decision-level fusion to solve the technical problems of measurement redundancy and low resource utilization efficiency caused by the use of fixed and non-feedback measurement strategies in existing physical rotation sparse array methods.

[0005] To address the aforementioned technical problems, this application adopts the following technical solution: a sparse array direction-of-arrival estimation method based on adaptive decision-level fusion, comprising:

[0006] S1. Establish a steering vector model to describe the received signal and channel transmission characteristics of the sparse array;

[0007] S2. Based on the guide vector model, perform exploratory rotation measurement and construct an initial set of historical candidate angles through incoherent decision-level fusion;

[0008] S3. Enter the adaptive verification decision stage, perform cluster analysis on the initial historical candidate angle set to determine the optimal candidate angle, and calculate its global consistency cost to determine whether the angle estimation algorithm has converged.

[0009] S4. Repeat the above steps to iterate and update the output. If convergence is achieved, output the current best candidate angle. If convergence is not achieved, perform targeted rotations to update the historical candidate angle set and continue iterating. If the iteration terminates due to reaching the maximum number of rotations, output the angle with the lowest cost in the historical candidate set.

[0010] Furthermore, the method in step S1 includes:

[0011] Based on formula (1), a received signal model is constructed; where formula (1) is:

[0012] (1);

[0013] in, This is the channel response vector. It is a zero-mean baseband signal. It is a complex Gaussian white noise vector;

[0014] Based on formula (2), a channel response vector model is constructed; where formula (2) is:

[0015] (2);

[0016] Where K is the number of paths. Let k be the direction of arrival for the k-th path. Let be the complex gain coefficient of the k-th path. This is the guiding vector corresponding to the k-th path;

[0017] Based on formula (3), a guiding vector model is constructed; where formula (3) is:

[0018] (3);

[0019] Where j is the imaginary unit and d is the array spacing.

[0020] Furthermore, the method in step S2 includes:

[0021] S21. Based on the sample covariance matrix, perform exploratory rotation and use the MUSIC algorithm to determine candidate angles;

[0022] S22. Perform eigenvalue decomposition on the covariance matrix and divide it into signal subspace and noise subspace;

[0023] S23. Based on the orthogonality between the signal subspace and the noise subspace, obtain the MUSIC spectrum from the l-th viewpoint;

[0024] S24. Perform spectral peak search for each viewpoint, and select multiple angles whose spectral peak values ​​exceed the preset cost threshold as candidate angles to obtain an initial set of historical candidate angles.

[0025] Furthermore, the method in step S21 includes:

[0026] Based on formula (4), the covariance matrix is ​​obtained; where formula (4) is:

[0027] (4);

[0028] Where T is the number of signal snapshots. It is the conjugate transpose of the channel response vector.

[0029] Furthermore, the method in step S22 includes:

[0030] Based on formula (5), the signal subspace and noise subspace are obtained; where formula (5) is:

[0031] (5);

[0032] in Let be the signal subspace matrix. The noise subspace matrix, The eigenvalue matrix of the signal. This is the noise eigenvalue matrix.

[0033] Furthermore, the method in step S23 includes:

[0034] Based on formula (6), the MUSIC spectrum under the l-th viewpoint is obtained; where formula (6) is:

[0035] (6);

[0036] in, The search test angle is defined in the global coordinate system. It is the conjugate transpose of the guide vector.

[0037] Furthermore, the method in step S3 includes:

[0038] S31. In each iteration, the initial historical candidate angle set is analyzed using a density-based one-dimensional clustering algorithm to obtain the largest cluster that meets the validity criteria, and the arithmetic mean of all candidate angles in the largest cluster is calculated as the optimal candidate angle.

[0039] S32. Based on the optimal candidate perspective, a globally consistent cost function is constructed;

[0040] S33. Compare the globally consistent cost function with the preset cost threshold to determine whether the angle estimation algorithm meets the convergence condition.

[0041] Furthermore, the method in step S4 includes:

[0042] S41. If the algorithm for angle estimation satisfies the convergence condition, the corresponding optimal candidate angle is output as the estimated angle.

[0043] S42. If the angle estimation algorithm does not meet the convergence condition, set the next rotation angle for targeted rotation, calculate the MUSIC spectrum under the new rotation angle, obtain the new candidate angle set and update it to the historical candidate angle set, and repeat step S3 to execute the iterative loop.

[0044] The beneficial effects of this application are as follows: By reconstructing the direction-of-arrival estimation into a sequential decision-making process and introducing an adaptive feedback mechanism to intelligently plan the measurement process, this application overcomes the inefficiency of existing fixed rotation strategies. This application can automatically adjust the measurement overhead according to the complexity of the scene, avoiding redundant measurements present in fixed strategies. While ensuring high accuracy, this application can complete the estimation with the fewest number of rotations, improving resource utilization efficiency and making it suitable for resource-constrained application platforms. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating an embodiment of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion of this application;

[0046] Figure 2 This is a schematic diagram of a system model of an embodiment of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion of this application;

[0047] Figure 3 This is a schematic diagram of the adaptive rotation algorithm flow of an embodiment of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion in this application;

[0048] Figure 4 This is a comparison of RMSE curves of different algorithms under different signal-to-noise ratios in a single-path scenario of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion in this application.

[0049] Figure 5 This is a comparison of the RMSE curves of the algorithm under different numbers of paths in a multipath scenario of an embodiment of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion in this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0051] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0052] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the sparse array direction-of-arrival estimation method based on adaptive decision-level fusion according to this application. The method includes:

[0053] S1. Establish a steering vector model to describe the received signal and channel transmission characteristics of the sparse array.

[0054] Specifically, perform system modeling, refer to Figure 2 The signal propagation scenario described in this application is a multipath environment, where a far-field, narrowband signal is emitted by a single transmitter (TX) and reaches the receiver (RX) via K paths. The signal receiver uses a rotatable N-element uniform linear array (ULA) with an array spacing of... ,in The wavelength is denoted as λ. All relevant geometric relationships are defined in a two-dimensional Cartesian coordinate system (XY plane), the origin of which is fixed at the center of rotation of the array.

[0055] The method of step S1 includes:

[0056] S11. When the array undergoes the l-th rotation, the rotation angle is... At that time, its received signal model The specific format is as follows:

[0057] (1);

[0058] in, This is the channel response vector. It is a zero-mean baseband signal. It is a complex Gaussian white noise vector;

[0059] S12. Its channel response vector model is:

[0060] (2);

[0061] Where K is the number of paths. Let k be the direction of arrival for the k-th path. Let be the complex gain coefficient of the k-th path. This is the guide vector corresponding to the k-th path.

[0062] S13. Its guiding vector can be defined as:

[0063] (3);

[0064] Where j is the imaginary unit and d is the array spacing.

[0065] S2. Based on the guide vector model, perform exploratory rotation measurements and construct an initial set of historical candidate angles through incoherent decision-level fusion.

[0066] Specifically, the method in step S2 includes:

[0067] S21. In the exploratory measurement phase, an initial set of historical candidate angles is constructed through incoherent decision-level fusion. (Reference) Figure 3 The core algorithm flow of this invention is shown in the figure. For each rotation, the sample covariance calculation formula is used to calculate the T snapshot signals. Calculating the covariance matrix The calculation formula is as follows:

[0068] (4);

[0069] Where T is the number of signal snapshots. It is the conjugate transpose of the channel response vector.

[0070] S22. Subsequently, the covariance matrix is ​​subjected to eigenvalue decomposition, dividing it into a signal subspace and a noise subspace. This process can be represented as:

[0071] (5);

[0072] in Let be the signal subspace matrix. The noise subspace matrix, The eigenvalue matrix of the signal. This is the noise eigenvalue matrix.

[0073] S23. Based on the orthogonality between the signal subspace and the noise subspace, the MUSIC spectrum at the l-th viewpoint is calculated using the following formula:

[0074] (6);

[0075] in, The search test angle is defined in the global coordinate system. It is the conjugate transpose of the guide vector.

[0076] S2.4 Performs a spectral peak search for each viewpoint, selecting the most... The peak value of each spectrum exceeds the preset cost threshold. The angle is used as a candidate angle to form a set of candidate angles under that viewpoint. ,in ,Will After summarizing, we obtain the initial set of historical candidate angles. .

[0077] S3. Enter the adaptive verification decision stage, perform cluster analysis on the initial historical candidate angle set to determine the optimal candidate angle, and calculate its global consistency cost to determine whether the steering vector model algorithm has converged.

[0078] Specifically, after completing the exploration and measurement, the algorithm enters the adaptive verification and decision-making phase. It performs cluster analysis on the historical candidate angle set to determine the optimal candidate angle and calculates its global consistency cost to determine whether the algorithm has converged. The detailed process is as follows: Figure 3 As shown.

[0079] The method in step S3 includes:

[0080] S31. In each iteration, a density-based one-dimensional clustering algorithm is used to process the historical candidate angle set. The analysis was performed, and the largest cluster that met the validity criteria was selected. Calculate all candidate angles within this cluster. The arithmetic mean is used as the optimal candidate angle. .

[0081] S32. Based on the optimal candidate perspective, a globally consistent cost function is constructed.

[0082] Specifically, based on the consistency of spectral peak positions across different viewpoints, this invention constructs a global consistency cost function to quantify the optimal candidate angle. The degree of consistency of the corresponding spectral peaks from various perspectives.

[0083] The principle of constructing the cost function: finding the optimal candidate angle. The candidate angle closest to it from the l-th viewpoint By accumulating the nearest peak distances from L different viewpoints, the optimal candidate angle can be obtained. The cost value. Its calculation formula:

[0084] (7);

[0085] in, Let L be the m-th candidate angle from the l-th observation viewpoint, and L be the total number of rotation measurements.

[0086] S33. Compare the global consistency cost function and preset cost thresholds To determine whether the algorithm for the directional vector model meets the convergence condition.

[0087] S4. Repeat the above steps to iterate and update the output. If convergence is achieved, output the current best candidate angle. If convergence is not achieved, perform targeted rotations to update the historical candidate angle set and continue iterating. If the iteration terminates due to reaching the maximum number of rotations, output the angle with the lowest cost in the historical candidate set.

[0088] For details, please refer to Figure 3 The method of step S4 includes:

[0089] S41. In step S3, if the cost value satisfies the convergence condition, i.e. At this point, the algorithm converges, and the corresponding optimal candidate angle is output as the estimated angle. .

[0090] S42. If the cost value does not meet the convergence condition, and the number of rotations l has not reached the preset maximum value. ,Right now If so, set the next rotation angle. Perform a targeted rotation, at which point the array sides face the optimal candidate angle. This maximizes the angular resolution in that direction, enabling efficient candidate angle verification and accurately eliminating spurious peaks caused by grating lobes. Subsequently, the MUSIC spectrum under the new rotational viewpoint is calculated to obtain a new set of candidate angles. and through Update the candidate angle obtained from this rotation to the set of historical candidate angles, and then return to step S3 to execute the iterative loop.

[0091] When the measurement cost reaches a preset upper limit (i.e., the number of array rotations reaches a preset maximum number of rotations), if the algorithm still does not meet the convergence condition, it selects the angle with the lowest cost value from the historical candidate angle set as the final estimated direction-of-arrival angle output. The determination formula is:

[0092] (8);

[0093] To verify the beneficial effects of the framework and method described in this invention, a series of Monte Carlo simulation experiments were conducted. The simulations used a rotatable sparse uniform linear array and employed root mean square error (RMSE) as the performance evaluation metric. The performance of the proposed fixed rotation strategy (FRS), adaptive rotation strategy (ARS), theoretical Cramer-Rao lower bound (CRLB), and the benchmark MUSIC algorithm applied to a traditional half-wavelength spacing ULA were compared. Furthermore, the robustness of the framework when extended to coherent multipath scenarios was evaluated. The specific parameter settings for each simulation are detailed in the subsequent results.

[0094] Figure 4 The simulation demonstrates the mean square error (RMSE) curves of the proposed Adaptive Rotation Strategy (ARS), Fixed Rotation Strategy (FRS), the baseline MUSIC algorithm, and the Cramer-Rao lower bound (CRLB) under different signal-to-noise ratios in a single-signal-source scenario, as a function of the number of rotations. In the simulation, ARS and FRS use sparse arrays with N=8 elements, array spacing d=5λ, and T=50 snapshots per rotation; the baseline MUSIC algorithm uses a conventional array with N=16 elements, array spacing d=0.5λ, and T=50 snapshots.

[0095] Depend on Figure 4 It can be seen that, in sparse array scenarios (element spacing greater than half a wavelength), the FRS strategy outperforms the benchmark MUSIC algorithm (element spacing half a wavelength) after approximately 7 rotations. The ARS algorithm within the incoherent decision-level fusion framework proposed in this invention exhibits even better performance, with lower measurement overhead than the FRS algorithm. It surpasses the benchmark MUSIC algorithm with only 5 rotations, and its RMSE is consistently lower than that of the FRS algorithm, especially showing a significant advantage at moderate rotation counts. The estimation accuracy of both methods improves significantly with increasing signal-to-noise ratio and gradually approaches CRLB, demonstrating the high accuracy characteristics of the proposed method.

[0096] To further verify the robustness of the framework of this invention, this embodiment also tested the method of this invention in a coherent multipath scenario. In this scenario, the method of this invention integrates spatial smoothing techniques to smooth the rank-deficient covariance matrix caused by signal coherence. Reconstruction is performed to restore its full-rank characteristic, enabling subsequent MUSIC algorithms to effectively distinguish between signal and noise subspaces. A prerequisite for effective spatial smoothing is that the number of subarrays Q must be greater than or equal to the number of paths K.

[0097] The formula for spatial smoothing is:

[0098] ;

[0099] in, Let J be the sample covariance matrix of the m-th subarray, J be the exchange matrix, and Q be the number of subarrays.

[0100] Figure 5 The paper demonstrates the variation curves of the overall root mean square error of each path in the algorithm with the number of rotations under coherent multipath scenarios, with multipath numbers K=2, 3, and 4. The simulation parameters were: number of array elements N=32, array spacing d=5λ, signal-to-noise ratio SNR=5dB, and number of snapshots T=50.

[0101] Depend on Figure 5 It can be seen that for different numbers of paths, the RMSE decreases with the increase of the number of rotations. As the number of paths increases, the algorithm requires more rotations to converge to a lower level. This proves the robustness of the proposed method in solving the grating lobe ambiguity problem in complex propagation environments.

[0102] This application overcomes the inefficiency of existing fixed rotation strategies by reconstructing direction-of-arrival estimation as a sequential decision-making process and introducing an adaptive feedback mechanism to intelligently plan the measurement process. This application can automatically adjust the measurement overhead according to the complexity of the scene, avoiding redundant measurements present in fixed strategies. While ensuring high accuracy, this application can complete the estimation with the fewest possible rotations, improving resource utilization efficiency and making it suitable for resource-constrained application platforms.

[0103] This application employs an incoherent decision-level fusion framework, evaluating multi-view candidate directions of arrival (DOAs) through a global consensus cost function, effectively improving the estimation accuracy. Simulation results show that the estimation performance of this invention outperforms several existing benchmark methods, and it closely matches the Cramer-Rao lower bound (CRLB) in regions with high signal-to-noise ratios, demonstrating the high-precision characteristics of the proposed method.

[0104] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A sparse array direction of arrival estimation method based on adaptive decision level fusion, characterized in that, Comprise: S1. Establish a steering vector model describing the received signal of a sparse array and channel transmission characteristics; S2. Based on the steering vector model, perform exploratory rotation measurement, construct an initial historical candidate angle set through non-coherent decision-level fusion; S3. Enter the adaptive verification decision stage, perform clustering analysis on the initial historical candidate angle set to determine the optimal candidate angle, and calculate its global consistency cost to determine whether the algorithm converges; S4. Repeat the above steps, perform iterative update and output, if it converges, output the current optimal candidate angle; if it does not converge, perform targeted rotation to update the historical candidate angle set and continue iteration, if the iteration is terminated due to reaching the maximum number of rotations, output the angle with the lowest cost in the historical candidate angle set.

2. The method of claim 1, wherein, The method of step S1 comprises: Based on formula (1), a received signal model is constructed; wherein the formula (1) is: （1）； wherein is a channel response vector, is a zero-mean baseband signal, is a complex Gaussian white noise vector; Based on formula (2), a channel response vector model is constructed; wherein the formula (2) is: （2）； wherein K is the number of paths, is the direction of arrival of the kth path, is the complex gain coefficient of the kth path, is the steering vector corresponding to the kth path. Based on formula (3), a steering vector model is constructed; wherein the formula (3) is: （3）； Wherein, j is the imaginary unit, d is the array spacing.

3. The method of claim 1, wherein, The method of step S2 comprises: S21. Based on the sample covariance matrix, perform exploratory rotation, and use the MUSIC algorithm to determine the candidate angle; S22. Perform eigenvalue decomposition on the covariance matrix and divide it into signal subspace and noise subspace; S23. Based on the orthogonality of the signal subspace and the noise subspace, obtain the MUSIC spectrum under the lth view angle; S24. Perform spectral peak search on each view angle, and take multiple angles with spectral peak values exceeding a preset cost threshold as candidate angles to obtain the initial historical candidate angle set.

4. The method of claim 3, wherein, The method of step S21 comprises: Based on formula (4), obtain the covariance matrix; wherein the formula (4) is: （4）； where T is the number of signal taps, is the conjugate transpose of the channel response vector.

5. The method of claim 3, wherein, The method of step S22 comprises: Based on formula (5), obtain the signal subspace and the noise subspace; wherein the formula (5) is: （5）； wherein is a signal subspace matrix, is a noise subspace matrix, is a signal eigenvalue matrix, is a noise eigenvalue matrix.

6. The method of claim 3, wherein, The method of step S23 comprises: Based on formula (6), obtain the MUSIC spectrum under the lth view angle; wherein the formula (6) is: （6）； wherein, is a search test angle defined in the global coordinate system, is a conjugate transpose of the steering vector.

7. The method of claim 1, wherein, The method of step S3 comprises: S31. In each iteration loop, use a one-dimensional clustering algorithm based on density to analyze the initial historical candidate angle set, obtain the largest cluster that meets the effectiveness standard, and calculate the arithmetic mean of all candidate angles in the largest cluster as the optimal candidate angle; S32. Based on the optimal candidate angle, construct a global consistency cost function; S33. Compare the global consistency cost function with a preset cost threshold to determine whether the angle estimation algorithm meets the convergence condition.

8. The method of claim 7, wherein, The method of step S4 comprises: S41. In response to the algorithm of the steering vector model meeting the convergence condition, output the corresponding optimal candidate angle as the estimated angle; S42. In response to the algorithm of the steering vector model not satisfying the convergence condition, the next rotation angle is set for targeted rotation, the MUSIC spectrum under the new rotation angle is calculated, a new candidate angle set is obtained and updated into the historical candidate angle set, and step S3 is repeated to perform an iteration loop.

Images