A method for submarine target sonar detection based on adversarial network

Abstract

The present invention discloses a method for detecting submarine targets using sonar based on an adversarial network. The following steps are involved: S1. Raw deep-sea sonar data is collected and preprocessed to obtain a standardized sonar tensor; S2. The standardized sonar tensor is used as training input and combined with an environmental parameter vector to train an improved StyleGAN-3 model; S3. The improved StyleGAN-3 model is called to output a weak target feature tensor; S4. A sonar detection network is constructed; S5. An optimal detection parameter set is output; S6. The optimal detection parameter set is loaded into the sonar detection network, target judgment is performed on the weak target feature tensor, and a target confidence value, a three-dimensional spatial positioning result, and a target motion track are output. In deep-sea scenarios, the present invention significantly reduces positioning error and target trajectory loss rate compared to traditional methods. The method has excellent online tracking and dynamic situational awareness capabilities, providing solid technical support for marine monitoring, security, and emergency command applications.

Description

Technical Field

[0001] The present invention relates to the technical field of sonar detection, and in particular to a seabed target sonar detection method based on an adversarial network. Background Art

[0002] With the continuous development of deep-sea exploration and intelligent sonar technology, sonar systems are playing an increasingly important role in marine environmental monitoring, deep-sea resource exploration, and underwater security. Detecting weak target echo signals in complex deep-sea environments has become a core challenge restricting the intelligent development of underwater sonar. Existing deep-sea weak echo target detection technologies mostly rely on traditional signal processing methods. These technologies suffer from low sensitivity, high false alarm rates, and severe parameter mismatch in environments with extremely low signal-to-noise ratios, non-stationary noise, and complex multipath interference.

[0003] Traditional deep-sea weak target detection algorithms are limited by the difficulty of sample acquisition, data scarcity, and extremely uneven distribution. Due to the complex conditions of deep-sea experiments, sea conditions, and confidentiality factors, the number of weak echo target samples that can actually be obtained is extremely small. This makes it difficult to fully train data-driven detection models based on deep learning, and detectors are prone to overfitting and reduced generalization capabilities. Furthermore, existing data augmentation and simulation sample synthesis methods, which primarily rely on parameterized physical modeling or empirical noise mixing, struggle to accurately reflect the joint distribution and complex variations of target signals and background noise in real deep-sea environments. Significant domain deviations exist between synthetic samples and measured samples, reducing the practical value of the models.

[0004] In terms of detection algorithm optimization, existing technologies generally adopt static decision thresholds, fixed filtering parameters or simple heuristic search strategies. General optimization algorithms are often unable to adapt to the highly time-varying and multi-scale characteristics of the noise field in deep-sea environments, resulting in parameter adaptation lags in the detection system under conditions of extremely low signal-to-noise ratio and drastic dynamic changes in signal and noise power, affecting the real-time detection rate and overall reliability of weak targets. In addition, existing sonar detection models usually lack the ability to jointly optimize the heterogeneity of deep-sea environments, noise non-stationarity and signal physical characteristics, making it difficult to achieve robust and efficient detection in multiple scenarios and conditions. Summary of the Invention

[0005] One purpose of the present invention is to propose a submarine target sonar detection method based on an adversarial network. In deep-sea scenarios, the positioning error and target trajectory loss rate of the present invention are greatly reduced compared with traditional methods. It has excellent online tracking and dynamic situation awareness capabilities, and provides solid technical support for marine monitoring, security and emergency command applications.

[0006] A method for detecting submarine targets using sonar based on an adversarial network according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect and preprocess raw deep-sea sonar data to obtain standardized sonar tensors;

[0008] S2. Using the standardized sonar tensor as training input, we improve the StyleGAN-3 model by combining it with environmental parameter vector training to complete the construction of the sonar scene generation network.

[0009] S3. Call the improved StyleGAN-3 model to generate a generated sample set. This sample set is proportionally fused with the raw deep-sea sonar data to construct a synthetic-measured hybrid dataset. The synthetic-measured hybrid dataset is then subjected to weak echo enhancement processing to output a weak target feature tensor.

[0010] S4. Construct a sonar detection network, using the weak target feature tensor as input and the target annotation information in the synthetic-measured hybrid dataset as the supervision signal, to obtain the set of detection parameters to be optimized. Based on the weak target feature tensor and the set of detection parameters to be optimized, establish a marine predator optimization algorithm framework and define an evaluation function with the core indicators of weak target energy spectrum difference, detection rate, and false alarm rate.

[0011] The target label information is a label obtained by manually or semi-automatically marking the category, spatial position or existence of weak targets in the synthetic-measured mixed data set after weak echo enhancement processing;

[0012] S5. Dynamically update the search strategy based on the evaluation function value within the framework of the marine predator optimization algorithm until the evaluation function converges and outputs the optimal detection parameter set;

[0013] S6. Load the optimal detection parameter set into the sonar detection network, perform target judgment on the weak target feature tensor, and output the target confidence value, three-dimensional spatial positioning result and target motion track.

[0014] Optionally, the S2 includes the following steps:

[0015] S21. Constructing a deep-sea weak echo sonar generation training dataset ,The deep-sea weak echo sonar generation training dataset consists of standardized sonar tensors and environmental parameter vectors, where the environmental parameter vectors are depth, salinity, temperature gradient and seabed type;

[0016] S22. Constructing an improved StyleGAN-3 model with adaptive spectral energy-phase joint constraints , improve the StyleGAN-3 model to integrate the space-time separable convolution structure and the complex domain residual gating structure in the feature generation network;

[0017] S23. Combine the target energy spectrum and background noise power spectrum density to dynamically construct the energy-phase joint adaptive loss function :

[0018] ;

[0019] in, and are the normalized energy spectra of generated and real weak echo samples, and are the phase matrices of the generated and real weak echo samples, 、 is the weight coefficient, j represents the imaginary unit, is the total number of training samples;

[0020] S24. In the improved StyleGAN-3 training process, the sample separability scores output by the discriminator based on the sonar detection network are introduced in real time As a joint loss function The adversarial feedback term dynamically adjusts the generator parameters to form a generation-discrimination collaborative feedback mechanism:

[0021] ;

[0022] in, Generate adversarial loss for the standard, To determine the feedback weight, Characterizes the sample separability score of the generated weak echo sample that is correctly identified as a weak target by the current sonar detection network;

[0023] S25. Using the training dataset End-to-end training improves the StyleGAN-3 model During the training process, the parameters are adaptively optimized according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism, and the improved StyleGAN-3 model is finally trained. .

[0024] Optionally, the improved StyleGAN-3 model includes a space-time separable convolution structure and a complex domain residual gating structure:

[0025] Space-time separable convolutional architecture to normalize sonar tensors As input, the one-dimensional convolution operation Conv1D is performed along the time dimension in combination with the time direction convolution kernel to extract the time direction convolution feature tensor of deep-sea weak targets in the time direction. ;

[0026] The time-direction convolution feature tensor is input into the one-dimensional convolution layer in the frequency dimension, and the one-dimensional convolution operation is performed in the frequency dimension in combination with the frequency-direction convolution kernel to extract the time-frequency joint convolution feature tensor of the weak target signal under the influence of Doppler diffusion and seabed scattering. ;

[0027] The time-frequency joint convolution feature tensor is input into the complex domain residual gating structure to establish the amplitude feature branch and the phase feature branch respectively:

[0028] The amplitude feature branch extracts the local amplitude information of the time-frequency joint convolution feature tensor through an amplitude convolution channel. The extracted result is batch normalized and then connected to the rectified linear unit function for nonlinear activation to obtain the amplitude feature tensor. ;

[0029] The phase feature branch extracts the phase change information of the time-frequency joint convolution feature tensor through another phase convolution channel. The extracted result is also batch normalized and then connected to the hyperbolic tangent function for nonlinear activation to obtain the phase feature tensor. ;

[0030] Based on the time-frequency joint convolution feature tensor, a residual signal gating channel is constructed to dynamically control the degree of fusion of target features and background information. The residual signal gating channel extracts the gated feature map through a separate gated channel convolution, and compresses the gated weight value to the [0,1] interval through the activation function, and finally outputs the gated weight tensor. ;

[0031] The magnitude eigentensor , phase eigentensor With the gated weight tensor Fusion is performed to form the final output complex domain residual feature tensor :

[0032] ;

[0033] in, represents the Hadamard product.

[0034] Optionally, S3 includes the following steps:

[0035] S31. Call the trained improved StyleGAN-3 model , input the environmental parameter vector of the sample to be generated, combined with the random latent variable , generate synthetic sample tensors under corresponding conditions ;

[0036] S32. Collect and generate measured deep-sea sonar tensors corresponding to samples , and estimate the weak target energy spectrum of the target signal in real time based on the target energy extraction function and the noise spectrum estimation function and background noise power spectral density ;

[0037] S33. Constructing dynamic fusion ratio coefficient based on weak target energy spectrum and background noise power spectrum density , the fusion ratio is used to control the superposition intensity of the generated samples and the measured data:

[0038] ;

[0039] in, represents the L2 norm;

[0040] S34. According to the dynamic fusion ratio coefficient For synthetic sample tensors and the measured sonar tensor Perform linear fusion to construct synthetic-measured mixed data tensor ;

[0041] S35. Synthetic-measured mixed data tensor Perform weak echo enhancement processing, use multi-scale convolution structure to build convolution pyramid, and extract multi-scale joint convolution feature tensors of weak targets at different time and frequency scales ;

[0042] S36. Input the multi-scale joint convolution feature tensor into the time-frequency attention module, and use the time attention weight and frequency attention weights Perform weighted fusion in the time dimension and frequency dimension to form the final weak target feature tensor ;

[0043] Optionally, the S4 includes the following steps:

[0044] S41. Building a sonar detection network , with weak target feature tensor As input, the target annotation information in the synthetic-measured mixed dataset As the supervision signal, the cross entropy loss function is used Training sonar detection network parameter set ;

[0045] S42. Establish a marine predator optimization algorithm framework based on weak target feature tensors and sonar detection network parameter sets Under the framework of the marine predator optimization algorithm, a global optimization is performed on the feature fusion weights, decision thresholds and filter parameters of the sonar detection network to initialize the marine predator group parameter set. , is the number of individuals in the group;

[0046] S43. Define the evaluation function based on the difference in weak target energy spectrum, detection rate and false alarm rate as indicators , used to measure the detection performance of the sonar detection network under the current detection parameter configuration:

[0047] ;

[0048] in, To measure the energy spectrum of weak targets, To detect the parameters of the network The detected target energy spectrum output below is For detection parameters The detection rate under For detection parameters The false alarm rate under is the balance coefficient.

[0049] Optionally, the S5 includes the following steps:

[0050] S51. Algorithm framework for optimizing marine predators In the process, the parameter set of the marine predator group is initialized, and each iteration After that, for each individual based on the current detection parameters , calculate the evaluation function value ;

[0051] S52. For each individual, based on the current synthetic-measured mixed data tensor , target signal energy spectrum and background noise power spectral density , calculate the individual adaptability confidence :

[0052] ;

[0053] in, To detect the parameters of the network The detected target energy spectrum output below is Indicates the sample expectations, is the L2 norm;

[0054] S53. Setting the individual adaptability confidence threshold , divide the group into global search subgroups and local optimization subgroups ;

[0055] S54. For individuals in the global search subgroup, a deep-sea adaptive Levy flight step-size operator is used. Perform global search strategy updates:

[0056] ;

[0057] in, For the Individuals in The detection parameter vector after rounds of iteration, all parameters are targeted at the current deep-sea weak echo sonar detection task, For the Individuals in The detection parameter vector at the round iteration, For the The individual detection parameter vector with the smallest evaluation function value among all individuals during the round iteration, is a random variable, is the passive tracking weight factor, is the active disturbance weight factor;

[0058] S55. For individuals in the local optimization subgroup, a group collaborative optimization strategy is adopted, integrating adaptive collaboration factors to fine-tune individual parameters:

[0059] ;

[0060] in, Before the evaluation function value Advantageous individuals, is the adaptive collaboration factor with the number of iterations, The top ranking of the evaluation function value in the group in each iteration The number of dominant individuals;

[0061] S56. After each round of iteration, update the global optimal detection parameter individual and its evaluation function value , perform a difference operation on the evaluation function value of the global optimal detection parameter individual and the evaluation function value of the global optimal detection parameter individual in the previous iteration. When the absolute value of the difference is less than the preset convergence threshold, terminate the optimization process;

[0062] S57. When the convergence condition is met, the globally optimal detection parameter individual obtained in the convergence round is used as the optimal detection parameter set.

[0063] Optionally, the deep-sea adaptive Levy flight step-size operator Based on the target energy feature complexity index in the deep sea weak echo target detection scenario Dynamically adjust step size distribution parameters:

[0064] ;

[0065] in, Characterization Individuals in The current detection parameter vector in iteration The corresponding target energy feature complexity index is calculated based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size adaptation factor, is the step stability index, is a stable exponential perturbation term, which represents the normal random variable randomly sampled from the standard normal distribution by the kth individual in the tth iteration, and is used to control the thickness of the distribution tail of the Lévy jump amplitude. is a random perturbation factor, which represents the normal random variable obtained by random sampling from the standard normal distribution of the kth individual in the tth iteration, and is used to control the scale change direction and perturbation amplitude of the Lévy step.

[0066] Optionally, the S6 includes the following steps:

[0067] S61. Loading the optimal detection parameter set into the sonar detection network , for the input weak target feature tensor Make target judgment and generate the corresponding weak target response tensor , where each element Indicates the time index With frequency index The confidence probability of detecting a weak target at ;

[0068] S62. Based on weak target response tensor Extract target confidence value ;

[0069] S63. Based on weak target response tensor The set of time-frequency locations with significant response probability in , combined with the corresponding synthetic-measured mixed data tensor The sensor array element information and sound propagation model parameters in the 3D space are used to calculate the target positioning and output the target positioning results. ,in, is the estimated horizontal distance to the target, is the estimated azimuth angle of the target, is the estimated depth of the target;

[0070] S64. Based on the target positioning results, a target space motion trajectory model is constructed, combining the multi-frame weak target feature tensor in the continuous time window The corresponding detection response tensor , perform temporal confidence-weighted trajectory association:

[0071] If continuous Existing in the frame The frame detection result satisfies , it is determined to be the same target track;

[0072] If the spatial distance between trajectory points changes , then linear interpolation smoothing is performed;

[0073] Output target motion trajectory as a sequence ;

[0074] in, is the confidence threshold for trajectory determination, is the maximum allowable jump distance, is the number of trajectory sliding window frames.

[0075] Optionally, the target confidence value is defined by the detection response intensity and the background response difference:

[0076] when , determine the existence of the target, and output the target confidence value ;

[0077] when , it is judged as not detected, and the target confidence value is recorded as 0;

[0078] in, is the target confidence threshold, which is set based on the optimal strategy based on experience in the training set.

[0079] The beneficial effects of the present invention are:

[0080] (1) The present invention adopts an improved StyleGAN-3 generative network that introduces space-time separable convolution and complex domain residual gating structure, combines the energy-phase dual-constraint loss function, and optimizes the feedback coupling with the detection network. It can dynamically generate high-fidelity, physically consistent weak target echoes, seabed reverberation and background noise samples according to the sea depth, salinity, temperature gradient and seabed type environmental parameters. By introducing the target energy spectrum and background noise power spectrum density to adaptively adjust the synthesis ratio, the domain deviation problem between traditional simulation samples and measured samples is significantly alleviated, which greatly enriches the training data type and achieves balanced coverage of samples in different scenarios.

[0081] (2) The present invention proposes an ocean predator optimization algorithm that combines the target signal energy spectrum and the noise power spectrum density with adaptive grouping. For group individuals with different adaptive confidence levels, a global search of deep-sea adaptive Levy flight steps and a local fine optimization of group collaborative optimization are used. The algorithm can automatically adjust the feature fusion weights, decision thresholds and filtering parameters of the detection network in highly non-stationary noise and signal dynamic change environments. In the context of extremely low signal-to-noise ratio and Doppler spread, the system false alarm rate is significantly reduced, and the detection rate and parameter stability are significantly better than the comparative technology.

[0082] (3) This invention achieves high-precision integrated output of target confidence, spatial positioning, and trajectory extraction in the sonar detection network through dynamic loading of the optimal detection parameter set and a time-frequency joint attention mechanism. Dynamic judgment is performed based on multi-frame, multi-scale, and multi-frequency weak target feature tensors. Combined with back-projection and spatiotemporal trajectory clustering algorithms, it effectively improves the three-dimensional spatial resolution of weak targets and the continuity of multi-target tracks in complex deep-sea environments. Actual engineering tests show that in deep-sea scenarios, the positioning error and target trajectory loss rate of this invention are significantly reduced compared to traditional methods. It has excellent online tracking and dynamic situational awareness capabilities, providing solid technical support for marine monitoring, security, and emergency command applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0084] Figure 1 This is a flow chart of a submarine target sonar detection method based on an adversarial network proposed by the present invention;

[0085] Figure 2 This is a schematic diagram of the structure of the improved StyleGAN-3 sonar sample generation and physical consistency constraints in the submarine target sonar detection method based on the adversarial network proposed in the present invention;

[0086] Figure 3 This is a schematic diagram of the adaptive global optimization and grouping coordination mechanism of detection parameters based on the marine predator optimization algorithm in the seabed target sonar detection method based on the adversarial network proposed by the present invention. DETAILED DESCRIPTION

[0087] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0088] refer to Figure 1-Figure 3 , a submarine target sonar detection method based on an adversarial network, comprising the following steps:

[0089] S1. Collecting raw deep-sea sonar data, and performing time-domain pulse compression filtering, matched filtering, and normalization based on noise statistical characteristics on the raw deep-sea sonar data to obtain a standardized sonar tensor;

[0090] S2. Using standardized sonar tensors as training input, combined with environmental parameter vectors consisting of depth, salinity, temperature gradient, and seafloor type, we trained an improved StyleGAN-3 model to construct a dedicated generative network for sonar scenarios. The improved StyleGAN-3 model introduced space-time separable convolution and complex-domain residual gating structures in the feature mapping layer to enhance the representation of phase continuity and spectral structure of weak echo signals.

[0091] S3. Call the improved StyleGAN-3 model to generate weak target echo samples, seabed reverberation samples, and ambient background noise samples based on the environmental parameter vector to form a generated sample set. This generated sample set is proportionally fused with the raw deep-sea sonar data. The fusion ratio is dynamically adjusted based on the difference in power spectrum density between the target energy spectrum and the background noise. A synthetic-measured hybrid dataset is constructed. Weak echo enhancement is performed on the synthetic-measured hybrid dataset. Weak target features are extracted using a multi-scale convolutional structure and a time-frequency attention mechanism, and the weak target feature tensor is output.

[0092] S4. Construct a sonar detection network, using the weak target feature tensor as input and the target annotation information in the synthetic-measured hybrid dataset as the supervision signal, to obtain the set of detection parameters to be optimized. Based on the weak target feature tensor and the set of detection parameters to be optimized, establish a marine predator optimization algorithm framework and define an evaluation function with the core indicators of weak target energy spectrum difference, detection rate, and false alarm rate.

[0093] The target label information is a label obtained by manually or semi-automatically marking the category, spatial position or existence of weak targets in the synthetic-measured mixed data set after weak echo enhancement processing;

[0094] S5. Within the framework of the marine predator optimization algorithm, perform a global search and iterative optimization of the feature fusion weights, decision thresholds, and filter parameters in the detection parameter set. Dynamically update the search strategy based on the evaluation function value until the evaluation function converges and output the optimal detection parameter set.

[0095] S6. Load the optimal detection parameter set into the sonar detection network, perform target judgment on the weak target feature tensor, and output the target confidence value, three-dimensional spatial positioning result and target motion track.

[0096] In this embodiment, S2 includes the following steps:

[0097] S21. Constructing a deep-sea weak echo sonar generation training dataset , the deep-sea weak echo sonar generation training dataset consists of standardized sonar tensor and environmental parameter vector:

[0098] ;

[0099] in, Indicates the Normalized sonar tensor samples, dimension , is the time domain length, is the frequency resolution; is the environmental parameter vector, respectively, depth ,salinity , temperature gradient and seabed types ; is the total number of training samples;

[0100] S22. Constructing an improved StyleGAN-3 model with adaptive spectral energy-phase joint constraints , improve the StyleGAN-3 model to integrate the space-time separable convolution structure and the complex domain residual gating structure in the feature generation network;

[0101] S23. Combining the target energy spectrum and the background noise power spectrum density, a dynamic energy-phase joint adaptive loss function is constructed to constrain the physical consistency of the generated samples in terms of spectral energy distribution and phase continuity:

[0102] ;

[0103] in, and are the normalized energy spectra of generated and real weak echo samples, and are the phase matrices of the generated and real weak echo samples, 、 is the weight coefficient, and j represents the imaginary unit.

[0104] Aiming at the physical nature of deep-sea weak echo sonar targets, the present invention proposes an energy-phase joint adaptive loss function. Its core idea is to explicitly introduce dual constraints on the energy spectrum characteristics and phase structure characteristics of the sonar target signal on the basis of the traditional generative adversarial network loss. The energy spectrum part is used to measure the similarity between the generated weak echo samples and the real samples in the frequency domain energy distribution. This part is described by the mean square error between the normalized energy spectra, which can constrain the generated data to be close to the real deep-sea target signal in frequency energy distribution, ensuring that the weak but real spectral energy peak of the weak target can be effectively reproduced and captured. The phase structure part is aimed at the multipath interference and phase distortion phenomena in the underwater acoustic propagation process. The Frobenius norm between the complex domain phase matrices is used as a measurement standard to explicitly constrain the generated samples and the measured samples to maintain a high degree of consistency at the microscopic level such as phase continuity and phase perturbation, thereby ensuring that the propagation law of the real sonar echo is consistent at the physical level.

[0105] The energy-phase joint adaptive loss function precisely aligns the subtle energy peaks and complex phase fluctuations of weak target signals in extremely low signal-to-noise ratio environments. The generated training samples not only improve the subsequent detection network's sensitivity to true weak targets, but also effectively suppress false signals and noise misjudgments. This addresses the existing problems of detector overfitting and reduced generalization capability caused by sample scarcity and imbalanced sample distribution. By dynamically adjusting energy and phase weights, the system can actively respond to changes in target and noise fields in different deep-sea sonar environments, achieving greater universality and environmental adaptability.

[0106] S24. In the improved StyleGAN-3 training process, the sample separability scores output by the discriminator based on the sonar detection network are introduced in real time As the adversarial feedback term of the joint loss function, the generator parameters are dynamically adjusted to form a generation-discrimination collaborative feedback mechanism. The joint loss function is:

[0107] ;

[0108] in, Generate adversarial loss for the standard, To determine the feedback weight, Characterizes the sample separability score of the generated weak echo sample that is correctly identified as a weak target by the current sonar detection network;

[0109] The present invention proposes a highly adaptive joint loss function design for the complex acoustic environment of the deep sea. Its core principle is that it not only imposes physical consistency constraints on the spectral energy and phase continuity of generated weak target echoes and background noise, but also introduces the sample separability score output by the detection network as a dynamic feedback indicator for generative model training, thereby constructing an end-to-end optimization mechanism with deep coupling between generation and detection tasks.

[0110] Based on the traditional StyleGAN-3 loss term, the joint loss function additionally introduces an energy spectrum difference term and a phase consistency term. The energy spectrum difference term constrains the distance in energy distribution between the generated samples and the measured weak target samples, ensuring the authenticity of the generated samples in the acoustic physical feature space; the phase consistency term imposes a Frobenius norm penalty on the complex domain phase matrix, so that the generated samples retain a continuous phase structure consistent with the real echo in specific deep-sea scenarios such as multipath interference and non-stationary noise. This breaks through the limitation of traditional GAN ​​that is only constrained by pixel / amplitude loss, and effectively enhances the signal-level availability and engineering transferability of synthetic data.

[0111] The joint loss function further incorporates the sample separability score output by the sonar detection network. The sample separability score reflects the probability that each generated sample will be correctly identified as an object (or non-object) after passing through the current detection network. Specifically, the separability score is constructed by combining the detection network's confidence in the generated sample's output and the degree of match with the annotation. A higher separability score indicates that the generated sample has greater discriminability and training value for the detection network. By incorporating the separability score into the loss function, the generative model not only improves physical consistency but also adaptively adjusts the generated distribution for downstream detection tasks, achieving the co-evolution of data synthesis and detection model performance.

[0112] The joint loss function couples sonar physical constraints, signal level indicators and detection network scores for the first time, thus achieving three breakthroughs: (1) At the data generation level, the generated samples not only pursue appearance similarity, but also ensure the authenticity and availability of key physical features in target detection; (2) In the model training mechanism, an information closed loop between generation, discrimination and detection is established, breaking the traditional isolated optimization mode of generation network and detection network, so that sample generation can directly optimize the detector performance and adaptively evolve; (3) The overall scalability and robustness of the system are improved. The sample separability score serves as a dynamic feedback signal, which effectively alleviates the problems of extreme imbalance of actual samples and severe noise interference, providing a sustainable evolution data and model foundation for complex deep-sea sonar target detection applications.

[0113] S25. Using the training dataset End-to-end training improves the StyleGAN-3 model During the training process, the parameters are adaptively optimized according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism, and the improved StyleGAN-3 model is finally trained. , the improved StyleGAN-3 model dynamically generates weak target echo samples, seabed reverberation samples and environmental background noise samples whose spectral structures and phase characteristics are highly consistent with the measured ones under different deep-sea environmental parameter vector conditions.

[0114] The present invention adopts an improved StyleGAN-3 generation network that introduces space-time separable convolution and complex domain residual gating structure, combines the energy-phase dual-constraint loss function, and optimizes feedback coupling with the detection network. It can dynamically generate high-fidelity, physically consistent weak target echoes, seabed reverberation and background noise samples according to the sea depth, salinity, temperature gradient and seabed type environmental parameters. By introducing the target energy spectrum and background noise power spectral density to adaptively adjust the synthesis ratio, the domain deviation problem between traditional simulation samples and measured samples is significantly alleviated, which greatly enriches the training data type and achieves balanced coverage of samples in different scenarios.

[0115] In this implementation, the improved StyleGAN-3 model includes a space-time separable convolutional structure and a complex domain residual gating structure:

[0116] Space-time separable convolutional architecture to normalize sonar tensors As input, the one-dimensional convolution operation Conv1D is performed along the time dimension in combination with the time direction convolution kernel to extract the time direction convolution feature tensor of deep-sea weak targets in the time direction. :

[0117] ;

[0118] in, Represents a one-dimensional convolution operation in the time dimension; is the time direction convolution kernel, is the time window length; it is used to identify the short-term energy fluctuation characteristics of weak targets. is the number of feature channels;

[0119] The time-direction convolution feature tensor is input into the one-dimensional convolution layer in the frequency dimension, and the one-dimensional convolution operation is performed in the frequency dimension in combination with the frequency-direction convolution kernel to extract the time-frequency joint convolution feature tensor of the weak target signal under the influence of Doppler diffusion and seabed scattering. :

[0120] ;

[0121] in, represents a one-dimensional convolution operation along the frequency dimension, is the frequency direction convolution kernel, is the frequency window length, which is used to extract the spectral structure details of deep-sea weak targets under the influence of Doppler spread and seabed scattering;

[0122] The time-frequency joint convolution feature tensor is input into the complex domain residual gating structure to establish the amplitude feature branch and the phase feature branch respectively:

[0123] The amplitude feature branch extracts the local amplitude information of the time-frequency joint convolution feature tensor through an amplitude convolution channel. The extracted result is batch normalized and then connected to the rectified linear unit function for nonlinear activation to obtain the amplitude feature tensor. , the amplitude characteristic tensor is used to characterize the energy fluctuation pattern of deep-sea weak targets in the local time-frequency window;

[0124] The phase feature branch extracts the phase change information of the time-frequency joint convolution feature tensor through another phase convolution channel. The extracted result is also batch normalized and then connected to the hyperbolic tangent function for nonlinear activation to obtain the phase feature tensor. , used to characterize the phase delay characteristics of deep-sea weak targets caused by multipath interference;

[0125] Based on the time-frequency joint convolution feature tensor, a residual signal gating channel is constructed to dynamically control the degree of fusion of target features and background information. The residual signal gating channel extracts the gated feature map through a separate gated channel convolution, and compresses the gated weight value to the [0,1] interval through the activation function, and finally outputs the gated weight tensor. ,Each element of the gating weight tensor is used to measure the strength of the residual signal that should be retained at the current position. The larger the gating weight, the more dependent the target information synthesized by the complex domain features, and the smaller the gating weight, the more input background features are retained;

[0126] The magnitude eigentensor , phase eigentensor With the gated weight tensor Fusion is performed to form the final output complex domain residual feature tensor :

[0127] ;

[0128] in, Represents the Hadamard product, the first term is the explicit target feature after phase modulation, and the second term is the spectral background structure retained in the residual channel, ensuring that the generated samples contain both weak target intrinsic features and are adaptable to noise suppression.

[0129] In modeling weak deep-sea sonar echo signals, target signals are often affected by multipath, scattering, and phase disturbances, resulting in complex variations in amplitude and phase. Traditional GANs and their variants generally process signal features in the real domain, making it difficult to accurately describe the phase information contained in sonar echoes. S23 introduces a complex-domain residual gating structure to dynamically fuse the amplitude and phase feature tensor branches with gating weights, outputting a complex-domain residual feature tensor that accurately restores the energy fluctuations of weak targets while effectively describing the phase continuity under multipath interference. The residual gating mechanism automatically adjusts the fusion weights of target features and background noise, taking into account both the physical characteristics of the signal and the diversity of data distribution.

[0130] Introducing the complex domain residual structure into the sonar generative adversarial model transcends the limitations of existing real-valued space feature extraction, achieves a high degree of expression of the physical intrinsic characteristics of deep-sea weak echoes, and greatly improves the realism and physical consistency of the generated samples.

[0131] In this embodiment, S3 includes the following steps:

[0132] S31. Call the trained improved StyleGAN-3 model , input the environmental parameter vector of the sample to be generated , combined with random latent variables , generate synthetic sample tensors under corresponding conditions :

[0133] ;

[0134] in, Indicates depth ,salinity , temperature gradient , seabed type The standardized sonar tensor generated under the constructed environmental conditions includes a composite structure of weak target echoes, seabed reverberation, and environmental background noise;

[0135] S32. Collect and generate measured deep-sea sonar tensors corresponding to samples , and estimate the weak target energy spectrum of the target signal in real time based on the target energy extraction function and the noise spectrum estimation function and background noise power spectral density , the target energy extraction function extracts the weak target energy spectrum based on the local energy peak and envelope fluctuation, and the noise spectrum estimation function estimates the background noise power spectrum density using spectrum smoothing and spectrum valley matching;

[0136] S33. Constructing dynamic fusion ratio coefficient based on weak target energy spectrum and background noise power spectrum density , the fusion ratio is used to control the superposition intensity of the generated samples and the measured data:

[0137] ;

[0138] in, Represents the L2 norm; proportional coefficient Controls the injection intensity of generated samples. The lower the background noise power spectrum density is, the dynamic fusion ratio coefficient is. Tends to 1 to enhance the target signal. The stronger the background noise power spectrum density is, the higher the dynamic fusion ratio coefficient is. tends to 0;

[0139] In deep-sea environments, the energy and noise power spectral densities of weak targets vary dramatically. To avoid domain shifts caused by fixed ratios, this paper proposes dynamically calculating the fusion ratio coefficient based on the sample energy and noise spectra. This adaptively adjusts the mixing ratio of synthetic and measured samples based on the relative intensities of the target and noise, enabling the training dataset to dynamically cover the complex range of variations in actual application scenarios. This breaks the rigid model of manually or empirically setting sample fusion ratios in existing technologies, enabling data-level adaptation to changes in the physical environment, ensuring consistency between the distribution of training samples and the test environment, and effectively alleviating the problem of insufficient model generalization.

[0140] S34. According to the dynamic fusion ratio coefficient For synthetic sample tensors and the measured sonar tensor Perform linear fusion to construct synthetic-measured mixed data tensor :

[0141] ;

[0142] The synthetic-measured hybrid data tensor retains both the measured background noise distribution and the weak features of the synthetic weak target;

[0143] S35. Synthetic-measured mixed data tensor Perform weak echo enhancement processing, use multi-scale convolution structure to build convolution pyramid, and extract multi-scale joint convolution feature tensors of weak targets at different time and frequency scales ;

[0144] S36. Input the multi-scale joint convolution feature tensor into the time-frequency attention module, and use the time attention weight and frequency attention weights Perform weighted fusion in the time dimension and frequency dimension to form the final weak target feature tensor ;

[0145] ;

[0146] in, , represents the time index and frequency index; the final weak target feature tensor represents the fused deep-sea weak target echo characteristics, and has the ability of multi-scale information aggregation and noise suppression.

[0147] In this embodiment, S4 includes the following steps:

[0148] S41. Building a sonar detection network , with weak target feature tensor As input, the target annotation information in the synthetic-measured mixed dataset As the supervision signal, the cross entropy loss function is used Training sonar detection network parameter set :

[0149] ;

[0150] in, Represents the sonar detection network at time index With frequency index The predicted value of the probability of weak target existence at the position, the cross entropy loss is used to fit the difference between the output of the sonar detection network and the annotation information, T represents the number of discrete sampling frames of the sonar data sample on the time axis, and F represents the number of discrete sampling points of the sonar data sample on the frequency axis;

[0151] S42. Establish a marine predator optimization algorithm framework based on weak target feature tensors and sonar detection network parameter sets Under the framework of the marine predator optimization algorithm, a global optimization is performed on the feature fusion weights, decision thresholds and filter parameters of the sonar detection network to initialize the marine predator group parameter set. , is the number of individuals in the group, and the detection parameter vector corresponding to each individual includes feature fusion weight, decision threshold and filter parameters;

[0152] S43. Define the evaluation function based on the difference in weak target energy spectrum, detection rate and false alarm rate as indicators , used to measure the detection performance of the sonar detection network under the current detection parameter configuration:

[0153] ;

[0154] in, To measure the energy spectrum of weak targets, To detect the parameters of the network The detected target energy spectrum output below is For detection parameters The detection rate under For detection parameters The false alarm rate under is the balance coefficient.

[0155] The evaluation function designed in this paper integrates three core metrics: target energy spectrum difference, detection rate, and false alarm rate. This function not only focuses on the physical accuracy of signal recovery through detection parameters, but also considers the actual detection capability and false alarm control of the detection network. Through weighted combination, the overall performance of the multi-target quantified detection system under different parameter configurations is evaluated, providing a comprehensive performance evaluation basis for optimization algorithms.

[0156] The evaluation function combines physical consistency with engineering performance indicators, can dynamically balance different optimization objectives, adapt to the detection-false alarm trade-off requirements in complex sea areas, and provide swarm intelligence algorithms with a global optimal search guide for practical engineering, which is different from the traditional evaluation system dominated by single precision or simple accuracy.

[0157] In this embodiment, S5 includes the following steps:

[0158] S51. Algorithm framework for optimizing marine predators In the process, the parameter set of the marine predator group is initialized, and each iteration After that, for each individual based on the current detection parameters , calculate its evaluation function value ;

[0159] S52. For each individual, based on the current synthetic-measured mixed data tensor , target signal energy spectrum and background noise power spectral density , calculate the individual adaptability confidence :

[0160] ;

[0161] in, To detect the parameters of the network The detected target energy spectrum output below is Indicates the sample expectations, is the L2 norm;

[0162] During the optimization process, the individual adaptability confidence is dynamically calculated based on the ratio of the recovery error of each individual detection parameter to the target energy spectrum and the actual signal-noise energy distribution of the sample. The individual adaptability confidence reflects the individual's adaptability to target modeling under the current mixed data distribution, and automatically drives the individual to the global search or local fine optimization stage.

[0163] This method realizes the adaptive division of signal-noise scenarios of individual behaviors in the optimization process, breaking through the traditional swarm intelligence algorithm stage allocation method that relies solely on rounds or static thresholds. It enables individuals in the group to always dynamically divide the work according to their own abilities, thus improving the overall optimization efficiency and the ability to adapt to complex environments.

[0164] S53. Setting the individual adaptability confidence threshold , divide the group into global search subgroups and local optimization subgroups ;

[0165] S54. For individuals in the global search subgroup, a deep-sea adaptive Levy flight step-size operator is used. Perform global search strategy updates:

[0166] ;

[0167] in, For the Individuals in The detection parameter vector after rounds of iteration, all parameters are targeted at the current deep-sea weak echo sonar detection task, For the Individuals in The detection parameter vector at the round iteration, For the The individual detection parameter vector with the smallest evaluation function value among all individuals during the round iteration, is a random variable, is the passive tracking weight factor, is the active disturbance weight factor;

[0168] S55. For individuals in the local optimization subgroup, a group collaborative optimization strategy is adopted, integrating adaptive collaboration factors to fine-tune individual parameters:

[0169] ;

[0170] in, Before the evaluation function value Advantageous individuals, is the adaptive collaboration factor with the number of iterations, The top ranking of the evaluation function value in the group in each iteration The number of dominant individuals;

[0171] The global search formula is based on the deep-sea adaptive Lévy flight step-size operator, and uses the individual's historical performance and target energy complexity to dynamically adjust the step size and perturbation direction, allowing individuals to make long-distance jumps in the parameter space and explore unknown extreme areas. The collaborative optimization formula uses the collaboration factor as the weight to integrate and adjust individuals towards the current dominant subgroup and the global optimal individual, thereby enhancing the local convergence and collaborative efficiency of high-confidence individuals.

[0172] Different from existing MPA, PSO and other algorithms that only rely on passive or active search, the present invention is based on a triple fusion mechanism of signal adaptability + complex dynamic perturbation + group collaboration, ensuring the adaptive global optimization and fine local convergence of optimal detection parameters in a non-stationary deep-sea environment with drastic fluctuations in the signal-to-noise ratio, greatly improving the optimization robustness and convergence speed.

[0173] S56. After each round of iteration, update the global optimal detection parameter individual and its evaluation function value , perform a difference operation on the evaluation function value of the global optimal detection parameter individual and the evaluation function value of the global optimal detection parameter individual in the previous iteration. When the absolute value of the difference is less than the preset convergence threshold, terminate the optimization process;

[0174] S57. When the convergence conditions are met, the globally optimal detection parameter individual obtained in the convergence round is used as the optimal detection parameter set. The optimal detection parameter set is used to configure the final sonar detection network and drive the subsequent target judgment and tracking modules. The optimal detection parameter set represents the detection parameter configuration with the best global detection performance in the current deep-sea weak echo sonar environment after the optimization is completed.

[0175] In this embodiment, the deep-sea adaptive Levy flight step-size operator Based on the target energy feature complexity index in the deep sea weak echo target detection scenario Dynamically adjust step size distribution parameters:

[0176] ;

[0177] in, Characterization Individuals in The current detection parameter vector in iteration The corresponding target energy feature complexity index is calculated based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size adaptation factor, is the step stability index, is a stable exponential perturbation term, which represents the normal random variable randomly sampled from the standard normal distribution by the kth individual in the tth iteration, and is used to control the thickness of the distribution tail of the Lévy jump amplitude. is a random perturbation factor, which represents the normal random variable obtained by random sampling from the standard normal distribution of the kth individual in the tth iteration, and is used to control the scale change direction and perturbation amplitude of the Lévy step.

[0178] The present invention proposes an ocean predator optimization algorithm that combines the target signal energy spectrum and the noise power spectrum density with adaptive grouping. For group individuals with different adaptive confidence levels, global search of deep-sea adaptive Levy flight step and local fine optimization of group collaborative optimization are adopted respectively. It can automatically adjust the feature fusion weights, decision thresholds and filtering parameters of the detection network in highly non-stationary noise and signal dynamic change environments. In the context of extremely low signal-to-noise ratio and Doppler spread, the system false alarm rate is significantly reduced, and the detection rate and parameter stability are significantly better than the comparative technology.

[0179] In this embodiment, S6 includes the following steps:

[0180] S61. Load the optimal detection parameter set into the sonar detection network , for the input weak target feature tensor Make target judgment and generate the corresponding weak target response tensor , where each element Indicates the time index With frequency index The confidence probability of detecting a weak target at ;

[0181] S62. Based on weak target response tensor Extract target confidence value ;

[0182] S63. Based on weak target response tensor The set of time-frequency locations with significant response probability in , combined with the corresponding synthetic-measured mixed data tensor The sensor array element information and sound propagation model parameters in the 3D space are used to calculate the target positioning and output the target positioning results. ,in, is the estimated horizontal distance to the target, is the estimated azimuth angle of the target, is the estimated depth of the target;

[0183] S64. Based on the target positioning results, a target space motion trajectory model is constructed, combining the multi-frame weak target feature tensor in the continuous time window The corresponding detection response tensor , perform temporal confidence-weighted trajectory association:

[0184] If continuous Existing in the frame The frame detection result satisfies , it is determined to be the same target track;

[0185] If the spatial distance between trajectory points changes , then linear interpolation smoothing is performed;

[0186] Output target motion trajectory as a sequence ;

[0187] in, is the confidence threshold for trajectory determination, is the maximum allowable jump distance, is the number of trajectory sliding window frames.

[0188] In this embodiment, the target confidence value is defined by the detection response intensity and the background response difference:

[0189] when , determine the existence of the target, and output the target confidence value ;

[0190] when , it is judged as not detected, and the target confidence value is recorded as 0;

[0191] in, is the target confidence threshold, which is set based on the optimal strategy based on experience in the training set.

[0192] Example 1

[0193] In the northern deepwaters (xx.xx°E, xx.xx°N), a research vessel deployed a 128-element broadband phased array sonar system approximately 180 kilometers from the coastline. The system operated at a 48kHz frequency and targeted a depth range of 2,450 to 2,820 meters. The sea state at the time was Class 2, with a peak background noise power spectral density of approximately 31.5dB / μPa² / Hz.

[0194] During the flight between 2:00 AM and 4:30 AM that day, the sonar system continuously collected six hours of raw channel data. The online detection algorithm initially identified 22 weak target events, 56 sets of measured weak echo samples, and 154 sets of ambient noise samples. Using traditional empirical model-based data augmentation, the research team relied on the Poisson scattering and Gaussian white noise superposition method to synthesize only 94 sets of weak echo training samples. The mean square error of the energy spectrum of the synthesized samples deviated from the measured mean by 0.23.

[0195] During the same time period, using the method of the present invention, the implementers constructed feature vectors based on the environmental parameters of 2450-2820 meters depth, 34.5ppt salinity, 1.9°C / 100m temperature gradient, and silt bottom type. The preprocessed sonar tensor and conditional parameters were input into the improved StyleGAN-3 generator. Within 3 minutes, the system automatically generated 920 groups of weak target echo samples with a high degree of match with the actual measurements, with a spectral structure mean square error of only 0.12, and covering multiple groups of typical biological noise and multipath reverberation scenarios. Each group of samples retained the corresponding synthetic environmental parameters and target labels to facilitate subsequent training and detailed evaluation.

[0196] At 3:25 a.m., the research vessel's detection system detected a suspected micro-target event with a signal-to-noise ratio of only -16.7dB. Traditional fusion methods linearly mix synthetic samples with actual measurements at a ratio of 1:1, resulting in a detection rate of 47.8% for extremely low signal-to-noise ratio samples after training. The method of the present invention automatically adjusts the fusion ratio based on the current target energy spectrum and noise power spectrum density, with the proportion of measured samples increasing to 36% and the proportion of synthetic samples increasing to 64%. After processing with a multi-scale convolutional pyramid and a time-frequency attention mechanism, the output weak target feature tensor clearly retains the subtle fluctuations of the target echo, and the noise component is significantly reduced.

[0197] At 3:48 am, the implementer invoked the marine predator optimization algorithm of the present invention to optimize the feature fusion weight, decision threshold and filter parameters of the detection network online. Assuming that the mean square error between the target energy spectrum and the true energy spectrum detected by parameter individual No. 12 at this moment is 0.073, and the mean square value of the noise spectrum is 0.132, after confidence calculation The system automatically assigns the individual to the global search subgroup, the Levy flight step is dynamically adjusted to 0.018, the confidence of individual No. 3 is as high as 0.74, and it is included in the local collaborative optimization subgroup. The weight adjustment speed is halved, and the collaboration factor .

[0198] Under the traditional method, at the same time, the parameters of individual No. 7 are updated using a static genetic algorithm. The mean square error of the target energy spectrum is 0.127. The detector's judgment confidence for target signals with extremely low signal-to-noise ratio is mostly between 0.51 and 0.58, and the false alarm rate is 28.3%.

[0199] At 4:02 am, after loading the optimal detection parameter set, the sonar detection network judged the collected weak target feature tensor frame by frame. In a certain frame (frame number #1298), the detection network Output target confidence , exceeds the decision threshold , and determine the presence of a weak target. Combined with the array data back-projection algorithm, the system outputs the spatial coordinates in real time , the judgment remains true for 14 consecutive frames, and the track tracking algorithm generates a motion track based on it and achieves clustering with the trajectory of the previous target within 5 minutes.

[0200] In the same frame of the control group, the confidence of the traditional detection network was only 0.62, which did not reach the judgment threshold of 0.70, resulting in the target being missed.

[0201] Throughout Example 1, the method of the present invention detected a total of 28 weak target events (target reflection cross section 0.4-1.0 m², signal-to-noise ratio -18.2-13.5 dB), with 4 false alarms and 2 consecutive track losses. This compares to 19 targets detected, 11 false alarms, and 7 track losses using the traditional method.

[0202] Taking the statistical results from 2:00-6:00 in the morning as an example: the mean square error of three-dimensional positioning of the method of the present invention is 9.8 meters, while that of the traditional method is 20.6 meters; the average detection confidence is 0.86, while that of the traditional method is 0.67; the target detection rate is increased to 76.1%, while that of the traditional method is only 53.9%; the false alarm rate is reduced to 12.5%, while that of the traditional method is 36.7%.

[0203] The method of the present invention effectively realizes high-fidelity weak target sample generation, online adaptive parameter optimization, three-dimensional target judgment, and continuous track tracking in an extremely low signal-to-noise ratio and deep-sea non-stationary noise environment, greatly improving the practicality and robustness of the detection system in extreme scenarios, and is significantly superior to existing traditional sonar target detection technology.

[0204] This invention achieves high-precision integrated output of target confidence, spatial positioning, and trajectory extraction in a sonar detection network through dynamic loading of the optimal detection parameter set and a joint time-frequency attention mechanism. Dynamic judgment is performed based on multi-frame, multi-scale, and multi-frequency weak target feature tensors. Combined with back-projection and spatiotemporal trajectory clustering algorithms, this method effectively improves the three-dimensional spatial resolution of weak targets and the continuity of multi-target tracks in complex deep-sea environments. Actual engineering tests have shown that in deep-sea scenarios, the positioning error and target trajectory loss rate of this invention are significantly reduced compared to traditional methods. The method possesses excellent online tracking and dynamic situational awareness capabilities, providing solid technical support for marine monitoring, security, and emergency command applications.

[0205] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for detecting submarine targets by sonar based on an adversarial network, characterized in that: The steps include: S1. Collecting raw deep-sea sonar data, and performing time-domain pulse compression filtering, matched filtering, and normalization based on noise statistical characteristics on the raw deep-sea sonar data to obtain a standardized sonar tensor; S2. Using standardized sonar tensors as training input, combined with environmental parameter vectors consisting of depth, salinity, temperature gradient, and seafloor type, we trained an improved StyleGAN-3 model to construct a dedicated generative network for sonar scenarios. The improved StyleGAN-3 model introduced space-time separable convolution and complex-domain residual gating structures in the feature mapping layer to enhance the representation of phase continuity and spectral structure of weak echo signals. S3. Call the improved StyleGAN-3 model to generate weak target echo samples, seabed reverberation samples, and ambient background noise samples based on the environmental parameter vector to form a generated sample set. This generated sample set is proportionally fused with the raw deep-sea sonar data. The fusion ratio is dynamically adjusted based on the difference in power spectrum density between the target energy spectrum and the background noise. A synthetic-measured hybrid dataset is constructed. Weak echo enhancement is performed on the synthetic-measured hybrid dataset. Weak target features are extracted using a multi-scale convolutional structure and a time-frequency attention mechanism, and the weak target feature tensor is output. S4. Construct a sonar detection network, using the weak target feature tensor as input and the target annotation information in the synthetic-measured hybrid dataset as the supervision signal, to obtain the set of detection parameters to be optimized. Based on the weak target feature tensor and the set of detection parameters to be optimized, establish a marine predator optimization algorithm framework and define an evaluation function with the core indicators of weak target energy spectrum difference, detection rate, and false alarm rate. The target label information is obtained by manually or semi-automatically annotating the weak target's category, spatial location, or presence in the synthetic-measured mixed data set after weak echo enhancement processing; S5. Within the framework of the marine predator optimization algorithm, perform a global search and iterative optimization of the feature fusion weights, decision thresholds, and filter parameters in the detection parameter set. Dynamically update the search strategy based on the evaluation function value until the evaluation function converges and output the optimal detection parameter set. S6. Load the optimal detection parameter set into the sonar detection network, perform target judgment on the weak target feature tensor, and output the target confidence value, three-dimensional spatial positioning result and target motion track.

2. The method for detecting submarine targets by sonar based on an adversarial network according to claim 1, wherein: The S2 comprises the following steps: S21. Constructing a deep-sea weak echo sonar generation training dataset ,The deep-sea weak echo sonar generation training dataset consists of standardized sonar tensors and environmental parameter vectors, where the environmental parameter vectors are depth, salinity, temperature gradient and seabed type; S22. Constructing an improved StyleGAN-3 model with adaptive spectral energy-phase joint constraints , improve the StyleGAN-3 model to integrate the space-time separable convolution structure and the complex domain residual gating structure in the feature generation network; S23. Combine the target energy spectrum and background noise power spectrum density to dynamically construct the energy-phase joint adaptive loss function ; S24. In the improved StyleGAN-3 training process, the sample separability scores output by the discriminator based on the sonar detection network are introduced in real time As a joint loss function The adversarial feedback item dynamically adjusts the generator parameters to form a generation-discrimination collaborative feedback mechanism; S25. Using the training dataset End-to-end training improves the StyleGAN-3 model During the training process, the parameters are adaptively optimized according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism, and the improved StyleGAN-3 model is finally trained. .

3. The method for detecting submarine targets by sonar based on an adversarial network according to claim 2, wherein: The improved StyleGAN-3 model includes a space-time separable convolutional structure and a complex domain residual gating structure: Space-time separable convolutional architecture to normalize sonar tensors As input, the one-dimensional convolution operation Conv1D is performed along the time dimension in combination with the time direction convolution kernel to extract the time direction convolution feature tensor of deep-sea weak targets in the time direction. ; The time-direction convolution feature tensor is input into the one-dimensional convolution layer in the frequency dimension, and the one-dimensional convolution operation is performed in the frequency dimension in combination with the frequency-direction convolution kernel to extract the time-frequency joint convolution feature tensor of the weak target signal under the influence of Doppler diffusion and seabed scattering. ; The time-frequency joint convolution feature tensor is input into the complex domain residual gating structure to establish the amplitude feature branch and the phase feature branch respectively: The amplitude feature branch extracts the local amplitude information of the time-frequency joint convolution feature tensor through an amplitude convolution channel. The extracted result is batch normalized and then connected to the rectified linear unit function for nonlinear activation to obtain the amplitude feature tensor. ; The phase feature branch extracts the phase change information of the time-frequency joint convolution feature tensor through another phase convolution channel. The extracted result is also batch normalized and then connected to the hyperbolic tangent function for nonlinear activation to obtain the phase feature tensor. ; Based on the time-frequency joint convolution feature tensor, a residual signal gating channel is constructed to dynamically control the degree of fusion of target features and background information. The residual signal gating channel extracts the gated feature map through a separate gated channel convolution, and compresses the gated weight value to the [0,1] interval through the activation function, and finally outputs the gated weight tensor. ; The magnitude eigentensor , phase eigentensor With the gated weight tensor Fusion is performed to form the final output complex domain residual feature tensor .

4. The method for detecting submarine targets by sonar based on an adversarial network according to claim 2, wherein: The S3 includes the following steps: S31. Call the trained improved StyleGAN-3 model , input the environmental parameter vector of the sample to be generated, combined with the random latent variable , generate synthetic sample tensors under corresponding conditions ; S32. Collect and generate measured deep-sea sonar tensors corresponding to samples , and estimate the weak target energy spectrum of the target signal in real time based on the target energy extraction function and the noise spectrum estimation function and background noise power spectral density ; S33. Constructing dynamic fusion ratio coefficient based on weak target energy spectrum and background noise power spectrum density ,The fusion ratio is used to control the superposition intensity of the generated samples and the measured data; S34. According to the dynamic fusion ratio coefficient For synthetic sample tensors and the measured sonar tensor Perform linear fusion to construct synthetic-measured mixed data tensor ; S35. Synthetic-measured mixed data tensor Perform weak echo enhancement processing, use multi-scale convolution structure to build convolution pyramid, and extract multi-scale joint convolution feature tensors of weak targets at different time and frequency scales ; S36. Input the multi-scale joint convolution feature tensor into the time-frequency attention module, and use the time attention weight and frequency attention weights Perform weighted fusion in the time dimension and frequency dimension to form the final weak target feature tensor .

5. The method for detecting submarine targets by sonar based on adversarial network according to claim 4, characterized in that: The S4 comprises the following steps: S41. Building a sonar detection network , with weak target feature tensor As input, the target annotation information in the synthetic-measured mixed dataset As the supervision signal, the cross entropy loss function is used Training sonar detection network parameter set ; S42. Establish a marine predator optimization algorithm framework based on weak target feature tensors and sonar detection network parameter sets Under the framework of the marine predator optimization algorithm, a global optimization is performed on the feature fusion weights, decision thresholds and filter parameters of the sonar detection network to initialize the marine predator group parameter set. , is the number of individuals in the group; S43. Define the evaluation function based on the difference in weak target energy spectrum, detection rate and false alarm rate as indicators , which is used to measure the detection performance of the sonar detection network under the current detection parameter configuration.

6. The method for detecting submarine targets by sonar based on adversarial network according to claim 1, characterized in that: The S5 comprises the following steps: S51. Algorithm framework for optimizing marine predators In the process, the parameter set of the marine predator group is initialized, and each iteration After that, for each individual based on the current detection parameters , calculate the evaluation function value ; S52. For each individual, based on the current synthetic-measured mixed data tensor , target signal energy spectrum and background noise power spectral density , calculate the individual adaptability confidence ; S53. Setting the individual adaptability confidence threshold , divide the group into global search subgroups and local optimization subgroups ; S54. For individuals in the global search subgroup, a deep-sea adaptive Levy flight step-size operator is used. Perform global search strategy updates ; S55. For individuals in the local optimization subgroup, a group collaborative optimization strategy is adopted to integrate adaptive collaborative factors and fine-tune individual parameters. ; S56. After each round of iteration, update the global optimal detection parameter individual and its evaluation function value , perform a difference operation on the evaluation function value of the global optimal detection parameter individual and the evaluation function value of the global optimal detection parameter individual in the previous iteration. When the absolute value of the difference is less than the preset convergence threshold, terminate the optimization process; S57. When the convergence condition is met, the globally optimal detection parameter individual obtained in the convergence round is used as the optimal detection parameter set.

7. The method for detecting submarine targets by sonar based on adversarial network according to claim 6, characterized in that: The deep-sea adaptive Levy flight step-size operator Based on the target energy feature complexity index in the deep sea weak echo target detection scenario Dynamically adjust step size distribution parameters: ； in, Characterization Individuals in The current detection parameter vector in iteration The corresponding target energy feature complexity index is calculated based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size adaptation factor, is the step stability index, is a stable exponential perturbation term, which represents the normal random variable obtained by random sampling from the standard normal distribution of the kth individual in the tth iteration, and is used to control the thickness of the distribution tail of the Levy jump amplitude. is a random perturbation factor, which represents the normal random variable obtained by random sampling from the standard normal distribution of the kth individual in the tth iteration, and is used to control the scale change direction and perturbation amplitude of the Levy step.

8. The method for detecting submarine targets by sonar based on adversarial network according to claim 1, characterized in that: The S6 comprises the following steps: S61. Loading the optimal detection parameter set into the sonar detection network , for the input weak target feature tensor Make target judgment and generate the corresponding weak target response tensor ; S62. Based on weak target response tensor Extract target confidence value ; S63. Based on weak target response tensor The set of time-frequency locations with significant response probability in , combined with the corresponding synthetic-measured mixed data tensor The sensor array element information and sound propagation model parameters in the system are used to perform three-dimensional spatial positioning calculations and output the target positioning results. S64. Based on the target positioning results, a target space motion trajectory model is constructed, combining the multi-frame weak target feature tensor in the continuous time window The corresponding detection response tensor ,perform temporal confidence weighted trajectory association.

9. The method for detecting submarine targets by sonar based on adversarial network according to claim 8, characterized in that: The target confidence value is defined by the detection response strength and the background response difference: when , determine the existence of the target, and output the target confidence value ; when , it is judged as not detected, and the target confidence value is recorded as 0; in, is the target confidence threshold, which is set based on the empirically optimal strategy on the training set.

Images