Submarine target sonar detection method based on adversarial network

Abstract

The invention discloses a seabed target sonar detection method based on an adversarial network, and the method comprises the steps: S1, collecting deep sea sonar original data, and carrying out the preprocessing of the data, and obtaining a standardized sonar tensor; s2, taking the standardized sonar tensor as training input, and training an improved StyleGAN-3 model in combination with an environmental parameter vector; s3, calling the improved StyleGAN-3 model, and outputting a weak target feature tensor; s4, constructing a sonar detection network; s5, outputting an optimal detection parameter set; and S6, loading the optimal detection parameter set to the sonar detection network, performing target judgment on the weak target feature tensor, and outputting a target confidence value, a three-dimensional space positioning result and a target motion track. Compared with a traditional method, the method has the advantages that in a deep sea scene, the positioning error and the target trajectory loss rate are greatly reduced, the method has excellent online tracking and dynamic situation awareness capabilities, and a solid technical support is provided for ocean monitoring, security and emergency command application.

Description

Technical Field

[0001] The present invention relates to the technical field of sonar detection, and in particular, to a method for detecting underwater targets by sonar based on a confrontation network. Background Art

[0002] With the continuous development of deep-sea exploration and intelligent sonar technologies, sonar systems play an increasingly important role in marine environmental monitoring, deep-sea resource exploration, and underwater security fields. Detecting the echo signals of weak targets in complex deep-sea environments has become one of the core problems restricting the intelligent level of underwater sonars. Most of the existing deep-sea weak echo target detection technologies rely on traditional signal processing means, and generally have problems such as low sensitivity, high false alarm rate, and serious parameter mismatch in extremely low signal-to-noise ratio, non-stationary noise, and complex multipath interference environments.

[0003] Traditional deep-sea weak target detection algorithms are limited by the difficulties in obtaining samples, scarce data, and extremely unbalanced distribution. Due to the complex deep-sea experimental conditions and the influence of sea conditions and confidentiality factors, the actual number of weak echo target samples that can be obtained is extremely small, resulting in the difficulty of fully training data-driven detection models based on deep learning, and the detectors are prone to overfitting and the problem of declining generalization ability. At the same time, the existing data augmentation and simulation sample synthesis methods mainly rely on parametric physical modeling or empirical noise mixing, and it is difficult to accurately reflect the joint distribution and complex changes of target signals and background noise in real deep-sea environments. There are significant domain biases between the synthesized samples and the measured samples, reducing the practical value of the model.

[0004] At the level of optimizing detection algorithms, the existing technologies generally adopt static decision thresholds, fixed filtering parameters, or simple heuristic search strategies. General optimization algorithms often cannot adapt to the high time-varying and multi-scale characteristics of the noise field in deep-sea environments, resulting in lagging parameter adaptation of detection systems under extremely low signal-to-noise ratio and conditions where the signal and noise power change dynamically, 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, and it is difficult to achieve robust and efficient detection under multiple scenarios and conditions. Summary of the Invention

[0005] An object of the present invention is to propose a method for detecting underwater targets by sonar based on a confrontation network. In deep-sea scenarios, the positioning error and the target trajectory loss rate of the present invention are significantly reduced compared with traditional methods, and it has excellent online tracking and dynamic situation awareness capabilities, providing solid technical support for marine monitoring, security, and emergency command applications.

[0006] A method for detecting underwater targets by sonar based on a confrontation network according to an embodiment of the present invention includes the following steps: S1. Collect the original deep-sea sonar data and perform preprocessing to obtain a standardized sonar tensor; S2. Use the standardized sonar tensor as the training input, and train an improved StyleGAN-3 model in combination with the environmental parameter vector to complete the construction of the sonar scene generation network; S3. Call the improved StyleGAN-3 model to form a generated sample set, fuse the generated sample set with the original deep-sea sonar data in proportion to construct a synthetic-measured mixed data set, perform weak echo enhancement processing on the synthetic-measured mixed data set, and output a weak target feature tensor; S4. Construct a sonar detection network, use the weak target feature tensor as the input, and use the target annotation information in the synthetic-measured mixed data set as the supervision signal to obtain a set of detection parameters to be optimized. Based on the weak target feature tensor and the set of detection parameters to be optimized, establish an ocean predator optimization algorithm framework, and define an evaluation function with the weak target energy spectrum difference, detection rate, and false alarm rate as the core indicators; The target label information is the label obtained by manually or semi-automatically annotating the features such as the category, spatial position, or presence of weak targets in the synthetic-measured mixed data set after weak echo enhancement processing; S5. Within the ocean predator optimization algorithm framework, dynamically update the search strategy based on the evaluation function value until the evaluation function converges, and output the optimal set of detection parameters; S6. Load the optimal set of detection parameters into the sonar detection network, perform target decision on the weak target feature tensor, and output the target confidence value, three-dimensional spatial positioning result, and target motion track.

[0007] Optionally, S2 includes the following steps: S21. Construct a deep-sea weak echo sonar generation training data set , and the deep-sea weak echo sonar generation training data set consists of a standardized sonar tensor and an environmental parameter vector. The environmental parameter vectors are depth, salinity, temperature gradient, and seabed type respectively; S22. Construct an improved StyleGAN-3 model with an adaptive spectrum energy-phase joint constraint , and the improved StyleGAN-3 model fuses a spatio-temporal separable convolution structure and a complex domain residual gating structure in the feature generation network; S23. Dynamically construct an energy-phase joint adaptive loss function in combination with the target energy spectrum and the background noise power spectral density : ; Where and are the normalized energy spectra of the generated and real weak echo samples respectively, and They are the phase matrices generated for the real weak echo samples respectively, , are the weight coefficients, j represents the imaginary unit, is the total number of training samples; S24. During the training process of the improved StyleGAN-3, the sample separability score output by the discriminator based on the sonar detection network is introduced in real time as the adversarial feedback term of the joint loss function ; Among them, is the standard generative adversarial loss, is the discriminant feedback weight, represents the sample separability score that the generated weak echo samples are correctly recognized as weak targets by the current sonar detection network; S25. Use the training dataset to train the improved StyleGAN-3 model end-to-end , and adaptively optimize the parameters according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism during the training process, and output the finally trained improved StyleGAN-3 model .

[0008] Optionally, the improved StyleGAN-3 model includes a spatio-temporal separable convolution structure and a complex domain residual gating structure: The spatio-temporal separable convolution structure takes the normalized sonar tensor as the input, and performs a one-dimensional convolution operation Conv1D along the time dimension in combination with the time-direction convolution kernel to extract the time-direction convolution feature tensor of the deep-sea weak target in the time direction ; Input the time-direction convolution feature tensor into the one-dimensional convolution layer in the frequency dimension, and perform a one-dimensional convolution operation 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 spread and seabed scattering ; Input the time-frequency joint convolution feature tensor into the complex domain residual gating structure, and establish an amplitude feature branch and a 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, and after the extraction result passes through the batch normalization operation, it is connected to the rectified linear unit function for non-linear 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 extraction result also undergoes batch normalization and then is followed by the hyperbolic tangent function for non-linear activation to obtain the phase feature tensor ; A residual signal gating channel is constructed based on the time-frequency joint convolution feature tensor to dynamically regulate the fusion degree of the target feature and the background information. The residual signal gating channel extracts the gating feature map through a separate gating channel convolution, and compresses the gating weight value to the interval [0, 1] through an activation function, and finally outputs the gating weight tensor ; The amplitude feature tensor , the phase feature tensor and the gating weight tensor are fused to form the final output complex-domain residual feature tensor : ; wherein represents the Hadamard product

[0009] Optionally, S3 includes the following steps: S31. Call the trained improved StyleGAN-3 model , input the environmental parameter vector of the sample to be generated, and combine it with the random latent variable to generate the synthetic sample tensor under the corresponding conditions; S32. Collect the measured deep-sea sonar tensor corresponding to the generated sample, and estimate the weak target energy spectrum of the target signal and the background noise power spectral density in real time according to the target energy extraction function and the noise spectrum estimation function; S33. Construct a dynamic fusion ratio coefficient based on the weak target energy spectrum and the background noise power spectral density. The fusion ratio is used to control the superposition intensity of the generated sample and the measured data: ; wherein represents the L2 norm; S34. Linearly fuse the synthetic sample tensor and the measured sonar tensor according to the dynamic fusion ratio coefficient to construct the synthetic-measured mixed data tensor ; S35. Process the synthetic-measured mixed data tensor Perform weak echo enhancement processing, construct a convolutional pyramid using a multi-scale convolutional structure, and extract the multi-scale joint convolutional feature tensors of weak targets at different time-frequency scales. ; S36. Input the multi-scale joint convolutional feature tensors into the time-frequency attention module, and perform weighted fusion in the time dimension and the frequency dimension respectively through the temporal attention weights and the frequency attention weights to form the final weak target feature tensors ; Optionally, the S4 includes the following steps: S41. Construct a sonar detection network , using the weak target feature tensors as the input, and the target annotation information in the synthetic-measured mixed dataset as the supervision signal, and train the sonar detection network parameter set using the cross-entropy loss function ; S42. Based on the weak target feature tensors and the sonar detection network parameter set, establish a Marine Predators Optimization (MPO) algorithm framework , and perform global optimization on the feature fusion weights, decision thresholds, and filter parameters of the sonar detection network under the MPO algorithm framework. Initialize the MPO population parameter set , where is the number of individuals in the population; S43. Define an evaluation function with the weak target energy spectrum difference, detection rate, and false alarm rate as indicators to measure the detection performance of the sonar detection network under the current detection parameter configuration: ; where, is the measured weak target energy spectrum, is the detected target energy spectrum output by the detection network under the detection parameter , is the detection rate under the detection parameter , is the false alarm rate under the detection parameter , is the balance coefficient.

[0010] Optionally, the S5 includes the following steps: S51. In the MPO algorithm framework , initialize the MPO population parameter set. After each iteration , for each individual, based on the current detection parameter , calculate the evaluation function value ; S52. For each individual, based on the current synthetic-measured hybrid data tensor , the target signal energy spectrum and the background noise power spectral density , calculate the individual adaptability confidence : ; where is the detected target energy spectrum output by the detection network under the parameter , represents the expectation of the sample , is the L2 norm; S53. Set the individual adaptability confidence threshold , and divide the population into a global search subgroup and a local optimization subgroup ; S54. For the individuals in the global search subgroup, adopt a global search strategy update based on the deep-sea adaptive Lévy flight step operator : ; where is the detection parameter vector of the -th individual after the -th iteration. All parameters are for the current deep-sea weak echo sonar detection task, is the detection parameter vector of the -th individual at the -th iteration, is the detection parameter vector of the individual with the minimum evaluation function value among all individuals at the -th iteration, is a random variable, is the passive tracking weight factor, is the active perturbation weight factor; S55. For the individuals in the local optimization subgroup, adopt a population collaborative optimization strategy, fuse the adaptive collaboration factor, and finely adjust the individual parameters: ; where are the top dominant individuals in terms of the evaluation function value, is the adaptive collaboration factor that changes with the number of iterations, is the number of dominant individuals ranked among the top in terms of the evaluation function value within the population at each iteration; S56. After each iteration, update the global optimal detection parameter individual and its evaluation function value , perform a difference operation on the evaluation function value of the globally optimal detection parameter individual and the evaluation function value of the globally 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, use the globally optimal detection parameter individual obtained in the convergence round as the optimal detection parameter set.

[0011] Optionally, the deep-sea adaptive Lévy flight step size operator dynamically adjusts the step size distribution parameter according to the target energy feature complexity index in the deep-sea weak echo target detection scenario : ; where represents the target energy feature complexity index corresponding to the current detection parameter vector of the -th individual in the -th iteration, which is calculated based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size scale adaptive factor, is the step size stability index, is the stability index perturbation term, representing a normal random variable randomly sampled from the standard normal distribution by the -th individual in the -th iteration, which is used to control the distribution tail thickness of the Lévy jump amplitude.

[0012] is the random perturbation factor, representing a normal random variable randomly sampled from the standard normal distribution by the -th individual in the -th iteration, which is used to control the scale change direction and perturbation amplitude of the Lévy step size. Optionally, S6 includes the following steps: S61. Load the optimal detection parameter set into the sonar detection network and perform target decision on the input weak target feature tensor to generate the corresponding weak target response tensor , where each element represents the confidence probability of detecting a weak target at the time index and the frequency index ; S62. Extract the target confidence value from the weak target response tensor Perform three-dimensional spatial positioning calculations on the sensor array element information and acoustic propagation model parameters in it, and output the target positioning result , where is the estimated horizontal distance of the target, is the estimated azimuth angle of the target, is the estimated depth of the target; S64. Based on the target positioning result, construct a target space motion trajectory model, and combine the multi-frame weak target feature tensors and the corresponding detection response tensors in the continuous time window to perform temporal confidence weighted trajectory association: If there are frames in the continuous frame detection results that satisfy , it is determined to be the same target track; If the spatial distance between trajectory points changes , perform linear interpolation smoothing; Output the target motion track as a sequence ; where is the trajectory determination confidence threshold, is the maximum allowable jump distance, is the number of frames in the trajectory sliding window.

[0013] Optionally, the target confidence value is jointly defined by the detection response intensity and the background response difference: When , it is determined that there is a target, and the output target confidence value is ; When , it is determined that it is not detected, and the target confidence value is recorded as 0; where is the target confidence threshold, which is set according to the empirical optimal strategy on the training set.

[0014] The beneficial effects of the present invention are: (1) The present invention adopts an improved StyleGAN-3 generation network that introduces a spatio-temporal separable convolution and a complex domain residual gating structure, combines an energy-phase dual-constraint loss function, and is optimized by feedback coupling with the detection network. It can dynamically generate high-fidelity and physically consistent weak target echoes, seabed reverberations, and background noise samples according to the environmental parameters such as sea area depth, salinity, temperature gradient, and seabed type. By introducing the target energy spectrum and the 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, the types of training data are greatly enriched, and the balanced coverage of samples in different scenarios is achieved.

[0015] (2) The present invention proposes an ocean predator optimization algorithm that combines the energy spectrum of the target signal and the noise power spectral density for adaptive grouping. For population individuals with different adaptation confidence levels, global search with a deep-sea adaptive Levy flight step size and local fine optimization of population cooperation optimization are respectively adopted. It can automatically adjust the feature fusion weight, decision threshold, and filtering parameters of the detection network in a highly non-stationary noise and signal dynamic change environment. In an extremely low signal-to-noise ratio and Doppler spread background, the false alarm rate of the system is significantly reduced, and the detection rate and parameter stability are significantly better than the comparison technology.

[0016] (3) Through the dynamic loading of the optimal detection parameter set and the time-frequency joint attention mechanism, the present invention realizes the high-precision integrated output of target confidence, spatial positioning, and trajectory extraction in the sonar detection network. Based on multi-frame, multi-scale, and multi-frequency weak target feature tensors for dynamic decision-making, combined with the back-projection and spatio-temporal trajectory clustering algorithms, the three-dimensional spatial resolution of weak targets and the continuity of multi-target trajectories in a complex deep-sea environment are effectively improved. Actual engineering tests show that in a deep-sea scenario, the positioning error and target trajectory loss rate of the present invention are both significantly reduced compared with traditional methods, and it has excellent online tracking and dynamic situation awareness capabilities, providing solid technical support for ocean monitoring, security, and emergency command applications. Brief Description of the Drawings

[0017] The 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 to the present invention. In the drawings: Figure 1 is a flowchart of a sonar detection method for underwater targets based on an adversarial network proposed by the present invention; Figure 2 is a schematic structural diagram of improving the StyleGAN-3 sonar sample generation and physical consistency constraint in a sonar detection method for underwater targets based on an adversarial network proposed by the present invention; Figure 3 is a schematic diagram of the detection parameter adaptive global optimization and grouping cooperation mechanism based on the ocean predator optimization algorithm in a sonar detection method for underwater targets based on an adversarial network proposed by the present invention. Detailed Embodiments

[0018] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0019] Refer to Figures 1-3 , a sonar detection method for underwater targets based on an adversarial network, includes the following steps: S1. Collect the original deep-sea sonar data, and perform time-domain pulse compression filtering, matched filtering, and normalization based on the noise statistical characteristics on the original deep-sea sonar data to obtain a standardized sonar tensor; S2. Use the standardized sonar tensor as the training input, and combine it with the environmental parameter vector composed of depth, salinity, temperature gradient, and seabed type to train the improved StyleGAN-3 model to complete the construction of the sonar scene-specific generation network. The improved StyleGAN-3 model introduces spatio-temporal separable convolution and complex-domain residual gating structures in the feature mapping layer to enhance the expression ability of weak echo signals in terms of phase continuity and spectral structure; S3. Invoke the improved StyleGAN-3 model to generate weak target echo samples, seabed reverberation samples, and environmental background noise samples according to the environmental parameter vector to form a generated sample set. Combine the generated sample set with the original deep-sea sonar data in proportion. The fusion ratio is dynamically adjusted according to the difference between the target energy spectrum and the background noise power spectral density. Construct a synthetic-measured mixed data set, perform weak echo enhancement processing on the synthetic-measured mixed data set, and use a multi-scale convolution structure and time-frequency attention mechanism to extract weak target features, and output a weak target feature tensor; S4. Construct a sonar detection network, use the weak target feature tensor as the input, and use the target annotation information in the synthetic-measured mixed data set as the supervision signal to obtain a 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 Predators Optimization algorithm framework, and define an evaluation function with the weak target energy spectrum difference, detection rate, and false alarm rate as the core indicators; The target label information is the label obtained by manually or semi-automatically annotating the features such as the category, spatial position, or existence of weak targets in the synthetic-measured mixed data set after weak echo enhancement processing; S5. Within the Marine Predators Optimization algorithm framework, perform global search and iterative optimization on the feature fusion weights, decision thresholds, and filter parameters in the set of detection parameters, and dynamically update the search strategy based on the evaluation function value until the evaluation function converges, and output the optimal set of detection parameters; S6. Load the optimal set of detection parameters into the sonar detection network, perform target decision on the weak target feature tensor, and output the target confidence value, three-dimensional space positioning result, and target motion track.

[0020] In this embodiment, S2 includes the following steps: S21. Construct a deep-sea weak echo sonar generation training data set , and the deep-sea weak echo sonar generation training data set is composed of a standardized sonar tensor and an environmental parameter vector: ; Among them, represents the A standardized sonar tensor sample, with dimensions of , being the time domain length, and being the frequency resolution; The is the environmental parameter vector, including depth , salinity , temperature gradient and seabed type ; The improved StyleGAN-3 model incorporates a spatio-temporal separable convolutional structure and a complex domain residual gating structure in the feature generation network. S23. Combine the target energy spectrum and the background noise power spectral density to dynamically construct an energy-phase joint adaptive loss function, which constrains the physical consistency of the generated samples in terms of spectral energy distribution and phase continuity: ; Among them, and are the normalized energy spectra of the generated and real weak echo samples respectively, and are the phase matrices of the generated and real weak echo samples respectively, , are weight coefficients, and j represents the imaginary unit.

[0021] In view of the physical essence of deep-sea weak echo sonar targets, the present invention proposes an energy-phase joint adaptive loss function. The core idea is to explicitly introduce dual constraints on the energy spectrum characteristics and phase structure characteristics of sonar target signals 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 terms of 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 terms of frequency energy distribution, ensuring that the weak but real spectral energy peaks of weak targets can be effectively reproduced and captured. The phase structure part faces the multi-path interference and phase distortion phenomena in the underwater acoustic propagation process, and uses the Frobenius norm between the complex domain phase matrices as the measurement standard to explicitly constrain the generated samples to be highly consistent with the measured samples at the microscopic levels of phase continuity and phase perturbation, so as to ensure consistency with the propagation law of real sonar echoes at the physical level.

[0022] The energy-phase joint adaptive loss function can accurately align the weak energy peaks and complex phase fluctuations of weak target signals in an extremely low signal-to-noise ratio environment. The generated training samples can not only improve the sensitivity of the subsequent detection network to real weak targets, but also effectively suppress false signals and noise misjudgments, solving the problems of detector overfitting and degraded generalization ability caused by sample scarcity and unbalanced sample distribution in the prior art. By dynamically adjusting the energy and phase weights, it can actively respond to the changes in the target / noise field in different deep-sea sonar environments, and has stronger universality and environmental adaptability.

[0023] S24. During the improvement of the StyleGAN-3 training process, the sample separability score output by the discriminator based on the sonar detection network is 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: ; where is the standard generative adversarial loss, is the discriminant feedback weight, represents the sample separability score that the generated weak echo samples are correctly recognized as weak targets by the current sonar detection network; The present invention proposes a design of a joint loss function that is highly adaptive to the complex deep-sea acoustic environment. The core principle lies in: not only imposing physical consistency constraints on the spectral energy and phase continuity of the generated weak target echoes and background noise, but also introducing the sample separability score output by the detection network as a dynamic feedback index for the training of the generation model, thereby constructing an end-to-end optimization mechanism with deep coupling between the generation and detection tasks.

[0024] On the basis of 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 between the generated samples and the measured weak target samples in terms of energy distribution, 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, enabling the generated samples to maintain a continuous phase structure consistent with the real echoes in specific deep-sea scenarios such as multipath interference and non-stationary noise, breaking through the limitation of traditional GANs that only use pixel / magnitude loss constraints, and effectively enhancing the signal-level usability and engineering transferability of the synthetic data.

[0025] The combined loss function further integrates the sample separability scores output by the sonar detection network. The sample separability scores reflect the probability that each generated sample is correctly identified as a target (or non-target) after passing through the current detection network. Specifically, the separability scores are jointly constructed by the output confidence of the detection network for the generated samples and the degree of matching with the annotations. The higher the separability score, the stronger the distinguishability and training value of the current generated sample for the detection network. By incorporating the separability scores into the loss function, it not only prompts the generative model to improve physical consistency but also adaptively adjusts the generation distribution more towards the downstream detection task, achieving the co-evolution of data synthesis and the performance of the detection model.

[0026] The combined loss function first couples the sonar physical constraints, signal-level metrics, and detection network scores, thus achieving three breakthroughs: (1) At the data generation level, the generated samples not only pursue appearance similarity but also ensure the authenticity and usability of key physical features in target detection; (2) In the model training mechanism, an information closed-loop among generation, discrimination, and detection is established, breaking the traditional isolated optimization mode of generative networks and detection networks, enabling sample generation to adaptively evolve directly towards optimizing the detector's efficiency; (3) It improves the overall scalability and robustness of the system. The sample separability scores, as dynamic feedback signals, effectively alleviate problems such as extremely unbalanced actual samples and severe noise interference, providing a sustainable evolutionary data and model foundation for complex deep-sea sonar target detection applications.

[0027] S25. Using the training dataset End-to-end training to improve the StyleGAN-3 model , adaptively optimize the parameters during the training process according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism, and output the finally trained improved StyleGAN-3 model , under different deep-sea environmental parameter vectors, the improved StyleGAN-3 model dynamically generates weak target echo samples, seabed reverberation samples, and environmental background noise samples with spectral structures and phase characteristics highly consistent with the measured values.

[0028] The present invention adopts an improved StyleGAN-3 generative network introducing spatio-temporal separable convolution and complex-domain residual gating structure, combines the energy-phase dual-constraint loss function, and optimizes it by coupling with the detection network feedback. It can dynamically generate weak target echoes, seabed reverberation, and background noise samples with high fidelity and strong physical consistency according to environmental parameters such as sea area depth, salinity, temperature gradient, and seabed type. By introducing the target energy spectrum and the background noise power spectral density to adaptively adjust the synthesis ratio, it significantly alleviates the domain deviation problem between traditional simulation samples and measured samples, greatly enriches the types of training data, and achieves balanced coverage of samples in different scenarios.

[0029] In this embodiment, the improved StyleGAN-3 model includes a spatio-temporal separable convolution structure and a complex domain residual gating structure: The spatio-temporal separable convolution structure takes the normalized sonar tensor as the input, and performs a one-dimensional convolution operation Conv1D along the time dimension in combination with the time-direction convolution kernel to extract the time-direction convolution feature tensor of the deep-sea weak target in the time direction : ; where represents the 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-time energy fluctuation characteristics of weak targets, is the number of feature channels; The time-direction convolution feature tensor is input into the one-dimensional convolution layer in the frequency dimension, and a 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 spread and seabed scattering : ; where represents the 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 the deep-sea weak target under the influence of Doppler spread and seabed scattering; The time-frequency joint convolution feature tensor is input into the complex domain residual gating structure, and an amplitude feature branch and a phase feature branch are established respectively: The amplitude feature branch extracts the local amplitude information of the time-frequency joint convolution feature tensor through an amplitude convolution channel. After the extraction result passes through the batch normalization operation, it is connected to the rectified linear unit function for non-linear activation to obtain the amplitude feature tensor , and the amplitude feature tensor is used to characterize the energy fluctuation pattern of the deep-sea weak target in the local time-frequency window; The phase feature branch extracts the phase change information of the time-frequency joint convolution feature tensor through another phase convolution channel. After the extraction result also passes through the batch normalization operation, it is connected to the hyperbolic tangent function for non-linear activation to obtain the phase feature tensor , which is used to depict the phase delay characteristics of the deep-sea weak target caused by multipath interference; Construct a residual signal gating channel based on the time-frequency joint convolution feature tensor to dynamically regulate the fusion degree of target features and background information. The residual signal gating channel extracts gating feature maps through a separate gating channel convolution, and compresses the gating weight values to the interval [0, 1] through an activation function, and finally outputs a gating weight tensor , each element of the gating weight tensor is used to measure the intensity that the residual signal should retain at the current position. The larger the gating weight, the more it depends on the target information synthesized from the complex domain features. The smaller the gating weight, the more input background features are retained; Multiply the amplitude feature tensor , the phase feature tensor and the gating weight tensor to form the final output complex domain residual feature tensor : ; wherein, 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 eigenfeatures and the adaptability of noise suppression.

[0030] In the modeling of deep-sea weak echo sonar signals, the target signal is often affected by multipath, scattering and phase perturbations, resulting in complex changes in its amplitude and phase. Traditional GANs and their variants generally process signal features in the real domain and are difficult to accurately describe the phase information contained in sonar echoes. S23 introduces a complex domain residual gating structure to dynamically fuse the amplitude feature tensor and the phase feature tensor branches with the gating weight, and outputs a complex domain residual feature tensor, which can not only accurately restore the energy fluctuations of weak targets, but also effectively describe the phase continuity under multipath interference. The residual gating mechanism automatically adjusts the fusion weight of target features and background noise, taking into account both the physical characteristics of the signal and the diversity of data distribution.

[0031] Introducing the complex domain residual structure into the sonar generative adversarial model overcomes the limitations of existing real-valued space feature extraction and realizes a high-level expression of the physical eigenfeatures of deep-sea weak echoes, greatly improving the realism and physical consistency of the generated samples.

[0032] In this embodiment, S3 includes the following steps: S31. Invoke the trained improved StyleGAN-3 model , input the environmental parameter vector of the sample to be generated, and combine it with the random latent variable to generate a synthetic sample tensor under the corresponding conditions: ; wherein, Indicating at a depth , salinity , temperature gradient , seabed type The standardized sonar tensor generated under the environmental conditions composed of, including the composite structure of weak target echoes, seabed reverberation and environmental background noise; S32. Collect and generate the measured deep-sea sonar tensor corresponding to the sample , and estimate the weak target energy spectrum of the target signal in real time according to the target energy extraction function and the noise spectrum estimation function and the 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 spectral density by spectral smoothing and spectral valley matching; S33. Construct a dynamic fusion ratio coefficient based on the weak target energy spectrum and the background noise power spectral density . The fusion ratio is used to control the superposition intensity of the generated sample and the measured data: ; wherein, represents the L2 norm; the ratio coefficient controls the injection intensity of the generated sample. When the background noise power spectral density is lower, the dynamic fusion ratio coefficient tends to 1 to enhance the target signal. When the background noise power spectral density is stronger, the dynamic fusion ratio coefficient tends to 0; In the deep-sea environment, the weak target energy and the noise power spectral density change violently. To avoid the domain shift problem caused by a fixed ratio, the present invention proposes to dynamically calculate the fusion ratio coefficient according to the sample energy spectrum and the noise spectrum, and adaptively adjust the mixing ratio of the synthetic sample and the measured sample according to the relative intensity of the target and the noise, so that the training data set can dynamically cover the complex change interval in the actual application scenario. It breaks the rigid mode of manually setting or empirically setting the sample fusion ratio in the prior art, realizes the adaptability of the data layer to the physical environment change, ensures the consistency of the distribution of the training sample and the test environment, and effectively alleviates the problem of insufficient model generalization ability.

[0033] S34. Linearly fuse the synthetic sample tensor and the measured sonar tensor according to the dynamic fusion ratio coefficient to construct a synthetic-measured mixed data tensor : ; The synthetic-measured mixed data tensor retains both the measured background noise distribution and the weak features of the synthetic weak target; S35. Perform weak echo enhancement processing on the synthetic-measured mixed data tensor to construct a convolutional pyramid using a multi-scale convolutional structure, and extract the multi-scale joint convolutional feature tensor of weak targets at different time-frequency scales ; S36. Input the multi-scale joint convolutional feature tensor into the time-frequency attention module, and perform weighted fusion in the time dimension and frequency dimension respectively through the time attention weight and the frequency attention weight to form the final weak target feature tensor ; ; wherein represents the time index and the frequency index; the final weak target feature tensor characterizes the echo features of deep-sea weak targets after fusion, and has the ability of multi-scale information aggregation and noise suppression.

[0034] In this embodiment, S4 includes the following steps: S41. Construct a sonar detection network with the weak target feature tensor as the input and the target annotation information in the synthetic-measured mixed dataset as the supervision signal, and train the sonar detection network parameter set using the cross-entropy loss function : ; wherein represents the predicted value of the probability of the existence of a weak target by the sonar detection network at the time index and the frequency index 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; S42. Based on the weak target feature tensor and the sonar detection network parameter set, establish a Marine Predators Optimization Algorithm framework , and perform global optimization on the feature fusion weight, decision threshold, and filter parameters of the sonar detection network under the Marine Predators Optimization Algorithm framework. Initialize the Marine Predators population parameter set , is the number of individuals in the population, and each individual corresponds to a detection parameter vector including the feature fusion weight, decision threshold, and filter parameters; S43. Define an evaluation function with the weak target energy spectrum difference, detection rate, and false alarm rate as indicators to measure the detection performance of the sonar detection network under the current detection parameter configuration: ; Among them, is the measured weak target energy spectrum, is the detected target energy spectrum output by the detection network under the detection parameter ; is the detection rate under the detection parameter ; is the false alarm rate under the detection parameter ; is the balance coefficient.

[0035] The evaluation function designed in the present invention integrates three core indicators: target energy spectrum difference, detection rate, and false alarm rate. It not only pays attention to the physical accuracy of signal recovery by detection parameters but also takes into account the actual detection ability and false alarm control of the detection network. Through weighted combination, the overall performance of the multi-target quantization detection system under different parameter configurations is provided as a basis for comprehensively evaluating the performance of the optimization algorithm.

[0036] 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 under complex sea areas, provides a globally optimal search guidance for the swarm intelligence algorithm facing actual engineering, and is different from the traditional evaluation system dominated by single precision or simple accuracy.

[0037] In this embodiment, S5 includes the following steps: S51. In the framework of the Marine Predators Optimization Algorithm , initialize the set of parameters of the marine predator population. After each round of iteration , for each individual, based on the current detection parameter , calculate its 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 : ; Among them, is the detected target energy spectrum output by the detection network under the parameter , represents the expectation of the sample , is the L2 norm; During the optimization process, the individual adaptability confidence is dynamically calculated based on the ratio of the recovery error of the target energy spectrum to the actual signal-noise energy distribution of the sample for each individual detection parameter. The individual adaptability confidence reflects the adaptability of the individual to the target modeling under the current mixed data distribution, and automatically drives the individual to be diverted to the global search or local fine optimization stage.

[0038] This method realizes the signal-noise scenario adaptive division of individual behaviors during the optimization process, breaking through the traditional practice of stage allocation in swarm intelligence algorithms that relies only on rounds or static thresholds, enabling individuals within the swarm to always dynamically divide labor according to their own capabilities, and improving the overall optimization efficiency and the ability to adapt to complex environments.

[0039] S53. Set the demarcation threshold of individual adaptability confidence , and divide the swarm into a global search subgroup and a local optimization subgroup ; S54. For individuals within the global search subgroup, use the deep-sea adaptive Lévy flight step operator to update the global search strategy: ; where is the detection parameter vector of the th individual after the th iteration, and all parameters are for the current deep-sea weak echo sonar detection task, is the detection parameter vector of the th individual at the th iteration, is the detection parameter vector of the individual with the smallest evaluation function value among all individuals at the th iteration, is a random variable, is the passive tracking weight factor, is the active perturbation weight factor; S55. For individuals within the local optimization subgroup, use the swarm collaborative optimization strategy, fuse the adaptive collaboration factor, and finely adjust the individual parameters: ; where are the top dominant individuals in terms of the evaluation function value, is the adaptive collaboration factor that changes with the number of iterations, is the number of dominant individuals ranked among the top in terms of the evaluation function value within the swarm at each iteration; The global search formula is based on the deep-sea adaptive Lévy flight step operator, which dynamically adjusts the step size and perturbation direction using the individual's historical performance and the complexity of the target energy, enabling the individual to make long-distance jumps within the parameter space to explore unknown extreme value regions. The collaborative optimization formula uses the collaboration factor as the weight to fuse and adjust the individual towards the current dominant subgroup and the global optimal individual, strengthening the local convergence and collaboration efficiency of high-confidence individuals.

[0040] Different from existing algorithms such as MPA and PSO that rely only on passive or active search, the present invention is based on a triple fusion mechanism of signal adaptability + complexity dynamic perturbation + group collaboration, ensuring adaptive global optimization and fine local convergence of the optimal detection parameters in a deep-sea environment with non-stationary and severely fluctuating signal-to-noise ratio, greatly improving the optimization robustness and convergence speed.

[0041] S56. After each iteration, update the individual of the global optimal detection parameter 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, use the global optimal detection parameter individual obtained in the convergence rounds 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 decision-making and tracking modules. The optimal detection parameter set represents the detection parameter configuration with the optimal global detection performance in the current deep-sea weak echo sonar environment after optimization.

[0042] In this embodiment, the deep-sea adaptive Lévy flight step operator dynamically adjusts the step size distribution parameter according to the complexity index of the target energy characteristics in the deep-sea weak echo target detection scenario : ; where represents the complexity index of the target energy characteristics corresponding to the current detection parameter vector of the th individual in the th iteration, which is calculated based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size scale adaptive factor, is the step size stability exponent, is the stability exponent perturbation term, representing the normal random variable randomly sampled from the standard normal distribution by the kth individual in the tth iteration, used to control the distribution tail thickness of the Lévy jump amplitude. is a random perturbation factor, which represents a normal random variable randomly sampled from a standard normal distribution by the k-th individual in the t-th iteration, and is used to control the scale change direction and perturbation amplitude of the Lévy step size.

[0043] The present invention proposes an ocean predator optimization algorithm that combines the target signal energy spectrum and the noise power spectral density for adaptive grouping. For population individuals with different adaptation confidence levels, global search with a deep-sea adaptive Levy flight step size and local fine optimization of population cooperation are respectively adopted. It can automatically adjust the feature fusion weights, decision thresholds, and filtering parameters of the detection network in an environment with highly non-stationary noise and dynamic signal changes. In an environment with extremely low signal-to-noise ratio and Doppler spread, the false alarm rate of the system is significantly reduced, and the detection rate and parameter stability are significantly better than those of the comparative technology.

[0044] In this embodiment, S6 includes the following steps: S61. Load the optimal detection parameter set into the sonar detection network , and perform target decision on the input weak target feature tensor to generate a corresponding weak target response tensor , where each element represents the confidence probability of detecting a weak target at the time index and the frequency index ; S62. Extract the target confidence value based on the weak target response tensor ; S63. According to the set of time-frequency positions with significant response probabilities in the weak target response tensor , combined with the sensor array element information and the acoustic propagation model parameters in the corresponding synthetic-measured mixed data tensor , perform three-dimensional space positioning calculation and output the target positioning result , where is the estimated value of the horizontal distance of the target, is the estimated value of the azimuth angle of the target, is the estimated value of the depth of the target; S64. Based on the target positioning result, construct a target space motion trajectory model, and combine the multi-frame weak target feature tensors within a continuous time window and the corresponding detection response tensors , and perform time-sequence confidence weighted trajectory association: If there are frames satisfying in consecutive frames, it is determined to be the same target track; If the spatial distance between trajectory points changes , linear interpolation smoothing is performed; The output target motion track is a sequence ; Among them, is the track determination confidence threshold, is the maximum allowable jump distance, is the number of frames in the track sliding window.

[0045] In this embodiment, the target confidence value is jointly defined by the detection response intensity and the background response difference: When , it is determined that there is a target, and the output target confidence value is ; When , it is determined that no target is detected, and the target confidence value is recorded as 0; Among them, is the target confidence threshold, which is set according to the empirical optimal strategy on the training set.

[0046] Example 1 In the deep water area in the north (xxx.xx° east longitude, xx.xx° north latitude), a research ship deployed a 128-element broadband phased array sonar device in the deep water area about 180 kilometers away from the coastline. The working frequency of the device is 48 kHz, and the target water depth range is 2450 - 2820 meters. The sea condition at that time was level 2, and the peak value of the background noise power spectral density was about 31.5 dB re 1 μPa² / Hz.

[0047] During the voyage from 2:00 to 4:30 in the early morning of that day, the sonar system continuously collected 6 hours of channel raw data. The online detection algorithm initially determined 22 weak target events, with 56 groups of measured weak echo samples and 154 groups of environmental noise samples. Using traditional data enhancement based on empirical models, the research team relied on the method of Poisson scattering + Gaussian white noise superposition and only synthesized 94 groups of weak echo training samples. The mean square deviation of the energy spectrum of the synthesized samples deviated from the measured mean value by 0.23.

[0048] During the same period, using the method of the present invention, the implementer formed a feature vector with environmental parameters such as a depth of 2450 - 2820 meters, a salinity of 34.5 ppt, a temperature gradient of 1.9 °C / 100 m, and a silty 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 high matching degree to the actual measurement. The mean square deviation of the spectral structure was only 0.12, and it covered multiple groups of typical biological noise and multipath reverberation scenarios. Each group of samples retained the corresponding synthesized environmental parameters and target labels, facilitating subsequent training and refined evaluation.

[0049] At 3:25 in the early morning, the detection system of the research ship detected a suspected micro-target event with a signal-to-noise ratio of only -16.7 dB. The traditional fusion method linearly mixes the synthetic samples and the measured samples at a ratio of 1:1, resulting in a detection rate of 47.8% for samples with extremely low signal-to-noise ratio after the detection network is trained. The method of the present invention automatically adjusts the fusion ratio according to the current target energy spectrum and the noise power spectral density. The proportion of measured samples is 36%, and the proportion of synthetic samples is increased to 64%. After processing by the multi-scale convolutional pyramid and the time-frequency attention mechanism, the output weak target feature tensor clearly retains the weak fluctuations of the target echo, and the noise component is significantly reduced.

[0050] At 3:48 in the early morning, the implementer calls the Marine Predator Optimization Algorithm of the present invention to online optimize the feature fusion weight, decision threshold, and filter parameters of the detection network. Assume that the mean square error between the target energy spectrum detected by the 12th parameter individual at this moment and the true energy spectrum is 0.073, and the mean square value of the noise spectrum is 0.132. After confidence calculation , the system automatically assigns this individual to the global search subgroup, and the Levy flight step size is dynamically adjusted to 0.018. The confidence of the 3rd individual is as high as 0.74 and is included in the local collaborative optimization subgroup, and the weight adjustment speed is halved, and the cooperation factor .

[0051] Under the traditional method, at the same moment, the parameters of the 7th individual are updated using the static genetic algorithm. The mean square error of the target energy spectrum is 0.127, and the decision confidence of the detector 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%.

[0052] At 4:02 in the early morning, after the sonar detection network loads the optimal detection parameter set, it makes frame-by-frame decisions on the collected weak target feature tensors. In a certain frame (frame number #1298), the detection network outputs the target confidence in the time-frequency unit , which exceeds the decision threshold , and it is determined that there is a weak target. Combining with the array element data backprojection algorithm, the system real-time outputs the spatial coordinates , and continuously maintains the decision as true for 14 frames. The track tracking algorithm generates a motion track based on this and realizes trajectory clustering with the previous target within 5 minutes.

[0053] In the same frame of the control group, the confidence of the traditional detection network is only 0.62, which does not reach the decision threshold of 0.70, resulting in the omission of this target.

[0054] During the whole process of this Embodiment 1, the method of the present invention detected a total of 28 weak target events (target radar cross section 0.4 - 1.0 m², signal-to-noise ratio -18.2 - 13.5 dB), 4 false alarms, and 2 consecutive track losses. Compared with the traditional method, 19 targets were detected, 11 false alarms, and 7 track losses.

[0055] Taking the statistical results from 2:00 to 6:00 in the early 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 level 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 decreased to 12.5%, while that of the traditional method is 36.7%.

[0056] Under the extremely low signal-to-noise ratio and deep-sea non-stationary noise environment, the method of the present invention effectively realizes the generation of high-fidelity weak target samples, online adaptive parameter optimization, three-dimensional target decision-making, and continuous track tracking, greatly improving the practicability and robustness of the detection system in extreme scenarios, and significantly superior to the existing traditional sonar target detection technologies.

[0057] Through the dynamic loading of the optimal detection parameter set and the time-frequency joint attention mechanism, the present invention realizes the high-precision integrated output of target confidence, spatial positioning, and trajectory extraction in the sonar detection network, makes dynamic decisions based on multi-frame, multi-scale, and multi-frequency weak target feature tensors, and combines the back-projection and spatio-temporal trajectory clustering algorithms, effectively improving 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 track loss rate of the present invention are both significantly reduced compared with the traditional method, and it has excellent online tracking and dynamic situation awareness capabilities, providing a solid technical support for marine monitoring, security, and emergency command applications.

[0058] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A sonar detection method for underwater targets based on adversarial networks, characterized in that, It includes the following steps: S1. Collect the original deep-sea sonar data and perform preprocessing to obtain a standardized sonar tensor; S2. Use the standardized sonar tensor as the training input, and train the improved StyleGAN-3 model in combination with the environmental parameter vector; S3. Invoke the improved StyleGAN-3 model to form a generated sample set, fuse the generated sample set with the original deep-sea sonar data in proportion, construct a synthetic-measured mixed data set, perform weak echo enhancement processing on the synthetic-measured mixed data set, and output a weak target feature tensor; S4. Construct a sonar detection network, use the weak target feature tensor as the input, and use the target annotation information in the synthetic-measured mixed data set as the supervision signal to obtain a set of detection parameters to be optimized. Based on the weak target feature tensor and the set of detection parameters to be optimized, establish an ocean predator optimization algorithm framework, and define an evaluation function with the weak target energy spectrum difference, detection rate, and false alarm rate as the core indicators; The target label information is the label obtained by manually or semi-automatically annotating the features such as the category, spatial position, or presence of weak targets in the synthetic-measured mixed data set after weak echo enhancement processing; S5. Within the ocean predator optimization algorithm framework, dynamically update the search strategy based on the evaluation function value until the evaluation function converges, and output the optimal set of detection parameters; S6. Load the optimal set of detection parameters into the sonar detection network, perform target decision on the weak target feature tensor, and output the target confidence value, three-dimensional spatial positioning result, and target motion track.

2. The sonar detection method for underwater targets based on adversarial network according to claim 1, characterized in that, S2 includes the following steps: S21. Construct a training dataset for deep - sea weak echo sonar , The training dataset for deep - sea weak echo sonar consists of a standardized sonar tensor and an environmental parameter vector. The environmental parameter vectors are depth, salinity, temperature gradient, and seabed type respectively; S22. Construct an improved StyleGAN-3 model with adaptive spectral energy-phase joint constraints , and the improved StyleGAN-3 model integrates a spatio-temporal separable convolution structure and a complex-domain residual gating structure in the feature generation network; S23. Dynamically construct an energy-phase joint adaptive loss function by combining the target energy spectrum and the background noise power spectral density ; S24. During the improvement of the StyleGAN-3 training process, the sample separability score output by the discriminator based on the sonar detection network is introduced in real time as an adversarial feedback term of the joint loss function to dynamically adjust the generator parameters and form a generation-discrimination collaborative feedback mechanism; S25. Utilize the training dataset End-to-end training to improve the StyleGAN-3 model , adaptively optimize the parameters according to the energy-phase joint constraint and the generation-discrimination collaborative feedback mechanism during the training process, and output the finally trained improved StyleGAN-3 model .

3. The method for sonar detection of underwater targets based on adversarial network according to claim 2, wherein The improved StyleGAN-3 model includes a spatio-temporal separable convolution structure and a complex-domain residual gating structure: Space-time separable convolutional structure with normalized sonar tensor Using as the input, perform a one-dimensional convolutional operation Conv1D along the time dimension with a time-direction convolutional kernel to extract the time-direction convolutional feature tensor of deep-sea weak targets in the time direction ; Input the time-direction convolutional feature tensor into the one-dimensional convolutional layer in the frequency dimension, perform a one-dimensional convolutional operation in the frequency dimension in combination with the frequency-direction convolutional kernel, and extract the time-frequency joint convolutional feature tensor of the weak target signal under the influence of Doppler spread and seabed scattering. ; Input the time-frequency joint convolution feature tensor into the complex-domain residual gating structure, and establish an amplitude feature branch and a 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. After the extraction result undergoes batch normalization operation, it is followed by a rectified linear unit function for non-linear 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 extraction result also undergoes batch normalization and then is followed by the hyperbolic tangent function for non-linear activation to obtain the phase feature tensor ; Construct a residual signal gating channel based on the time-frequency joint convolutional feature tensor to dynamically regulate the fusion degree of target features and background information. The residual signal gating channel extracts gating feature maps through a separate gating channel convolution, compresses the gating weight values to the [0, 1] interval through an activation function, and finally outputs a gating weight tensor ; Fuse the amplitude feature tensor , phase feature tensor and gating weight tensor to form the final output complex domain residual feature tensor.

4. The sonar detection method for underwater targets based on adversarial network according to claim 2, characterized in that S3 includes the following steps: S31. Invoke the improved StyleGAN-3 model that has been trained, input the environmental parameter vector of the sample to be generated, and combine it with a random latent variable to generate a synthetic sample tensor under the corresponding conditions ; S32. Collect and generate the measured deep-sea sonar tensor corresponding to the sample , and estimate the weak target energy spectrum of the target signal and the background noise power spectral density in real time according to the target energy extraction function and the noise spectrum estimation function ; ; S33. Construct a dynamic fusion ratio coefficient based on the weak target energy spectrum and the background noise power spectral density , and 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 perform linear fusion on the synthetic sample tensor and the measured sonar tensor to construct a synthetic-measured hybrid data tensor ; S35. Perform weak echo enhancement processing on the synthetic-measured mixed data tensor and construct a convolutional pyramid using a multi-scale convolutional structure to extract the multi-scale joint convolutional feature tensor of weak targets at different time-frequency scales ; S36. Input the multi-scale joint convolution feature tensor into the time-frequency attention module, and perform weighted fusion in the time dimension and frequency dimension respectively through the time attention weight and the frequency attention weight to form the final weak target feature tensor .

5. A sonar detection method for underwater targets based on adversarial network according to claim 4, characterized in that S4 includes the following steps: S41. Construct a sonar detection network , using the weak target feature tensor as the input, and the target annotation information in the synthetic-measured mixed dataset as the supervision signal, and use the cross-entropy loss function to train the sonar detection network parameter set ; S42. Establish an ocean predator optimization algorithm framework based on the weak target feature tensor and the sonar detection network parameter set , globally optimize the feature fusion weights, decision thresholds, and filter parameters of the sonar detection network under the ocean predator optimization algorithm framework, and initialize the ocean predator population parameter set , is the number of individuals in the population; S43. Define an evaluation function with the differences in weak target energy spectra, detection rates, and false alarm rates as indicators , which is used to measure the detection performance of the sonar detection network under the current detection parameter configuration.

6. The sonar detection method for underwater targets based on adversarial network according to claim 1, wherein S5 includes the following steps: S51. In the framework of the Marine Predators Optimization Algorithm Initialize the parameter set of the marine predator population. After each iteration Based on the current detection parameters for each individual Calculate the evaluation function value ; S52. For each individual, based on the current synthetic-measured hybrid data tensor , the target signal energy spectrum and the background noise power spectral density , calculate the individual adaptability confidence ; S53. Set the confidence threshold for individual adaptability , divide the population into a global search subgroup and a local optimization subgroup ; For the individuals in the global search subgroup, a global search strategy update is carried out using the deep-sea adaptive Lévy flight step operator ;​ For the individuals within the locally optimized subgroup, adopt the group collaborative optimization strategy, integrate the adaptive collaboration factor, and finely adjust the individual parameters ; After each 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, use the global optimal detection parameter individual obtained in the convergence round as the optimal set of detection parameters.

7. A sonar detection method for underwater targets based on an adversarial network according to claim 6, characterized in that The deep-sea adaptive Lévy flight step operator dynamically adjusts the step distribution parameter according to the target energy feature complexity index in the deep-sea weak echo target detection scenario: ； Among them, characterizes the th individual in the th iteration for the current detection parameter vector corresponding target energy feature complexity index, which is obtained based on the local gradient entropy value of the weak target feature tensor output by the detection network. is the step size scale adaptive factor, is the step size stability index, is the stability index perturbation term, representing a normal random variable randomly sampled from the standard normal distribution by the kth individual in the tth iteration, used to control the distribution tail thickness of the Lévy jump amplitude. is the random perturbation factor, representing a normal random variable randomly sampled from the standard normal distribution by the kth individual in the tth iteration, used to control the scale change direction and perturbation amplitude of the Lévy step size.

8. The sonar detection method for underwater targets based on adversarial network according to claim 1, characterized in that S6 includes the following steps: S61. Load the optimal detection parameter set into the sonar detection network , and perform target decision on the input weak target feature tensor to generate a corresponding weak target response tensor ; S62. Based on the weak target response tensor Extract the target confidence value ; S63. Based on the time-frequency position set with significant response probability in the weak target response tensor and combining the sensor array element information and acoustic propagation model parameters in the corresponding synthetic-measured mixed data tensor , three-dimensional space positioning calculation is performed to output the target positioning result; ​ S64. Based on the target positioning result, construct a target space motion trajectory model, and combine the multi-frame weak target feature tensors within a continuous time window with the corresponding detection response tensors , and perform temporal confidence weighted trajectory association.

9. A sonar detection method for underwater targets based on an adversarial network according to claim 8, characterized in that, The target confidence value is jointly defined by the detection response intensity and the background response difference: When , it is determined that there is a target, and the output target confidence value is ; When is detected, it is determined that it is not detected, and the target confidence value is recorded as 0; wherein, is the target confidence threshold, which is set according to the empirically optimal policy on the training set.

Images