Active sonar streaming method based on physical sensing and GPU acceleration

Abstract

This application provides an active sonar streaming processing method based on physical sensing and GPU acceleration, comprising: acquiring temporal beam data of the current frame in the current beam direction; constructing a physical sensing weighted matrix; constructing a physical sensing weighted robust principal component analysis (PCA) optimization model based on the PCA weighted matrix; inputting the temporal beam data into the PCA weighted robust PCA optimization model and solving it using a GPU-accelerated randomized singular value decomposition (RSD) method to obtain a sparse target matrix; and determining the azimuth history data result of the current frame based on the sparse target matrix. The method proposed in this application utilizes the low-rank sparse separation characteristics of weighted robust principal component analysis, combined with the physical sensing weighted matrix, to accurately identify and remove reverberant backgrounds with strong time-frequency correlations, effectively suppress random noise interference, and significantly recover submerged target signals under low signal-to-mixing ratio conditions.

Description

Technical Field

[0001] This application relates to the field of underwater acoustic signal processing technology, and in particular to an active sonar streaming method based on physical sensing and GPU acceleration. Background Technology

[0002] With the advancement of maritime rights protection and the construction of underwater defense systems, active sonar, as a core means of underwater target detection, faces increasingly stringent performance requirements. However, in shallow waters or areas with complex topography, signals emitted by active sonar generate strong reverberation interference after reflection from the sea surface, seabed, and uneven water bodies. Reverberation exhibits time-frequency characteristics similar to the emitted signal and is often far more energetic than weak target echoes. This leads to a significant performance degradation of traditional matched filter (MF) and constant false alarm rate (CFAR) detectors under low signal-to-clutter ratio (SCR) conditions, making target features easily masked by reverberation. In recent years, Robust Principal Component Analysis (RPCA) has been introduced into underwater acoustic signal processing for reverberation suppression due to its excellent low-rank and sparsity separation capabilities. Its basic principle is to utilize the strong correlation (low-rank characteristics) of the reverberant background in the time-frequency or beam domain and the sparsity characteristics of moving target echoes, separating the two through optimization algorithms.

[0003] However, existing RPCA-based sonar signal processing methods still face the following severe challenges in practical engineering applications: (1) Lack of physical mechanism constraints, resulting in limited detection performance. Traditional RPCA models typically only utilize general L1 norm sparsity constraints, which fall under the category of "blind separation." They ignore the unique physical properties of active sonar echoes (such as the specific time-frequency shape of linear frequency modulated signals, spatial correlation between beams, instantaneous frequency stability of phase, etc.). In extremely low signal-to-noise ratio environments, simple mathematical sparsity constraints can easily misjudge random noise as targets (false alarms) or filter out weak target signals as noise (missed alarms). (2) Extremely high computational complexity, making it difficult to meet real-time requirements. The solution process of the RPCA algorithm relies on iterative singular value decomposition (SVD), and its computational complexity increases cubically with the increase of matrix dimension. For high-dimensional time-frequency data generated by high-resolution array sonar, traditional CPU serial computing architecture cannot complete the processing in a short time and can usually only perform offline analysis, which cannot meet the "streaming" and "real-time" requirements of tactical sonar for target detection. (3) Weak parameter adaptation capability. Existing methods mostly use fixed regularization parameters, which cannot adapt to dynamically changing marine environmental noise levels, resulting in insufficient robustness under non-steady sea states.

[0004] Therefore, how to improve the separation accuracy of weak targets by deeply integrating the physical prior knowledge of sonar echoes while ensuring the processing speed of the algorithm to adapt to streaming data is a technical problem that urgently needs to be solved in the field of underwater acoustic signal processing. Summary of the Invention

[0005] This application provides an active sonar streaming processing method based on physical sensing and GPU acceleration. To solve the above-mentioned technical problems, this application adopts the following technical methods: In a first aspect, this application provides an active sonar streaming processing method based on physical sensing and GPU acceleration, including: Obtain the temporal beam data of the current frame in the current beam direction; Construct a weighted matrix of physical perception; Based on the aforementioned physical perception weighted matrix, a physical perception weighted robust principal component analysis optimization model is constructed. The time-domain beam data is input into the physical sensing weighted robust principal component analysis optimization model, and solved using a GPU-accelerated randomized singular value decomposition method to obtain a sparse target matrix. Based on the sparse target matrix, the orientation history map data result of the current frame is determined.

[0006] Optionally, the process of determining the time-domain beam data includes the following steps: Acquire multi-channel raw signals received by the sonar array; Based on the multi-channel raw signals, construct the sub-array observation matrix within the current processing frequency band; Obtain the optimal weight vector; Based on the subarray observation matrix and the optimal weight vector, the temporal beam data of the current frame in the current beam direction is determined.

[0007] Optionally, constructing the subarray observation matrix within the current processing frequency band based on the multi-channel raw signals includes: After segmenting and cutting the original signal of the channel, multi-channel time-domain frame data is obtained; Perform a Fast Fourier Transform on the multi-channel time-domain frame data to obtain full-band frequency-domain data; For the full-band frequency domain data, extract the frequency point data within the effective signal frequency band and construct the sub-array observation matrix within the current processing frequency band.

[0008] Optionally, determining the temporal beam data of the current frame in the current beam direction based on the subarray observation matrix and the optimal weight vector includes: Based on the subarray observation matrix and the optimal weight vector, the beam response vector within the effective frequency band is obtained; Construct a full-band spectrum vector with all initial vectors set to zero; The beam response vector is mapped and filled into the full-band spectrum vector to generate a filled spectrum vector; The filled spectrum vector is sequentially subjected to inverse fast Fourier transform and truncation to obtain the time-domain beam data of the current frame in the current beam direction.

[0009] Optionally, the process of constructing the physical perception weighting matrix includes the following steps: Construct signal morphology feature weights, spatial correlation feature weights, statistical distribution feature weights, phase stability feature weights, and local contrast feature weights; A physical perception weighted matrix is ​​constructed based on the signal morphology feature weight, the spatial correlation feature weight, the statistical distribution feature weight, the phase stability feature weight, and the local contrast feature weight.

[0010] Optionally, the step of inputting the time-domain beam data into the physically-sensing weighted robust principal component analysis optimization model and solving it using a GPU-accelerated randomized singular value decomposition method to obtain a sparse target matrix includes: Step S201: Use short-time Fourier transform to construct the time-domain beam data to generate a time-frequency domain complex observation matrix; Step S202: Determine whether the current frame corresponding to the time-frequency domain complex observation matrix is ​​the first frame; If so, initialize the low-rank background matrix, sparse target matrix, and Lagrange multiplier matrix to zero, and initialize the penalty parameter to be greater than zero. Step S203: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the sparse target matrix, calculate the low-rank background auxiliary matrix; Step S204: Using a GPU-accelerated randomized singular value decomposition method, perform singular value thresholding on the low-rank background auxiliary matrix to update the low-rank background matrix; Step S205: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the updated low-rank background matrix, calculate the sparse target auxiliary matrix; Step S206: Using the physical perception weighting matrix, perform a weighted soft thresholding operation on the sparse target auxiliary matrix to update the sparse target matrix; Step S207: Update the Lagrange multiplier matrix and the penalty parameter; Step S208: Determine whether the current iteration count has reached the maximum iteration count or whether the reconstruction error is less than the convergence tolerance; If so, output the sparse target matrix; If not, proceed to steps S202-S208.

[0011] Optionally, if the determination result of step S202 is negative, the following operations are performed: The low-rank background matrix and sparse target matrix obtained from the convergence of the previous frame are used as priors.

[0012] Optionally, determining the azimuth time map data result of the current frame based on the sparse target matrix includes: Perform inverse short-time Fourier transform and matched filtering on the sparse target matrix to output the azimuth history map data of the current frame.

[0013] Secondly, this application also provides a computer system, comprising: Memory is used to store instructions that can be executed by the processor; A processor for executing the instructions to implement the method as described in any of the first aspects.

[0014] Thirdly, this application also provides a computer-readable medium storing computer program code that, when executed by a processor, implements the method as described in any of the first aspects.

[0015] This application has the following beneficial effects: The method proposed in this application utilizes the low-rank sparse separation characteristics of weighted robust principal component analysis, combined with a physically-perceived weighted matrix, to accurately identify and remove reverberant backgrounds with strong time-frequency correlations, effectively suppress random noise interference, and significantly recover submerged target signals under low signal-to-mixing ratio conditions. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the active sonar streaming processing method based on physical perception and GPU acceleration provided in this application embodiment; Figure 2 This is a schematic diagram of the results provided in the embodiments of this application; Figure 2 (a) is a schematic diagram of the results based on the classic CPU serial RPCA sparse decomposition method. Figure 2 (b) is a schematic diagram of the result of the method proposed in this application. Detailed Implementation

[0017] To facilitate understanding by those skilled in the art, the present application will be further described below in conjunction with embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present application.

[0018] To solve the above technical problems, such as Figure 1 As shown, this application proposes an active sonar streaming processing method based on physical sensing and GPU acceleration, including: Step S101: Obtain the temporal beam data of the current frame in the current beam direction; All steps in this application are performed in a CPU-GPU heterogeneous computing environment, with the CPU handling data acquisition and control, and the GPU handling intensive computation. To meet the real-time processing requirements of massive sonar data, the system adopts a streaming overlapping frame segmentation strategy: The CPU acquires the multi-channel raw signals received by the sonar array, then converts the raw signals into single-precision floating-point format, and segments them sequentially to extract the t-th frame data, thus obtaining the multi-channel time-domain frame data. This step sets the single-frame time length to [value missing]. (corresponding physical duration) The frame shift step size is (corresponding physical duration) ),satisfy There is a length of [missing information] between two adjacent frames. The overlapping area.

[0019] Multi-channel temporal frame data is transferred to the GPU memory using a high-speed bus. Subsequent processing of the multi-channel temporal frame data, as well as steps S102 to S105, are all executed within the GPU based on the current frame data.

[0020] The number of execution points for multi-channel time-domain frame data is The Fast Fourier Transform (FFT) is used to obtain full-band frequency domain data. Then, based on the frequency and bandwidth characteristics of the active sonar transmitted signal, frequency point data within the effective signal band are extracted to construct the sub-array observation matrix for the current processing frequency band. The specific construction process is as follows: For the frequency domain snapshot data after framing, at each processing frequency point, the full array data is divided into... Multiple overlapping subarrays; the data from each subarray are rearranged into a subarray observation matrix within the current processing frequency band. ,in The number of frequency points involved in the calculation; using the formula Calculate the sampling covariance matrix, where This indicates the conjugate transpose.

[0021] Perform iterative diagonal loading based on white noise gain constraints for each preset beam scanning direction. Calculate the corresponding subarray steering vector The optimal weight vector is then solved iteratively according to the following logic. : (a) Initialization: Set the initial diagonal load. Set the increment step for the load. Set the maximum number of iterations. ; (b) Solving for the temporary weight vector: In the k-th iteration, construct the loaded covariance matrix. Where I is the identity matrix; calculate the weight vector. And after normalization, we obtain ; (c) Robustness constraint decision: Calculate the white noise gain corresponding to the current weight vector. ; (d) Iterative update: If WNG is lower than the preset robustness threshold and If the limit is not reached, update the diagonal load. If the current condition is not met, return to step (b) to continue the calculation; otherwise, determine the current condition. The optimal weight vector .

[0022] Based on the above subarray observation matrix and optimal weight vector The frequency domain subarray data is converted into a time domain beam sequence through the following calculations: Using the conjugate of the optimal weight vector Subarray observation matrix within the current processing frequency band By performing a weighted summation, the beam response vector within the effective frequency band is obtained. : (1) in, Represents the complex conjugate operation; Build length is Full-band spectrum vector The vector is initialized to all zeros; the set of frequency indices of the effective signal in the full frequency band is determined according to the preset signal frequency range. The calculated effective frequency band beam response vector Mapping input The corresponding index position Generate a filled spectrum vector The remaining frequency points are kept at zero to restore the original sampling rate characteristics of the signal.

[0023] For the filled spectrum vector Execution points are The inverse fast Fourier transform is used to obtain the time-domain transform result; based on the preset frame length... Extract the first part of the time-domain transform result The data points are used as the final time-domain beam sequence for the current beam direction in the t-th frame.

[0024] Step S102: Construct the physical perception weighted matrix; Establish a weighted matrix for physical perception The calculation formula is as follows: (2) in Represents the Hadamard product. , (i=1,2,3) are the preset balance coefficients, and the feature weights of each component are calculated as follows: Signal morphological feature weights : (3) in, The amplitude spectrum of the current sparse matrix. The two-dimensional convolution kernel is constructed based on the time-frequency distribution of the transmitted signal. Represents two-dimensional convolution. It is a tiny positive number; Spatial correlation feature weights : (4) (5) in, This represents the amplitude spectrum of the k-th beam. , , These are spatial weighting coefficients; Statistical distribution characteristic weights : (6) (7) in, For the mean filter kernel, These are parameters that are adaptively adjusted based on the global noise floor. Local contrast feature weights : (8) (9) in, , These are the local inner window kernel and the outer window kernel, respectively. , These are the adjustment parameters for the Sigmoid function; Phase stability feature weights : (10) (11) Where j is, The phase matrix of the observed data, For the first-order difference operator along the time axis, is the sensitivity adjustment coefficient, and stdfilt is the local standard deviation filtering operation.

[0025] Step S103: Based on the physical perception weighted matrix, construct a physical perception weighted robust principal component analysis optimization model; The time-domain beam data is converted into a time-frequency domain complex observation matrix D using the Short-Time Fourier Transform (STFT). This is based on a physically sensed weighting matrix. A physical perception-weighted robust principal component analysis (RPCA) optimization model is constructed, with the goal of decomposing the time-frequency domain complex observation matrix D into a low-rank background matrix L and a sparse target matrix S. The objective function of the physical perception-weighted RPCA optimization model is defined as follows: (12) (13) in, Denotes the nuclear norm of a low-rank matrix L. Describing the L1 norm, This is the regularization parameter.

[0026] By constructing a physical perception weighted matrix that incorporates five dimensions of features—signal morphology, spatial correlation, statistical distribution, local contrast, and phase stability—the prior physical knowledge of active sonar echoes is introduced into the weighted robust principal component analysis model. This matrix acts as a spatial adaptive regularization constraint during mathematical optimization, achieving a leap from purely data-driven blind separation to physical mechanism-guided separation. It effectively resolves the contradiction between suppressing non-stationary clutter interference in strong reverberation backgrounds and preserving weak target features in low signal-to-mixing ratio scenarios. Specifically, its advantages are reflected in four aspects: First, it significantly reduces the false alarm rate and overcomes "false sparsity" interference, compared to existing technologies. Traditional sparse recovery algorithms, which rely solely on signal amplitude sparsity for separation and are prone to misidentifying high-energy, short-term, non-stationary reverberation spikes as targets, are addressed in this invention by applying differentiated penalty weights to time-frequency units. This significantly penalizes amplitude-sparse but phase-divergent and waveform-mismatched reverberation clutter, forcing the algorithm to zero out or classify it into low-rank background during iterative solving. This eliminates "pseudo-sparse" reverberation interference at the physical mechanism level, greatly reducing the false alarm rate. Secondly, it enables directional protection of weak targets and improves low signal-to-mixing ratio detection capabilities. A threshold contraction strategy based on spatial variation is implemented through a physical sensing weighting matrix, targeting targets that conform to the physical characteristics of active sonar echoes. The signal components are assigned minimal weights, significantly reducing the truncation threshold at that position in the soft threshold operator. This allows weak target signals, whose energy is submerged in background noise, to be fully preserved in the sparse matrix S through a lower threshold, effectively solving the target miss problem caused by excessively high thresholds under low signal-to-mixing conditions in traditional methods, and improving the system's detection probability. Thirdly, it enhances environmental robustness and solves the statistical model mismatch problem. Traditional constant false alarm rate (CFAR) detection relies on specific assumptions about the statistical distribution of background noise, which are prone to failure in the heavy-tailed reverberation of complex shallow marine environments. This invention, however, integrates multi-dimensional physical features, does not rely on a single statistical distribution assumption, and is based on the relationship between the target and the environment. The algorithm makes judgments based on deterministic differences in physical properties. Regardless of different seabed materials such as sandy bottoms or rocky bottoms, or different water depths, as long as the physical continuity of the target echo exists, the algorithm can maintain stable separation performance and has extremely strong environmental adaptability. Fourth, it improves the convergence speed and reconstruction accuracy of the algorithm. From the perspective of optimization theory, the physical perception weighted matrix introduced in this invention constitutes a reweighted norm minimization problem. Compared with the standard norm, it provides a more accurate gradient descent direction, which can approach the optimal solution faster with less observation data. Combined with the hot start strategy, it significantly reduces the number of iterations and the reconstruction error, providing theoretical support for the real-time application of the algorithm in engineering.

[0027] Step S104: Input the time-domain beam data into the physical sensing weighted robust principal component analysis optimization model, and solve it using a GPU-accelerated randomized singular value decomposition method to obtain a sparse target matrix; A combined Augmented Lagrange Multiplier Method (ALM) and streaming hot-start strategy are employed for the solution. This strategy uses the convergence result of the previous frame as the initial value for the current frame, allowing the optimization algorithm to skip the initial exploration phase and start directly within the neighborhood of the optimal solution. This dual approach of reducing the time per iteration and decreasing the total number of iterations significantly improves the overall processing speed of the algorithm, enabling the previously offline-only high-precision RPCA algorithm to perform real-time streaming processing of massive amounts of sonar data. The specific solution process includes the following steps: Step S201: Use short-time Fourier transform to construct the time-domain beam data to generate a time-frequency domain complex observation matrix; This step uses short-time Fourier transform to construct the time-domain beam data, thereby generating a time-frequency domain complex observation matrix, which facilitates subsequent calculations.

[0028] Step S202: Determine whether the current frame corresponding to the time-frequency domain complex observation matrix is ​​the first frame; If so, initialize the low-rank background matrix, sparse target matrix, and Lagrange multiplier matrix to zero, and initialize the penalty parameter to be greater than zero. Initialization is required here. First, determine the current frame corresponding to the time-frequency domain complex observation matrix to see if it is the first frame (t=1). Then, set the low-rank background matrix accordingly. sparse target matrix and Lagrange multiplier matrix All are initialized to zero matrices, with penalty parameters... For a very small constant greater than 0 (e.g.) ).

[0029] If the current frame is not the first frame but a subsequent frame (t>1), then the converged low-rank background matrix of the previous frame is used as the basis for solving. sparse target matrix and Lagrange multiplier matrix As a priori, the penalty parameter ,in A relaxation coefficient greater than 1 (e.g.) Specifically, the overlapping data segment (of length t-1) at the end of the result of frame t-1 is... The data is shifted forward along the time axis to form the beginning of the initial values ​​for frame t; the newly added data in frame t is initialized to zero. This step utilizes the temporal correlation of the sonar background field, which can significantly accelerate convergence.

[0030] Step S203: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the sparse target matrix, calculate the low-rank background auxiliary matrix; In the k-th iteration, based on the penalty parameter Lagrange multiplier matrix Time-frequency domain complex observation matrix Sparse target matrix The low-rank background auxiliary matrix is ​​calculated. As shown in the following formula: (14) Step S204: Using a GPU-accelerated randomized singular value decomposition method, perform singular value thresholding on the low-rank background auxiliary matrix to update the low-rank background matrix; The low-rank background auxiliary matrix is ​​calculated. Then, GPU-accelerated randomized singular value decomposition (rSVD) was used to... Perform a singular value thresholding (SVT) operation to update the low-rank matrix. As shown in the following formula: (15) in, For singular value thresholding operators, The specific implementation of rSVD is as follows: generating a Gaussian random matrix. Calculate projection Perform QR decomposition on Z to obtain the orthogonal basis Q, and calculate the small-scale core matrix. Economic SVD decomposition of B yields , , Ultimately reconstructed , .

[0031] Step S205: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the updated low-rank background matrix, calculate the sparse target auxiliary matrix; According to the penalty parameters Lagrange multiplier matrix Time-frequency domain complex observation matrix and the updated low-rank background matrix The sparse target auxiliary matrix is ​​calculated. As shown in the following formula: (16) Step S206: Using the physical perception weighting matrix, perform a weighted soft thresholding operation on the sparse target auxiliary matrix to update the sparse target matrix; This step mainly utilizes the constructed physical perception weighted matrix. ,right Perform a weighted soft thresholding operation to update the sparse target matrix. : (17) in, For the weighted soft thresholding operator, for each element in the matrix and their corresponding thresholds The more specific calculation formula is as follows: (18) in, For regularization parameters; Step S207: Update the Lagrange multiplier matrix and the penalty parameter; Based on the current reconstruction error H, update the Lagrange multiplier matrix and penalty parameter as shown in the following equation: (19) (20) (twenty one) in, To penalize the parameter growth step size, This is the upper limit of the penalty parameter; Step S208: Determine whether the current iteration count has reached the maximum iteration count or whether the reconstruction error is less than the convergence tolerance; If so, output the sparse target matrix; If not, proceed to steps S202-S208.

[0032] Determine if the current iteration count has reached the maximum iteration count. Or whether the reconstruction error is less than the convergence tolerance. If any of the above conditions are met, then the sparse target matrix is ​​output. If the conditions are not met, repeat steps S202-S208.

[0033] Step S104 migrates computationally intensive matrix operations to the GPU for massively parallel processing. In dynamically changing shallow sea environments, the traditional frame-by-frame independent "cold start" processing method is prone to causing jumps or flickering in the estimation of the low-rank background matrix on the time axis. The streaming warm start strategy of this application implicitly introduces a smooth constraint in the time dimension. By passing the low-rank subspace structure of the previous frame, it ensures a smooth transition of background estimation between consecutive frames. This not only suppresses the fluctuation of the solution caused by random initialization of the algorithm, but also significantly enhances the continuity and stability of the system in tracking weak target trajectories under non-stationary reverberation conditions.

[0034] In addition, this application reduces the average number of convergence iterations from the conventional 40-50 to less than 15 through streaming warm-start, significantly reducing the total floating-point operations and memory read / write frequency of the GPU. With the same hardware power consumption and configuration, this architecture significantly improves the system's data throughput, enabling it to support broadband sonar signal processing with higher sampling rates and more beam channels, or, while meeting real-time performance requirements, freeing up more computing resources for subsequent advanced target recognition tasks, greatly improving the overall performance ratio of engineering equipment. Step S105: Based on the sparse target matrix, determine the orientation history map data result of the current frame.

[0035] The sparse target matrix obtained by decomposition is subjected to inverse short-time Fourier transform (ISTFT) and matched filtering to output the azimuth history map data result for frame t. The CPU control module receives the results returned by the GPU and uses the overlap addition method to synthesize the current frame result with the historical results in the overlapping area to output continuous azimuth history map data.

[0036] Validation Analysis: Comparative experiments were conducted on the multi-channel raw signals received by the same sonar array using the classical CPU serial RPCA sparse decomposition method and the method proposed in this application. The experimental results are shown in the figure below. Figure 2 As shown in Table 1, the experimental hardware and computation time are as follows: Table 1 ; from Figure 2 (a) It can be seen that the classical RPCA sparse decomposition method judges sparsity only based on amplitude. Shallow sea reverberation contains a large number of non-stationary strong spike signals, which are also mathematically sparse. Therefore, they are misjudged as targets and preserved by traditional algorithms. As a result, there are large areas of high-energy residue in the azimuth history map, and a large number of random noise points are scattered throughout the space. The target edges are blurred and almost submerged by the background. Figure 2 (b) It can be seen that, due to the introduction of weighted multiple physical sensing features in the algorithm, even if there is high-energy reverberation and noise in the signal, but the phase is disordered and the waveform is mismatched, the algorithm identifies that these do not conform to physical laws and assigns a large penalty weight through the physical sensing weighting matrix. At the same time, the algorithm detects that the target has a very high degree of physical sensing feature matching, thus reducing the threshold at that location. Therefore, the background of the azimuth history map is almost pure dark blue and low-energy, and the originally stubborn bottom reverberation and noise are removed. At the same time, the target is clear and is not broken or lost due to strong denoising.

[0037] In summary, the method proposed in this application utilizes the low-rank sparse separation characteristics of Weighted Robust Principal Component Analysis (RPCA) combined with a physically-perceived weighted matrix to accurately identify and remove reverberant background (low-rank components) with strong time-frequency correlation, effectively suppressing random noise interference and significantly recovering submerged target signals under low signal-to-mixing ratio (SCR) conditions. Furthermore, the CUDA-accelerated randomized singular value decomposition (rSVD) algorithm reduces the computational complexity of traditional RPCA's high-dimensional matrix operations from cubic to near-linear levels, significantly reducing computation time and overcoming the bottleneck of complex sparse recovery algorithms being unable to meet the real-time processing requirements of massive sonar data.

[0038] In some embodiments, this application also provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0039] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.

[0040] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.

[0041] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.

Claims

1. An active sonar streaming processing method based on physical sensing and GPU acceleration, characterized in that, include: Obtain the temporal beam data of the current frame in the current beam direction; Construct a weighted matrix of physical perception; Based on the aforementioned physical perception weighted matrix, a physical perception weighted robust principal component analysis optimization model is constructed. The time-domain beam data is input into the physical sensing weighted robust principal component analysis optimization model, and solved using a GPU-accelerated randomized singular value decomposition method to obtain a sparse target matrix. Based on the sparse target matrix, the orientation history map data result of the current frame is determined.

2. The method according to claim 1, characterized in that, The process of determining the time-domain beam data includes the following steps: Acquire multi-channel raw signals received by the sonar array; Based on the multi-channel raw signals, construct the sub-array observation matrix within the current processing frequency band; Obtain the optimal weight vector; Based on the subarray observation matrix and the optimal weight vector, the temporal beam data of the current frame in the current beam direction is determined.

3. The method according to claim 2, characterized in that, The construction of the sub-array observation matrix within the current processing frequency band based on the multi-channel raw signals includes: After segmenting and cutting the original multi-channel signal, multi-channel time-domain frame data is obtained; Perform a Fast Fourier Transform on the multi-channel time-domain frame data to obtain full-band frequency-domain data; For the full-band frequency domain data, extract the frequency point data within the effective signal frequency band and construct the sub-array observation matrix within the current processing frequency band.

4. The method according to claim 3, characterized in that, The step of determining the temporal beam data of the current frame in the current beam direction based on the subarray observation matrix and the optimal weight vector includes: Based on the subarray observation matrix and the optimal weight vector, the beam response vector within the effective frequency band is obtained; Construct a full-band spectrum vector with all initial vectors set to zero; The beam response vector is mapped and filled into the full-band spectrum vector to generate a filled spectrum vector; The filled spectrum vector is sequentially subjected to inverse fast Fourier transform and truncation to obtain the time-domain beam data of the current frame in the current beam direction.

5. The method according to claim 1, characterized in that, The construction process of the physical perception weighted matrix includes the following steps: Construct signal morphology feature weights, spatial correlation feature weights, statistical distribution feature weights, phase stability feature weights, and local contrast feature weights; A physical perception weighted matrix is ​​constructed based on the signal morphology feature weight, the spatial correlation feature weight, the statistical distribution feature weight, the phase stability feature weight, and the local contrast feature weight.

6. The method according to claim 1, characterized in that, The step involves inputting the time-domain beam data into the physically-sensing weighted robust principal component analysis optimization model, and solving it using a GPU-accelerated randomized singular value decomposition method to obtain a sparse target matrix; including: Step S201: Use short-time Fourier transform to construct the time-domain beam data to generate a time-frequency domain complex observation matrix; Step S202: Determine whether the current frame corresponding to the time-frequency domain complex observation matrix is ​​the first frame; If so, initialize the low-rank background matrix, sparse target matrix, and Lagrange multiplier matrix to zero, and initialize the penalty parameter to be greater than zero. Step S203: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the sparse target matrix, calculate the low-rank background auxiliary matrix; Step S204: Using a GPU-accelerated randomized singular value decomposition method, perform singular value thresholding on the low-rank background auxiliary matrix to update the low-rank background matrix; Step S205: Based on the penalty parameter, the Lagrange multiplier matrix, the time-frequency domain complex observation matrix, and the updated low-rank background matrix, calculate the sparse target auxiliary matrix; Step S206: Using the physical perception weighting matrix, perform a weighted soft thresholding operation on the sparse target auxiliary matrix to update the sparse target matrix; Step S207: Update the Lagrange multiplier matrix and the penalty parameter; Step S208: Determine whether the current iteration count has reached the maximum iteration count or whether the reconstruction error is less than the convergence tolerance; If so, output the sparse target matrix; If not, proceed to steps S202-S208.

7. The method according to claim 6, characterized in that, If the judgment result of step S202 is negative, the following operations are performed: The low-rank background matrix and sparse target matrix obtained from the convergence of the previous frame are used as priors.

8. The method according to claim 1, characterized in that, The determination of the azimuth time map data result for the current frame based on the sparse target matrix includes: Perform inverse short-time Fourier transform and matched filtering on the sparse target matrix to output the azimuth history map data of the current frame.

9. A computer system, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing the instructions to implement the method as described in any one of claims 1 to 8.

10. A computer-readable medium, characterized in that, The system stores computer program code that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images