A Non-stationary Noise Suppression Method for Vector Hydrophones Based on Deep Unfolded Networks
By combining deep unfolded networks with multi-channel signal processing of vector hydrophones, non-stationary noise suppression in complex marine environments is achieved, solving the tracking lag and signal coherence problems of traditional methods, and improving the extraction accuracy and adaptability of target signals.
Patent Information
- Application Number
- CN202611316224.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-28
AI Technical Summary
Existing technologies are difficult to adapt to non-stationary noise in complex marine environments. Traditional adaptive filtering schemes suffer from tracking lag and insufficient gain in low signal-to-noise ratio environments. Deep learning schemes lack physical constraints and are prone to destroying signal coherence, which limits the detection range and recognition accuracy of underwater acoustic equipment.
A noise suppression method based on deep unfolded networks is adopted. Through multi-channel signal acquisition and synchronization of vector hydrophones, preprocessing, iterative denoising of deep unfolded networks, and physical manifold residual diversion, non-stationary noise and target signal are decoupled and separated, thus ensuring the spatial coherence of the signal.
It improves the target signal extraction gain and scene adaptation capability in low signal-to-noise ratio environments, and ensures the spatial coherence and signal processing accuracy of multi-channel signals.
Smart Images

Figure CN122835538B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater acoustic signal processing technology, and more specifically, to a method for suppressing non-stationary noise in vector hydrophones based on a depth unfolded network. Background Technology
[0002] Underwater acoustic detection is a core technological support for marine resource exploration, underwater target monitoring, and underwater acoustic communication. With the upgrading of the marine equipment industry, vector hydrophones have become the mainstream hardware architecture due to their multi-dimensional sound field synchronous sensing capabilities. Noise suppression is the core link in vector acoustic signal processing, which directly determines the accuracy of subsequent signal analysis and is a core and essential direction for improving the performance of underwater acoustic equipment.
[0003] Currently, the noise suppression technologies deployed on a large scale in the industry mainly fall into two categories. The first is adaptive filtering solutions, which construct adaptive filters based on the minimum mean square error criterion and achieve noise cancellation by iteratively updating the filter weights. This solution has low computational overhead, mature hardware deployment, and stable engineering performance in stable noise scenarios. The second is end-to-end deep learning denoising solutions, which fit the mapping relationship between noisy and clean signals using deep neural networks. These solutions have better fitting capabilities for complex noise patterns than traditional algorithms and offer superior adaptability to various noise types.
[0004] Existing technologies still face common technical bottlenecks in engineering applications, making them difficult to adapt to non-stationary noise scenarios in complex marine environments. Firstly, traditional adaptive filtering schemes rely on the assumption of noise stationarity, resulting in inherent lag in tracking time-varying non-stationary noise. In low signal-to-noise ratio environments, the denoising gain is insufficient, failing to effectively extract weak target signals. Secondly, purely data-driven deep learning schemes lack vector acoustic physical constraints, and the denoising process easily disrupts the spatial coherence of multi-channel signals, leading to decreased accuracy in subsequent processing such as azimuth estimation and target recognition. These problems directly limit the detection range and recognition accuracy of underwater acoustic equipment in complex sea areas, restricting its adaptability to various scenarios. Therefore, we propose a non-stationary noise suppression method for vector hydrophones based on deep unfolded networks. Summary of the Invention
[0005] This invention addresses the technical problems in existing technologies, such as tracking lag in non-stationary noise, insufficient denoising gain in low signal-to-noise ratio environments, and the lack of physical constraints and easy destruction of spatial coherence of vector signals in pure data-driven denoising schemes. It provides a non-stationary noise suppression method for vector hydrophones based on deep unfolded networks. This method integrates a deep unfolded iterative denoising framework with a vector acoustic physical constraint mechanism. Through closed-loop dynamic updating of noise covariance and residual diversion of target propagation physical manifold, it can efficiently suppress non-stationary noise while ensuring the spatial coherence of multi-channel signals, thereby improving the extraction accuracy and scene adaptability of target signals in complex underwater acoustic environments.
[0006] To address the aforementioned technical problems, this invention provides a method for suppressing non-stationary noise in vector hydrophones based on deep unfolded networks, comprising the following steps: S1. Multi-channel signal acquisition and synchronization of vector hydrophone: The original underwater sound field signal is acquired by the sound pressure sensing unit and the multi-dimensional vibration velocity sensing unit of the vector hydrophone. The time synchronization correction of the multi-channel original sound signal is performed to obtain the synchronized original multi-channel vector sound signal containing non-stationary noise. S2. Preprocessing of the signal to be processed and construction of network input: Perform channel amplitude and phase calibration and acoustic scale normalization on the synchronous raw multi-channel vector acoustic signal, extract multi-dimensional complex time-frequency acoustic features, and construct standardized feature data; S3. Non-stationary noise separation processing based on deep unfolded network: Standardized feature data is input into deep unfolded network for iterative denoising. The target propagation physical manifold is dynamically constructed based on the recovered target signal. After physical splitting of the network residual, a safe noise covariance is constructed. Non-stationary noise and target signal are decoupled and separated through non-stationary noise covariance closed-loop state feedback. The denoised target signal feature data is output. S4. Vector acoustic signal reconstruction and output: Combining the multi-channel acoustic physical characteristics of the vector hydrophone, the time-frequency inverse transformation and time-domain reconstruction are performed on the denoised target signal feature data to output the target vector acoustic signal after non-stationary noise suppression.
[0007] As a further improvement to this technical solution, the multi-channel signal acquisition and synchronization of the vector hydrophone in S1 includes the following steps: S1.1 The original acoustic signal of the underwater sound field is acquired in parallel by the sound pressure sensing unit and the multi-dimensional vibration velocity sensing unit of the vector hydrophone, and the original time-domain acoustic signal of each channel is output respectively. S1.2. Using a unified hardware trigger clock as the time reference, the sampling start time of the original sound signals of each channel is aligned to eliminate sampling start deviation between channels; S1.3 Based on the aligned time reference, timestamp matching and time delay interpolation correction are performed on the sampling points of the original acoustic signals of each channel to eliminate the transmission delay deviation between channels; S1.4 Integrate the time-synchronized sound pressure channel and the multi-dimensional vibration velocity channel signals to obtain a synchronous original multi-channel vector sound signal containing non-stationary noise.
[0008] As a further improvement to this technical solution, the signal preprocessing and network input construction in S2 include the following steps: S2.1 Based on the preset channel calibration coefficients, amplitude gain correction and phase shift correction are performed on each channel of the synchronous original multi-channel vector acoustic signal to eliminate amplitude and phase inconsistencies between channels. S2.2 Based on the acoustic impedance parameters of the water medium, pressure equivalent scale conversion is performed on the vibration velocity channel signals to complete the acoustic scale normalization processing of the multi-channel signals. S2.3 Perform short-time Fourier transform on the multi-channel vector acoustic signal that has completed amplitude and phase calibration and acoustic scale normalization to obtain the complex time-frequency coefficients corresponding to each channel, thus forming a multidimensional complex time-frequency acoustic feature. S2.4 Perform dimensional regularization and numerical normalization on the multidimensional complex time-frequency acoustic features to generate standardized feature data that is adapted to the input format of the depth unfolded network.
[0009] As a further improvement to this technical solution, the non-stationary noise separation processing based on deep unfolded networks in S3 includes the following steps: S3.1 Load the non-stationary noise covariance matrix obtained from the previous frame update to complete the initial configuration of the deep unfolding network; S3.2 Input the standardized feature data into the deep unwrap network to perform multi-layer iterative operations and recover the target signal estimation result of the current frame; S3.3 Dynamically construct the target propagation physical manifold based on the target signal estimation results of the current frame; S3.4. Based on the target propagation physical manifold, the network residuals are physically split to obtain two types of residual components; S3.5 Construct a safety noise covariance matrix based on the residual components after splitting, update the non-stationary noise covariance matrix and feed it back to the next frame for processing, and obtain the denoised target signal feature data of the current frame.
[0010] As a further improvement to this technical solution, the initial configuration of the deep unfolded network in step S3.1, which involves loading the non-stationary noise covariance matrix updated in the previous frame, includes the following steps: S3.11. Read the non-stationary noise covariance matrix of the previous frame's processing output. When processing the first frame, read the preset initial noise covariance matrix. S3.12. Map the non-stationary noise covariance matrix to the fidelity term weighting coefficients of the deep unfolded network to complete the initial configuration of the network iteration rules.
[0011] As a further improvement to this technical solution, step S3.2, which involves inputting standardized feature data into a deep unrolled network to perform multi-layer iterative operations and recovering the target signal estimation result for the current frame, includes the following steps: S3.21. Based on the non-stationary noise covariance matrix loaded and initialized in S3.1, use its inverse matrix as the weighting coefficients for the fidelity terms to construct the gradient update rules for the deep unfolded network. S3.22. Perform multi-level iterative operations using the proximal gradient descent framework, update signal features layer by layer, and output the target signal estimation result for the current frame.
[0012] As a further improvement to this technical solution, the dynamic construction of the target propagation physical manifold based on the target signal estimation result of the current frame in S3.3 includes the following steps: S3.31. Based on the target signal estimation results of the current frame, calculate the cross-correlation between the sound pressure channel and the complex time-frequency coefficients of each vibration velocity channel within the local time-frequency window; S3.32. Extract the target propagation direction vector based on the cross-correlation quantity, and construct a normalized target propagation guidance vector; S3.33. Generate a physical projection matrix based on the target propagation guidance vector to form a target propagation physical manifold.
[0013] As a further improvement to this technical solution, the physical splitting of network residuals based on the target propagation physical manifold in S3.4 to obtain two types of residual components includes the following steps: S3.41. Calculate the difference between the standardized feature data observed in the current frame and the target signal estimation result to obtain the network residual; S3.42. The network residuals are orthogonally projected and decomposed using the physical projection matrix to obtain the parallel components that conform to the target propagation law and the perpendicular components that do not conform to the target propagation law. The calculation relationship is as follows: ; ; In the formula, For network residuals, For the physical projection matrix, It is the identity matrix. For parallel components, The vertical component is represented by the above variables, which are all complex vectors / matrices in the complex time-frequency domain.
[0014] As a further improvement to this technical solution, the steps in S3.5, which involve constructing a secure noise covariance matrix based on the residual components after splitting, updating the non-stationary noise covariance matrix and feeding it back to the next frame for processing, and outputting the denoised target signal feature data, include the following: S3.51. Based on the target signal estimation result of the current frame, and combined with the target propagation consistency of the current frame and the short-term continuity of the target propagation direction between adjacent frames, calculate the target physical credibility factor. The target physical credibility factor has a value range of 0 to 1. S3.52. Using the target physical reliability factor as the weight, perform a weighted fusion of the outer product of the conjugate transposes of the two types of residual components to generate the security noise covariance matrix. The calculation relationship is as follows: ; In the formula, For safety noise covariance, For the target physical credibility factor, This is the conjugate transpose of the parallel components. The conjugate transpose of the vertical component; superscript This represents the conjugate transpose operation; S3.53. A first-order recursive update method is adopted, with the safety noise covariance matrix as the update innovation quantity, and the non-stationary noise covariance matrix is updated frame by frame. S3.54. Save the updated non-stationary noise covariance matrix as the initialization input for the depth unrolling network in the next frame, and output the denoised target signal feature data of the current frame.
[0015] As a further improvement to this technical solution, the vector acoustic signal reconstruction and output in S4 includes the following steps: S4.1 Perform short-time inverse Fourier transform on each channel of the denoised target signal feature data to obtain the time-domain initial signal corresponding to each channel; S4.2 Based on the multi-channel acoustic physical characteristics of the vector hydrophone, the phase consistency of the initial time domain signal of each channel is corrected to maintain the spatial coherence of the multi-channel signal. S4.3 Integrate the multi-channel time-domain signals that have completed phase correction by channel dimension, complete time-domain reconstruction, and output the target vector acoustic signal after non-stationary noise suppression.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention adopts a deep unfolded network iterative denoising architecture, combined with a non-stationary noise covariance closed-loop frame-by-frame update mechanism, which breaks through the limitations of traditional algorithms that rely on the assumption of noise stationarity, enhances the dynamic adaptation capability to time-varying non-stationary noise, and effectively improves the target signal extraction gain in low signal-to-noise ratio environments. 2. This invention introduces a physical manifold to guide the physical splitting of residuals during target propagation, and constructs a safe noise covariance by weighting physical reliability. It strictly follows the physical laws of vector acoustics during noise suppression, effectively preserves the spatial coherence of multi-channel signals, and ensures the accuracy and stability of subsequent signal processing. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the overall process flow of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0019] like Figure 1 As shown, this embodiment provides a method for suppressing non-stationary noise in vector hydrophones based on a deep unfolded network. This method is suitable for enhancing and recovering multi-channel acoustic signals from vector hydrophones in non-stationary noise environments during underwater target detection. The method can be implemented in a hardware environment consisting of a vector hydrophone sound pressure and multi-dimensional vibration velocity sensing unit, a multi-channel synchronous acquisition module, and a signal processing platform. Standardized complex time-frequency characteristic data is obtained by performing amplitude-phase calibration and time-frequency transformation on the synchronously acquired raw signals. Its core processing mechanism is as follows: using a deep unfolded network as the iterative denoising agent, in each frame processing, the network fidelity term is weighted using the non-stationary noise covariance matrix estimated in the previous frame, and the target propagation physical manifold is dynamically constructed based on the estimation results of the current frame. Based on this physical manifold, the network residuals are orthogonally projected and decomposed to generate a safe noise covariance and recursively update the noise statistics, forming a closed-loop state feedback of the non-stationary noise covariance, thereby achieving decoupling and separation of the target signal and non-stationary noise. The steps are linked by the frame-by-frame transfer of the noise covariance matrix, enabling signal acquisition, preprocessing, iterative separation, and signal reconstruction to work collaboratively. The specific steps of this method are described in detail below: S1. Multi-channel signal acquisition and synchronization of vector hydrophone: The raw underwater sound field signal is acquired through the sound pressure sensing unit and multi-dimensional vibration velocity sensing unit of the vector hydrophone. The time synchronization correction of the multi-channel raw sound signal is performed to obtain a synchronized raw multi-channel vector sound signal containing non-stationary noise. Specifically, it includes the following steps: S1.1 The original acoustic signal of the underwater sound field is acquired in parallel by the sound pressure sensing unit and the multi-dimensional vibration velocity sensing unit of the vector hydrophone, and the original time-domain acoustic signal of each channel is output respectively. Specifically, the sound pressure sensing unit of the vector hydrophone and three spatially orthogonally arranged particle velocity sensing units operate in parallel underwater. Each sensing unit independently converts the sensed sound field physical quantities into analog electrical signals. These analog electrical signals are then conditioned by a preamplifier and an anti-aliasing low-pass filter before being sent to the input channels of the multi-channel synchronous data acquisition module. The data acquisition module is internally equipped with a unified hardware trigger clock. All channels' analog-to-digital converters synchronously start sampling under the control of the same trigger edge of this clock, with each channel operating at the same sampling rate. Perform equal-interval sampling. After sampling, the sound pressure channel outputs the original digital sequence in the time domain. The three vibration velocity channels output the original digital sequence in the time domain. ,in These are the sampling point numbers. These raw time-domain signals contain both the signal components contributed by the target sound source and the non-stationary noise interference components from the environment.
[0020] S1.2. Using a unified hardware trigger clock as the time reference, the sampling start time of the original acoustic signal of each channel is aligned to eliminate sampling start deviation between channels. Specifically, the trigger edge of the unified hardware trigger clock is used as the common time reference for each channel. The time corresponding to the first sampling point in the data stream of the sound pressure channel and each vibration velocity channel is uniformly calibrated to zero time, so that the signals of each channel are macroscopically aligned at the beginning position of the sample sequence. This eliminates the sampling start deviation between channels caused by differences in the transmission path of the trigger signal and differences in device response, and provides a consistent time reference origin for subsequent fine time delay correction.
[0021] S1.3 Based on the aligned time reference, timestamp matching and time delay interpolation correction are performed on the sampling points of the original acoustic signals of each channel to eliminate the transmission delay deviation between channels; Specifically, after completing the initial time alignment, this step performs fine-grained time delay correction on each channel signal to compensate for differences in analog conditioning links, transmission cable lengths, and inter-channel group delay deviations accumulated during analog-to-digital conversion. The sound pressure level channel is selected as the time reference channel, with a sampling period of [missing information]. , No. Each sampling point corresponds to an absolute time. For each vibration velocity channel (as indicated by the channel subscript) (Distinguishing between them), the fixed transmission delay deviation relative to the sound pressure reference channel is obtained in advance through offline calibration. The deviation is measured in seconds and can be positive or negative. A positive value indicates that the vibration velocity channel signal lags behind the reference channel, while a negative value indicates that it leads.
[0022] Let the first The original sampling sequence of the vibration velocity channel is After time delay correction, the channel is in the 1st... signal value at each sampling time Calculated by band-limited interpolation: ; in, Indicates the number after time delay correction Vibration velocity channel in the first The signal value at each sampling point; Indicates the first The first sampling sequence of the vibration velocity channel Each sample value; The normalized sinc function is defined as follows: ,in It is the independent variable of the function; This refers to the sampling point number of the output sequence after time delay correction; The sampling period is measured in seconds. For the first The fixed transmission delay deviation of the vibration velocity channel relative to the sound pressure reference channel, in seconds; The sampling point number of the original sampling sequence is used as the summation variable.
[0023] This interpolation process utilizes the ideal low-pass characteristic of the sinc function to reconstruct the value of a continuous-time signal at any given moment from the original discrete samples without distortion, while satisfying the sampling theorem. This is achieved by setting the interpolation time to... This allows the sampled value of the vibration velocity channel signal to be moved from the original sampling time to a time that is completely consistent with the reference channel.
[0024] In practical engineering implementation, the aforementioned infinitely long sinc interpolation kernel is truncated using a finite-length window to control the computational load. Specifically, a window length of [missing information] can be used. ( The interpolation kernel is truncated (where the value is a positive integer, chosen based on the required interpolation precision), and a smoothing window function (such as a Hamming window or Hanning window) is applied to suppress the Gibbs effect caused by truncation. After truncation and windowing, the summation range becomes... to Each channel signal is interpolated and mapped onto a unified time grid. This process effectively compensates for transmission delay deviations on the order of sub-sampling points between channels, ensuring that each channel sampling point precisely corresponds to the same absolute time.
[0025] The offline calibration can be achieved as follows: In an anechoic pool or free-field environment, place the vector hydrophone in the far field of a standard sound source at a known orientation. Using the sound pressure channel as a reference, perform cross-correlation calculations on the received signals of each vibration velocity channel and the sound pressure channel. Read the relative time delay value between each channel from the peak position of the cross-correlation function. Take multiple measurements with the sound source orientation changed and average them to obtain the fixed transmission time delay deviation of each vibration velocity channel relative to the sound pressure channel. This calibration process can be completed before the system is put into use, and the obtained deviation parameters are directly called as known prior parameters in the embodiments. This fixed delay calibration is suitable for scenarios with slowly varying channel delays, and can compensate for delay drift introduced by factors such as temperature drift and device aging through periodic offline recalibration.
[0026] S1.4 Integrate the time-synchronized sound pressure channel and the multi-dimensional vibration velocity channel signals to obtain a synchronous original multi-channel vector sound signal containing non-stationary noise.
[0027] Specifically, the sound pressure channel signal after S1.3 time delay interpolation correction... (The sound pressure channel serves as the reference channel, with zero time delay deviation, and its sequence remains unchanged, i.e.) (and the three time-delay corrected vibration velocity channel signals) , , Arranged and combined along the channel dimension, a four-channel synchronous raw vector acoustic signal is formed: ; In the formula: Indicates the first The synchronous raw multi-channel vector acoustic signal at each sampling point is a four-dimensional column vector. Indicates the delay correction of the first The sound pressure channel signal value at each sampling point; , , These represent the time delay correction values for the [number]th [unit]. The signal values of the three orthogonal vibration velocity channels at each sampling point.
[0028] This vector signal The signals are used as input to subsequent signal preprocessing modules, with frames as the unit. Since S1.2 and S1.3 have already completed sampling start time alignment and sub-sampling point level transmission delay correction, the signals of the four channels are precisely synchronized in the time dimension, with each channel sharing the same index. The sample values at each point correspond to the underwater acoustic field information at the same physical moment. This strict time alignment condition provides a reliable data synchronization foundation for the subsequent multi-channel amplitude and phase calibration and complex time-frequency feature extraction in S2, as well as for the decoupling and separation of target signals and noise based on the propagation physical manifold in S3.
[0029] S2. Signal Preprocessing and Network Input Construction: Channel amplitude and phase calibration and acoustic scale normalization are performed on the synchronous raw multi-channel vector acoustic signal to extract multi-dimensional complex time-frequency acoustic features and construct standardized feature data. Specifically, this includes the following steps: S2.1 Based on the preset channel calibration coefficients, amplitude gain correction and phase shift correction are performed on each channel of the synchronous original multi-channel vector acoustic signal to eliminate amplitude and phase inconsistencies between channels. Specifically, let the synchronous original multi-channel vector acoustic signal output by S1 be... Indexed by channel Corresponding to the sound pressure channel and the three vibration velocity channels respectively, for the first The channel has a preset channel calibration coefficient that is complex. Its model The gain is for amplitude compensation and is dimensionless; its argument is... This is the phase compensation amount, in radians. This calibration focuses on equivalent amplitude and phase correction within the operating frequency band of interest, and is performed point-by-point in the time domain. ; In the formula: Indicates the first The channel is in The signal values of each sampling point after amplitude and phase calibration; , The imaginary unit; Indicates the first The channel is in The signal value before calibration at each sampling point, i.e. The Each component.
[0030] Calibration coefficient Obtained through offline calibration: The vector hydrophone is placed in a standard sound field. Using the sound pressure channel as a reference, the calibration coefficients for each channel are determined by the amplitude ratio and phase difference of the transfer function between each velocity channel and the sound pressure channel within the frequency band of interest. After calibration, each coefficient is stored as a priori parameters and directly called upon during actual operation. The calibrated multi-channel vector acoustic signal is denoted as: ; In the formula, Indicates the first The multi-channel vector acoustic signal after amplitude and phase calibration at each sampling point is a four-dimensional column vector. The signal value after calibration of the sound pressure channel, i.e. time ; , , These are the calibrated signal values for the three vibration velocity channels, respectively. time .
[0031] S2.2 Based on the acoustic impedance parameters of the water medium, pressure equivalent scale conversion is performed on the vibration velocity channel signals to complete the acoustic scale normalization processing of the multi-channel signals. Specifically, the calibrated first Vibration velocity channel signal Multiply by the characteristic impedance of the water medium The equivalent sound pressure signal is obtained. : ; In the formula: Indicates the first Vibration velocity channel in the first The equivalent sound pressure signal value at each sampling point, in Pascals; This indicates the density of the water medium, expressed in kilograms per cubic meter. The speed of sound in water is expressed in meters per second. This represents the characteristic impedance of water, measured in Rayles.
[0032] The sound pressure channel signal is preserved as The normalized multi-channel vector acoustic signal is denoted as: ; In the formula: Indicates the first The multi-channel vector acoustic signal after acoustic scaling of each sampling point is a four-dimensional column vector. This represents the normalized signal value from the sound pressure channel. ; , , These are the equivalent sound pressure signal values of the three vibration velocity channels after pressure equivalence conversion, corresponding to... time .
[0033] S2.3 Perform short-time Fourier transform on the multi-channel vector acoustic signal that has completed amplitude and phase calibration and acoustic scale normalization to obtain the complex time-frequency coefficients corresponding to each channel, thus forming a multidimensional complex time-frequency acoustic feature. Specifically, for the first Normalized signal of the channel Let the frame length be Each sampling point, frame shift is sampling points ( The window function is . No. Frame, First Complex time-frequency coefficients at each frequency point for: ; In the formula: Indicates the first The channel is in Time frame, first Complex time-frequency coefficients at each frequency point; For time frame indexing, , Total number of frames; For frequency point index, .
[0034] After performing the above transformation on each of the four channels, the complex time-frequency coefficient matrices of each channel are stacked along the channel dimension to form a multidimensional complex time-frequency acoustic feature. For a given time frame... All four channels of this frame The complex time-frequency coefficients of each frequency point are integrated according to the rule of frequency point priority and channel following. 3D complex column vector , which serves as the input observation for the current frame of the deep unfolded network.
[0035] S2.4 Perform dimensional regularization and numerical normalization on the multidimensional complex time-frequency acoustic features to generate standardized feature data that is adapted to the input format of the depth unfolded network.
[0036] Specifically, the complex time-frequency coefficients of each channel and frequency point in S2.3 are arranged in the dimensional order required by the depth of the network input layer, forming a fixed-dimensional feature vector. For each dimensional component of the normalized feature vector, Z-score normalization is performed based on preset statistical parameters: the mean of each dimensional component is pre-calculated on the training set. and standard deviation ( (For feature dimension index), during processing, the first... The complex time-frequency coefficients of each dimension are subtracted by their real and imaginary parts respectively. Divide by The normalization parameter is adapted to underwater acoustic environments with characteristics similar to the training set. If the noise characteristics of the deployment environment differ significantly from those of the training set, the normalization parameter can be recalculated by collecting pure noise data from the deployment environment to ensure feature distribution matching. The normalized components of each dimension have zero mean and unit variance. After the above processing, standardized feature data in frames is generated and fed into S3 for iterative denoising.
[0037] S3. Non-stationary noise separation processing based on deep unfolded network: Standardized feature data is input into deep unfolded network for iterative denoising. The target propagation physical manifold is dynamically constructed based on the recovered target signal. After physical splitting of the network residual, a safe noise covariance is constructed. Non-stationary noise and target signal are decoupled and separated through non-stationary noise covariance closed-loop state feedback. The denoised target signal feature data is output. In this embodiment, S3 processes the standardized feature data output by S2 in frames. Simultaneously, this embodiment is designed based on a typical underwater detection scenario of a single far-field plane wave target, utilizing the spatial coherence of target acoustic propagation to achieve noise separation under physical constraints. For complex scenarios such as multiple targets and near-field scenarios, adaptation can be achieved by expanding the subspace dimension, while maintaining a consistent core processing flow.
[0038] Let the current frame number be... The normalized feature vector of this frame is Its dimensions are The sorting order is as follows: first by frequency index. Arranged from smallest to largest, the complex time-frequency coefficients of the sound pressure channel and the three vibration velocity channels are sequentially arranged at each frequency point, i.e. ; The overall processing flow of S3 is as follows: load the non-stationary noise covariance matrix updated from the previous frame, complete the depth unrolling network initialization; and... The input network undergoes multiple iterative denoising processes to obtain the target signal estimate for the current frame. ;based on The target propagation physical manifold is dynamically constructed, and the network residuals are orthogonally projected and decomposed. A secure noise covariance matrix is constructed and noise statistics are recursively updated. The updated noise covariance matrix is fed back to the next frame, and simultaneously output... This serves as the target signal feature data after denoising in the current frame. The following steps will be explained in detail: S3.1 Load the non-stationary noise covariance matrix obtained from the previous frame update to complete the initial configuration of the depth unrolling network; including the following steps: S3.11. Read the non-stationary noise covariance matrix of the previous frame's processing output. When processing the first frame, read the preset initial noise covariance matrix. In S3.11, let the current frame number be... ( ).when At this point, the noise covariance matrix from the previous frame is not yet available; therefore, the preset initial noise covariance matrix is read. . Based on the pre-defined system operating environment, one approach is as follows: Before system deployment, collect a segment of pure noise data under a typical non-stationary noise environment. Then, normalize the complex time-frequency feature vector of this data segment using the Z-score normalization parameter, which is identical to that used in S2.4. Calculate its sample covariance matrix as the initial noise covariance matrix, ensuring that the noise statistics and network input features are in the same numerical space. When At that time, read the first The updated non-stationary noise covariance matrix output in frame S3.5 The noise covariance matrix loaded in the current frame is uniformly denoted as... : ; In the formula: The non-stationary noise covariance matrix loaded in the current frame, with dimension [missing information]. Complex matrix; The preset initial noise covariance matrix is obtained from the environmental noise data before deployment. For the first Frame S3.5 updates the output non-stationary noise covariance matrix.
[0039] S3.12. Map the non-stationary noise covariance matrix to the fidelity term weighting coefficients of the deep unfolded network to complete the initial configuration of the network iteration rules.
[0040] In S3.12, calculate The inverse matrix is used as the weighting matrix for the fidelity terms. To avoid Inverse calculation is unstable near singularity; regularization is used to address this. ; In the formula: For the fidelity-preserving weighted matrix, the dimension is... Complex matrix; Let be the regularization parameter, and set it to . 10 of the average of the diagonal elements -6 times; for 3D identity matrix.
[0041] S3.2. The standardized feature data is input into the deep unrolled network to perform multi-layer iterative operations, recovering the target signal estimation result of the current frame. Specifically, the deep unrolled network maps each iteration of the near-end gradient descent algorithm to a layer of the network, and the signal is gradually refined through multi-layer stacking. In this embodiment, the network has a total of layer, The integer is a positive integer, and each layer performs one proximal gradient descent update. This includes the following steps: S3.21. Based on the non-stationary noise covariance matrix loaded and initialized in S3.1, use its inverse matrix as the weighting coefficients for the fidelity terms to construct the gradient update rules for the deep unfolded network. In S3.21, let the first... The input signal of the layer is estimated as follows (when hour, Set as ), No. The output signal of the layer is estimated as follows Each layer first updates the signal along the negative gradient direction of the fidelity term. The fidelity term is defined as... The gradient update rule is as follows: ; In the formula: For the first Intermediate variables obtained after the gradient descent step. 3D complex vector; For the first Target signal estimation from layer output, 3D complex vector; For the first The gradient update step size of the layer is a positive real number and a learnable parameter. This is the weighted matrix for the fidelity terms in S3.12; This is the normalized feature vector of the current frame.
[0042] S3.22. Perform multi-level iterative operations using the proximal gradient descent framework, update signal features layer by layer, and output the target signal estimation result for the current frame.
[0043] In S3.22, Through proximal operators A structured prior constraint is introduced for the target signal, which is the local smoothness of the target signal in the frequency domain, with the real and imaginary parts respectively satisfying this smoothness constraint. In one specific implementation, the proximal operator employs a residual block structure consisting of convolutional layers, batch normalization, and ReLU activation functions. First, the... Rearranged to size The tensor (four channels, each channel divided into real and imaginary components, frequency dimension length is) The real and imaginary parts participate in the convolution operation separately, without cross-component mixing; a one-dimensional convolution is performed on the frequency dimension (with a kernel size of 3 and the number of channels remaining unchanged), which is then passed through two convolutional layers. Each convolutional layer is followed by batch normalization and ReLU activation. The convolution output is added to the rearranged input tensor through an identity skip connection, and the result is restored to the original state. In the form of a complex vector, we obtain . No. A complete update of a layer is represented as: ; In the formula: For the first Target signal estimation from layer output, 3D complex vector.
[0044] The network executes sequentially. Layer iteration, the first The layer output is the final estimate of the target signal in the current frame: ; In the formula: For the first Target signal estimation results after frame denoising via a deep unfolding network 3D complex vector.
[0045] Step size of each layer Learnable parameters, such as convolution weights within the proximal operators, are obtained through end-to-end supervised training. During training, synthetic noisy data containing known clean target signals are used. The dataset covers typical underwater target signal morphologies and various types of non-stationary noise, with labels representing the standardized complex time-frequency features of the corresponding clean signals. The mean square error between the network output and the clean target signal is used as the loss function, and the network parameters are optimized through backpropagation. After training, the parameters are fixed, and inference is performed frame-by-frame.
[0046] S3.3. Dynamically construct the target propagation physical manifold based on the target signal estimation results of the current frame; specifically, the target propagation physical manifold is a signal subspace spanned by the target propagation steering vector and conforming to the propagation direction of the target sound source. This step utilizes... Extract the target propagation direction information and generate the corresponding physical projection matrix. This includes the following steps: S3.31. Based on the target signal estimation results of the current frame, calculate the cross-correlation between the sound pressure channel and the complex time-frequency coefficients of each vibration velocity channel within the local frequency window; In S3.31, Converted into complex time-frequency coefficients according to the aforementioned arrangement rules: The... Channel, First The target signal estimation complex time-frequency coefficients at frequency points are denoted as follows: ,in 0 corresponds to the sound pressure channel. Each corresponds to one of the three vibration velocity channels.
[0047] Defined in the frequency dimension as the first Local frequency window centered on a frequency point The window contains the first Each frequency point and each of its two sides There are 2Q+1 adjacent frequency points, totaling 2Q+1 frequency points. It is a positive integer. The value is selected based on the signal bandwidth and frequency resolution; the wider the bandwidth and the higher the frequency resolution, the better. The value can be increased accordingly, typically ranging from 2 to 5. When Being at the spectral boundary leads to insufficient neighboring points on one side. At that time, only the actual frequency points are used, that is Calculate the cross-correlation between the sound pressure channel and the estimated complex time-frequency coefficients of each vibration velocity channel within the window: ; In the formula: For the first Frame, First The sound pressure channel at the frequency point and the first Cross-correlation of vibration velocity channels correspond direction; For the sound pressure channel inside the window Estimating the complex time-frequency coefficients of the target signal at a frequency point; For the first The vibration velocity channel is inside the window. Estimating the complex time-frequency coefficients of the target signal at a frequency point; superscript This indicates the complex conjugate operation.
[0048] S3.32. Extract the target propagation direction vector based on the cross-correlation quantity, and construct a normalized target propagation guidance vector; In S3.32, the target propagation direction vector is composed of three cross-correlation quantities. : ; In the formula: For the first Frame, First The target propagation direction vector at a frequency point, a three-dimensional complex vector; , , Corresponding to time .
[0049] calculate of norm .when Greater than the preset threshold At that time, construct a normalized target propagation steering vector. : ; when If the direction estimation at the current frequency point is unreliable, the steering vector of the previous frame at that frequency point should be used. Substitute; if If the corresponding point in the previous frame is also unavailable, then set the guide vector to the default omnidirectional uniform vector. The single-frame fallback bias, after first-order recursive smoothing, will not significantly affect the noise statistics results. If the default vector fallback is triggered for more than three consecutive frames at the same frequency point, the noise covariance sub-block corresponding to that frequency point will pause updates and use historical valid statistical values to avoid continuous contamination of noise statistics due to directional estimation distortion. (The replaced text appears to be incomplete and requires further context.) It still serves as the guiding vector for the current frame in the construction of subsequent projection matrices and is saved to the buffer for use in the next frame. Threshold Based on the noise level setting, the physical meaning is the minimum confidence threshold for direction estimation. The lower the signal-to-noise ratio (SNR), the higher the threshold should be to filter out unreliable estimates. For example, in typical low to medium SNR scenarios, it should be set to 10. -3 .
[0050] In the formula: For the first Frame, First The normalized target propagation steering vector at a frequency point is a four-dimensional complex vector that satisfies... Its first element corresponds to the sound pressure channel.
[0051] S3.33. Generate a physical projection matrix based on the target propagation guidance vector to form a target propagation physical manifold.
[0052] In S3.33, the normalized guide vector Constructing the physical projection matrix : ; because Normalized, denominator However, this embodiment adopts the above form in order to maintain consistency with the general definition of the physical projection matrix.
[0053] In the formula: For the first Frame, First The physical projection matrix at a frequency point is a 4×4 complex matrix, whose column space is the target propagation physical manifold at that frequency point, and it satisfies idempotency. ; superscript This indicates the conjugate transpose operation.
[0054] S3.4. Based on the target propagation physical manifold, perform physical splitting of the network residuals to obtain two types of residual components; including the following steps: S3.41. Calculate the difference between the standardized feature data observed in the current frame and the target signal estimation result to obtain the network residual; In S3.41, the current frame observation feature vector is... With target signal estimation Subtracting them yields the network residual. : In the formula: For the first Network residuals of frames 3D complex vector.
[0055] S3.42. The network residuals are orthogonally projected and decomposed using the physical projection matrix to obtain the parallel components that conform to the target propagation law and the perpendicular components that do not conform to the target propagation law. The calculation relationship is as follows: ; ; In the formula, For network residuals, For the physical projection matrix, It is the identity matrix. For parallel components, The vertical component is represented by the above variables, which are all complex vectors / matrices in the complex time-frequency domain.
[0056] Specifically, for the first Each frequency point, from The residual components of the four channels are extracted according to the arrangement rules to form a four-dimensional residual vector. Use the projection matrix for that frequency point in S3.33. Perform orthogonal projection decomposition on it: ; ; In the formula: For the first Frame, First The network residual vector at frequency points is a four-dimensional complex vector. The physical projection matrix constructed for S3.33; It is a 4×4 identity matrix; As a parallel component, located on the physical manifold of target propagation, it is a four-dimensional complex vector representing the part of the residual that conforms to the target propagation law; The vertical component is orthogonal to the target propagation physical manifold. It is a four-dimensional complex vector representing the part of the residual that does not conform to the target propagation law.
[0057] For all After performing the above decomposition on each frequency point, the parallel and vertical components of each frequency point are reassembled according to the channel-frequency arrangement order in S2.3 to obtain the parallel component vector of the entire frequency band. and vertical component vector Both are A complex vector. Where, Because it is orthogonal to the propagation law of the target, it can be reliably attributed to non-stationary noise; It may contain components of the target signal leakage and noise in the same direction of propagation.
[0058] S3.5. Construct a safety noise covariance matrix based on the residual components after splitting, update the non-stationary noise covariance matrix, and feed it back to the next frame for processing to obtain the denoised target signal feature data of the current frame. This includes the following steps: S3.51. Based on the target signal estimation result of the current frame, and combined with the target propagation consistency of the current frame and the short-term continuity of the target propagation direction between adjacent frames, calculate the target physical credibility factor. The target physical credibility factor has a value range of 0 to 1. In S3.51, the target physical credibility factor It is synthesized by weighted sum of two factors.
[0059] The first term is the target propagation consistency factor in the current frame. This is used to measure the consistency of the steering vectors at various frequency points in the current frame. When At that time, directly set ;when At that time, calculate all of the current frame The average value of the inner product modulus between each pair of steering vectors at each frequency point: ; In the formula: For the first The current frame target propagation consistency factor of the frame, with a value range of [0,1]; The first Normalized steering vectors at different frequency points in a frame.
[0060] The second term is the short-time continuity factor of the target propagation direction between adjacent frames. , used to measure the first Frame and the The stability of the target propagation direction between frames. When At that time, set ;when At that time, calculate the average value of the normalized steering vector inner product magnitude of all corresponding frequency points in two adjacent frames: ; In the formula: For the first The short-term continuity factor of the target propagation direction between adjacent frames, with a value range of [0,1]. For the first Frame number The guide vector at the frequency point.
[0061] The two factors are weighted and combined to form the target physical credibility factor. : ; In the formula: For the first The target physical credibility factor of the frame, with a value range of [0,1]; This is a weighting coefficient, ranging from [0,1], used to adjust the relative contribution of the current frame consistency factor and the inter-frame continuity factor. When the target motion is stable and noise fluctuations are small, this coefficient can be increased. The weight is chosen to enhance intra-frame consistency; when the target has high maneuverability and the noise has significant time-varying characteristics, the weight can be reduced. The values are chosen to enhance the continuity between frames; typical values are as follows: .
[0062] S3.52. Using the target physical reliability factor as the weight, perform a weighted fusion of the outer product of the conjugate transposes of the two types of residual components to generate the security noise covariance matrix. The calculation relationship is as follows: ; In the formula, For safety noise covariance, For the target physical credibility factor, This is the conjugate transpose of the parallel components. The conjugate transpose of the vertical component; superscript This represents the conjugate transpose operation; In S3.52, with As weights, the vertical components obtained from S3.4 Parallel components We construct the safety noise covariance matrix by weighted fusion of the conjugate transpose outer product. : ; In the formula: For the first The security noise covariance matrix of the frame. 3D complex matrix; This is the full-band vertical component vector output by S3.4, which is the residual part that does not conform to the target propagation law; This is the full-band parallel component vector output by S3.4, which is the residual part that conforms to the target propagation law; superscript This represents the conjugate transpose operation; The target physical credibility factor.
[0063] The meaning of this weighted fusion is: vertical component Always include the full contribution in the noise covariance estimation; parallel components Then The proportion included. When When the value is close to 1, the target signal estimation has high reliability, and the parallel components may contain fewer leaked target signal components, resulting in a lower inclusion ratio and thus protecting the target signal from false cancellation; when When the noise level is low, the reliability of the target signal estimation is low, and the parallel components may contain a lot of unmodeled noise, so the inclusion ratio should be increased accordingly. In scenarios with limited computing resources, the security noise covariance matrix can be constructed in a block diagonal form, retaining only the 4×4 cross-correlation blocks within each frequency point and ignoring inter-frequency correlations to reduce computational complexity. This scheme has negligible accuracy loss in scenarios with narrowband signals and weak inter-frequency correlations, and is suitable for embedded low-computing-power deployment scenarios; for broadband high-coherence scenarios, it is recommended to use the full matrix form to ensure denoising performance.
[0064] S3.53. A first-order recursive update method is adopted, using the safety noise covariance matrix as the update innovation quantity, to update the non-stationary noise covariance matrix frame by frame. This update method is suitable for slow time-varying non-stationary noise scenarios. For sudden strong pulses and other fast-change noise, the smoothing coefficient can be dynamically adjusted in conjunction with noise change detection.
[0065] In S3.53, the security noise covariance matrix generated in S3.52 is... To update the innovation amount of the current frame, perform a first-order recursive smoothing update on the non-stationary noise covariance matrix: ; In the formula: For the first The non-stationary noise covariance matrix after frame update. 3D complex matrix; The initial noise covariance matrix of the current frame loaded into S3.11 (i.e., the update result or initial value of the previous frame). It is a recursive smoothing coefficient with a value range of [0,1], used to control the fusion ratio between historical noise statistics and the current frame update. The larger the value, the smoother the noise covariance matrix update, and the slower the response to transient noise. It is suitable for weak non-stationary and slowly changing noise scenarios. The smaller the value, the more sensitive the update, but the estimation variance may increase. It is suitable for strongly non-stationary, rapidly changing noise scenarios. In practical applications, it can be selected according to the time-varying rate of the environmental noise. For typical stationary scenarios, it can be 0.95~0.98, and for strongly non-stationary scenarios, it can be 0.85~0.9.
[0066] S3.54. Save the updated non-stationary noise covariance matrix as the initialization input for the depth unrolling network in the next frame, and output the denoised target signal feature data of the current frame.
[0067] In S3.54, the non-stationary noise covariance matrix obtained from S3.53 is updated. Saved to the noise state buffer for later use. The data is loaded as input to S3.1 during frame processing. Simultaneously, the normalized guide vectors for all frequency points of the current frame are... Saved to the guide vector cache for use by the first... Frame S3.51 calculates the continuity factor This is used to form a cross-frame closed-loop state feedback. The target signal estimation result output from S3.2 is then used. The denoised target signal feature data of the current frame is output to S4 for inverse time-frequency transformation and time-domain reconstruction. S3's processing of the current frame is now complete, and the system proceeds to the next step. The S3 processing flow of frames.
[0068] S4. Vector Acoustic Signal Reconstruction and Output: Combining the multi-channel acoustic physical characteristics of the vector hydrophone, inverse time-frequency transform and time-domain reconstruction are performed on the denoised target signal feature data, outputting the target vector acoustic signal after non-stationary noise suppression. Specifically, the input of S4 is the denoised target signal feature data output frame by frame from S3. Let the total number of frames processed by S3 be... , No. The target signal estimation result of the frame is This vector is organized according to the channel-frequency arrangement determined in S2.3, that is, for the first... aisle( (corresponding to sound pressure, x-velocity, y-velocity, z-velocity respectively) The frequency point, its denoised complex time-frequency coefficients are denoted as The following steps involve performing temporal reconstruction and physical characteristic recovery on each channel. S4.1 Perform short-time inverse Fourier transform on each channel of the denoised target signal feature data to obtain the time-domain initial signal corresponding to each channel; Specifically, for the first Channel, all Complex time-frequency coefficients of a frame Perform a short-time inverse Fourier transform. The inverse transform uses the same frame length as S2.3. Frame shift and analysis window function And select a composite window function. To satisfy the overlapping and addition reconstruction condition. The specific process is as follows: For the first... Frame, first Performing an inverse discrete Fourier transform on the complex time-frequency coefficients at each frequency point yields a result of length [missing value]. Temporal frames : ; In the formula: Indicates the first Channel 1 The first frame after inverse transformation Individual sample point values; For the first Channel 1 Frame number The target signal complex time-frequency coefficients at each frequency point.
[0069] Then multiply the frame by the composite window. and shift by frame The values are accumulated into the reconstructed temporal buffer and then summed using an overlap-addition method. For indices belonging to the interval [0, N-1], the corresponding frame sample values are included in the summation; indices outside the interval are considered zero. This yields the result... Time-domain reconstructed signal of the channel : ; In the formula, only when At that time, the frame sample value and the window function value are non-zero. For any sampling time... The analysis window and the synthesis window satisfy the constant overlap and addition condition. This ensures complete reconstruction without processing. When using a Hamming or Hanning window and the frame shift is 50% of the frame length, this condition is approximately met, and the reconstruction amplitude flatness deviation is within an acceptable range for engineering applications.
[0070] After the above inverse transformation, the result is The initial time-domain signal of the sound pressure channel is expressed in Pascals. These are the initial time-domain equivalent sound pressure signals for the three vibration velocity channels (due to pressure equivalent scale conversion in S2.2).
[0071] S4.2 Based on the multi-channel acoustic physical characteristics of the vector hydrophone, the phase consistency of the initial time domain signal of each channel is corrected to maintain the spatial coherence of the multi-channel signal. Specifically, under plane wave far-field conditions, sound pressure and particle velocity are in phase in the propagation direction. After S3 denoising and S4.1 inverse transform, a slight relative phase shift between channels may be introduced, or boundary deviations may occur during the overlapping and addition process in the framing process, causing the channels to no longer strictly satisfy the above coherence relationship. This step utilizes the inherent sound pressure-velocity coherence characteristics of the vector hydrophone to perform phase consistency correction on the initial time-domain signals of each channel and restore the original physical dimensions of the velocity channel.
[0072] First, the equivalent sound pressure signal of the vibration velocity channel is restored to the dimension of particle velocity. This is because in S2.2, the vibration velocity signal is multiplied by the characteristic impedance of the water medium. To perform acoustic scale normalization, the equivalent sound pressure signals of each velocity channel obtained in S4.1 are divided by... : ; In the formula: Indicates the first The initial time-domain signal of the vibration velocity channel after restoring the vibration velocity dimension, in meters per second; The density of the water medium, This is the speed of sound in water.
[0073] The initial signal in the time domain of the sound pressure channel is directly denoted as: .
[0074] Then, using the sound pressure channel For reference, the three vibration velocity channels were analyzed respectively. Phase consistency correction is performed. One implementation method is adaptive relative time delay correction: the sound pressure channel and the... Cross-correlation function of vibration velocity channels The summation range is limited to the effective overlap interval of the two signals, that is: ; In the formula: The time delay variable is expressed in sampling periods. This represents the total length of the signal.
[0075] The time delay corresponding to the peak position of the cross-correlation function is taken as the relative time delay estimate for that channel. If the peak amplitude of the cross-correlation is lower than the preset signal-to-noise ratio threshold, or the main-sidelobe ratio is less than the threshold, the time delay estimation is deemed unreliable, and the correction amount of the previous effective segment is used to avoid introducing secondary phase deviation. Using the estimated time delay, the vibration velocity channel is then processed using the band-limited interpolation method described in S1.3. Fractional time delay compensation is performed to obtain the phase-corrected vibration velocity channel signal. When the relative motion between the target sound source and the vector hydrophone is slow, and the relative time delay between channels can be considered constant within the signal processing segment, the above-mentioned global time delay correction can effectively restore the in-phase coherence relationship between channels. If there is a significant time-varying deviation, the time delay estimation and correction can be performed frame by frame, or the relative phase information between channels in the target propagation steering vector estimated by S3 can be used in the frequency domain for frequency-by-frequency compensation before inverse transformation.
[0076] The corrected sound pressure channel signal is denoted as Thus, a four-channel signal with consistent phase and correct dimensions was obtained.
[0077] S4.3 Integrate the multi-channel time-domain signals that have completed phase correction by channel dimension, complete time-domain reconstruction, and output the target vector acoustic signal after non-stationary noise suppression.
[0078] Specifically, the four-channel time-domain signals after phase correction are arranged along the channel dimension to obtain the final output target vector acoustic signal. : ; In the formula: For the first The target vector acoustic signal after non-stationary noise suppression is output from each sampling point, and is a four-dimensional real vector. This is the sound pressure level channel output signal, measured in Pascals. , , These are three orthogonal vibration velocity channels outputting signals, in meters per second.
[0079] The output signal fully preserves the spatial information and correct physical dimensions of the multi-channel vector hydrophone, and can be used for subsequent target detection, orientation, recognition and other processing.
[0080] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.
[0081] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A method for non-stationary noise suppression of a vector hydrophone based on a deep unfolding network, characterized in that, Includes the following steps: S1. Multi-channel signal acquisition and synchronization of vector hydrophone: The original underwater sound field signal is acquired by the sound pressure sensing unit and the multi-dimensional vibration velocity sensing unit of the vector hydrophone. The time synchronization correction of the multi-channel original sound signal is performed to obtain the synchronized original multi-channel vector sound signal containing non-stationary noise. S2. Preprocessing of the signal to be processed and construction of network input: Perform channel amplitude and phase calibration and acoustic scale normalization on the synchronous raw multi-channel vector acoustic signal, extract multi-dimensional complex time-frequency acoustic features, and construct standardized feature data; S3. Non-stationary noise separation processing based on deep unfolded network: Standardized feature data is input into deep unfolded network for iterative denoising. The target propagation physical manifold is dynamically constructed based on the recovered target signal. After physical splitting of the network residual, a safe noise covariance is constructed. Non-stationary noise and target signal are decoupled and separated through non-stationary noise covariance closed-loop state feedback. The denoised target signal feature data is output. Includes the following steps: S3.1 Load the non-stationary noise covariance matrix obtained from the previous frame update to complete the initial configuration of the deep unfolding network; S3.2 Input the standardized feature data into the deep unwrap network to perform multi-layer iterative operations and recover the target signal estimation result of the current frame; S3.3 Dynamically construct the target propagation physical manifold based on the target signal estimation results of the current frame; S3.
4. Based on the target propagation physical manifold, perform physical splitting of the network residuals to obtain two types of residual components; including the following steps: S3.
41. Calculate the difference between the standardized feature data observed in the current frame and the target signal estimation result to obtain the network residual; S3.
42. The network residuals are orthogonally projected and decomposed using the physical projection matrix to obtain the parallel components that conform to the target propagation law and the perpendicular components that do not conform to the target propagation law. The calculation relationship is as follows: ; ; wherein is the network residual, is the physical projection matrix, is the identity matrix, is the parallel component, is the perpendicular component; S3.5 Construct a safety noise covariance matrix based on the residual components after splitting, update the non-stationary noise covariance matrix and feed it back to the next frame for processing, and obtain the target signal feature data after denoising in the current frame. Includes the following steps: S3.
51. Based on the target signal estimation result of the current frame, and combined with the target propagation consistency of the current frame and the short-term continuity of the target propagation direction between adjacent frames, calculate the target physical credibility factor. The target physical credibility factor has a value range of 0 to 1. S3.
52. Using the target physical reliability factor as the weight, perform a weighted fusion of the outer product of the conjugate transposes of the two types of residual components to generate the security noise covariance matrix. The calculation relationship is as follows: ; wherein is a safety noise covariance, is a target physical plausibility factor, is a conjugate transpose of the parallel component, is a conjugate transpose of the vertical component; superscript denotes a conjugate transpose operation; S3.
53. A first-order recursive update method is adopted, with the safety noise covariance matrix as the update innovation quantity, and the non-stationary noise covariance matrix is updated frame by frame. S3.
54. Save the updated non-stationary noise covariance matrix as the initialization input for the depth unrolling network in the next frame, and output the denoised target signal feature data of the current frame. S4. Vector acoustic signal reconstruction and output: Combining the multi-channel acoustic physical characteristics of the vector hydrophone, the time-frequency inverse transformation and time-domain reconstruction are performed on the denoised target signal feature data to output the target vector acoustic signal after non-stationary noise suppression.
2. The deep unfolding network based vector hydrophone non-stationary noise suppression method according to claim 1, characterized in that, The multi-channel signal acquisition and synchronization of the vector hydrophone in S1 includes the following steps: S1.1 The original acoustic signal of the underwater sound field is acquired in parallel by the sound pressure sensing unit and the multi-dimensional vibration velocity sensing unit of the vector hydrophone, and the original time-domain acoustic signal of each channel is output respectively. S1.
2. Using a unified hardware trigger clock as the time reference, the sampling start time of the original sound signals of each channel is aligned to eliminate sampling start deviation between channels; S1.3 Based on the aligned time reference, timestamp matching and time delay interpolation correction are performed on the sampling points of the original acoustic signals of each channel to eliminate the transmission delay deviation between channels; S1.4 Integrate the time-synchronized sound pressure channel and the multi-dimensional vibration velocity channel signals to obtain a synchronous original multi-channel vector sound signal containing non-stationary noise.
3. The deep unfolding network based vector hydrophone non-stationary noise suppression method according to claim 1, wherein, The signal preprocessing and network input construction in S2 include the following steps: S2.1 Based on the preset channel calibration coefficients, amplitude gain correction and phase shift correction are performed on each channel of the synchronous original multi-channel vector acoustic signal to eliminate amplitude and phase inconsistencies between channels. S2.2 Based on the acoustic impedance parameters of the water medium, pressure equivalent scale conversion is performed on the vibration velocity channel signals to complete the acoustic scale normalization processing of the multi-channel signals. S2.3 Perform short-time Fourier transform on the multi-channel vector acoustic signal that has completed amplitude and phase calibration and acoustic scale normalization to obtain the complex time-frequency coefficients corresponding to each channel, thus forming a multidimensional complex time-frequency acoustic feature. S2.4 Perform dimensional regularization and numerical normalization on the multidimensional complex time-frequency acoustic features to generate standardized feature data that is adapted to the input format of the depth unfolded network.
4. The deep unfolding network based vector hydrophone non-stationary noise suppression method of claim 1, wherein, The initial configuration of the deep unrolled network in step S3.1, which involves loading the non-stationary noise covariance matrix updated from the previous frame, includes the following steps: S3.
11. Read the non-stationary noise covariance matrix of the previous frame's processing output. When processing the first frame, read the preset initial noise covariance matrix. S3.
12. Map the non-stationary noise covariance matrix to the fidelity term weighting coefficients of the deep unfolded network to complete the initial configuration of the network iteration rules.
5. The method for suppressing non-stationary noise in vector hydrophones based on deep unfolded networks according to claim 1, characterized in that, The steps in S3.2, which involve inputting standardized feature data into a deep unwrap network to perform multi-layer iterative operations and recovering the target signal estimation result for the current frame, include the following: S3.
21. Based on the non-stationary noise covariance matrix loaded and initialized in S3.1, use its inverse matrix as the weighting coefficients for the fidelity terms to construct the gradient update rules for the deep unfolded network. S3.
22. Perform multi-level iterative operations using the proximal gradient descent framework, update signal features layer by layer, and output the target signal estimation result for the current frame.
6. The method for suppressing non-stationary noise in vector hydrophones based on deep unfolded networks according to claim 1, characterized in that, The dynamic construction of the target propagation physical manifold based on the target signal estimation result of the current frame in S3.3 includes the following steps: S3.
31. Based on the target signal estimation results of the current frame, calculate the cross-correlation between the sound pressure channel and the complex time-frequency coefficients of each vibration velocity channel within the local time-frequency window; S3.
32. Extract the target propagation direction vector based on the cross-correlation quantity, and construct a normalized target propagation guidance vector; S3.
33. Generate a physical projection matrix based on the target propagation guidance vector to form a target propagation physical manifold.
7. The method for suppressing non-stationary noise in vector hydrophones based on deep unfolded networks according to claim 1, characterized in that, The vector acoustic signal reconstruction and output in S4 includes the following steps: S4.1 Perform short-time inverse Fourier transform on each channel of the denoised target signal feature data to obtain the time-domain initial signal corresponding to each channel; S4.2 Based on the multi-channel acoustic physical characteristics of the vector hydrophone, the phase consistency of the initial time domain signal of each channel is corrected to maintain the spatial coherence of the multi-channel signal. S4.3 Integrate the multi-channel time-domain signals that have completed phase correction by channel dimension, complete time-domain reconstruction, and output the target vector acoustic signal after non-stationary noise suppression.
Citation Information
Patent Citations
Sound source orientation estimation method, system and device based on sound vector field
CN115267674A
Single-vector hydrophone target tracking method and system based on adaptive square root cubature Kalman
CN118011400A
