Underwater target material attribute identification method and system based on acoustic-electromagnetic intermodulation characteristics

Abstract

The invention discloses an underwater target material attribute identification method and system based on acoustic electromagnetic intermodulation characteristics, and relates to the technical field of target material identification. The method comprises the following steps: transmitting sound waves underwater to excite a target to vibrate; synchronously acquiring an acoustic-magnetic mixed signal; acoustic echoes are separated, and magnetic field segments with aligned time sequences are intercepted; reconstructing theoretical magnetic field frequency domain energy characteristics by using a cross-modal generation model and taking acoustic echoes as conditions, and generating acoustic electromagnetic intermodulation characteristics; fusing the acoustic multi-dimensional features and the acoustic electromagnetic intermodulation features, and constructing a multi-modal fusion feature vector of each to-be-detected target; and inputting the multi-modal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification judgment, thereby realizing identification and marking of material attributes of different to-be-detected targets. Through the generative artificial intelligence and machine learning classification technology, the problems that underwater environment signals are weak and materials are difficult to distinguish are solved.

Description

Technical Field

[0001] This invention relates to the field of target material identification technology, specifically to a method and system for identifying underwater target material properties based on acoustic-electromagnetic intermodulation characteristics. Background Technology

[0002] Underwater target detection and identification are core technologies in marine engineering, underwater security, and resource development. Currently, commonly used detection methods primarily rely on sonar technology, which detects targets by emitting sound waves and receiving the reflected echoes. However, in the complex seabed environment, man-made targets made of metallic materials and natural backgrounds made of non-metallic materials often have extremely similar geometric shapes and acoustic reflection intensities. Relying solely on the intensity or time-of-flight characteristics of the acoustic echo is insufficient to effectively distinguish the material properties of the target, resulting in a high false alarm rate and severely impacting the accuracy and efficiency of detection.

[0003] To address the challenge of material identification, detection techniques utilizing the acoustic-electromagnetic intermodulation mechanism have gained increasing attention. This technology is based on the physical principle that the vibration of a conductor in a magnetic field generates an induced electromagnetic field, theoretically capable of directly reflecting the conductive material properties of a target. However, in practical underwater applications, this technology faces significant challenges: First, the acoustic-magnetic field signals generated by underwater targets are extremely weak and are often submerged in the electromagnetic noise of the marine environment. More importantly, due to the unknown attitude of non-cooperative underwater targets relative to the Earth's magnetic field, and the significant differences in the propagation characteristics of sound waves and electromagnetic waves, the phase of the received magnetic field signal has a high degree of uncertainty. Traditional methods based on time-domain waveform matching or simple frequency-domain feature extraction struggle to extract stable and discriminative features when the phase is unknown, leading to a sharp decline in recognition performance.

[0004] Secondly, most existing identification methods rely on single-dimensional features for judgment, lacking systematic learning of the feature changes of targets under different angles and environments. In the complex and ever-changing underwater environment, single features are often not robust enough to form a stable and reliable basis for judgment. The lack of an effective mechanism to construct a feature sample library containing rich material information and to utilize intelligent methods such as machine learning to deeply mine and classify multi-dimensional features is another major bottleneck limiting the accurate identification of underwater target materials.

[0005] Therefore, how to obtain high-quality acoustic-magnetic correlation features under underwater conditions of unknown phase and weak signal, and establish a systematic sample learning and classification mechanism to achieve high-precision identification and labeling of target material properties, is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To address the challenges of distinguishing between metallic and non-metallic targets in underwater environments using existing technologies, as well as the technical difficulty of weak signals in acoustic-electromagnetic intermodulation detection, this invention provides a method and system for identifying the material properties of underwater targets based on acoustic-electromagnetic intermodulation characteristics.

[0007] In a first aspect, the present invention provides a method for identifying the material properties of underwater targets based on acoustic-electromagnetic intermodulation characteristics, comprising the following steps: S1. Transmit acoustic signals into the underwater detection area to acoustically excite multiple targets within the detection area, causing the targets to generate mechanical vibration responses. S2. Use an underwater composite sensor array to synchronously acquire mixed observation signals in the detection area. The mixed observation signals include acoustic aliasing signals containing multiple target echoes and environmental magnetic field signals. S3. Perform spatiotemporal separation processing on the acoustic aliasing signal to detect the acoustic independent echo segment corresponding to each target and its acoustic delay time relative to the emission time; based on the acoustic delay time, backtrack and extract the magnetic field segment to be analyzed that is time-aligned with the excitation vibration time of the target in the environmental magnetic field signal. S4. For each target to be tested, extract the acoustic multidimensional features of the independent acoustic echo segments; at the same time, map the independent acoustic echo segments into conditional vectors, and combine them with random noise vectors conforming to a preset distribution to input into a pre-trained cross-modal generation model to reconstruct the theoretical magnetic field frequency domain energy features; calculate the spectral similarity between the theoretical magnetic field frequency domain energy features and the measured frequency domain energy features of the magnetic field segment to be analyzed, as the acoustic-electromagnetic intermodulation features; S5. The acoustic multidimensional features are fused with the acoustic-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each target under test; S6. Input the multimodal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification and judgment, thereby realizing the identification and labeling of material properties of different test targets.

[0008] Specifically, step S3 includes the following steps: S31. Spatial filtering is performed on the acoustic aliasing signal to separate the acoustic wave beams from different directions. S32. Perform matched filtering detection on each acoustic beam, and extract multiple acoustic pulse segments using a sliding time window based on the order of echo arrival times. S33. Mark each acoustic pulse segment as a corresponding acoustic independent echo segment, and determine the acoustic delay time of the acoustic independent echo segment relative to the transmitted signal; based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves, calculate the excitation time at which the target under test generates a mechanical vibration response; in the environmental magnetic field signal, extract the magnetic field data centered on the excitation time and with a duration corresponding to the acoustic independent echo segment, as the magnetic field segment to be analyzed.

[0009] By employing the above technical solutions and signal processing techniques that combine spatial filtering and matched filtering, it is possible to accurately separate independent acoustic echoes belonging to specific targets from aliased signals in underwater environments with multiple coexisting targets and strong reverberation. Furthermore, based on precise timestamps, time-series aligned extraction of magnetic field data is achieved, laying a reliable data foundation for subsequent feature extraction and correlation analysis.

[0010] Specifically, step S4, the acoustic multidimensional features, includes the envelope amplitude features of the acoustic independent echo segments in the time domain and the Mel frequency cepstral coefficients in the frequency domain.

[0011] Specifically, the cross-modal generative model is trained based on the diffusion model, and the specific training steps include: A training dataset containing paired acoustic echo samples and real magnetic field amplitude spectrum samples is constructed; random noise of different intensities is superimposed on the real magnetic field amplitude spectrum samples to generate noise samples; the noise samples and acoustic echo samples are used as conditional inputs to the model to train the model to predict the random noise components contained in the noise samples; the error between the predicted random noise components and the actual superimposed random noise components is calculated, and the model parameters are updated based on the objective of minimizing the error.

[0012] By employing the above technical solutions and leveraging the powerful denoising and generation capabilities of the diffusion model, the model is forced to learn and recover the acoustically constrained magnetic field signal structure from Gaussian noise. Even when the measured acoustic signal contains a certain amount of environmental noise, the trained model can still robustly reconstruct high-fidelity theoretical magnetic field characteristics based on the learned probability distribution, thus enhancing the physical interpretability of the features.

[0013] Specifically, step S4 maps independent acoustic echo segments into conditional vectors, combines them with random noise vectors conforming to a preset distribution, inputs them into a pre-trained cross-modal generation model, and reconstructs the theoretical magnetic field frequency domain energy features corresponding to the target under test. The specific steps include: S41. Map the acoustic independent echo segments to latent space condition vectors and sample to generate Gaussian random noise vectors with the same dimension as the frequency domain energy characteristics of the theoretical magnetic field. S42. Input the Gaussian random noise vector and the latent space condition vector into the cross-modal generation model. Use the cross-modal generation model to transform the Gaussian random noise vector into a theoretical magnetic field amplitude spectrum constrained by the latent space condition vector. The theoretical magnetic field amplitude spectrum serves as the frequency domain energy feature of the theoretical magnetic field. S43. Perform spectral transformation on the magnetic field segment to be analyzed to obtain the measured amplitude spectrum, and normalize the theoretical magnetic field amplitude spectrum and the measured amplitude spectrum respectively, and calculate the cosine similarity and energy ratio characteristics of the two in the normalized frequency domain distribution. S44. The vector composed of cosine similarity and energy ratio characteristics is determined as the acoustic-electromagnetic intermodulation characteristic.

[0014] The above technical solutions transform the problem of matching time-domain waveforms into a similarity measurement problem of frequency-domain amplitude spectra, effectively avoiding the problem of phase ambiguity of magnetic field signals caused by unknown attitude of underwater non-cooperative targets, and achieving high-sensitivity feature matching under conditions without reference phase.

[0015] Specifically, step S6, the machine learning model includes at least one of support vector machine, convolutional neural network, or random forest; the pre-trained machine learning model is trained through the following steps: acquiring and fusing the acoustic multidimensional features and acoustic-electromagnetic intermodulation features of known materials to construct a material feature sample library containing multimodal fusion feature vectors of known materials and corresponding material attribute labels; importing the material feature sample library into the machine learning model for classification training to construct a classification decision boundary; the specific steps of identification and labeling include: using the pre-trained machine learning model to classify the multimodal fusion feature vectors extracted in real time and outputting material attribute labels; and labeling and outputting the attributes of targets determined to be electromagnetic response materials based on the material attribute labels.

[0016] By using the above technical solutions and machine learning models with different algorithm architectures to classify and train the sample database, the potential bias of a single decision criterion can be effectively reduced.

[0017] Secondly, the present invention also provides an underwater target material property identification system based on acoustic-electromagnetic intermodulation characteristics, comprising: The acoustic excitation module is configured to transmit acoustic signals into the underwater detection area to acoustically excite multiple targets within the detection area, causing the multiple targets to generate mechanical vibration responses. The synchronous acquisition module is configured to synchronously acquire mixed observation signals of the detection area using an underwater composite sensor array. The mixed observation signals include acoustic aliasing signals containing multiple target echoes and environmental magnetic field signals. The signal separation and interception module is configured to perform spatiotemporal separation processing on the acoustic aliasing signal, detect the acoustic independent echo segment corresponding to each target under test and its acoustic delay time relative to the emission time; based on the acoustic delay time, backtrack and intercept the magnetic field segment to be analyzed from the environmental magnetic field signal that is time-aligned with the excitation vibration time of the target under test. The cross-modal reconstruction and feature extraction module is configured to extract the acoustic multidimensional features of each acoustic independent echo segment for each target under test. Simultaneously, the acoustic independent echo segments are mapped into conditional vectors, and combined with random noise vectors conforming to a preset distribution, they are input into a pre-trained cross-modal generation model to reconstruct the theoretical magnetic field frequency domain energy features. The spectral similarity between the theoretical magnetic field frequency domain energy features and the measured frequency domain energy features of the magnetic field segment to be analyzed is calculated as the acoustic-electromagnetic intermodulation feature. The multimodal fusion module is configured to fuse acoustic multidimensional features with acousto-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each target under test. The classification decision module is configured to input multimodal fusion feature vectors into a machine learning model pre-trained based on a preset material feature sample library for classification decisions, thereby enabling the identification and labeling of material properties of different test targets.

[0018] Specifically, the signal separation and interception module is configured to perform spatial filtering on the acoustic aliasing signal to separate acoustic wave beams from different directions; perform matched filtering detection on each acoustic wave beam, and intercept multiple acoustic pulse segments using a sliding time window according to the order of echo arrival times; mark each acoustic pulse segment as a corresponding acoustic independent echo segment, and determine the acoustic delay time of the acoustic independent echo segment relative to the transmitted signal; calculate the excitation time at which the target under test generates a mechanical vibration response based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves; and intercept magnetic field data centered on the excitation time and with a duration corresponding to the acoustic independent echo segment from the environmental magnetic field signal as the magnetic field segment to be analyzed.

[0019] Specifically, the cross-modal reconstruction and feature extraction module is configured to map acoustic independent echo segments into latent space condition vectors and sample and generate Gaussian random noise vectors with the same dimension as the theoretical magnetic field frequency domain energy features. The Gaussian random noise vector and the latent space condition vector are input into the cross-modal generation model, which uses the cross-modal generation model to transform the Gaussian random noise vector into a theoretical magnetic field amplitude spectrum constrained by the latent space condition vector. The theoretical magnetic field amplitude spectrum serves as the theoretical magnetic field frequency domain energy feature. The magnetic field segment to be analyzed is subjected to spectral transformation to obtain the measured amplitude spectrum. The theoretical magnetic field amplitude spectrum and the measured amplitude spectrum are normalized respectively, and the cosine similarity and energy ratio features of the two in the normalized frequency domain distribution are calculated. The vector composed of the cosine similarity and energy ratio features is determined as the acoustic-electromagnetic intermodulation feature.

[0020] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0021] The present invention achieves the following beneficial effects: This invention introduces a cross-modal generation model as a pre-feature enhancement method, reconstructing high-quality theoretical magnetic field frequency domain energy features by learning physical mapping relationships. This not only avoids the phase ambiguity problem caused by the unknown attitude of underwater targets, but also significantly enhances the robustness of the features. It achieves fully automated processing from signal acquisition and feature reconstruction to classification and labeling, and can directly output detection results with clear material attribute labels, greatly improving the intelligence level and decision-making efficiency of underwater detection. Attached Figure Description

[0022] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0023] Figure 1 This is a flowchart illustrating the underwater target material property identification method based on acoustic-electromagnetic intermodulation characteristics provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the specific process of step S3 provided in the embodiment of the present invention; Figure 3 This is a schematic diagram of the underwater target material property identification system based on acoustic-electromagnetic intermodulation characteristics provided in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment provides a method for identifying the material properties of underwater targets based on acoustic-electromagnetic intermodulation characteristics. This method utilizes active acoustic excitation to induce acoustic-electromagnetic intermodulation signals in the target, reconstructs theoretical magnetic field characteristics through a cross-modal generation model, and combines a sample database with machine learning techniques to achieve accurate identification of the material properties of underwater targets. The method specifically includes the following steps: S1. Transmit acoustic signals into the underwater detection area to acoustically excite multiple targets within the detection area, causing them to generate mechanical vibration responses.

[0026] Specifically, in order to achieve long-distance propagation and effectively excite target vibration in an underwater environment, the transmitted acoustic signal is preferably a high-power broadband signal.

[0027] In a preferred embodiment, the source level of the emitted acoustic signal should be sufficient to induce adequate mechanical vibration in the target. When the acoustic wave impacts the underwater target, the target's shell or structure experiences minute mechanical vibrations. For targets made of electromagnetically responsive materials, as good conductors, high-frequency minute vibrations against the Earth's magnetic field background will cut magnetic field lines, thereby generating a weak induced alternating magnetic field around the target. This magnetic field is the acousto-electromagnetic intermodulation signal, and its frequency characteristics are highly correlated with the acoustic excitation frequency. For targets made of non-electromagnetically responsive materials, due to their extremely poor conductivity or insulator nature, even if vibration occurs, no significant induced magnetic field will be generated.

[0028] S2. Use an underwater composite sensor array to synchronously acquire mixed observation signals in the detection area. The mixed observation signals include acoustic aliasing signals containing echoes from multiple targets and environmental magnetic field signals.

[0029] In a preferred embodiment, an underwater composite sensor array consisting of multiple nodes can be constructed, such as a tetrahedral array or a cross-shaped array structure, containing at least four composite sensing nodes. Each node integrates a broadband underwater acoustic transducer for receiving acoustic signals and a high-sensitivity magnetic sensor for receiving magnetic signals. To achieve precise alignment of cross-modal signals, all sensor nodes are connected to a unified high-precision clock source to ensure strict synchronization of the timestamps of the acoustic and magnetic field signals.

[0030] S3. Perform spatiotemporal separation processing on the acoustic aliasing signal to detect the acoustic independent echo segment corresponding to each target and its acoustic delay time relative to the emission time; based on the acoustic delay time, backtrack and extract the magnetic field segment to be analyzed that is time-aligned with the excitation vibration time of the target in the environmental magnetic field signal.

[0031] Specifically, such as Figure 2 As shown, this step aims to solve the problem of multi-target interference by using the strong directionality of acoustic signals to guide the interception of magnetic field signals. Specifically, it includes: S31. Spatial filtering of acoustic aliasing signals. Utilizing the phase difference of each array element, a minimum variance distortionless response beamforming algorithm is employed to perform an all-around scan of the detection space, separating acoustic beams from different directions. This isolates acoustic echoes from different targets in the spatial domain and significantly suppresses uncorrelated environmental background noise through coherent accumulation technology, greatly improving the signal-to-noise ratio. Consequently, the detection and identification range of the system is effectively extended without increasing the transmission power.

[0032] S32. Perform matched filtering detection on each acoustic beam. Use the transmitted signal as a reference template to perform cross-correlation calculation with the beam signal and detect the correlation peak. Based on the order of echo arrival times, i.e., the distance to the target, use a sliding time window to extract multiple acoustic pulse segments.

[0033] S33. Mark each acoustic pulse segment as a corresponding independent acoustic echo segment, and determine the acoustic delay time of the independent acoustic echo segment relative to the transmitted signal. Based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves, calculate the excitation time at which the target under test generates a mechanical vibration response. In the environmental magnetic field signal, extract the magnetic field data centered on the excitation time and corresponding to the duration of the independent acoustic echo segment as the magnetic field segment to be analyzed.

[0034] S4. For each target to be tested, extract the acoustic multidimensional features of the independent acoustic echo segments; at the same time, map the independent acoustic echo segments into conditional vectors, and combine them with random noise vectors conforming to a preset distribution to input into a pre-trained cross-modal generation model to reconstruct the theoretical magnetic field frequency domain energy features; calculate the spectral similarity between the theoretical magnetic field frequency domain energy features and the measured frequency domain energy features of the magnetic field segment to be analyzed, as the acoustic-electromagnetic intermodulation features.

[0035] Specifically, the acoustic multidimensional features include the envelope amplitude characteristics of independent acoustic echo segments in the time domain and the Mel frequency cepstral coefficients in the frequency domain. These features reflect information about the target's geometry and surface roughness.

[0036] Specifically, the extraction process of acoustic-electromagnetic intermodulation features aims to solve the problem of weak magnetic fields without reference signals. The specific process is as follows: S41. The acoustic independent echo segments are mapped into latent space conditional vectors using a feature encoder. Simultaneously, a standard Gaussian random noise vector with the same dimension as the theoretical magnetic field frequency domain energy feature is sampled and generated in computer memory.

[0037] S42. Input the Gaussian random noise vector and the latent space condition vector into the cross-modal generation model. The model is constructed based on the denoising diffusion probability model. Using the inverse denoising process, the Gaussian random noise vector is transformed into a theoretical magnetic field amplitude spectrum constrained by the latent space condition vector. This amplitude spectrum represents the magnetic field energy distribution that should theoretically occur, assuming that the current acoustic echo is generated by an electromagnetically responsive material target.

[0038] Specifically, the model is built based on a denoised diffusion probability model. The main network of the model can adopt a U-Net network structure with a cross-attention mechanism or a Transformer architecture. During the training phase, a training dataset containing paired acoustic echo samples and real magnetic field amplitude spectrum samples is constructed. The training dataset is generated based on the settings of physics simulation software. Specifically, a three-dimensional finite element model of the underwater target is constructed using multiphysics simulation software. The target material parameters are set, and the surface vibration velocity of the target under acoustic excitation is calculated through the acoustic-structure interaction equation. Then, the induced magnetic field generated by the vibrating conductor cutting magnetic field lines in the background magnetic field is calculated, thereby obtaining paired acoustic echo and magnetic field data. Random noise of different intensities is superimposed on the real magnetic field amplitude spectrum samples to generate noise samples. The noise samples and acoustic echo samples are used as conditional inputs to the model, and the model is trained to predict the random noise components contained in the noise samples. The resonant frequency and damping coefficient of the target shell in electromagnetic response class reflect the mechanical properties of the target, while the acoustic-electromagnetic intermodulation characteristics reflect the electrical properties of the target. The cross-modal generative model learns from massive samples to uncover the intrinsic mapping and joint distribution patterns of target mechanical vibration characteristics and electromagnetic response characteristics in multidimensional physical space. It calculates the error between the predicted random noise components and the actual superimposed random noise components, and updates the model parameters based on minimizing this error.

[0039] S43. Perform spectral transformation on the magnetic field segment to be analyzed to obtain the measured amplitude spectrum. Normalize both the theoretical and measured magnetic field amplitude spectra.

[0040] S44. Calculate the cosine similarity and energy ratio characteristics of the two components in the normalized frequency domain distribution. The vector composed of these characteristics is defined as the acoustic-electromagnetic intermodulation characteristic. A high similarity indicates that the measured magnetic field contains a component of the same frequency excited by the sound wave, and the target is very likely a metallic material.

[0041] S5. The acoustic multidimensional features are fused with the acoustic-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each target under test.

[0042] Specifically, the acoustic features are combined with the acoustic-electromagnetic intermodulation features by feature splicing and normalization processing is performed to form a high-dimensional multimodal fusion feature vector.

[0043] S6. Input the multimodal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification and judgment, thereby realizing the identification and labeling of material properties of different test targets.

[0044] Specifically, step S6 includes: 1. Establish a material feature sample library: A sample library is constructed by aggregating multimodal fusion feature vectors of known materials (such as copper, aluminum, iron, rock, and coral) obtained through experiments or high-fidelity simulations. Each data entry in the sample library contains a feature vector and its corresponding material label.

[0045] 2. Machine learning model training: Classification algorithms such as Support Vector Machines, Convolutional Neural Networks, or Random Forests are selected. The model is trained using data from a sample database, enabling it to learn the distribution patterns of different materials in a multi-dimensional feature space and constructing a classification decision boundary.

[0046] 3. Real-time identification and tagging: During the real-time detection phase, the newly acquired and processed multimodal fusion feature vectors are input into the trained machine learning model.

[0047] The model outputs the material attribute labels of the target (e.g., Label=1 for electromagnetic response / metal; Label=0 for non-electromagnetic response / rock).

[0048] It should be noted that, in the specific embodiments of the present invention, the internal basic topology and specific training algorithm processes of the feature encoder, cross-modal generative model (including the U-Net main network and diffusion probability model), and convolutional neural network, etc., can all adopt conventional and well-known architectures and open-source algorithms in the art. For example, the cross-modal generative model can adopt the standard denoising diffusion probability model (DDPM) architecture, the U-Net network can adopt a standard symmetric structure including multiple layers of downsampling and upsampling, and the loss function can use the conventional mean squared error (MSE) for network parameter updates.

[0049] like Figure 3 As shown, this embodiment provides an underwater target material property identification system based on acoustic-electromagnetic intermodulation characteristics. The system includes: The acoustic excitation module 301 is configured to transmit acoustic signals to the underwater detection area to acoustically excite multiple targets under test within the detection area, causing the multiple targets under test to generate mechanical vibration responses. The synchronous acquisition module 302 is configured to synchronously acquire mixed observation signals of the detection area using an underwater composite sensor array. The mixed observation signals include acoustic aliasing signals containing multiple target echoes and environmental magnetic field signals. The signal separation and interception module 303 is configured to perform spatiotemporal separation processing on the acoustic aliasing signal, detect the acoustic independent echo segment corresponding to each target under test and its acoustic delay time relative to the emission time; based on the acoustic delay time, backtrack and intercept the magnetic field segment to be analyzed that is time-aligned with the excitation vibration time of the target under test in the environmental magnetic field signal. Specifically, the signal separation and interception module 303 is configured to perform spatial filtering on the acoustic aliasing signal to separate acoustic wave beams from different directions; perform matched filtering detection on each acoustic wave beam, and intercept multiple acoustic pulse segments using a sliding time window according to the order of echo arrival times; mark each acoustic pulse segment as a corresponding independent acoustic echo segment, and determine the acoustic delay time of the independent acoustic echo segment relative to the transmitted signal; calculate the excitation time at which the target under test generates a mechanical vibration response based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves; and intercept magnetic field data centered on the excitation time and corresponding to the duration of the independent acoustic echo segment from the environmental magnetic field signal as the magnetic field segment to be analyzed.

[0050] The cross-modal reconstruction and feature extraction module 304 is configured to extract acoustic multidimensional features of independent acoustic echo segments for each target under test; simultaneously, the independent acoustic echo segments are mapped into conditional vectors, and combined with random noise vectors conforming to a preset distribution, they are input into a pre-trained cross-modal generation model to reconstruct the theoretical magnetic field frequency domain energy features; the spectral similarity between the theoretical magnetic field frequency domain energy features and the measured frequency domain energy features of the magnetic field segment to be analyzed is calculated as the acoustic-electromagnetic intermodulation feature; The multimodal fusion module 305 is configured to fuse acoustic multidimensional features with acoustic-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each target under test. The classification decision module 306 is configured to input the multimodal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification decision, thereby realizing the identification and labeling of the material properties of different test targets.

[0051] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the following: emitting acoustic signals into an underwater detection area to acoustically excite multiple targets within the detection area, causing the multiple targets to generate mechanical vibration responses; synchronously acquiring mixed observation signals of the detection area using an underwater composite sensor array, the mixed observation signals including acoustic aliasing signals containing echoes from multiple targets and environmental magnetic field signals; performing spatiotemporal separation processing on the acoustic aliasing signals to detect the acoustic independent echo segments corresponding to each target and their acoustic delay time relative to the emission time; based on the acoustic delay time, retrospectively extracting the magnetic field segments to be analyzed from the environmental magnetic field signals that are time-aligned with the excitation vibration time of the target; and for each... For each target under test, acoustic multidimensional features are extracted from independent acoustic echo segments. Simultaneously, these segments are mapped to conditional vectors and combined with random noise vectors conforming to a preset distribution, input into a pre-trained cross-modal generation model to reconstruct the theoretical magnetic field frequency domain energy features. The spectral similarity between the theoretical magnetic field frequency domain energy features and the measured frequency domain energy features of the magnetic field segments under analysis is calculated and used as the acoustic-electromagnetic intermodulation feature. The acoustic multidimensional features and the acoustic-electromagnetic intermodulation feature are fused to construct a multimodal fusion feature vector for each target. Feature processing algorithms are used to process the multimodal fusion feature vectors, establishing a material feature sample library for different targets. This material feature sample library is then imported into a machine learning model for classification training and processing, thereby enabling the identification and labeling of the material properties of different targets.

[0052] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The simple fact that certain measures are recited in mutually different dependent claims does not indicate that combinations of these measures cannot be used for improvement. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A method for identifying the material properties of underwater targets based on acoustic-electromagnetic intermodulation characteristics, characterized in that, Includes the following steps: S1. Transmit acoustic wave signals to the underwater detection area to acoustically excite multiple targets under test within the detection area, causing the multiple targets under test to generate mechanical vibration response; S2. Simultaneously acquire mixed observation signals of the detection area using an underwater composite sensor array. The mixed observation signals include acoustic aliasing signals containing multiple target echoes and environmental magnetic field signals. S3. Perform spatiotemporal separation processing on the acoustic aliasing signal to detect the acoustic independent echo segment corresponding to each target under test and its acoustic delay time relative to the transmission time. Based on the acoustic delay time, a magnetic field segment to be analyzed is extracted from the environmental magnetic field signal, which is time-aligned with the moment of the excited vibration of the target under test. S4. For each target to be tested, extract the acoustic multidimensional features of the acoustic independent echo segments; simultaneously, map the acoustic independent echo segments into latent space conditional vectors, and sample and generate Gaussian random noise vectors with the same dimension as the theoretical magnetic field frequency domain energy features; input the Gaussian random noise vectors and the latent space conditional vectors into a pre-trained cross-modal generation model, and use the cross-modal generation model to transform the Gaussian random noise vectors into a theoretical magnetic field amplitude spectrum constrained by the latent space conditional vectors, with the theoretical magnetic field amplitude spectrum serving as the theoretical magnetic field frequency domain energy feature; The measured amplitude spectrum is obtained by performing a spectral transformation on the magnetic field segment to be analyzed. The theoretical magnetic field amplitude spectrum and the measured amplitude spectrum are normalized respectively, and the cosine similarity and energy ratio characteristics of the two in the normalized frequency domain distribution are calculated. The vector composed of the cosine similarity and the energy ratio characteristics is determined as the acoustic-electromagnetic intermodulation characteristics. S5. The acoustic multidimensional features are fused with the acoustic-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each target under test; S6. Input the multimodal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification and judgment, thereby realizing the identification and labeling of the material properties of different test targets.

2. The method according to claim 1, characterized in that, The specific steps of step S3 include: S31. Spatial filtering is performed on the acoustic aliasing signal to separate the acoustic wave beams from different directions. S32. Perform matched filtering detection on each acoustic beam, and extract multiple acoustic pulse segments using a sliding time window based on the order of echo arrival times. S33. Mark each acoustic pulse segment as the corresponding acoustic independent echo segment, and determine the acoustic delay time of the acoustic independent echo segment relative to the transmitted signal; calculate the excitation time when the target under test generates a mechanical vibration response based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves; extract the magnetic field data centered on the excitation time and with a duration corresponding to the acoustic independent echo segment from the environmental magnetic field signal, and use it as the magnetic field segment to be analyzed.

3. The method according to claim 1, characterized in that, The acoustic multidimensional features described in step S4 include the envelope amplitude features of the acoustic independent echo segments in the time domain and the Mel frequency cepstral coefficients in the frequency domain.

4. The method according to claim 1, characterized in that, The cross-modal generation model is trained based on a diffusion model, and the specific training steps include: A training dataset containing paired acoustic echo samples and real magnetic field amplitude spectrum samples is constructed; random noise of different intensities is superimposed on the real magnetic field amplitude spectrum samples to generate noise samples; the noise samples and the acoustic echo samples are used as conditional input models to train the model to predict the random noise components contained in the noise samples; the error between the predicted random noise components and the actual superimposed random noise components is calculated, and the model parameters are updated based on the objective of minimizing the error.

5. The method according to claim 1, characterized in that, The machine learning model mentioned in step S6 includes at least one of support vector machine, convolutional neural network, or random forest; the pre-trained machine learning model is trained through the following steps: acquiring and fusing the acoustic multidimensional features and acoustic-electromagnetic intermodulation features of known materials to construct a material feature sample library containing multimodal fusion feature vectors of known materials and corresponding material attribute labels; importing the material feature sample library into the machine learning model for classification training to construct a classification decision boundary; the specific steps of identification and labeling include: using the pre-trained machine learning model to classify the multimodal fusion feature vectors extracted in real time and outputting material attribute labels; and labeling and outputting the attributes of targets determined to be electromagnetic response materials based on the material attribute labels.

6. A system for identifying the material properties of underwater targets based on acoustic-electromagnetic intermodulation characteristics, characterized in that, include: An acoustic excitation module is configured to transmit acoustic signals into an underwater detection area to acoustically excite multiple targets within the detection area, causing the multiple targets to generate mechanical vibration responses. The synchronous acquisition module is configured to synchronously acquire mixed observation signals of the detection area using an underwater composite sensor array. The mixed observation signals include acoustic aliasing signals containing multiple target echoes and environmental magnetic field signals. The signal separation and interception module is configured to perform spatiotemporal separation processing on the acoustic aliasing signal, and detect the acoustic independent echo segment corresponding to each target under test and its acoustic delay time relative to the transmission time. Based on the acoustic delay time, a magnetic field segment to be analyzed is extracted from the environmental magnetic field signal, which is time-aligned with the moment of the excited vibration of the target under test. A cross-modal reconstruction and feature extraction module is configured to extract acoustic multidimensional features of the acoustic independent echo segments for each target under test; simultaneously, the acoustic independent echo segments are mapped to latent space conditional vectors, and Gaussian random noise vectors with the same dimension as the theoretical magnetic field frequency domain energy features are sampled and generated; the Gaussian random noise vector and the latent space conditional vector are input into a pre-trained cross-modal generation model, and the cross-modal generation model is used to transform the Gaussian random noise vector into a theoretical magnetic field amplitude spectrum constrained by the latent space conditional vector, and the theoretical magnetic field amplitude spectrum is used as the theoretical magnetic field frequency domain energy feature; The measured amplitude spectrum is obtained by performing a spectral transformation on the magnetic field segment to be analyzed. The theoretical magnetic field amplitude spectrum and the measured amplitude spectrum are normalized respectively, and the cosine similarity and energy ratio characteristics of the two in the normalized frequency domain distribution are calculated. The vector composed of the cosine similarity and the energy ratio characteristics is determined as the acoustic-electromagnetic intermodulation characteristics. A multimodal fusion module is configured to fuse the acoustic multidimensional features with the acousto-electromagnetic intermodulation features to construct a multimodal fusion feature vector for each of the targets under test; The classification decision module is configured to input the multimodal fusion feature vector into a machine learning model pre-trained based on a preset material feature sample library for classification decision, thereby realizing the identification and labeling of the material properties of different test targets.

7. The system according to claim 6, characterized in that, The signal separation and interception module is configured to perform spatial filtering on the acoustic aliasing signal to separate acoustic beams from different directions; perform matched filtering detection on each acoustic beam; and intercept multiple acoustic pulse segments using a sliding time window according to the order of echo arrival times; mark each acoustic pulse segment as a corresponding acoustic independent echo segment, and determine the acoustic delay time of the acoustic independent echo segment relative to the transmitted signal. Based on the difference between the propagation speed of sound waves in water and the propagation speed of electromagnetic waves, the excitation time at which the target under test generates a mechanical vibration response is calculated; in the environmental magnetic field signal, magnetic field data centered on the excitation time and with a duration corresponding to the acoustic independent echo segment are extracted as the magnetic field segment to be analyzed.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images