A method for simulating ocean reverberation based on a physical information condition generative adversarial network and a related device

Abstract

This invention provides a marine reverberation simulation method and related apparatus based on a physically-informed conditional generative adversarial network (GAN). The method includes: using acquired sonar pulse signals as input to the GAN to obtain a simulated reverberation time-series signal; the GAN comprises a generator network structure and a discriminator network structure, wherein the discriminator network structure embeds a physical consistency loss function. By embedding the physical consistency loss function into the discriminator network, this invention ensures that the GAN is simultaneously constrained by the real data distribution and the physical laws of ocean sound propagation during adversarial training. This solves the technical problems of existing marine reverberation simulation methods, which struggle to meet real-time simulation requirements in complex environments and lack adaptability to environmental changes. It can generate reverberation signals with high physical consistency and high fidelity, while meeting real-time requirements, and can be widely applied to the design and verification of underwater active sonar systems.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustics engineering, specifically relating to a marine reverberation simulation method and related devices based on physical information condition generative adversarial networks. Background Technology

[0002] Developing next-generation active sonar systems for ships, submarines, and autonomous underwater vehicles (AUVs / ROVs) requires extensive reverberation data for algorithm and system-level verification during both the design and validation phases. However, real-world sea trials are constrained by high costs associated with chartering vessels, launching, and long timelines, making high-fidelity reverberation simulators particularly crucial.

[0003] Marine environmental reverberation is a major background disturbance in the operation of underwater acoustic systems. It primarily originates from surface waves, seabed sediment, marine life, and distant vessels. Its power spectrum varies across a wide frequency band and exhibits significant spatial correlation characteristics. In marine active sonar systems, emitted sound waves are repeatedly scattered by the seawater medium and the seabed interface, forming reverberant signals with long durations. Traditional broadband reverberation waveform simulation methods based on normal mode theory model the reverberant signal using the propagation characteristics of normal modes, calculating the seabed scattering coefficient, and simulating various seabed reflection mechanisms, which can simulate reverberation waveforms relatively accurately. However, this method still has limitations in modeling the dispersion and attenuation characteristics of broadband signals and struggles to meet the real-time simulation requirements of complex environments, lacking adaptability to environmental changes.

[0004] In recent years, with the rapid development of artificial intelligence, deep learning methods have also been applied to marine reverberation simulation. Generative Adversarial Networks (GANs) can learn the temporal and frequency domain characteristics of real reverberation signals through adversarial training between the generator and discriminator, thereby improving the realism of the simulation signal. However, existing GAN-based marine reverberation simulation methods often fail to effectively incorporate the physical constraints of marine sound propagation, resulting in deviations in simulation results under specific complex marine environments and making it difficult to fully adapt to dynamically changing physical conditions. Therefore, it is urgent to design a deep learning model that incorporates physical constraints, ensuring that the simulation results closely approximate real reverberation signals while improving the model's environmental adaptability and generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide a marine reverberation simulation method and related apparatus based on physical information conditions generative adversarial networks (GANs). This solves the problems of existing marine reverberation simulation methods, which struggle to meet real-time simulation requirements in complex environments and lack adaptability to environmental changes. This invention enables the generation of high-precision reverberation signals based on actual collected environmental data and physical conditions, while also meeting real-time requirements.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a marine reverberation simulation method based on physical information conditional generative adversarial networks, comprising the following steps: The acquired sonar pulse signal is used as the input to the physical information conditional generative adversarial network to obtain the simulated reverberation time sequence signal; The physical information conditional generative adversarial network includes a generator network structure and a discriminator network structure, wherein the discriminator network structure is embedded with a physical consistency loss function.

[0007] Preferably, the expression for the physical consistency loss function is... for:

[0008] in, Let this be the energy loss function; The autocorrelation loss function; The spectral envelope loss function; This is the reverberation intensity loss function.

[0009] Preferably, the expression for the reverberation intensity loss function is: (6) In the formula, The intensity of the scattering unit. Indicates the transmission of a signal. For the bidirectional time delay from the platform to the scattering unit, The center frequency of the transmitted signal. For relative Doppler frequency shift, This represents the number of scattering rings. This represents the number of scattering units.

[0010] Preferably, the training process of the physical information conditional generative adversarial network is as follows: Constructing physical constraints for ocean reverberation, the physical constraints for ocean reverberation include sound wave propagation path parameters, the sound wave propagation path parameters include ocean interface scattering intensity, path delay and attenuation; The physical constraints of ocean reverberation, seabed parameters, and acquired historical sonar pulse signals are used as inputs to construct a physical information conditional generative adversarial network (PEGAN). The PEGAN is then trained to obtain the trained PEGAN.

[0011] Preferably, the acquired historical sonar pulse signals are preprocessed before being used to construct the physical information conditional generative adversarial network, specifically including: The acquired historical sonar pulse signals were weighted using the Hanning window function to obtain the weighted pulse signals. The weighted pulse signal is subjected to Fourier transform to obtain the short-time power spectral density; The short-time power spectral density is preprocessed by low-pass filtering and gain normalization to obtain the normalized pulse signal; The normalized pulse signal obtained is used as the input to the constructed physical information conditional generative adversarial network.

[0012] Secondly, the present invention provides a marine reverberation simulation system based on a physical information condition generative adversarial network, comprising: The simulated reverberation signal acquisition unit is used to take the acquired sonar pulse signal as the input of the physical information conditional generative adversarial network to obtain the simulated reverberation timing signal. A generative adversarial network (GAN) unit is used to deploy a physical information conditional GAN, wherein the physical information conditional GAN ​​includes a generator network structure and a discriminator network structure, and the discriminator network structure embeds a physical consistency loss function.

[0013] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the electronic device performs the method described thereon.

[0014] Fourthly, the present invention provides a computing device cluster, comprising at least one computing device, each computing device including a processor and a memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device to cause the cluster of computing devices to perform the method according to the method.

[0015] Fifthly, the present invention provides a computer program product, the computer program product including computer-executable instructions, wherein the computer-executable instructions implement a method when executed.

[0016] In a sixth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the method described herein.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a marine reverberation simulation method based on physical information conditions and generative adversarial networks (GANs). Addressing the technical problems of existing marine reverberation simulation methods being unable to meet the real-time simulation requirements of complex environments and lacking adaptability to environmental changes, this application embeds a physical consistency loss function into the discriminator network structure. This allows the GAN to not only learn the time-domain and frequency-domain statistical characteristics of real reverberation signals during adversarial training but also be constrained by the physical laws of marine sound propagation. This solves the problem of physical deviations in simulation results from purely data-driven models under specific complex marine environments, significantly improving the physical consistency between simulated signals and real reverberation in terms of energy attenuation, autocorrelation characteristics, spectral envelope, and reverberation intensity. Furthermore, this application uses sonar pulse signals as direct input to generate simulated reverberation time-series signals end-to-end, eliminating the need for complex propagation path calculations and scattering analysis as required by traditional normal mode theory methods. The method significantly reduces computational complexity through meta-accumulation, meeting the requirements of real-time simulation. Furthermore, due to the conditional generation capability of generative adversarial networks (GANs), when marine environmental parameters (such as sound velocity profiles, seabed type, and sea surface wind speed) change, only the input conditions need to be adjusted to quickly generate reverberation signals for the corresponding environment, without the need to remodel or calibrate physical parameters. This effectively solves the problem of insufficient adaptability of existing technologies to environmental changes. In addition, compared with traditional deep learning methods, the introduction of a physical consistency loss function allows the model to maintain adherence to physical laws even with limited training samples, reducing reliance on large-scale sea trial data and lowering data acquisition costs. Finally, the simulated reverberation time-series signal output by this method can be directly used for algorithm verification and performance evaluation of active sonar systems, providing high-fidelity, low-cost, and repeatable reverberation data support for the design of sonar systems for underwater platforms such as ships, submarines, and autonomous underwater vehicles. Attached Figure Description

[0018] Figure 1 This is the overall structural design diagram of the marine reverberation simulation software and hardware system of the present invention; Figure 2 This is a block diagram of the physical information conditional adversarial network structure in this invention; Figure 3 This is a diagram of the generator network structure in this invention; Figure 4 This is a diagram of the discriminator network structure in this invention; Figure 5 The image shows the time-domain signal and spectrum of the ocean reverberation signal. Figure 6 This is a probability distribution diagram of the envelope of the ocean reverberation signal. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [the described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [the described condition or event] is detected," or "in response to detection of [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] Example 1 This embodiment provides a marine reverberation simulation method based on physical information conditional generative adversarial networks, which includes the following steps: Includes the following steps: The acquired sonar pulse signal is used as the input to the physical information conditional generative adversarial network to obtain the simulated reverberation time sequence signal; The physical information conditional generative adversarial network includes a generator network structure and a discriminator network structure, wherein the discriminator network structure is embedded with a physical consistency loss function.

[0026] Example 2 This embodiment provides a marine reverberation simulation method based on physical information condition generative adversarial networks, applicable to reverberation signal simulation and monitoring in underwater active sonar environments. The method specifically includes the following steps: Step 1: Pulse signal acquisition stage.

[0027] First, the sampling frequency is set to 100kHz, and sonar pulse signals are continuously acquired in real time according to this sampling frequency; Secondly, the Hanning window function is used to weight the acquired sonar pulse signal. The adjacent frames are overlapped with a 50% overlap rate, that is, 2500 points are left at the beginning and end of each frame to overlap with the previous and next frames, so as to obtain the weighted pulse signal.

[0028] In this embodiment, the Hanning window function is used to weight the acquired pulse signal. The specific method is as follows: The first window of the signal spans from 300ms to 400ms, and this signal is truncated. Truncating means saving the signal according to time scales from 300ms to 400ms in formats such as CSV or MAT. The second window spans from 0 to 100ms, with the 0-100ms portion overlapping.

[0029] The Hanning window function is as follows: (1) In the formula, N represents the length of the window function.

[0030] The weighted pulse signal is input to the Fast Fourier Transform (FFT) module, and a 4096-point FFT is used for frequency domain transformation. The FFT result is logarithmically calculated by squared amplitude to generate short-time power spectral density (STFT), with a frequency resolution of approximately 12.2 Hz (50 kHz / 4096) and a time-domain frame shift of 50 ms (5000 × (1 – overlap rate) / 50 kHz).

[0031] Finally, low-pass filtering and gain normalization preprocessing are performed on the short-time power spectral density to eliminate analog front-end noise and DC components, resulting in a normalized pulse signal.

[0032] In this embodiment, the low-pass filtering refers to a basic operation in signal processing. Its function is to suppress signal components higher than a preset cutoff frequency fc in the time or frequency domain, and only retain components lower than or equal to fc. The resulting filtered signal is transformed back to the time domain signal by inverse FFT, and the result is a smooth signal that removes high-frequency noise and prevents aliasing.

[0033] The normalization process involves uniformly scaling the signal amplitude to eliminate gain differences between different acquisition channels or time segments, ensuring that the input signal has a consistent dynamic range and statistical characteristics for subsequent processing (such as feature extraction and model training). Peak normalization is used here, which is a common data preprocessing method. It uses the maximum and minimum values ​​of the data to perform a linear transformation on the original data, aiming to scale the range of data values ​​to [0, 1].

[0034] The normalized pulse signal is sent to the artificial intelligence computing platform via a high-speed Ethernet or fiber optic interface to provide the conditional input space for the training samples.

[0035] Step 2: Obtaining physical constraints for ocean reverberation. This step uses MATLAB 2021.

[0036] In this embodiment, the physical constraints of ocean reverberation include sound wave propagation path parameters, which include ocean interface scattering intensity, path delay, and attenuation.

[0037] The specific method for obtaining the physical constraints of ocean reverberation is as follows: First, based on traditional empirical formulas and relevant literature, the seabed parameters required to determine the parameters of the sound wave propagation path in the marine environment based on the small slope approximation model were determined.

[0038] In this embodiment, the seabed parameters include sound velocity profile (5 layers × 2 parameters, sound velocity value and layer thickness of each layer), seabed depth (50-200 m), water temperature (2-30 °C), salinity (30-35‰), signal frequency bandwidth (10 kHz-100 kHz), incident grazing angle (0°-90°), scattered grazing angle (0°-180°), scattered azimuth angle (-180°-180°), sea surface wind speed (0-10 m / s), and seabed sediment type (muddy, sandy, rocky).

[0039] In practical applications, if the system detects new background or echo conditions, it should process these new background or echo conditions.

[0040] In this embodiment, the new background or echo conditions are, for example, sound waves emitted by marine animals. These signals propagate differently from conventional underwater acoustic signals and may have unique frequency ranges, scattering characteristics, and attenuation behavior.

[0041] This embodiment uses the following method: Incident angle: The propagation path and incident angle of the signal are determined based on the relative orientation of the marine animal signal source.

[0042] Scattering glancing angle and scattering azimuth angle: Considering the directionality of marine animal vocalizations and possible echo paths, and taking into account environmental noise and the underwater medium, the scattering glancing angle and scattering azimuth angle are calculated.

[0043] Signal Classification and Categorization: This embodiment also incorporates a marine animal acoustic database, classifying and processing animal acoustic signals based on characteristics such as incident angle and frequency range, distinguishing them from conventional marine echo signals. In this way, this embodiment can process newly emerging animal signals and accurately reflect their unique propagation characteristics when generating simulated signals, improving the accuracy and adaptability of the simulation system in dynamic environments.

[0044] The sound velocity profile supports up to 5 layers of partitioned description. The thickness and sound velocity value of each layer are accurately reconstructed through linear or high-order interpolation to ensure a smooth transition of the sound velocity profile between water layers.

[0045] The roughness of the seabed bottom surface is described by the root mean square roughness (HRMS), and the type of seabed sediment is determined by the roughness of the seabed bottom surface.

[0046] Based on the obtained sound speed profile, water temperature, salinity, seabed depth, incident grazing angle, and scattered grazing angle, the path length and path delay corresponding to the sound wave propagation path are obtained. The signal frequency bandwidth is used to calculate attenuation parameters at different frequencies; Sea surface wind speed is used to calculate bubble scattering parameters; Calculate the scattering intensity at the ocean interface based on the incident grazing angle and the scattered grazing angle; In this embodiment, the ocean interface scattering intensity mainly consists of seabed scattering intensity and sea surface scattering intensity. Seabed scattering comprises specular scattering and non-spectral scattering; therefore, the seabed scattering intensity is defined as follows: (2) In the formula, For underwater mirror scattering, This is non-reflective scattering from the seabed. It is obtained from equations (3) and (4) respectively.

[0047] (3) (4) In the formula, The boundary condition scattering coupling coefficient for scattering from the rough seabed interface is related to the seabed density. The sound velocity attenuation of seabed compression waves c p Seafloor shear wave sound velocity attenuation c s It depends on the boundary conditions of the seabed; h rms denoted by root mean square roughness, it is related to the spectral intensity ω² and spectral index γ² of the seabed roughness spatial spectrum; P represents the absolute value of the horizontal amplitude of the difference between the incident wave vector and the scattered wave vector; P z δ represents the vertical amplitude of the difference between the incident wave vector and the scattered wave vector; η represents the Dirac function; I represents the integral containing the roughness space spectrum.

[0048] The expression for sea surface scattering intensity is: (5) In the formula, The intensity of scattering at the sea surface undulation interface; The intensity of bubble scattering; Indicates the incident glancing angle. ; Indicates the grazing angle of scattering. ; Indicates the azimuth angle of the scattering. .

[0049] MATLAB was used to calculate the propagation path of reverberant sound waves and obtain the scattering intensity of the ocean interface corresponding to multiple sound propagation paths.

[0050] The obtained marine reverberation physical constraints are sent to the condition input space of the artificial intelligence computing platform via high-speed Ethernet or fiber optic interface, realizing the seamless connection between environmental modeling and signal simulation data, and providing high-precision environmental priors for subsequent physical consistency generation of reverberation signals and online adaptive training.

[0051] Step 3: Construct and train a physical information-conditional generative adversarial network. This step is as follows: Figure 2 As shown, the normalized pulse signal obtained in step 1 and the ocean reverberation physical constraints obtained in step 2 are used to train the physical information conditional generative adversarial network.

[0052] The physical information conditional generative adversarial network structure includes a generator network structure, a discriminator network structure, and a physical consistency loss function.

[0053] Construction of the Conditional Input Space: The normalized pulse signal and the physical constraints of ocean reverberation are used to construct the conditional input space of the training samples. The space includes the sound velocity profile (5 layers × 2 parameters, sound velocity value and layer thickness of each layer, 10 dimensions in total), seabed depth (50-200 m, 1 dimension), water temperature (2-30 ℃, 1 dimension), salinity (30-35‰, 1 dimension), signal frequency bandwidth (10 kHz-100 kHz, 1 dimension), incident grazing angle (0°-90°, 1 dimension each for upper and lower limits, 2 dimensions in total), scattered grazing angle (0°-180°, 1 dimension each for upper and lower limits, 2 dimensions in total), scattered azimuth angle (-180°-180°, 1 dimension each for upper and lower limits, 2 dimensions in total), scattering intensity (1 dimension), path delay (1 dimension), path attenuation (1 dimension), normalized pulse signal (1 dimension), and seabed sediment type (muddy, sandy, and rock are encoded using one-hot encoding, 2 dimensions). The above parameters total 26 dimensions.

[0054] Generator network structure: This module is as follows Figure 3 As shown, the input is first concatenated with a 100-dimensional Gaussian random noise vector and 26-dimensional physical condition information to form a 126-dimensional feature vector. The network adopts a 5-layer deconvolutional residual structure. The first layer is a fully connected layer, which maps the 126-dimensional input to 4096-dimensional features (which can be regarded as 256×4×4) and reshapes it into a 4×4 256-channel tensor. The second to fifth layers are deconvolutional layers with 256, 128, 64 and 1 channels respectively. The kernel size is 4×4 and the stride is 2. After each deconvolution, a Conditional BatchNorm and LeakyReLU activation function (negative slope 0.2) are applied. After activation, an embedding vector is generated by a fully connected network that maps the 26-dimensional conditions to the number of channels in the current layer. This embedding vector is added to the features of the layer channel by channel, thereby achieving deep fusion of physical condition information and generating an initial simulated reverberation time series signal that conforms to the specified marine environmental conditions and physical constraints of sound propagation. The generator network structure outputs the final simulated reverberation timing signal based on the output of the discriminator network structure.

[0055] Discriminator network structure: This module is as follows Figure 4 As shown, a four-layer one-dimensional convolutional neural network is used, with the input being an initial simulated reverberation time-series signal of length 16384, where: The pulse signal branch in the initial simulated reverberation timing signal is used to extract features through four layers of convolution, with the number of channels being 64, 128, 256, and 512 respectively, and the convolution kernel size being 4×1 and the stride being 2.

[0056] The marine reverberation physical constraints in the initial simulated reverberation time-series signal are encoded into 64-dimensional conditional embeddings through two fully connected layers (128 and 64 hidden layer nodes), and then concatenated and fused with the signal features along the channel dimension after the third convolutional layer. The fused features are flattened and then passed through two fully connected layers (1024→1), and finally the Sigmoid activation function is used to output the realism probability. When the output probability is 0.8 or higher, it is determined to be true, and the output realism probability is fed back to the generator network structure.

[0057] The discriminator network structure embeds a physical consistency loss function, which quantifies the deviation between the generated data and physical laws by calculating the residual between the generated data and the physical model.

[0058] The physical consistency loss function introduces multiple loss constraints based on the physical model, including: Reverberation intensity calculation: (6) In the formula, The intensity of the scattering unit. Indicates the transmission of a signal. For the bidirectional time delay from the platform to the scattering unit, The center frequency of the transmitted signal. For relative Doppler frequency shift, This represents the number of scattering rings. This represents the number of scattering units.

[0059] Total loss function: (7) In the formula: , and These are the weighting coefficients; Let this be the energy loss function; The autocorrelation loss function; The spectral envelope loss function; This is the reverberation intensity loss function.

[0060] In this embodiment, each weight coefficient is determined through pre-experimentation, grid search, or empirical parameter tuning. Each weight coefficient is a non-negative number and satisfies λ1+λ2+λ3+λ4=1.

[0061] The energy loss function is used to constrain the consistency of the generated reverberant signal with the real reverberant signal or the signal calculated by the physical model in terms of energy attenuation characteristics. Its expression is: (8) In the formula, This represents the generated reverberation timing signal. Represents the actual reverberation signal or the reference reverberation signal calculated from a physical model; Indicates the number of time windows; Indicates the generation of the reverberation signal at the th Short-term energy within a time window; Indicates the reference reverberation signal at the 1st Short-term energy within a time window.

[0062] The short-time energy can be expressed as: (9) In the formula, Indicates the signal to be calculated. Indicates the first A time window.

[0063] The autocorrelation loss function is used to constrain the consistency of the generated reverberant signal with the real reverberant signal or the signal calculated by the physical model in terms of time-related structure, and its expression is: (10) In the formula, Indicates the number of time delay points selected; Indicates the first One time delay; This indicates that the generated reverberation signal has a time delay. The normalized autocorrelation function under the given conditions; This indicates the reference reverberation signal in time delay The normalized autocorrelation function under the given conditions.

[0064] The normalized autocorrelation function can be expressed as: (11) The spectral envelope loss function is used to constrain the consistency of the generated reverberant signal with the real reverberant signal or the signal calculated by the physical model in the frequency domain envelope shape, and its expression is: (12) In the formula, Indicates the number of frequency sampling points; Represents the spectral envelope of the generated reverberant signal; This represents the spectral envelope of the reference reverberation signal. The spectral envelope is obtained by performing a Fast Fourier Transform, amplitude spectrum calculation, and smoothed envelope extraction on either the generated or reference reverberation signal.

[0065] In this embodiment, the following weight combinations were determined through preliminary experiments for different seabed types:

[0066] In this embodiment, the residual between the generated signal and the physical model calculation results is first incorporated into existing constraints such as energy attenuation loss, spectral envelope loss, and reverberation intensity loss. Specifically, the introduction of a new scattering mechanism or noise model will decompose the new loss term into a form related to the existing loss function. Examples of the two new constraints are as follows: When seabed topography changes, necessitating the introduction of a complex seabed topography scattering model: Assume the new physical model incorporates a more complex seabed topography scattering model that considers the impact of different types of seabed sediments and topographic undulations on sound waves. In this case, the error between the generated signal and the scattering intensity calculated by the model will be added to the reverberation intensity loss. For example, the scattering intensity of the generated signal differs from the scattering intensity calculated using this new scattering model. This error increases the value of the loss function, prompting the generator to adjust the network weights to reduce the error and make the generated signal more consistent with the actual impact of complex seabed topography on sound wave scattering.

[0067] When sea surface conditions change and a bubble scattering model needs to be introduced: if the new model involves a more accurate modeling of bubble scattering from the sea surface, and considering the influence of bubbles on signals of different frequencies, this new mechanism can be incorporated into the energy attenuation loss. In this context, bubble scattering affects the signal propagation speed and attenuation characteristics. The error between the energy attenuation characteristics of the generated signal and those calculated by the new bubble model is added to the energy attenuation loss term. For example, the attenuation of the generated signal in densely bubble regions deviates from the ideal attenuation calculated by the new model. This deviation will be reflected in the energy attenuation loss term. Optimizations were performed to ensure that the energy decay behavior of the generated signal under the influence of bubbles is more in line with physical laws.

[0068] These examples demonstrate that new physical constraints, when appropriately decomposed into existing loss functions, can help models more accurately simulate and generate reverberant signals that conform to complex physical conditions, while maintaining physical consistency and high signal quality.

[0069] After completing the architecture of the physical information-conditional generative adversarial network, the network's input and output are defined. The network input consists of physical information and noise. First, a 100-dimensional Gaussian random noise vector is concatenated with 26-dimensional physical condition information to form a 126-dimensional feature vector. The network output consists of simulated reverberation time-series data that conforms to specified physical conditions. Based on the network's input-output relationship, a training dataset is constructed. This training dataset includes 2000 samples of real ocean reverberation data and 10000 samples of physical modeling simulation data.

[0070] After defining the input and output content of the physical information conditional generative adversarial network (PEA), the PEA is trained on an artificial intelligence computing platform. The training strategy employed is mixed-precision training (FP16+FP32), with a batch size of 16 and a generator learning rate of 2×10⁻⁶. -4 The discriminator learning rate is 1×10. -4 The training process consists of 200 rounds (the number of training rounds is preset to 10,000 rounds, and an early stopping mechanism is used to monitor the dynamic changes of the generator loss index during training. When the generator loss index does not show a smaller value for 20 consecutive rounds, training is stopped, and the weights and architecture of the model with the smallest generator loss index at that point are saved as the optimal model).

[0071] Step 4: Deploy a physical information conditional generative adversarial network (PEA) to augment data. The Physical Information Conditional Generative Adversarial Network (PIGAN) refers to the optimal model of the PIGAN trained in step 3. This step deploys this optimal model on an artificial intelligence computing platform.

[0072] The model deployment employs a dynamic memory pool and thread pool reuse strategy, utilizing heterogeneous coprocessors of ARM Cortex-A78 and NVIDIA Jetson Orin NX for parallel acceleration. Low-latency computation (below 20ms) is achieved through mixed-precision (FP16+FP32) and multi-stream GPU pipelines. A single reverberation signal generation takes only about 2ms, meeting the requirements of real-time online simulation. Simultaneously, the multi-threading mechanism supports the concurrent generation of signal samples under multiple environmental conditions, ensuring rapid response to control strategies. This stage employs a dynamic memory pool and thread pool reuse strategy to minimize peak inference power consumption and memory allocation overhead, ensuring long-term stable system operation.

[0073] There are two triggering methods for the data augmentation process. One is online triggering, which means that when the discriminator's ability to distinguish the distribution of real samples from expanded samples is lower than a preset threshold, or when insufficient diversity of the training set is detected, the system automatically starts a data augmentation. The other is timed triggering, which is configured to perform batch augmentation tasks every 12 hours to continuously supplement samples under different sea conditions and working conditions, so as to meet the model's adaptive needs to new environments.

[0074] In each triggering method, the same steps are followed to invoke the deployed Physical Information Conditional Generative Adversarial Network (PEGAN) module to generate extended samples. Using the aforementioned 26-dimensional conditional parameters and 100-dimensional random noise as input, an artificial intelligence computing platform is used to generate reverberation signals in real time. The generation operation refers to receiving the physical conditional parameter vector and Gaussian random noise vector from the conditional input, concatenating them, and then feeding them into the PEGAN. The physical conditional parameters include sound velocity profile, seabed depth, water temperature, salinity, and incident angle, totaling 26 dimensions, which are deeply fused with the features of each layer of the network through a conditional batch normalization layer; the random noise vector has a dimension of 100. During the network's forward inference process, the generator sequentially passes through 5 layers of deconvolutional residual structures, mapping the 126-dimensional input to a time-domain reverberation signal, and outputting a preliminary reverberation waveform as an extended sample in real time. The discriminator then distinguishes between the extended samples and the real signal in parallel.

[0075] After obtaining the generated extended samples, the signal is processed to obtain the final frequency domain features of the extended samples. A short-time energy envelope extraction algorithm, a well-known algorithm in the field of signal processing, is used. The time-domain statistical instantaneous value versus envelope value curve is calculated. The curve has a frequency resolution of 12.2 Hz, with obvious peak-valley characteristics, and can be used for subsequent path analysis and Doppler estimation. The entire post-processing pipeline takes less than 10 ms on the edge platform, and the results are directly written to the user interaction module via zero-copy DMA.

[0076] The formula for calculating the instantaneous reverberation value is: (8) In the formula, W is the probability density. Let V be the instantaneous variance of the reverberation signal, and let V be the instantaneous value of the reverberation signal, which follows a Gaussian distribution.

[0077] The formula for calculating the reverberation envelope is: (9) In the formula: W is the probability density. Let be the variance of the reverberation signal amplitude, and E be the envelope amplitude of the reverberation signal, which follows a Rayleigh distribution.

[0078] After obtaining the statistical properties of the expanded sample, an online quality assessment was performed. The dot product ratio of the signal statistical result curve PDF was used as the evaluation criterion to assess the similarity between the generated signal and the real signal. Under various typical shallow sea environments and moving platform conditions, the prediction error was less than 5%, and good consistency was demonstrated in both large-scale offline evaluation and a small number of online field tests.

[0079] After generating and evaluating qualified extended samples, the data augmentation stage integrates the new samples with the original observation data into the training set through a database synchronization service. These new samples are then prioritized for use in the next round of incremental model training to improve the model's fitting ability and robustness to complex shallow-sea reverberation signals. Robustness refers to the ability to maintain key performance indicators without significant degradation under conditions such as input disturbances and environmental changes. This stage forms a closed loop with modules such as signal acquisition, environmental modeling, and predictive inference, continuously optimizing network weights and physical consistency. This enables the system to achieve real-time reverberation simulation, online incremental learning, and dynamic suppression in practical applications, meeting the high-precision and high-reliability operational requirements of marine active sonar.

[0080] Step 5: Export the analysis results. In this embodiment, the following functional modules are included: Plot the time-domain waveforms, spectrum, and envelope statistics curves output from steps 1 and 4. (Corresponding to...) Figures 5-6 ) The pulse signals acquired in Step 1 are used in MATLAB to plot time-domain waveforms and spectrum diagrams with time on the horizontal axis and amplitude on the vertical axis. The data is then saved using MATLAB in common data table formats (CSV), bitmap image formats (PNG), vector graphics formats (SVG), and PDF formats that can be embedded in reports. The envelope statistics obtained in Step 5 are also saved using MATLAB in common data table formats (CSV), bitmap image formats (PNG), vector graphics formats (SVG), and PDF formats that can be embedded in reports.

[0081] Users can choose to export spectrum analysis results, time-frequency graphs, and envelope distribution curves, and can customize the file name, storage path, and data field order in the export settings.

[0082] Example 3 This embodiment provides a marine reverberation simulation system based on a physical information condition generative adversarial network, comprising: The simulated reverberation signal acquisition unit is used to take the acquired sonar pulse signal as the input of the physical information conditional generative adversarial network to obtain the simulated reverberation timing signal. A generative adversarial network (GAN) unit is used to deploy a physical information conditional GAN, wherein the physical information conditional GAN ​​includes a generator network structure and a discriminator network structure, and the discriminator network structure embeds a physical consistency loss function.

[0083] Example 4 This embodiment also provides a computing device. The computing device includes a bus, a processor, a memory, and a communication interface. The processor, memory, and communication interface communicate with each other via the bus. The computing device can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memory in the computing device.

[0084] A bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, a bus can include a path for transmitting information between various components of a computing device (e.g., memory, processor, communication interfaces).

[0085] The processor may include any one or more of the following: central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), application specific integrated circuit (ASIC), field-programmable gate array (FPGA), microprocessor (MP), or digital signal processor (DSP).

[0086] Memory can include volatile memory, such as random access memory (RAM). Processors can also include non-volatile memory. volatile memory, such as read-only memory (ROM). ROM (memory only), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0087] The memory stores executable program code, which the processor executes to implement the functions of the aforementioned units, thereby achieving, for example, the method described in Embodiment 1. That is, the memory may store instructions for the methods and functions relating to the computing device in any of the above embodiments.

[0088] The communication interface uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between computing devices and other devices or communication networks.

[0089] Example 5 This embodiment also provides a computing device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device such as a desktop computer, a laptop computer, or a smartphone.

[0090] The computing device cluster includes at least one computing device. The memory of one or more computing devices in the computing device cluster may store the same instructions for performing the methods and functions related to the computing devices in any of the above embodiments.

[0091] In some possible implementations, the memories of one or more computing devices in the computing device cluster may also each store partial instructions for executing the methods and functions of the computing devices involved in any of the above embodiments. In other words, a combination of one or more computing devices can jointly execute the instructions for performing the methods and functions of the computing devices.

[0092] It should be noted that the memory in different computing devices within a computing device cluster can store different instructions, which are used to execute parts of the device's functions.

[0093] In some possible implementations, one or more computing devices in a computing device cluster can be connected via a network. This network can be a wide area network (WAN) or a local area network (LAN), etc. Two computing devices are connected via the network. Specifically, they connect to the network through communication interfaces on each computing device.

[0094] Embodiments of this disclosure also provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions related to a computing device in any of the above embodiments.

[0095] Example 6 This embodiment also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, cause the processor to perform the methods and functions of the computing device involved in any of the above embodiments.

[0096] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0097] Example 7 This embodiment provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods as described above with reference to the accompanying drawings. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0098] Computer program code used to implement the methods of this disclosure may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the computer or other programmable data processing apparatus, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be performed. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0099] In the context of this disclosure, computer program code or related data may be carried on any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and so on. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0100] Computer-readable media can be any tangible medium that contains or stores programs for or relating to an instruction execution system, apparatus, or device, or a data storage device such as a data center containing one or more available media. Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. More detailed examples of computer-readable storage media include electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0101] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A marine reverberation simulation method based on physical information condition generative adversarial networks, characterized in that, Includes the following steps: The acquired sonar pulse signal is used as the input to the physical information conditional generative adversarial network to obtain the simulated reverberation time sequence signal; The physical information conditional generative adversarial network includes a generator network structure and a discriminator network structure, wherein the discriminator network structure is embedded with a physical consistency loss function.

2. The ocean reverberation simulation method based on physical information condition generative adversarial networks according to claim 1, characterized in that, The expression for the physical consistency loss function for: in, Let this be the energy loss function; The autocorrelation loss function; The spectral envelope loss function; This is the reverberation intensity loss function.

3. The ocean reverberation simulation method based on physical information condition generative adversarial networks according to claim 2, characterized in that, The expression for the reverberation intensity loss function is: （6） In the formula, The intensity of the scattering unit. Indicates the transmission of a signal. For the bidirectional time delay from the platform to the scattering unit, The center frequency of the transmitted signal. For relative Doppler frequency shift, This represents the number of scattering rings. This represents the number of scattering units.

4. The ocean reverberation simulation method based on physical information condition generative adversarial networks according to claim 1, characterized in that, The training process of the physical information-conditional generative adversarial network: Constructing physical constraints for ocean reverberation, the physical constraints for ocean reverberation include sound wave propagation path parameters, the sound wave propagation path parameters include ocean interface scattering intensity, path delay and attenuation; The physical constraints of ocean reverberation, seabed parameters, and acquired historical sonar pulse signals are used as inputs to construct a physical information conditional generative adversarial network (PEGAN). The PEGAN is then trained to obtain the trained PEGAN.

5. The ocean reverberation simulation method based on physical information conditional generative adversarial networks according to claim 4, characterized in that, The acquired historical sonar pulse signals are input into the physical information conditional generative adversarial network (GAN) for preprocessing, specifically including: The acquired historical sonar pulse signals were weighted using the Hanning window function to obtain the weighted pulse signals. The weighted pulse signal is subjected to Fourier transform to obtain the short-time power spectral density; The short-time power spectral density is preprocessed by low-pass filtering and gain normalization to obtain the normalized pulse signal; The normalized pulse signal obtained is used as the input to the constructed physical information conditional generative adversarial network.

6. A marine reverberation simulation system based on physical information condition generative adversarial networks, characterized in that, include: The simulated reverberation signal acquisition unit is used to take the acquired sonar pulse signal as the input of the physical information conditional generative adversarial network to obtain the simulated reverberation timing signal. A generative adversarial network (GAN) unit is used to deploy a physical information conditional GAN, wherein the physical information conditional GAN ​​includes a generator network structure and a discriminator network structure, and the discriminator network structure embeds a physical consistency loss function.

7. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer instructions that, when executed by the processor, cause the electronic device to perform the method of any one of claims 1 to 5.

8. A computing device cluster, characterized in that, It includes at least one computing device, each computing device including a processor and memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device to cause the cluster of computing devices to perform the method according to any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes computer-executable instructions that, when executed, implement the method of any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when executed by a processor, implement the method of any one of claims 1 to 5.

Images