Single-sample neural network beam forming method combining formation and target orientation estimation and related equipment

Abstract

The invention relates to the technical field of beam forming, in particular to a single-sample neural network beam forming method combining formation and target orientation estimation and related equipment. According to the method, underwater acoustic signals are obtained based on a hydrophone array and are subjected to Fourier transform to form multi-frequency-point spectrum data, and a closed-loop neural beam forming frame is constructed. A sample covariance matrix is generated by using the multi-frequency-point data, and an empirical covariance matrix is obtained through an exponential weighted moving average algorithm. And inputting the time sequence convolutional network to extract cross-frequency covariance features, and generating a feature domain interference suppression control quantity which is used for constructing a robust covariance matrix and completing beam forming. And optimizing the broadband azimuth spectrum and updating the steering vector in combination with a formation adaptive mechanism, and finally obtaining the broadband azimuth spectrum and a target direction estimation result. According to the method, robust covariance estimation is realized, broadband interference is effectively suppressed, and formation errors are adaptively corrected, so that the accuracy and reliability of broadband target orientation estimation are remarkably improved.

Description

A single-sample neural network beamforming method and related equipment for joint array and target azimuth estimation Technical Field

[0001] This application relates to the field of beamforming technology, and in particular to a single-sample neural network beamforming method and related equipment for joint array and target azimuth estimation. Background Technology

[0002] In related technologies, with the rapid development of underwater acoustic detection, ocean surveillance, and underwater target identification tasks, array-based target location estimation (DOA) has become a key means of acquiring underwater acoustic information. Hydrophone arrays typically acquire broadband underwater acoustic signals from multiple array elements and use spatial filtering, frequency domain feature extraction, and azimuth spectrum analysis to infer the target's location, applying it to underwater search, navigation safety, and monitoring of marine equipment operations. Among these, array beamforming technology, as a core component, constructs directional gain to improve the signal-to-noise ratio and angular resolution, decisively influencing the final DOA estimation performance.

[0003] However, existing underwater acoustic beamforming and DOA estimation methods face significant robustness challenges in real-world marine environments. Complex and variable sea conditions cause small deformations in element positions due to current disturbances, platform attitude changes, and hydrodynamic influences. Simultaneously, marine noise exhibits strong multipath characteristics, non-stationarity, and frequency correlation, leading to inaccurate estimation of the sample covariance matrix under weak snapshot or low signal-to-noise ratio conditions. Furthermore, traditional beamforming methods often rely on large amounts of snapshot data to estimate the covariance matrix, which is prone to covariance matrix degradation under single-sample or small-sample conditions, resulting in a significant decrease in direction estimation accuracy and making them unsuitable for real-time positioning requirements in dynamic or sudden missions.

[0004] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0005] The main objective of this application is to propose a single-sample neural network beamforming method and related equipment for joint array formation and target azimuth estimation. This method can achieve robust covariance estimation, effectively suppress broadband interference, and adaptively correct array errors under single-sample conditions, thereby significantly improving the accuracy and reliability of broadband target azimuth estimation.

[0006] To achieve the above objectives, one aspect of this application proposes a single-sample neural network beamforming method for joint array configuration and target azimuth estimation. The method includes the following steps: acquiring underwater acoustic signals based on a hydrophone array; performing Fourier transform on the underwater acoustic signals to generate multi-frequency spectral data; constructing a closed-loop neural beamforming framework; constructing a sample covariance matrix based on the multi-frequency spectral data; smoothly updating the sample covariance matrix using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix; inputting the empirical covariance matrix into a temporal convolutional network to extract cross-frequency covariance features and generate a feature domain interference suppression control quantity; constructing a robust covariance matrix based on the feature domain interference suppression control quantity and the closed-loop neural beamforming framework; performing beamforming on the robust covariance matrix to obtain a broadband azimuth spectrum; optimizing the broadband azimuth spectrum based on an array self-adaptation mechanism to update the array steering vector, forming an array adaptive correction result; and outputting the final broadband azimuth spectrum and the corresponding target arrival direction estimation result based on the array adaptive correction result.

[0007] In some embodiments, the closed-loop neural beamforming framework includes a robust covariance construction layer and a beamforming output layer; the robust covariance construction layer is used to generate a robust covariance matrix for broadband beam optimization based on the sample covariance matrix of the multi-frequency spectral data and constraint information provided by the feature domain interference suppression control quantity; the beamforming output layer is used to perform broadband beamforming based on the robust covariance matrix.

[0008] In some embodiments, the step of using an exponentially weighted moving average algorithm to smoothly update the sample covariance matrix to obtain an empirical covariance matrix includes: at each snapshot, obtaining the sample covariance matrix of the current frame, using the empirical covariance matrix of the previous frame as the historical covariance matrix; weightedly fusing the sample covariance matrix obtained in the current snapshot with the historical covariance matrix, and using an exponentially weighted moving average algorithm to complete the smooth update, thereby generating the empirical covariance matrix corresponding to the current snapshot.

[0009] In some embodiments, the formula for the empirical covariance matrix is ​​as follows: ;in, Indicates the first The empirical covariance matrix of the frame; Indicates the previous time corresponding to the first The empirical covariance matrix of the frame; This represents the decay factor of the exponentially weighted moving average. This indicates the weight of the current snapshot outer product term; Indicates the first Frame, frequency point Array observation vectors; express The conjugate transpose of; Indicates the current snapshot frequency. The quick snapshot covariance matrix.

[0010] In some embodiments, the feature domain interference suppression control quantity includes eigenvalue soft suppression control quantity, isotropic noise contraction control quantity, and diagonal loading control quantity.

[0011] In some embodiments, constructing a robust covariance matrix based on the feature domain interference suppression control and the closed-loop neural beamforming framework includes: in the closed-loop neural beamforming framework, performing soft suppression processing on the principal feature subspace of the sample covariance matrix based on the eigenvalue soft suppression control to obtain eigenvalue soft suppression components; in the closed-loop neural beamforming framework, performing weighted fusion on the sample covariance matrix based on the isotropic noise contraction control to obtain isotropic contraction coefficients; in the closed-loop neural beamforming framework, performing loading correction on the diagonal terms of the sample covariance matrix based on the diagonal loading control to obtain diagonal loading factors; and structurally fusing the eigenvalue soft suppression components, the isotropic contraction coefficients, and the diagonal loading factors to form a robust covariance matrix for broadband beamforming.

[0012] In some embodiments, the formula for the robust covariance matrix is ​​as follows: ;in, Represents a robust covariance matrix; The eigenvalue soft suppression component is represented by M; M represents the number of array elements. Indicates the isotropic contraction coefficient; This represents the total energy of the covariance matrix; Represents the identity matrix; This represents the diagonal loading factor.

[0013] In some embodiments, the formula for calculating the broadband azimuth spectrum is as follows: ;in, Indicates the broadband azimuth spectrum value; Indicates at the angle of incidence Frequency The array guide vector below; express The conjugate transpose of; Representing the robust covariance matrix The inverse matrix.

[0014] In some embodiments, the formation self-adaptation mechanism is constructed based on the Rayleigh distribution criterion and optimizes the formation by shifting array elements, delaying channels, and rotating the entire array.

[0015] To achieve the above objectives, another aspect of this application proposes a single-sample neural network beamforming system. The system includes: an underwater acoustic signal acquisition module for acquiring underwater acoustic signals based on a hydrophone array; a data preprocessing module for performing Fourier transform on the underwater acoustic signals to generate multi-frequency spectral data; a first construction module for constructing a closed-loop neural beamforming framework; a second construction module for constructing a sample covariance matrix based on the multi-frequency spectral data; a first calculation module for smoothly updating the sample covariance matrix using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix; and a feature extraction module for... An empirical covariance matrix is ​​input into a temporal convolutional network to extract cross-frequency covariance features and generate a feature domain interference suppression control quantity. A third construction module is used to construct a robust covariance matrix based on the feature domain interference suppression control quantity and the closed-loop neural beamforming framework. A second calculation module is used to perform beamforming on the robust covariance matrix to obtain a broadband azimuth spectrum. An adaptive correction module is used to optimize the broadband azimuth spectrum based on the array self-adaptation mechanism to update the array steering vector and form an array adaptive correction result. An output module is used to output the final broadband azimuth spectrum and the corresponding target arrival direction estimation result based on the array adaptive correction result.

[0016] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0018] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0019] The embodiments of this application include at least the following beneficial effects: This application provides a single-sample neural network beamforming method and related equipment for joint array and target azimuth estimation. This scheme effectively enhances the adaptability to complex interference sources in underwater acoustic environments by introducing a closed-loop neural beamforming framework and combining it with feature domain interference suppression control. Especially under the condition of only a single sample, by using a temporal convolutional network to mine the cross-frequency feature relationship in the frequency domain covariance matrix, high azimuth estimation accuracy can be maintained under conditions of wide bandwidth and low signal-to-noise ratio, significantly improving the robustness and practicality of the system. By using the exponentially weighted moving average (EWMA) algorithm to smoothly update the sample covariance matrix, the instability of covariance estimation caused by the scarcity of samples can be effectively suppressed. On this basis, a feature domain control is introduced to construct a robust covariance matrix, making the beamforming process more robust and able to actively suppress spatially correlated interference sources, thereby obtaining a clearer broadband azimuth spectrum. After obtaining the initial broadband azimuth spectrum, the array steering vector is adaptively updated through an array self-adaptation mechanism. This helps correct direction estimation deviations caused by array distortion or actual array errors, thereby outputting a more accurate target arrival direction. This closed-loop adaptive feedback mechanism enables the system to continuously optimize and is suitable for underwater target detection and localization tasks in dynamic marine environments. Attached Figure Description

[0020] Figure 1 is a flowchart illustrating a single-sample neural network beamforming method for joint array and target azimuth estimation provided in an embodiment of this application; Figure 2 shows the variation of statistical error with signal-to-noise ratio for the three methods provided in an embodiment of this application; Figure 3 shows the variation of DOA estimation error with element position perturbation variance for different methods provided in an embodiment of this application; Figure 4 shows the azimuth history of the full-time processing results of (top) Bartlett, (middle) MVDR, and (bottom) Neural-MVDR provided in an embodiment of this application; Figure 5 shows the comparison results of predicted trajectories and GPS trajectories for the three methods provided in an embodiment of this application; Figure 6 shows the robustness provided in an embodiment of this application. Figure 7 shows the azimuth history of the full-time processing results of (top) Bartlett, (middle) MVDR, and (bottom) Neural-MVDR under the random bias mismatch of the array with a mean of 0 and a variance of 2 provided in the embodiment of this application; Figure 8 shows the comparison between the full-time processed DOA prediction value and GPS under the random bias mismatch of the array with a mean of 0 and a variance of 2 provided in the embodiment of this application; Figure 9 shows the Bartlett processing results of the mismatched array pattern and the predicted array pattern under the random bias mismatch of the array with a mean of 0 and a variance of 2 provided in the embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0022] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0023] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] This application provides a single-sample neural network beamforming method and related equipment for joint array configuration and target azimuth estimation. This scheme effectively enhances adaptability to complex interference sources in underwater acoustic environments by introducing a closed-loop neural beamforming framework and combining it with feature-domain interference suppression control. Especially under conditions with only a single sample, utilizing a temporal convolutional network to mine cross-frequency characteristic relationships in the frequency domain covariance matrix maintains high azimuth estimation accuracy even with wide bandwidth and low signal-to-noise ratio, significantly improving the system's robustness and practicality. Using the exponentially weighted moving average (EWMA) algorithm to smoothly update the sample covariance matrix effectively suppresses the instability in covariance estimation caused by the scarcity of samples. Based on this, a robust covariance matrix is ​​constructed by introducing feature-domain control, making the beamforming process more robust and enabling proactive suppression of spatially correlated interference sources, thereby obtaining a clearer broadband azimuth spectrum. After obtaining the preliminary broadband azimuth spectrum, an array self-adaptation mechanism is used to adaptively update the array steering vector, helping to correct direction estimation deviations caused by array distortion or actual array configuration errors, thus outputting a more accurate target arrival direction. This closed-loop adaptive feedback mechanism enables the system to continuously optimize, making it suitable for underwater target detection and localization tasks in dynamic marine environments.

[0026] This application provides a single-sample neural network beamforming method and related equipment for joint array and target azimuth estimation, relating to the field of beamforming technology. The single-sample neural network beamforming method for joint array and target azimuth estimation provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the server can also be a node server in a blockchain network; the software can be an application implementing a single-sample neural network beamforming method for joint array and target azimuth estimation, but is not limited to the above forms.

[0027] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0028] Figure 1 is an optional flowchart of a single-sample neural network beamforming method for joint array and target azimuth estimation provided in an embodiment of this application. The method in Figure 1 may include, but is not limited to, steps S1 to S10: S1: Acquiring underwater acoustic signals based on a hydrophone array; In this embodiment, a multi-channel hydrophone array is deployed on the seabed, towed platform, or other water areas to synchronously receive radiated noise or reflected echoes from underwater targets. Each array element undergoes amplitude and phase consistency calibration before being put into use, and the DC bias is removed before processing to ensure the consistency of the data source.

[0029] During the method execution, all channels of the multi-channel hydrophone array are sampled using a unified sampling clock, with the sampling rate set in the thousands of hertz range to cover the predetermined processing frequency band. The acquired raw time-domain data enters the processing flow in frames, with each frame containing several snapshots, and each snapshot consisting of synchronous sampling points from all array elements.

[0030] To improve the stability of subsequent frequency domain characteristics, several preprocessing operations can be performed on the received signal at the initial stage of the data processing flow. These include bandpass filtering to suppress low-frequency ocean current noise and high-frequency electronic noise, time alignment to eliminate potential minor synchronization offsets, and fast amplitude consistency checks to identify abnormal array elements with excessive amplitude-phase deviations. If an abnormal array element is detected, the data weight of that channel can be appropriately reduced in subsequent steps to avoid adverse effects on the overall processing results.

[0031] S2: Perform a Fourier transform on the underwater acoustic signal to form multi-frequency spectral data. In this embodiment, after obtaining the underwater acoustic signal, a short-time Fourier transform can be performed on it to form array frequency domain features for subsequent beamforming. Specifically, an analysis window with good sidelobe suppression capability (such as a Hanning window or a Blackman window) can be used to segment the time domain data, and an appropriate window length can be set according to the sampling rate, signal bandwidth, and real-time requirements, for example, using a length of 8192 points, and using a window overlap rate of about 50% or more, so that the frequency domain representation has continuity and sufficient time resolution.

[0032] After performing a Fourier transform on each segment of time-domain data, a complex matrix of array elements × frequency points is obtained, which describes the amplitude and phase response of each array element in that frequency band. When the processing frequency band is defined (e.g., 150–350 Hz), a set of stable and effective frequency points can be selected from the transform results as input for subsequent processing. These frequency points maintain a high signal-to-noise ratio and can reflect the energy distribution of the target and interference sources at different frequencies.

[0033] In complex sea conditions or under strong interference, to avoid sudden spikes or anomalous jumps in frequency domain energy, the complex spectrum of adjacent frequency points can be lightly smoothed to reduce the impact of discrete frequency domain noise on subsequent covariance estimation. Furthermore, extreme anomalous frequency points can be removed based on energy thresholds to ensure more robust subsequent empirical covariance updates and feature learning processes.

[0034] Ultimately, the complex frequency domain data of all effective frequency points can be arranged in frequency order to form continuous multi-frequency spectrum data, providing consistent and stable input for the next step of sample covariance construction and feature learning.

[0035] S3: Construct a closed-loop neural beamforming framework; wherein, the closed-loop neural beamforming framework includes a robust covariance construction layer and a beamforming output layer; the robust covariance construction layer is used to generate a robust covariance matrix for broadband beam optimization based on the sample covariance matrix of the multi-frequency spectral data and the constraint information provided by the feature domain interference suppression control quantity; the beamforming output layer is used to perform broadband beamforming based on the robust covariance matrix.

[0036] In this embodiment, in the robust covariance construction layer, the cross-frequency statistical structure is first extracted based on the sample covariance matrix and input into a temporal convolutional network for encoding to obtain potential features describing the main interference subspace, the equal-power noise structure, and the matrix condition number state. Subsequently, the feature domain interference suppression control quantity output by the network is generated through a multi-branch structure, including soft suppression information for strong interference feature directions, shrinkage adjustment information for overall noise energy, and diagonal loading information to ensure the numerical stability of the matrix. The robust covariance construction layer combines these three types of information with the original statistics, so that the final robust covariance matrix can not only retain the structural features of the real target, but also effectively reduce spectral leakage and ill-conditionedness caused by array errors, strong interference, or insufficient snapshots, greatly improving its usability in broadband beamforming.

[0037] In the beamforming output layer, the broadband beam weights are solved and the spectrum is generated based on the robust covariance matrix. Since the previous layer has explicitly suppressed interference energy and improved the positive definiteness of the covariance matrix, the beamforming output layer can achieve higher main lobe focusing and lower sidelobe leakage while maintaining distortion-independent characteristics. After the broadband azimuth spectrum is generated, the frame's feedback path adaptively fine-tunes the sub-element positions, channel delays, and global small rotations using the energy distribution in the reliable azimuth region of the spectrum. The corrected steering vector is then fed back to the covariance construction layer, forming a closed-loop optimization within a single sample, causing the beam output to continuously converge towards a direction that better matches the actual beamformation.

[0038] S4: Construct a sample covariance matrix based on multi-frequency spectral data. In this embodiment, for a given frequency point, the array complex frequency domain vector of the corresponding frame can be combined with its own complex conjugate vector by an outer product to form the instantaneous covariance estimate of that frame. Since the instantaneous covariance is easily affected by insufficient snapshots and random noise, in order to improve the estimation stability, the instantaneous covariance formed by multiple consecutive snapshots can be initially averaged to serve as the basis for the sample covariance of that frequency point.

[0039] In complex sea conditions or scenarios with strong interference, the sample covariance matrix is ​​prone to energy distribution distortion. For example, some array elements may be affected by fluid disturbances or biological noise, causing their corresponding covariance terms to suddenly increase. To avoid these anomalies affecting the overall estimation, channel consistency detection can be performed before constructing the sample covariance. For example, by comparing the energy fluctuations of each array element at nearby frequency points, possible abnormal channels can be identified, and they can be given lower weights during covariance calculation.

[0040] Furthermore, to further enhance the robustness of the covariance matrix, anomalous snapshots with energy significantly higher than the background (such as short-term high-energy disturbances caused by the sudden passing of vessels on the water) can be removed during the calculation process, retaining only statistically stable snapshots to form the sample covariance. After the above processing, a preliminary sample covariance matrix reflecting the current sea state, disturbance structure, and formation status can be obtained.

[0041] S5: The sample covariance matrix is ​​smoothly updated using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix; wherein, the smooth update of the sample covariance matrix using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix includes: at each snapshot, obtaining the sample covariance matrix of the current frame, and using the empirical covariance matrix of the previous frame as the historical covariance matrix; the sample covariance matrix obtained in the current snapshot is weighted and fused with the historical covariance matrix, and the exponentially weighted moving average algorithm is used to complete the smooth update to generate the empirical covariance matrix corresponding to the current snapshot.

[0042] In this embodiment, the sample covariance matrix corresponding to the current snapshot can be obtained at the beginning of each snapshot processing, and the empirical covariance matrix updated in the previous snapshot or frame can be used as a historical covariance reference. In sea trial environments or highly time-varying sea states, the sample covariance matrix will exhibit significant randomness due to the limited number of snapshots, fluctuations in array element noise, or sudden changes in interference intensity. Therefore, it is necessary to introduce historical statistics to enhance the stability of the current covariance estimation. By using this frame-by-frame approach, the current covariance estimation can simultaneously reflect the characteristics of the latest snapshot and the background trend of previous frames.

[0043] Subsequently, after obtaining the current sample covariance matrix, an exponentially weighted fusion can be performed on the current sample covariance and the historical covariance. The core idea of ​​this fusion method is to adjust the proportion of "historical information" and "current information" according to a pre-set attenuation factor, so that the method can maintain long-term statistical stability in the face of slowly changing ocean noise environments, while responding quickly to new features when encountering sudden changes such as target appearance or increased interference. If the current sample covariance contains random strong noise or occasional anomalous energy, exponentially weighted fusion can effectively weaken its influence, thereby avoiding the introduction of incorrect interference subspace estimation in subsequent beamforming.

[0044] After weighting, the empirical covariance matrix corresponding to the current snapshot is obtained. This matrix is ​​numerically more robust and will not exhibit drastic jumps due to the limited number of snapshots or sudden changes in noise statistics. Simultaneously, this empirical covariance matrix retains the characteristics of the current snapshot, enabling it to accurately reflect the structural features of the real signal under the current sea state. The final empirical covariance matrix will serve as the input to the temporal convolutional network in the next step, allowing the network to simultaneously learn noise trends, interference patterns, and array mismatch characteristics across frequency and time dimensions, thus enabling the generation of subsequent feature domain interference suppression control parameters.

[0045] Specifically, the formula for the empirical covariance matrix is ​​as follows: In the formula, Indicates the first The empirical covariance matrix of the frame; Indicates the number corresponding to the previous moment. Frame empirical covariance matrix; This represents the decay factor of the exponentially weighted moving average. This indicates the weight of the current snapshot outer product term; Indicates the first Frame, frequency point Array observation vectors; express The conjugate transpose of; Indicates the current snapshot frequency. The quick snapshot covariance matrix.

[0046] in, The calculation formula is as follows: In the formula, Indicates the first Frame, frequency point Array observation vectors; Indicates the first Frame, frequency point The target signal complex amplitude; Indicates the incident azimuth of the actual target; Indicates the first Frame, frequency point Interference and noise vectors; Indicates parameters with small deformation (micro-displacement of array elements) Channel micro-delay Global small rotation The guide vector of ).

[0047] in, The calculation formula is as follows: In the formula, , , It is a planar rotation matrix. Speed ​​of sound; Indicates the first The nominal position coordinate vector of each array element.

[0048] S6: Input the empirical covariance matrix into a temporal convolutional network, extract cross-frequency covariance features, and generate feature domain interference suppression control quantities; the feature domain interference suppression control quantities include eigenvalue soft suppression control quantities, isotropic noise contraction control quantities, and diagonal loading control quantities.

[0049] In this embodiment, after the empirical covariance matrix is ​​constructed, it can be used as a cross-frequency sequence input to a temporal convolutional network (TCN) with a causal structure for feature extraction. This network uses multi-layer dilated convolutions to model the correlation between adjacent frequency points, enabling the method to identify recurring interference patterns and noise structure trends at different frequencies, thereby accurately modeling the energy distribution within the covariance matrix. The TCN has 5 layers with a kernel size of 3, and uses dilated convolutions with dilation rates of 1, 2, 4, 8, and 16 to achieve a larger receptive field and higher expressive power. The activation function is ReLU. The TCN possesses both memory and causality. Compared to traditional symmetric non-causal CNNs and recurrent networks such as LSTM, it has a smaller data dependency, faster computation speed, and lower latency.

[0050] The empirical covariance matrix is ​​input into the network sequentially according to frequency points. The first few convolutional layers of the network can extract low-frequency, slowly changing features determined by the marine environment, such as phase consistency shifts caused by multipath propagation; the middle convolutional layers can identify the coherence characteristics of interference sources at multiple frequency points, such as the directional behavior of strong interference subspaces; and the deeper convolutional layers can capture higher-order local inconsistencies related to the amplitude and phase errors of array elements.

[0051] In this way, the network ultimately generates the following three types of interpretable feature-domain interference suppression control variables: 1. Eigenvalue soft suppression control variables (for strong interference principal subspaces): Eigenvalue soft suppression control variables are used to dynamically weaken the most significant eigenvalues ​​in the covariance matrix, softening the energy-dominant strong interference directions without completely destroying the original signal subspace structure. This type of control variable is particularly effective when strong interference sources are persistent and exhibit consistent cross-frequency behavior, significantly mitigating the binding effect on the directional spectrum caused by interference subspace expansion.

[0052] 2. Isotropic Noise Shrinkage Control Variable (Mitigating the Influence of Insufficient Snapshots and Random Noise): The isotropic noise shrinkage control variable is used to convexly combine the covariance matrix and the isotropic noise model within a certain range, thereby resisting the ill-conditioned nature of the covariance matrix when snapshots are severely insufficient. This control variable is particularly important in sea trial environments, as it can maintain the positive definiteness and stability of the covariance matrix when random noise levels change rapidly.

[0053] 3. Diagonal Loading Control (Improving Numerical Stability): The diagonal loading control appropriately enhances the main diagonal terms of the covariance matrix, increasing the minimum eigenvalue of the matrix and thus improving the stability of matrix inversion during subsequent beamforming. This approach is particularly effective for sea states with low signal-to-noise ratios, strong multipath propagation, or significant formation errors.

[0054] All three types of control variables are automatically generated based on the current input samples and do not require any offline training. They can adaptively reflect environmental changes within the samples. After obtaining the three types of control variables, they can be fused with the input covariance matrix in the next step to form a robust covariance matrix.

[0055] S7: Construct a robust covariance matrix based on the feature domain interference suppression control and the closed-loop neural beamforming framework; wherein, constructing a robust covariance matrix based on the feature domain interference suppression control and the closed-loop neural beamforming framework includes: in the closed-loop neural beamforming framework, performing soft suppression processing on the principal feature subspace of the sample covariance matrix based on the eigenvalue soft suppression control to obtain eigenvalue soft suppression components; in the closed-loop neural beamforming framework, performing weighted fusion on the sample covariance matrix based on the isotropic noise contraction control to obtain isotropic contraction coefficients; in the closed-loop neural beamforming framework, performing loading correction on the diagonal terms of the sample covariance matrix based on the diagonal loading control to obtain diagonal loading factors; and structurally fusing the eigenvalue soft suppression components, the isotropic contraction coefficients, and the diagonal loading factors to form a robust covariance matrix for broadband beamforming.

[0056] In this embodiment, after generating the eigendomain interference suppression control, the sample covariance matrix can be progressively modified through multiple components during closed-loop neural beamforming to mitigate the impact of strong interference, noise imbalance, and array mismatch on beamforming performance. First, based on the eigenvalue soft suppression control, the principal eigenspace of the sample covariance matrix can be flexibly weakened. Especially when a stable and energy-dominant interference source exists, weakening the intensity of this principal subspace prevents the energy corresponding to the interference direction from dominating the covariance structure, thus avoiding the interference spread effect commonly found in traditional beamforming methods. This process preserves the structure of the signal subspace, ensuring that subsequent azimuth estimation can still accurately utilize target-related information.

[0057] Furthermore, this embodiment can also utilize isotropic noise shrinkage control to fuse the sample covariance matrix and the equal-power noise model. This shrinkage operation can introduce necessary smoothness into the covariance structure under conditions of low snapshot, low signal-to-noise ratio, or unstable data statistics. By weighted fusion of the two types of matrices, the high-variance noise terms in the covariance matrix are effectively weakened, and the local anomalous covariance terms related to array errors are also smoothed, making the overall covariance matrix structurally closer to an ideal positive definite distribution, which is beneficial for maintaining numerical stability during the beamforming stage.

[0058] Furthermore, this embodiment can also enhance the main diagonal terms of the sample covariance matrix based on the diagonal loading control. This step improves the matrix condition number by increasing the minimum eigenvalue, providing further compensation when the covariance matrix is ​​close to ill-conditioned or has significant element amplitude and phase errors. Finally, by synergistically fusing the three types of components—eigenvalue soft suppression component, isotropic contraction coefficient, and diagonal loading factor—a robust covariance matrix with stronger stability, greater interference suppression capability, and cross-frequency consistency can be obtained, providing reliable input for subsequent broadband beamforming.

[0059] The formula for the robust covariance matrix is ​​as follows: ;in, Represents a robust covariance matrix; The eigenvalue soft suppression component is represented by M; M represents the number of array elements. Indicates the isotropic contraction coefficient; This represents the total energy of the covariance matrix; Represents the identity matrix; This represents the diagonal loading factor.

[0060] S8: Beamforming is performed on the robust covariance matrix to obtain a broadband azimuth spectrum. In this embodiment, after the robust covariance matrix is ​​constructed, the corresponding frequency domain azimuth response can be calculated at each effective frequency point according to the distortion-independent constraint. Specifically, in this step, the output power of the array in that direction is calculated angle by angle based on the current robust covariance matrix and the preset candidate angular domain. Since the robust covariance has undergone eigenvalue soft suppression, noise reduction, and diagonal loading, the frequency domain beam output at this time can still maintain good main lobe focusing capability and side lobe suppression capability under strong interference, multipath effect, or slight array mismatch conditions. A narrowband azimuth spectrum with angle-power distribution is obtained for each frequency point, which is used to describe the possible target radiation direction at that frequency point.

[0061] After obtaining the narrowband azimuth spectra of all frequencies, broadband convergence processing can be further performed to fuse azimuth information at different frequencies in the same angular domain. To avoid numerical bias and noise amplification caused by direct superposition, the azimuth spectra of each frequency point are made to have similar energy scales before fusion. Through cross-frequency fusion, the stability of target direction estimation can be effectively improved: if some frequencies are affected by noise, array perturbation, or multipath interference, their offset will be balanced by the stable response of the other frequencies, thus making the final broadband azimuth spectrum present a clearer main lobe peak and lower side lobe noise.

[0062] After the broadband azimuth spectrum is formed, continuous scanning of the entire angular domain can be performed to identify the most significant peak directions. The main lobe of the broadband azimuth spectrum is more spatially concentrated and has a clear advantage, enabling the peak positions obtained in this step to accurately reflect the true arrival direction of the target. Furthermore, since this method is a closed-loop structure, the broadband azimuth spectrum not only serves as the final output but also participates in the subsequent array self-adaptation process, allowing its peak information to be used in reverse to guide the correction of the steering vector and the updating of array deformation parameters, providing a more accurate geometric consistency basis for the next round of beamforming.

[0063] The calculation formula for the broadband azimuth spectrum is as follows: ;in, Indicates the broadband azimuth spectrum value; Indicates at the angle of incidence Frequency The array guide vector below; express The conjugate transpose of; Representing the robust covariance matrix The inverse matrix.

[0064] S9: The broadband azimuth spectrum is optimized based on the array self-adaptation mechanism to update the array steering vector and form an array adaptive correction result; the array self-adaptation mechanism is constructed based on the Rayleigh distribution criterion and optimizes the array by array element translation, channel delay and global rotation.

[0065] In this embodiment, after obtaining the broadband azimuth spectrum, the steering vector can be further corrected using an array self-adaptation mechanism, making the azimuth spectrum output more focused and reducing the mismatch effect caused by array errors. In specific implementation, reliable angular regions can first be selected from the current broadband azimuth spectrum. A soft masking method is used to assign higher weights to azimuth angles near the target main lobe, while automatically reducing the weights of sidelobe regions and uncertain regions. This soft-weight-based focusing strategy allows the method to extract stable geometric consistency information from the spectral peak shape, providing a reliable basis for subsequent array adjustments.

[0066] Subsequently, the degree of matching between the current steering vector and the azimuth spectrum energy distribution can be evaluated using the Rayleigh consistency criterion. To achieve fine-grained correction of array errors, three types of physically interpretable array deformations can be introduced: micro-displacement of array elements, micro-delay of channels, and small global rotations within the plane. Each type of deformation parameter can be updated incrementally with minimal step sizes, allowing the steering vector to move slowly along the direction that improves Rayleigh consistency. In this way, a continuous iterative process from "current steering vector → azimuth spectrum matching evaluation → micro-deformation compensation → new steering vector" can be achieved.

[0067] During the iteration process, the steering vector can be recalculated and written back into the closed-loop beamforming step after each deformation update, allowing the new steering vector to participate in the next round of azimuth spectrum estimation. If the focus of the azimuth spectrum main lobe significantly improves, it indicates that the array geometry error has been effectively corrected, and the number of deformation steps can be increased further. If the Rayleigh consistency does not improve significantly after multiple consecutive updates, it can be determined that the array has approached the optimal form, and the iteration can be terminated early. In this way, the array can still achieve dynamic geometric self-calibration even without any training data, significantly improving the reliability of the beamforming output.

[0068] Specifically, the formation is adaptively configured using the Rayleigh distribution criterion, and the calculation formula is as follows: In the formula, This represents the Rayleigh Quotient function.

[0069] This is a similarity measure for array geometric deformation. To avoid being affected by background noise, a reliable angular domain mask is used. The geometric cost is obtained by performing soft maximum aggregation and cross-frequency summation, as shown in the following formula: In the formula, The objective function for optimizing the array geometry is denoted as . This represents a temperature-weighted softmax enhancement of Rayleigh consistency; Indicates the first The positional displacement vectors of each array element; Indicates the first The channel micro-delay deviation of each array element; This represents the array's global small rotation angle; This represents the regularization coefficient for perturbation of array element positions; This represents the time delay deviation regularization coefficient; This represents the regularization coefficient for rotational deformation.

[0070] Used to limit coordinate deformation Used to limit time delay error, Used to limit the global rotation angle The distillation temperature controls the smoothness of the distillation process. The distillation temperature is typically set to... A larger value indicates a smoother surface, while a smaller value indicates a more prominent peak. , , The typical setting range is... Between, in the embodiments of the present invention, respectively set as , as well as This indicates a preference for optimizing the array manifold based on position and time delay differences over global small rotations. The optimization convergence criterion is the geometric cost of two iterations. .

[0071] right , , The optimization can be achieved using the gradient descent method, where the gradient can be obtained in a closed form using the chain rule, denoted as... , , , ,make In the formula, Indicates Rayleigh consistency The gradient of the guide vector; Represents the steering vector energy, used for Rayleigh normalization; This indicates that the covariance matrix acts on the steering vector and is used to calculate the energy projection in the direction of a.

[0072] Then the array element micro-displacement: In the formula, This represents the m-th component of the steering vector; express The m-th element is used for chain-recursive differentiation; Represents the unit propagation direction vector; Represents frequency wavenumber; This represents the operator for taking the real part.

[0073] Channel micro-delay: In the formula, Indicates the time delay deviation of the steering vector phase. The derivative of .

[0074] Global small rotation: In the two dimensions Aggregated to When calculating the gradient, soft maximum weights are used in the angle domain, therefore ;in, , , , Single-frame, multi-step, small-step projection gradient ascent can achieve online optimization of array geometry. The optimization step size is typically set to... The updated version Immediately write back the steering vector; closed-loop feedback affects azimuth spectrum estimation.

[0075] S10: Based on the array adaptive correction results, output the final broadband azimuth spectrum and the corresponding target arrival direction estimation results.

[0076] In this embodiment, once the optimization iteration of the array self-adaptation reaches a preset convergence threshold, the steering vector corresponding to the current deformation parameters can be used as the final steering vector, and the final broadband azimuth spectrum can be regenerated accordingly. In this azimuth spectrum, the main lobe region typically exhibits a peak with highly concentrated energy, while the side lobe regions are significantly reduced due to the combined effect of robust covariance construction and array calibration, resulting in a spectrum with good directivity and resolution. This final azimuth spectrum can serve as a stable output of this method for the current input sample, used for subsequent acoustic situation analysis or target localization.

[0077] To determine the target's final direction of arrival, the azimuth angle can be directly extracted from the main lobe peak position of the final broadband azimuth spectrum. Since the aforementioned steps have already performed adaptive covariance correction, feature domain interference suppression, and array geometry optimization within a single sample, the final azimuth angle is no longer significantly affected by array element micro-displacements, steep interference subspaces, or time-varying noise, exhibiting higher robustness. In typical sea states, the azimuth angle output by this method typically maintains close consistency with the actual trajectory, and can maintain high accuracy even under conditions of low signal-to-noise ratio or array perturbations.

[0078] Furthermore, if the method needs to process continuous frame data, the final formation deformation parameters obtained in the current step can be used as the initial formation state for the next frame, enabling the entire processing to have cross-frame continuity and fast convergence capability. When the target trajectory changes gradually, this cross-frame inheritance strategy can significantly reduce the number of optimizations for formation self-adaptation and accelerate real-time processing. When environmental noise changes abruptly or interference occurs, the method can also readapt within a few iterations, forming a long-term stable closed-loop azimuth estimation system.

[0079] In this embodiment, a comparative experiment was conducted on the publicly available underwater acoustic dataset SWellEx-96. During the experiment, the surface vessel R / V Sproul navigated along a track located south of the array, moving northward at a radial speed of 2.5 m / s, while towing two sets of acoustic sources, one for deep water and one for shallow water; the distance between the acoustic sources was determined using GPS. The receiving array was a horizontal line array HLANorth deployed on the seabed at a depth of approximately 213 m, with a total aperture of 240 m, 27 channels, an array sampling rate of 3276.8 Hz, a processing signal frequency range of 150 Hz-350 Hz, 8192 Fast Fourier Transform points, a 50% overlap rate per snapshot, and 150 processing frequencies selected per snapshot. The experimental analysis of this invention was completed on a computing platform equipped with an NVIDIA GeForce RTX 4090 GPU. In this invention, the optimization step size was set to... distillation temperature , , , Set to respectively , as well as This indicates a preference for optimizing the array manifold based on array geometry and time delay differences over global small rotations. Attenuation coefficient Set as This is roughly equivalent to 40 frames per second, making it a common choice for SWellEx-96 data.

[0080] First, based on the array configuration, the effectiveness and robustness of the method are demonstrated through data simulation. The statistical indicators are mean absolute error (MAE), median error, and 90th percentile (P90). Signal-to-noise ratio (SNR) is defined as the average SNR of the array bandwidth. Under a fixed azimuth of 240°, simulation data of different array SNRs were generated based on the HLA-North non-uniform array, and 1000 Monte Carlo experiments were conducted to compare and evaluate CBF (Bartlett linear beamforming), adaptive beamforming algorithm (MVDR), and closed-loop neural beamforming framework (Neural-MVDR). The results, shown in Figure 2, indicate that the MAE, Median, and P90 indices of the three methods all decrease monotonically with increasing SNR, accompanied by a significant threshold effect: CBF enters the effective operating range first on the low SNR side (MAE / Median is close to 0° at approximately 0-2 dB, and P90 drops to near 0° at 2 dB), demonstrating its robustness to covariance estimation errors; the threshold of traditional MVDR is furthest to the right (approximately +3). The fact that it only achieves stable convergence at +6 dB indicates its sensitivity to covariance matrix estimation. Neural-MVDR significantly outperforms MVDR in the low to medium SNR range, with its threshold shifted approximately 2–3 dB to the left compared to MVDR, and it approaches 0° earlier on MAE / Median / P90, demonstrating that the learned robust covariance estimation effectively suppresses estimation noise and improves robustness. In the high SNR range, the performance of all three converges gradually.

[0081] Furthermore, the robustness of this method to array perturbations was verified by adding noise to the array. The results are shown in Figure 3.

[0082] In Figure 3, (a) subplot represents the mean absolute error curve: this subplot shows the variance of perturbations of various beamforming methods under different array configurations. The mean absolute error (MAE) is used to measure the overall estimation accuracy. MAE represents the average deviation between the predicted azimuth and the true azimuth. As element position perturbations increase, the MAE of all methods shows an upward trend, but the increase and fluctuations in the curves reflect the differences in robustness of different algorithms under array mismatch scenarios.

[0083] (b) Subplot showing the Median Error Curve: This subplot illustrates the median error for each method under different array perturbation conditions. The median measures the error level in typical scenarios and, compared to MAE, better reflects the stability of the method across most frames. With... As the median error increases, the median error of each method also increases, but the magnitude of the increase varies, reflecting the method's ability to resist disturbances caused by the positional deviation of the array elements.

[0084] (c) Subplot showing the P90 (90th percentile error) curve: This subplot shows the P90 (90th percentile error) of each method under different matrix perturbation variances, representing the error level in the worst 10% of cases, reflecting the reliability of the algorithm in harsh scenarios. P90 is particularly sensitive to mismatch and covariance estimation errors. The rate of increase of the curve can intuitively show the tail risk and robustness of different methods under strong perturbation and mismatch conditions.

[0085] With the variance of formation perturbation As the perturbation increases, the statistical errors of several methods generally show an upward trend and exhibit significant robustness stratification: Neural-MVDR consistently maintains the lowest error and the slowest increase across all three metrics, demonstrating the strongest robustness to array manifold mismatch; CBF is relatively less affected by manifold mismatch, with its error increasing slowly with variance, and is generally superior to the traditional MVDR series. In contrast, MVDR, MVDR-ES, and MVDR-CMT show significant degradation with increasing perturbation, especially with a significant amplification of tail errors at P90, reflecting sensitivity to covariance estimation errors and steering vector mismatch; MVDR-OAS improves in the medium perturbation range through contraction estimation, but is still inferior to CBF and Neural-MVDR overall, indicating that learning-driven covariance robust estimation solutions and array manifold optimization can effectively suppress systematic errors caused by mismatch.

[0086] The performance of the method was further analyzed using sea trial data measured by SWellEx-96. The azimuth history diagram of the full-time data processing is shown in Figure 4, and the comparison between the predicted trajectory and GPS is shown in Figure 5. In Figure 4, the magenta "*" indicates sparse sampling of the predicted peak, and the white dashed line represents the GPS trajectory. The Bartlett prediction results exhibit strong sidelobes in the symmetrical azimuth of the linear array. This can also be seen in the trajectories in Figure 5. Between 23:55 and 00:00, the sidelobes of the Bartlett prediction results are even higher than the target azimuth, leading to prediction errors. Strong interference also exists near the 60° azimuth. MVDR is significantly affected by sidelob interference near 60°. Compared to the classic linear Bartlett beamforming and diagonally loaded MVDR beamforming, the Neural-MVDR proposed in this invention produces a clearer trajectory in the azimuth history diagram, and the error statistics are significantly better than other methods. The average errors of the Bartlett, MVDR, and Neural-MVDR methods are 8.2°, 51.7°, and 4.6°, respectively. Neural-MVDR takes 0.3 seconds to execute a single round of optimization and inference on a GPU. For comparison, Bartlett and MVDR take 0.0075 seconds and 0.035 seconds respectively for a single sample computation, on an Intel Core i9-14900K (3.20GHz) platform. Although Neural-MVDR takes significantly longer than Bartlett and MVDR, this time is still sufficient for real-time computing requirements in practical engineering.

[0087] The performance of this invention is compared with other robust MVDR methods, including condition number-constrained focus loading MVDR-CondNum, covariance shrinkage MVDR-OAS, feature domain hard projection MVDR-ES, and array geometry robust MVDR-CMT. The statistical errors are 50.7°, 51.5°, 53.9°, and 45.8°, respectively. The performance improvement of the model-based robust adaptive method is very limited.

[0088] To further verify the DOA estimation performance of Neural-MVDR under array geometric perturbation, random biases with a mean of 0 and a variance of 2 were added to the east-west and north-south distances of each element in the hydrophone array. The azimuth history and DOA prediction results are shown in Figures 7 and 8. In Figure 7, the magenta "*" represents sparse sampling of the predicted peak, and the white dashed line represents the GPS trajectory. Under significant array mismatch, MVDR essentially fails, failing to accurately estimate the target azimuth. The Bartlett gain decreases significantly, while the sidelobes increase significantly. The Neural-MVDR method, however, shows no significant performance degradation and maintains a relatively clear trajectory. The average errors of the Bartlett, MVDR, and Neural-MVDR methods are 18.0°, 82.3°, and 4.0°, respectively.

[0089] Traditional Bartlett beamforming was performed using both the calibrated element spacing and the predicted element spacing, and the results are shown in Figure 9. It can be seen that, under the same back-end beamforming processing, the sidelobes in the predicted element azimuth history are significantly reduced, resulting in better DOA performance. The statistical errors are 18.0° and 10.1°, respectively, further demonstrating the effectiveness of the Neural-MVDR method in array geometry optimization. However, the effect of changing the back-end beamforming to Bartlett is still inferior to the results of directly using Neural-MVDR processing.

[0090] This application also provides a single-sample neural network beamforming system, the system comprising: an underwater acoustic signal acquisition module for acquiring underwater acoustic signals based on a hydrophone array; a data preprocessing module for performing Fourier transform on the underwater acoustic signals to form multi-frequency spectral data; a first construction module for constructing a closed-loop neural beamforming framework; a second construction module for constructing a sample covariance matrix based on the multi-frequency spectral data; a first calculation module for smoothly updating the sample covariance matrix using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix; and a feature extraction module for processing the empirical covariance matrix. The system employs a temporal convolutional network to extract cross-frequency covariance features and generate a feature domain interference suppression control quantity. A third construction module constructs a robust covariance matrix based on the feature domain interference suppression control quantity and the closed-loop neural beamforming framework. A second calculation module performs beamforming on the robust covariance matrix to obtain a broadband azimuth spectrum. An adaptive correction module optimizes the broadband azimuth spectrum based on an array self-adaptation mechanism to update the array steering vector, resulting in an array adaptive correction result. An output module outputs the final broadband azimuth spectrum and the corresponding target arrival direction estimation result based on the array adaptive correction result.

[0091] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0092] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0093] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0094] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0095] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0096] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0097] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0098] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0099] This application provides a single-sample neural network beamforming method and related equipment for joint array configuration and target azimuth estimation. This scheme effectively enhances adaptability to complex interference sources in underwater acoustic environments by introducing a closed-loop neural beamforming framework and combining it with feature-domain interference suppression control. Especially under conditions with only a single sample, utilizing a temporal convolutional network to mine cross-frequency characteristic relationships in the frequency domain covariance matrix maintains high azimuth estimation accuracy even with wide bandwidth and low signal-to-noise ratio, significantly improving the system's robustness and practicality. Using the exponentially weighted moving average (EWMA) algorithm to smoothly update the sample covariance matrix effectively suppresses the instability in covariance estimation caused by the scarcity of samples. Based on this, a robust covariance matrix is ​​constructed by introducing feature-domain control, making the beamforming process more robust and enabling proactive suppression of spatially correlated interference sources, thereby obtaining a clearer broadband azimuth spectrum. After obtaining the preliminary broadband azimuth spectrum, an array self-adaptation mechanism is used to adaptively update the array steering vector, helping to correct direction estimation deviations caused by array distortion or actual array configuration errors, thus outputting a more accurate target arrival direction. This closed-loop adaptive feedback mechanism enables the system to continuously optimize, making it suitable for underwater target detection and localization tasks in dynamic marine environments.

[0100] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0101] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0104] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification 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 so that the embodiments of this application described herein can be implemented in orders other than those illustrated or 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 apparatus 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 apparatus.

[0105] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A single-sample neural network beamforming method for joint array formation and target azimuth estimation, characterized in that, The method includes the following steps: acquiring underwater acoustic signals based on a hydrophone array; performing Fourier transform on the underwater acoustic signals to form multi-frequency spectral data; constructing a closed-loop neural beamforming framework; constructing a sample covariance matrix based on the multi-frequency spectral data; smoothing and updating the sample covariance matrix using an exponentially weighted moving average algorithm to obtain an empirical covariance matrix; inputting the empirical covariance matrix into a temporal convolutional network to extract cross-frequency covariance features and generate a feature domain interference suppression control quantity; constructing a robust covariance matrix based on the feature domain interference suppression control quantity and the closed-loop neural beamforming framework; performing beamforming on the robust covariance matrix to obtain a broadband azimuth spectrum; optimizing the broadband azimuth spectrum based on an array adaptive mechanism to update the array steering vector and form an array adaptive correction result; and outputting the final broadband azimuth spectrum and the corresponding target arrival direction estimation result based on the array adaptive correction result.

2. The method according to claim 1, characterized in that, The closed-loop neural beamforming framework includes a robust covariance construction layer and a beamforming output layer. The robust covariance construction layer is used to generate a robust covariance matrix for broadband beam optimization based on the sample covariance matrix of the multi-frequency spectral data and the constraint information provided by the feature domain interference suppression control. The beamforming output layer is used to perform broadband beamforming based on the robust covariance matrix.

3. The method according to claim 1, characterized in that, The step of using an exponentially weighted moving average algorithm to smoothly update the sample covariance matrix to obtain the empirical covariance matrix includes: at each snapshot, obtaining the sample covariance matrix of the current frame, using the empirical covariance matrix of the previous frame as the historical covariance matrix; weightedly fusing the sample covariance matrix obtained in the current snapshot with the historical covariance matrix, and using an exponentially weighted moving average algorithm to complete the smooth update, thereby generating the empirical covariance matrix corresponding to the current snapshot.

4. The method according to claim 3, characterized in that, The formula for the empirical covariance matrix is ​​as follows: ;in, Indicates the first The empirical covariance matrix of the frame; Indicates the previous time corresponding to the first The empirical covariance matrix of the frame; This represents the decay factor of the exponentially weighted moving average; This indicates the weight of the current snapshot outer product term; Indicates the first Frame, frequency point Array observation vectors; express The conjugate transpose of; This indicates the current snapshot frequency. The quick snapshot covariance matrix.

5. The method according to claim 1, characterized in that, The feature domain interference suppression control quantity includes eigenvalue soft suppression control quantity, isotropic noise contraction control quantity, and diagonal loading control quantity.

6. The method according to claim 5, characterized in that, The construction of a robust covariance matrix based on the feature domain interference suppression control and the closed-loop neural beamforming framework includes: in the closed-loop neural beamforming framework, performing soft suppression processing on the principal feature subspace of the sample covariance matrix based on the eigenvalue soft suppression control to obtain eigenvalue soft suppression components; in the closed-loop neural beamforming framework, performing weighted fusion on the sample covariance matrix based on the isotropic noise contraction control to obtain isotropic contraction coefficients; in the closed-loop neural beamforming framework, performing loading correction on the diagonal terms of the sample covariance matrix based on the diagonal loading control to obtain diagonal loading factors; and structurally fusing the eigenvalue soft suppression components, the isotropic contraction coefficients, and the diagonal loading factors to form a robust covariance matrix for broadband beamforming.

7. The method according to claim 6, characterized in that, The formula for the robust covariance matrix is ​​as follows: ;in, Represents a robust covariance matrix; The eigenvalue soft suppression component is represented by M; M represents the number of array elements. Indicates the isotropic contraction coefficient; This represents the total energy of the covariance matrix; Represents the identity matrix; This represents the diagonal loading factor.

8. The method according to claim 1, characterized in that, The formula for calculating the broadband azimuth spectrum is as follows: ;in, Indicates the broadband azimuth spectrum value; Indicates at the angle of incidence Frequency The array guide vector below; express The conjugate transpose of; Representing the robust covariance matrix The inverse matrix.

9. The method according to claim 1, characterized in that, The array self-adaptation mechanism is built on the Rayleigh distribution criterion and optimizes the array by shifting array elements, delaying channels, and rotating the array globally.

10. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; when said at least one program is executed by said at least one processor, said at least one processor implements the method as described in any one of claims 1 to 9.

Images