Water surface micro-wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology

Abstract

The application provides a water surface micro-wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology. The method is used for water surface micro-wave detection tasks. The method introduces a time-frequency domain constraint module and a blind source separation precision feedback module in the U-Net adversarial network to perform nonlinear restoration on the micro-wave signal, then performs feature extraction based on a subspace translation covariance matrix, introduces a non-metric scaling technology to improve a mixed matrix estimation, and then restores a source signal through an optimal sub-matrix in a sparse space. Efficient and accurate water surface micro-wave detection is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing of millimeter wave radar, in particular to a water surface micro-amplitude wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology. BACKGROUND

[0002] The ocean is rich in natural resources, providing an important material basis for the development of modern industry. In the process of ocean resource development, the demand for information exchange between water and underwater is frequent, so realizing water-air cross-medium communication has important engineering application value. Traditional cross-medium communication relies on autonomous underwater vehicles (AUVs), buoys and other intermediate devices to realize the forwarding of acoustic signals, but AUVs need to frequently dive to collect information, resulting in large communication delay, and buoys are easily affected by sea waves, resulting in unstable communication.

[0003] An emerging water-air cross-medium communication link implementation scheme is to detect the micro-amplitude wave changes on the water surface caused by underwater acoustic sources through millimeter wave radar, realizing information transmission without intermediaries. Underwater acoustic waves propagate to the water surface in the form of pressure waves, which will excite the micro-displacement of the water surface. Millimeter wave radar transmits electromagnetic waves of a specific modulation frequency through high-frequency circuits, has advantages such as high-resolution imaging, excellent penetration ability, and high-precision ranging. Therefore, by analyzing the changes in the modulation characteristics of the radio frequency signal caused by water surface disturbances during propagation, the underwater acoustic source information can be perceived and reconstructed. However, the actual underwater environment of the ocean is complex (e.g., thermocline, halocline, water bubbles, etc.), and underwater information transmission is affected by the communication channel, which is prone to nonlinear distortion. Secondly, the micro-amplitude wave excited by the actual acoustic signal hitting the water surface presents significant dispersion characteristics, and its amplitude decays with increasing acoustic signal frequency, which will introduce additional nonlinear distortion components to the radar received signal. Therefore, how to efficiently and accurately detect the water surface micro-amplitude wave in a complex ocean environment is a problem to be solved. SUMMARY

[0004] The purpose of the present application is to solve the problem of nonlinear distortion of water surface micro-amplitude wave, and a water surface micro-amplitude wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology is proposed.

[0005] The present application is realized by the following technical scheme, the present application proposes a water surface micro-amplitude wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology, the method comprises:

[0006] Step 1: divide the water surface micro-amplitude wave data and underwater mixed audio data into training set, validation set and test set according to the proportion; based on the U-NET adversarial network improved by the time-frequency domain constraint module and the blind source separation precision feedback module, the water surface micro-amplitude wave signal is nonlinearly restored;

[0007] Step 2: Perform eigenvector extraction on the reduced water surface micro-wave signal, divide the micro-wave time-frequency vector into different subspaces, and extract the eigenvectors of the covariance matrix of the sub-space time-frequency domain vector by eigenvalue decomposition method;

[0008] Step 3: Introduce non-metric scaling technology feature matrix dimension reduction, and perform blind source signal mixing matrix estimation based on the feature matrix composed of eigenvectors;

[0009] Step 4: Based on the optimal sub-matrix, the source signal is recovered, the local inverse transformation is performed on each time-frequency point to recover the source signal strength, and finally the inverse short-time Fourier transform is performed to recover the source signal.

[0010] Further, in step 1, a time-frequency domain constraint module is added to the generator of the U-NET adversarial network, which maps the water surface micro-wave data to the time-frequency domain, so that the adversarial training can retain the time domain and frequency domain features at the same time, and the adversarial network generator loss function is updated; wherein the time-frequency domain regularization loss term in the time-frequency domain constraint module And the updated generator loss function are defined as follows, respectively:

[0011]

[0012]

[0013] wherein, is the time-frequency domain regularization loss term, and are the regularization weights of the norm and the norm, respectively, and are the time-frequency domain representations of the source signal and the generated sample, respectively, is the loss function of the generator, is the updated generator loss function.

[0014] Further, in step 1, the U-NET adversarial network introduces a blind source separation accuracy feedback module, which performs blind source separation on the samples generated by the generator and verifies the correlation of the underwater source signals, which is used as feedback control of the adversarial network; wherein the blind source separation accuracy feedback module outputs the source signal correlation coefficient :

[0015]

[0016] wherein, is the source signal correlation coefficient output by the blind source separation feedback module, and are the initial source signal and the source signal estimated by the blind source separation algorithm, respectively, and when the correlation coefficient When the correlation coefficient is less than the preset threshold, the U-Net GAN deep feedback training is driven, and the loss function of the generator is updated as follows:

[0017]

[0018] wherein, is the final loss function of the generator, is the weight coefficient of the correlation coefficient.

[0019] Further, in step 2, in order to effectively separate the source signal in the micro-amplitude wave signal restored by the adversarial network, the eigenvector is extracted based on the subspace translation covariance matrix; the time-frequency domain vector matrix is divided into K non-overlapping subspaces in the time axis, and the translation covariance matrix of the subspace signal is calculated :

[0020]

[0021] wherein, represents the time-frequency vector which is not translated in the subspace, represents the translated part, is the translation amount;

[0022] The translation covariance matrix of each subspace signal is approximately calculated in the following form:

[0023]

[0024] wherein, represents the approximate value of the covariance matrix , and is the number of time stamps of each subspace.

[0025] Further, in step 2, the eigenvector of the source component is extracted by the EVD eigenvalue decomposition method, which represents the dominant direction of the source signal:

[0026]

[0027] wherein, represents the matrix composed of the eigenvector, represents the diagonal matrix composed of the eigenvalue;

[0028] The eigenvector is extracted for all subspace feature matrices, and the eigenvector in only the single-source dominant subspace is obtained, at this time, the rank is approximately 1, and the eigenvector corresponding to the maximum eigenvalue in is extracted , and the Q single-source eigenvectors obtained are composed into a feature matrix , which represents the statistical characteristics of the estimated spatial distribution of the observation signal.

[0029] Further, in step 3, non-metric scaling technique is introduced, and the feature vector matrix is reduced dimension by using non-metric scaling technique, and a two-dimensional space matrix of feature vectors is obtained :

[0030]

[0031] wherein, each row is the projection coordinate of the feature vector in 2-dimensional space, is the distance matrix, and is the feature vector in the corresponding feature matrix , wherein, ; the feature vector represented by each coordinate in is defined as the SFV independent feature vector .

[0032] Further, in step 3, first, the Euclidean distance between SFV vectors is calculated

[0033] ; then, the local density of each SFV vector is calculated :

[0034]

[0035] All local density values are arranged in descending order, and the minimum distance corresponding to each SFV vector and all other higher density points is calculated :

[0036]

[0037] By the local density value and the minimum distance , a decision diagram is drawn, and the coordinate point with larger product of the two is selected as the clustering center ; the corresponding feature vector is obtained by using the reverse projection of the clustering center, as the column vector of the mixing matrix, and thus the mixing matrix can be estimated.

[0038] Further, in step 4, for the estimated mixing matrix , all sub-matrices thereof are extracted; for each time-frequency point , the corresponding optimal sub-matrix is obtained by using norm:

[0039]

[0040] Optimal sub-matrix representing the sparse distribution of the source signal at this time-frequency point; The norm minimization makes the sparse solution maximization, and the observation signal is concentrated on the mixing matrix column vector at this time;

[0041] through the optimal sub-matrix Realize local inverse transformation Get the source signal strength at this time-frequency point The estimated value of the mth source signal at is:

[0042]

[0043] Wherein, m is the index of the mth source signal, and indicates the position of the mth source signal in all sub-matrices In all sub-matrices The mth element in the source signal strength is the mth column vector corresponding to the estimated mixing matrix The mth column vector corresponding to the estimated mixing matrix

[0044] After recovering the source signal strength at each TF point through the sub-matrix, the time domain source signal can be obtained through inverse Fourier transform .

[0045] The application further provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the water surface micro wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology when executing the computer program.

[0046] The application further provides a computer readable storage medium for storing computer instructions, which realize the steps of the water surface micro wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology when the processor executes the computer instructions.

[0047] The application has the following beneficial effects:

[0048] The application provides a water surface micro wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology, which is used for water surface micro wave detection tasks. The method introduces a time-frequency domain constraint module and a blind source separation precision feedback module in a U-NET adversarial network to nonlinearly restore the micro wave signal, then extracts features based on a subspace translation covariance matrix, introduces a non-metric scaling technology to improve the mixing matrix estimation, and then restores the source signal through the optimal sub-matrix in the sparse space. Efficient and accurate water surface micro wave detection is realized. BRIEF DESCRIPTION OF DRAWINGS ​​​​​

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on the accompanying drawings are within the scope of the present application.

[0050] Figure 1 The flow chart of the water surface micro-wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology according to the present application.

[0051] Figure 2 The overall structural framework diagram of the generative adversarial network.

[0052] Figure 3 The schematic diagram of the improved generator module.

[0053] Figure 4 The schematic diagram of the mixing matrix estimation module.

[0054] Figure 5 The schematic diagram of the micro-wave time-domain nonlinear distortion restoration.

[0055] Figure 6 The schematic diagram of the micro-wave time-frequency domain nonlinear distortion restoration amplitude spectrum.

[0056] Figure 7 The schematic diagram of the micro-wave simulated data set source signal detection result.

[0057] Figure 8 The schematic diagram of the actual radar echo data main frequency detection result. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of the present application.

[0059] The application proposes a water surface micro-amplitude wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology. According to the idea of adversarial training network, a time-frequency domain constraint module and a blind source separation accuracy feedback module are added to the U-NET adversarial network, which can guarantee the time-frequency domain characteristics of the generated signal and retain more source signal characteristics through feedback driving network. Secondly, the blind source separation UBSS model introduces the non-metric scaling (NMDS) technology to estimate the mixing matrix in the time-frequency domain. The eigenvectors of the time-frequency domain vector covariance matrix are extracted by the eigenvalue decomposition (EVD) method, the eigenvectors are reduced by the non-metric scaling NMDS technology, and the mixing matrix is estimated by the density peak value clustering, so that the mixing matrix estimation is more accurate. Finally, in order to solve the problem that the number of mixed signal sources is greater than the number of observation channels, all submatrices of the estimated mixing matrix are extracted and the optimal submatrix corresponding to each time-frequency point is calculated, and the optimal submatrix is used to perform local inverse transformation on each time-frequency point to solve the problem of irreversible underdetermined state mixing matrix.

[0060] Specifically, in combination with Figures 1-8 The application proposes a water surface micro-amplitude wave detection method based on nonlinear reconstruction and high-dimensional decoupling technology, and the purpose of the method is to solve the problem of cross-medium communication based on water surface micro-amplitude wave. The micro-amplitude wave signal distortion is restored by using the generative adversarial network, and the underwater sound source information is analyzed by using the blind source separation algorithm. The method comprises the following steps:

[0061] Step 1: The water surface micro-amplitude wave data and the underwater mixed audio data are divided into training set, validation set and test set according to the proportion; the U-NET adversarial network improved based on the time-frequency domain constraint module and the blind source separation accuracy feedback module is used to perform nonlinear restoration on the water surface micro-amplitude wave signal; the micro-amplitude wave multi-source simulation data set and the actual radar echo data set are used in step 1 to verify the effectiveness of the scheme in theory and actual measurement. The improved U-NET adversarial network of the data set is used for nonlinear distortion restoration, and the quality of the mixed signal is enhanced. Finally, the training parameters of the adversarial network are set.

[0062] The data set used in step 1 is the micro-amplitude wave simulation data set and the millimeter wave radar echo data set, which is based on the underwater sound signal propagation attenuation formula and the micro-amplitude wave generation formula The water surface micro-amplitude wave vibration model is constructed, wherein A is the micro-amplitude wave amplitude, is the underwater sound source sound pressure, is the attenuation coefficient of water, is the distance of the sound wave to the water surface, is the density of water, is the angular frequency of the sound wave, The sound speed in water is v. The audio data is used as the input sound pressure, and the multiple-source micro-amplitude waves interfering with each other are generated according to the micro-amplitude wave superposition theorem, and the micro-amplitude wave amplitude is output as the system training simulation data set, which is used to verify the proposed algorithm. In the experiment, the generation of the adversarial network epoch is set to 1000, the batch size is set to 100, the optimizer is Adam, and the initial learning rate is set to 1e-4.

[0063] To realize the de-nonlinearization of the observation signal, the U-Net GAN network with a skip connection is used to perform nonlinear reconstruction and feature optimization on the observation signal, and the separability of the signal is enhanced. The loss functions of the generator and the discriminator are obtained based on the least square loss (MSE), and the objective function of the U-Net GAN is defined as follows:

[0064]

[0065] wherein, represents the generator, represents the discriminator. The loss function of the generator is defined as , which is expressed as , and the prediction loss function of the discriminator is defined as , and is expressed as .

[0066] In step 1, the generator minimizes the loss to optimize the generated sample In order to further improve the convergence, a time-frequency domain constraint module is added to the generator of the U-NET adversarial network, which maps the water surface micro-amplitude wave data to the time-frequency domain, so that the adversarial training retains the time domain and frequency domain features at the same time, and the loss function of the generator of the adversarial network is updated; wherein the time-frequency domain regularization loss term in the time-frequency domain constraint module and the updated generator loss function are defined as follows:

[0067]

[0068]

[0069] wherein, is the time-frequency domain regularization loss term, and are the regularization weights of the norm and the norm respectively, and are the time-frequency domain representations of the source signal and the generated sample respectively, is the loss function of the generator, is the updated generator loss function.

[0070] In step 1, the U-NET adversarial network introduces a blind source separation accuracy feedback module to perform blind source separation on the samples generated by the generator and verify the correlation of the underwater source signals, which is used as feedback control of the adversarial network; wherein the blind source separation accuracy feedback module outputs the source signal correlation coefficient :

[0071]

[0072] wherein, is the source signal correlation coefficient output by the blind source separation feedback module, are the initial source signals and the source signals estimated by the blind source separation algorithm, when the correlation coefficient is less than the preset threshold (0.9 in the present application), the U-Net GAN deep feedback training is driven, and the training tends to retain the characteristics of the source signals to ensure the subsequent signal separation effect, and the loss function of the generator is updated as follows:

[0073]

[0074] wherein, is the final loss function of the generator, is the weight coefficient of the correlation coefficient, which is 0.8 in the present embodiment. The control balances the trade-off between the target to be optimized by the generator and the correlation coefficient, and is used to drive the observation signal generated by the adversarial network to have more source signal characteristics.

[0075] Step 2: Feature vector extraction is performed on the restored water surface micro-ripple signal, the micro-ripple time-frequency vector is divided into different subspaces, and the feature vectors of the sub-space time-frequency domain vector covariance matrix are extracted by the eigenvalue decomposition (EVD) method; Step 2 maps the restored micro-ripple data to the time-frequency domain, and divides the time-frequency domain vector matrix into K non-overlapping subspaces along the time axis. The translation covariance matrix of each subspace is calculated to extract the feature vectors, and the statistical characteristics of the source signal distribution in the time-frequency domain space are obtained.

[0076] In step 2, in order to effectively separate the source signals in the micro-ripple signal restored by the adversarial network, the feature vectors are extracted based on the translation covariance matrix of the subspace; the observation signal time-frequency domain vector matrix is obtained by the STFT transform of the observation signal generated by the adversarial network. The time-frequency domain vector matrix is divided into K non-overlapping subspaces along the time axis, and each subspace time-frequency domain vector matrix is represented as . By translating the subspace signal time-frequency domain vector, the translation covariance matrix of the subspace signal is calculated, the correlation in the time and space of the signal is obtained, and the translation covariance matrix :

[0077]

[0078] where, represents the unshifted time-frequency vectors in the subspace, represents the shifted part, is the shift amount;

[0079] the shift covariance matrix of each subspace signal is approximated by the following form:

[0080]

[0081] where, represents the approximation of the covariance matrix , is the number of time stamps for each subspace.

[0082] In step 2, the energy distribution of the source signal is usually locally concentrated in the time-frequency domain, and the rank of the covariance matrix of the subspace is 1. The eigenvectors of the source component are extracted by the EVD eigenvalue decomposition method, which represent the dominant direction of the source signal:

[0083]

[0084] where, represents the matrix composed of eigenvectors, represents the diagonal matrix composed of eigenvalues;

[0085] The eigenvectors of all subspace eigenvalue matrices are extracted to obtain the eigenvectors in the single-source dominant subspace, and at this time, the rank is approximately 1, and the eigenvector corresponding to the largest eigenvalue in is extracted , and the Q single-source eigenvectors obtained are composed into an eigenvalue matrix , which represents the statistical characteristics of the estimated spatial distribution of the observation signal.

[0086] Step 3: Introduce non-metric scaling technology to reduce the dimension of the eigenvalue matrix, and use the density peak clustering method to estimate the blind source signal mixing matrix based on the eigenvalue matrix composed of eigenvectors; Step 3 introduces non-metric scaling NMDS technology to reduce the high-dimensional eigenvectors to two dimensions, calculates the Euclidean distance and local density between two-dimensional eigenvectors, and performs clustering, and obtains the mixing matrix through the clustering center. The eigenvectors in the time-frequency domain enhance the expression of signal characteristics, making the mixing matrix estimation more accurate.

[0087] In step 3, since the feature vector corresponding to the cluster center and the column vector of the mixing matrix both represent the distribution characteristics of the signal in the space, the cluster center and the column vector of the mixing matrix have a corresponding relationship in structure. The number of cluster clusters corresponds to the number of source signals N. The application estimates the mixing matrix by dimension reduction of the subspace feature vector and based on the two-dimensional distribution of the subspace feature vector. The application introduces a non-metric scaling technique to reduce the dimension of the feature vector matrix by using the non-metric scaling technique to obtain a two-dimensional space matrix of the feature vector :

[0088]

[0089] wherein, each row is the projection coordinate of the feature vector in the 2-dimensional space, is a distance matrix (by calculating the Euclidean distance between the feature vectors , to form a distance matrix ), and is the feature vector in the feature matrix corresponding to the SFV vector, wherein, ; the feature vector represented by each coordinate in is defined as a SFV independent feature vector (Separate feature vectors) .

[0090] In step 3, first, the Euclidean distance between the SFV vectors is calculated;

[0091] Then, the local density of each SFV vector is calculated :

[0092]

[0093] All local density values are arranged in descending order, and the minimum distance corresponding to each SFV vector and all other higher density points is calculated:

[0094]

[0095] A decision graph is drawn by the local density value and the minimum distance , and the coordinate point with a large product of the two is selected as the cluster center ; the corresponding feature vector is obtained by back-projection using the cluster center as the column vector of the mixing matrix, and thus the mixing matrix can be estimated.

[0096] Step 4: source signal recovery based on the optimal sub-matrix, local inverse transformation is performed for each time-frequency point to recover the source signal strength, and finally the source signal is recovered through inverse short-time Fourier transform (ISTFT). Step 4 uses the optimal sub-matrix of the mixing matrix to recover the source, which solves the problem that the source signal cannot be effectively separated due to the irreversibility of the mixing matrix in the underdetermined state, and the mixing matrix is not full rank and irreversible in blind source separation.

[0097] In step 4, for the estimated mixing matrix , all sub-matrices of the mixing matrix are extracted; for each time-frequency point , the optimal sub-matrix corresponding to the time-frequency point is obtained by using the norm .

[0098]

[0099] The optimal sub-matrix represents the sparse distribution of the source signal at this time-frequency point; The norm minimization maximizes the sparsity, and the observation signal is concentrated on the column vector of the mixing matrix at this time;

[0100] The local inverse transformation is realized by using the optimal sub-matrix to obtain the source signal strength at this time-frequency point; the estimated value of the mth source signal at is:

[0101]

[0102] Wherein, is the index of the mth source signal, indicating the position of the mth source signal in the optimal sub-matrix in all sub-matrices . is the mth element in the source signal strength , which corresponds to the mth column vector of the estimated mixing matrix .

[0103] After recovering the source signal strength at each TF point through the sub-matrix, the time-domain source signal can be obtained through inverse Fourier transform.

[0104] The application also provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the water surface micro-amplitude wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology when executing the computer program.

[0105] ​The application further provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the water surface micro-wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology.

[0106] The memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM). It is to be noted that the memory described with the methods of the present application is intended to include, but not be limited to, these and any other suitable types of memory.

[0107] In the above embodiments, all or part of the method can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the method can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as high-density digital video disc (digital video disc, DVD)), or semiconductor media (such as solid state disc (solid state disc, SSD)) and the like.

[0108] In the implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor or instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution or executed by combination of hardware and software modules in the processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0109] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with a signal processing capability. In the implementation process, each step of the method embodiments can be completed by the integrated logic circuit of hardware or the instruction in the form of software in the processor. The processor mentioned above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method.

[0110] The above describes the water surface micro-amplitude wave detection method based on the nonlinear reconstruction and high-dimensional decoupling technology in detail. The principle and implementation manner of the present application are described by using specific examples. The above embodiment is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for detecting micro-amplitude waves on the water surface based on nonlinear reconstruction and high-dimensional decoupling techniques, characterized in that, The method includes: Step 1: Divide the surface micro-wave data and underwater mixed audio data into training set, validation set and test set according to the ratio; perform nonlinear reconstruction of the surface micro-wave signal based on the U-NET adversarial network improved by the time-frequency domain constraint module and the blind source separation accuracy feedback module; Step 2: Extract the feature vector from the restored water surface micro-amplitude wave signal, divide the micro-amplitude wave time-frequency vector into different subspaces, and extract the feature vector of the subspace time-frequency domain vector covariance matrix by eigenvalue decomposition method; Step 3: Introduce non-metric scaling techniques to reduce the dimensionality of the feature matrix, and estimate the blind source signal mixing matrix based on the feature matrix composed of feature vectors; Step 4: Recover the source signal based on the optimal submatrix, perform local inverse transform at each time frequency point to recover the source signal strength, and finally recover the source signal through inverse short-time Fourier transform; In step 1, a time-frequency domain constraint module is added to the generator of the U-NET adversarial network to map the surface micro-wave data to the time-frequency domain, so that adversarial training retains both time-domain and frequency-domain features and updates the loss function of the adversarial network generator; wherein, the time-frequency domain regularization loss term in the time-frequency domain constraint module and the updated generator loss function They are defined as follows: in, For time-frequency domain regularization loss term, and They are Norm and Regularized weights of the norm and These are the time-frequency domain representations of the source signal and the generated sample, respectively. Let the loss function of the generator be... The updated generator loss function; In step 1, the U-NET adversarial network introduces a blind source separation accuracy feedback module to perform blind source separation on the samples generated by the generator and to verify the correlation of underwater source signals, using it as a feedback control adversarial network; the output of the blind source separation accuracy feedback module is the correlation coefficient of the source signals. : in, It is the correlation coefficient of the source signal output by the blind source separation feedback module. , respectively the initial source signal and The source signal estimated by the blind source separation algorithm, when the correlation coefficient... When the value is less than a preset threshold, the U-Net GAN is driven to perform deep feedback training, and the generator's loss function is updated as follows: in, It is the generator's final loss function. It is the weighting coefficient of the correlation coefficient.

2. The method according to claim 1, characterized in that, In step 2, to effectively separate the source signal from the micro-amplitude wave signal reconstructed by the adversarial network, eigenvector extraction is performed based on the subspace translation covariance matrix. The time-frequency domain vector matrix is ​​divided into K non-overlapping subspaces along the time axis, and the translation covariance matrix of the subspace signal is calculated. : in, Represents the untranslated time-frequency vector in the subspace. Represents the translated portion. It is the translation amount; Translation covariance matrix of each subspace signal Approximate calculation is performed using the following method: in, Represents the covariance matrix Approximate value, The number of timestamps for each subspace.

3. The method according to claim 2, characterized in that, In step 2, the eigenvectors of the source components are extracted using the EVD eigenvalue decomposition method, representing the dominant direction of the source signal: in, The matrix representing the eigenvalue vectors, Represents a diagonal matrix composed of eigenvalues; Extract eigenvectors from the feature matrices of all subspaces to obtain eigenvectors that exist only within the single-source dominant subspace. The rank is approximately 1, and the extraction... The eigenvector corresponding to the largest eigenvalue in the middle The feature matrix is ​​composed of the Q single-source feature vectors obtained. , representing the statistical characteristics of the estimated spatial distribution of the observed signal.

4. The method according to claim 3, characterized in that, In step 3, a non-metric scaling technique is introduced to reduce the dimensionality of the feature vector matrix, thereby obtaining the two-dimensional space matrix of the feature vectors. : in, Each row represents the projection coordinates of the feature vector in 2D space. It is a distance matrix. and It is the corresponding characteristic matrix The eigenvectors in, where, ;Will The eigenvectors represented by each coordinate in the vector are defined as SFV independent eigenvectors. .

5. The method according to claim 4, characterized in that, In step 3, the Euclidean distance between SFV vectors is first calculated. Then, for each SFV vector Calculate local density : Sort all local density values ​​in descending order for each SFV vector. Calculate and all other higher density points Corresponding minimum distance : Through local density values and minimum distance Draw a decision graph and select the coordinate point with the larger product of the two as the cluster center. ; Obtain the corresponding feature vectors by backprojecting from the cluster centers. , as column vectors of the mixture matrix, can be used to estimate the mixture matrix. .

6. The method according to claim 5, characterized in that, In step 4, for the estimated mixing matrix Extract all its submatrices For each time frequency point ,use Find the optimal submatrix corresponding to the norm: Optimal Submatrix This represents the sparse distribution of the source signal at this frequency point; Minimizing the norm maximizes the sparsity of the solution, at which point the observed signal is concentrated on the column vector of the mixing matrix. Through the optimal submatrix Achieve local inverse transformation Obtain the source signal strength at this frequency point. ;No. Individual source signal in The estimated value at this location is: in, It is the index of the m-th source signal, indicating the optimal submatrix corresponding to the m-th source signal. In all submatrices The position in the middle; Source signal strength The m-th element in the matrix is ​​related to the estimated mixing matrix. The m-th column vector corresponds to; After recovering the source signal intensity at each TF point using a submatrix, the time-domain source signal can be obtained through inverse Fourier transform. .

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

Images