An underwater acoustic-magnetic combined sensing method

Abstract

The application discloses an underwater acoustic-magnetic combined sensing method and relates to the technical field of underwater target detection. The method synchronously collects original acoustic signals and original magnetic signals of an underwater target through a fiber-optic acoustic-magnetic composite sensor, adopts an algorithm combining adaptive filtering and phase tracking to perform noise suppression processing on the original acoustic signals, then extracts acoustic signal features and magnetic signal features respectively, completes initial target detection through a deep neural network to obtain rough position information of the target, based on the rough position information, obtains a first position estimation through asynchronous cyclic feedback fusion processing by using the magnetic signal features, drives an acoustic propagation model to generate theoretical acoustic features, optimizes a prediction model through matching residuals, and outputs target state information after iterative convergence. The application can fully exert the complementary advantages of acoustic and magnetic information, improve the accuracy and reliability of underwater target detection, and is suitable for target detection tasks of platforms such as submarines and autonomous underwater vehicles.

Description

Technical Field

[0001] This invention relates to the field of underwater target detection technology, specifically to an underwater acoustic-magnetic joint sensing method. Background Technology

[0002] The complex environment of nearshore waters means that the radiated noise signals and magnetic anomaly signals generated by the motion of underwater targets (such as submarines and unmanned underwater vehicles) are crucial physical field information for target detection, identification, and localization. Acoustic detection alone is susceptible to interference from complex hydrological environments and insensitive to silent targets, while magnetic anomaly detection alone has limited range. Therefore, comprehensively utilizing acoustic and magnetic physical field information, leveraging their complementary advantages, has become an important technological development direction for improving the overall performance of underwater sensing systems.

[0003] Currently, existing research has explored the fusion processing of acoustic and magnetic information. For example, patent publication number CN118349919B discloses a technical solution for surface-underwater target identification based on acoustic and magnetic information fusion. This method extracts prior features of the target's acoustic and magnetic field spectra, obtains autonomous features using a feature extraction network, and finally aggregates these features for target identification, aiming to improve the recognition rate. Other existing technologies propose an approach that utilizes sonar to provide initial target localization, and then combines this with magnetometer measurement data to optimize the target position calculation. In addition, patent publication number CN118500384B discloses a fusion localization method and device for underwater moving targets based on acoustic and magnetic information fusion. This method estimates the target's acoustic field information and loss function based on axis frequency magnetic field separately, and optimizes the solution using acoustic information as the initial value to improve positioning accuracy.

[0004] While existing technologies have recognized the advantages of acoustic-magnetic fusion and have attempted fusion at the feature level or decision level, these methods are typically based on the assumption that acoustic and magnetic signals are acquired synchronously or treated as synchronous data. In real-world complex underwater environments, acoustic and magnetic signals differ fundamentally in their generation mechanisms, propagation speeds (sound speed approximately 1500 m / s, magnetic field changes can be considered instantaneous propagation), and susceptibility to environmental disturbances. This results in significant temporal asynchrony and dynamic differences in the acoustic and magnetic signals excited by the same target event at the sensor end. Existing static or one-off fusion frameworks struggle to effectively model and handle this asynchronous data flow problem caused by inherent physical differences. Specifically, this manifests as an inability to dynamically coordinate the temporal relationship between the rapidly responding but perturbed acoustic information and the stable but slowly updating magnetic information. This leads to performance degradation in continuous target tracking and accurate state estimation, limiting further improvements in sensing capabilities. Therefore, designing a sensing mechanism that can adaptively process asynchronous acoustic and magnetic data flows and achieve dynamic and deep collaboration between the two is a pressing technical problem that needs to be solved.

[0005] In summary, existing underwater sensing methods based on acoustic-magnetic information fusion are insufficient in dealing with the asynchronous characteristics of acoustic and magnetic signals, which restricts breakthroughs in the ability to continuously, stably, and with high precision sense moving targets in complex environments. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an underwater acoustic-magnetic joint sensing method to effectively solve the problem of collaborative processing of heterogeneous acoustic and magnetic data streams, thereby achieving more reliable and accurate detection, identification and positioning of underwater moving targets.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an underwater acoustic-magnetic joint sensing method, the method comprising the following steps: Step 1: Simultaneously acquire the raw acoustic and magnetic signals of the underwater target using a fiber optic acousto-magnetic composite sensor; Step 2: Perform noise suppression processing on the original acoustic signal to obtain a noise-reduced acoustic signal. The noise suppression processing adopts an algorithm that combines adaptive filtering and phase tracking. Step 3: Extract features from the denoised acoustic signal and the original magnetic signal respectively to obtain acoustic signal features and magnetic signal features; Step four: Based on the acoustic signal characteristics and the magnetic signal characteristics, perform asynchronous cyclic feedback fusion processing to obtain the state information of the underwater target; The asynchronous cyclic feedback fusion processing includes: obtaining a first position estimate based on the magnetic signal features through a first prediction model; using the first position estimate as input to drive an acoustic propagation model to calculate theoretical acoustic features corresponding to the acoustic signal features; calculating the matching residual between the theoretical acoustic features and the actual acoustic features extracted from the denoised acoustic signal; generating a feedback correction amount based on the matching residual; and using the feedback correction amount to optimize the first prediction model to update the first position estimate; iteratively executing the steps of driving the acoustic propagation model, calculating the matching residual, and optimizing the first prediction model until a preset convergence condition is met, and outputting the final target state information.

[0008] Furthermore, the fiber optic acoustomagnetic composite sensor integrates a fiber optic vector hydrophone unit and a fiber optic integrated diamond NV color center vector magnetometer unit. The fiber optic vector hydrophone unit adopts a three-branch orthogonally distributed fiber optic Michelson interferometer structure to sense the particle velocity of the three-dimensional sound field. The fiber-integrated diamond NV color center vector magnetometer unit adopts a structure in which diamond containing nitrogen vacancy color centers is integrated on the end face of a tapered optical fiber, and is used to sense a three-dimensional vector magnetic field.

[0009] Furthermore, the noise suppression processing in step two specifically includes: The common-mode phase noise of the fiber optic acoustomagnetic composite sensor system is monitored using the output signal of a reference interferometer that is of the same origin as the sensing fiber in the fiber optic acoustomagnetic composite sensor. Establish an empirical noise model that includes platform vibration noise and flow-induced noise; The noise components related to the output signal of the reference interferometer are filtered out from the original acoustic signal by a normalized minimum mean square error adaptive filtering algorithm. Phase tracking technology is used to dynamically compensate for non-stationary residual noise in the filtered signal.

[0010] Furthermore, step three involves feature extraction of the denoised acoustic signal, specifically including: An improved empirical mode decomposition algorithm that incorporates sample entropy evaluation index is used to decompose the denoised acoustic signal into multiple intrinsic mode functions; Extract the instantaneous frequency, instantaneous amplitude, and energy distribution characteristics of each intrinsic mode function; Calculate the acoustic energy flow vector and phase information of the sound field.

[0011] Furthermore, the first prediction model is a Kalman filter. Based on the magnetic signal characteristics, a first position estimate is obtained through the first prediction model, specifically including: The target is equivalent to a magnetic dipole model; The three components of the vector magnetic field in the magnetic signal characteristics are used as the observation values ​​of the Kalman filter; The coarse location information provided by the initial target detection is used as the state initialization value for the Kalman filter; By recursively applying Kalman filtering, the target's position and state at the next moment are predicted, thus obtaining the first position estimate.

[0012] Furthermore, based on the matching residuals, a feedback correction amount is generated, and the first prediction model is optimized using the feedback correction amount, specifically as follows: The matched residuals are converted into adjustments to the observation noise covariance matrix of the Kalman filter; Using the adjusted observation noise covariance matrix, the Kalman filter recursion is re-executed to update the first position estimate.

[0013] Furthermore, the acoustic propagation model is a vector sound field propagation model based on normal mode theory. Driving the acoustic propagation model, theoretical acoustic features corresponding to the acoustic signal features are calculated, specifically including: Using the first location estimate as the sound source location, combined with current marine environmental parameters, including water temperature, salinity, depth profile, and seabed sediment; Run the vector sound field propagation model to calculate the theoretical propagation path, time delay, multipath structure, and phase relationship of the sound signal from the sound source location to the sensor location; Based on the calculation results, the theoretical acoustic features are generated.

[0014] Furthermore, calculating the matching residual between the theoretical acoustic features and the actual acoustic features extracted from the denoised acoustic signal specifically includes: The theoretical acoustic features are organized into theoretical feature vectors. ; The actual acoustic features extracted from the noise-reduced acoustic signal are organized into an actual feature vector. ; The matching residual Calculated using the following formula: in, Norm operations on vectors A diagonal weight matrix, pre-defined based on the signal-to-noise ratio or prior importance of each acoustic feature component, is used to adjust the contribution of different feature dimensions in the residual calculation. This indicates that the corresponding elements of the actual eigenvector are subtracted from the corresponding elements of the theoretical eigenvector.

[0015] Furthermore, step four, prior to the asynchronous cyclic feedback fusion process, also includes an initial target detection step: The acoustic signal features and the magnetic signal features are input into a deep neural network; The deep neural network uses a cross-stage partial connection structure as the backbone network and embeds an attention mechanism module. The deep neural network performs multi-scale feature fusion through a feature pyramid network and a path aggregation network. The deep neural network outputs the initial existence probability, category, and approximate location region of the target; The rough location region is used for initializing the first prediction model.

[0016] Furthermore, the method is applied to an underwater submersible platform equipped with the fiber optic acoustomagnetic composite sensor; The underwater submersible platform includes submarines or autonomous underwater vehicles; The final target status information includes the target's three-dimensional coordinates, speed, heading, and identity category.

[0017] Compared with existing technologies, this underwater acoustic-magnetic joint sensing method has the following advantages: I. This invention designs an asynchronous cyclic feedback fusion processing mechanism to obtain an initial position estimate based on magnetic signal characteristics, drive the acoustic propagation model to generate theoretical acoustic features, and use the matching residuals of theoretical and actual acoustic features to construct a feedback correction link to dynamically optimize the prediction model. This mechanism can adaptively handle the essential differences in the generation mechanism, propagation speed, and environmental disturbance characteristics of acoustic and magnetic signals, effectively solve the problem of collaborative processing of asynchronous data streams, thereby breaking through the limitations of existing static or one-time fusion frameworks, improving the stability of continuous underwater target tracking and accurate state estimation, and enhancing the comprehensive performance of acoustic and magnetic fusion sensing in complex nearshore water environments.

[0018] II. This invention achieves synchronous acquisition of acoustic and magnetic signals by integrating a fiber optic vector hydrophone and a fiber optic integrated diamond NV color center vector magnetometer. Combined with adaptive filtering and phase tracking noise suppression algorithms and improved feature extraction methods, it comprehensively captures multi-dimensional effective features in the acoustic and magnetic signals. It can give full play to the advantages of acoustic detection range and magnetic detection anti-interference, improve the signal-to-noise ratio and the integrity of feature representation, thereby enhancing the adaptability of the underwater sensing system to complex hydrological environments and improving the recognition accuracy of target three-dimensional coordinates, motion state and identity category.

[0019] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0021] Figure 1 This is a diagram illustrating the method steps of the present invention; Figure 2 This is a schematic diagram of the fiber optic acousto-magnetic composite sensor of the present invention; Figure 3 This is a schematic diagram illustrating the asynchronous cyclic feedback fusion processing principle of the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0023] Example 1 like Figures 1 to 3 As shown, this embodiment uses a fiber optic acousto-magnetic composite sensor mounted on an autonomous underwater vehicle as the core sensing unit, and focuses on underwater target detection scenarios in nearshore waters. The following is a detailed description of the specific implementation.

[0024] The underwater acoustic-magnetic joint sensing system in this embodiment mainly includes a fiber optic acoustic-magnetic composite sensor, a signal synchronous acquisition module, a signal processing module, a data storage module, and a power supply module. These modules work together to achieve synchronous acquisition, noise suppression, feature extraction, fusion processing, and target status output of underwater target acoustic-magnetic signals.

[0025] Fiber optic acoustomagnetic composite sensor: The fiber optic acoustomagnetic composite sensor integrates a fiber optic vector hydrophone unit and a fiber optic integrated diamond NV color center vector magnetometer unit. The two are arranged coaxially in structure to ensure the consistency of the detection direction. At the same time, it adopts pressure-resistant sealed packaging to adapt to the underwater operation environment.

[0026] Fiber Optic Vector Hydrophone Unit: This unit employs a three-branch orthogonally distributed fiber optic Michelson interferometer structure, with the optical axes of the three interferometers arranged along the X, Y, and Z axes of a Cartesian coordinate system. Each interferometer's two sensing arms are wound around a three-dimensional symmetrical elastic body centered on a mass block. The elastic body is made of titanium alloy, and the mass block is made of high-density tungsten alloy, ensuring high sensitivity to the velocity of sound field particles. When the acoustic signal generated by an underwater target acts on the sensor, the velocity of the sound field particles causes the mass block to move inertially, resulting in opposite deformations of the fiber optic arms of the interferometer. This leads to a change in the phase difference of the interference light, and the three-dimensional velocity information of the sound field particles can be obtained by detecting this phase difference.

[0027] Fiber-Integrated Diamond NV Center Vector Magnetometer Unit: This unit employs a structure integrating diamond containing nitrogen-vacancy centers into the end face of a tapered optical fiber. The tapered fiber is fabricated using a fused taper process, with optimized tapered length and tip diameter to improve fluorescence excitation and collection efficiency. Type IIa diamond synthesized under high temperature and pressure is selected, and its internal nitrogen-vacancy center concentration is controlled to ensure magnetic field detection sensitivity. The diamond is fixed to the end face of the tapered fiber using a low-melting-point glass powder fusion bonding process. The other end of the fiber is connected to a laser source and a fluorescence detector, forming a complete optical transmission link. When an external magnetic field is applied to the diamond, the electronic energy levels of the nitrogen-vacancy centers undergo Zeeman splitting. By detecting the fluorescence intensity versus microwave frequency curve (ODMR spectrum), the three-dimensional vector magnetic field information can be calculated.

[0028] Signal Synchronization Acquisition Module: The signal synchronization acquisition module includes an optical signal demodulation unit, a microwave drive unit, and a clock synchronization unit. The optical signal demodulation unit converts the interference optical signal output from the fiber optic vector hydrophone unit into an electrical signal, and simultaneously converts the fluorescence signal output from the fiber optic integrated diamond NV center vector magnetometer unit into an electrical signal. The microwave drive unit provides a stable microwave excitation signal for the diamond NV center to realize the detection of the ODMR spectrum. The clock synchronization unit provides a unified sampling clock for the two sensing units to ensure the synchronous acquisition of the original acoustic signal and the original magnetic signal. The sampling frequency is set according to the detection requirements; in this embodiment, the sampling frequency is set to a fixed value.

[0029] Signal processing module: The signal processing module adopts an embedded processor and integrates noise suppression algorithm, feature extraction algorithm, initial target detection algorithm and asynchronous cyclic feedback fusion algorithm. It is used to perform a series of processing on the acquired raw signals and finally output the state information of the underwater target.

[0030] Data storage module and power supply module: The data storage module is used to store raw acquired data, intermediate processed data and final target status information. It adopts industrial-grade solid-state drives and has vibration and shock resistance characteristics. The power supply module uses a lithium battery pack to provide stable power to the entire system and support long-term underwater operation.

[0031] In this embodiment, the underwater acoustic-magnetic joint sensing method is implemented sequentially according to the following steps: Synchronous acquisition of raw acoustic and magnetic signals: The autonomous underwater vehicle equipped with an optical fiber acoustic-magnetic composite sensor is deployed in the target detection area. After the system is started, the clock synchronization unit of the signal synchronous acquisition module generates a unified sampling clock signal, which controls the optical fiber vector hydrophone unit and the optical fiber integrated diamond NV color center vector magnetometer unit to start signal acquisition simultaneously.

[0032] The three orthogonal fiber Michelson interferometers of the fiber optic vector hydrophone unit sense the particle velocity of the sound field in the X, Y, and Z directions, respectively. The phase difference change of the interference light is converted into the corresponding electrical signal by the optical signal demodulation unit, which is the original sound signal. The original sound signal contains particle velocity components in the three directions.

[0033] In the fiber-optic integrated diamond NV center vector magnetometer unit, excitation light emitted from a laser source is transmitted to the diamond via a tapered optical fiber, exciting nitrogen-vacancy centers to produce fluorescence. The fluorescence returns via the same optical fiber and is received by a fluorescence detector. The microwave drive unit outputs a swept-frequency microwave signal, which acts on the diamond, causing Zeeman splitting of the electronic energy levels of the nitrogen-vacancy centers. The fluorescence intensity detected by the fluorescence detector changes with the microwave frequency, forming an ODMR spectrum. Through preliminary calculation of the ODMR spectrum, the original magnetic signal is obtained, which contains vector magnetic field components in three directions.

[0034] During the synchronous acquisition process, the raw acoustic and magnetic signals are transmitted in real time to the signal processing and data storage modules through the signal synchronous acquisition module. The data storage module backs up and stores the raw data.

[0035] Noise suppression processing: The signal processing module calls a noise suppression algorithm to perform noise suppression processing on the original sound signal, obtaining a noise-reduced sound signal. The specific process is as follows: Common-mode phase noise monitoring: A reference interferometer, derived from the same sensing fiber as the fiber optic acoustomagnetic composite sensor, is used. This reference interferometer employs the same fiber material and arrangement as the sensing interferometer, but does not directly contact the sound and magnetic fields; it is used solely to monitor the system's own common-mode phase noise. The optical signal output from the reference interferometer is converted into an electrical signal by an optical signal demodulation unit, which serves as the common-mode phase noise reference signal.

[0036] Empirical Noise Model Establishment: Based on preliminary tests of the vibration characteristics and underwater flow field characteristics of the autonomous underwater vehicle platform, an empirical noise model incorporating platform vibration noise and flow-induced noise was established. Platform vibration noise mainly originates from the vehicle's propulsion system, servo motors, and pump system; its frequency and amplitude characteristics were obtained through preliminary ground and tank tests. Flow-induced noise mainly originates from the interaction between the fluid and sensor surfaces during vehicle movement; its distribution characteristics were obtained based on fluid dynamics simulations and fitting of experimental data.

[0037] Adaptive filtering noise reduction: A normalized minimum mean square error adaptive filtering algorithm is used, with the common-mode phase noise reference signal as the input reference, to filter the original acoustic signal. The iterative formula of the normalized minimum mean square error adaptive filtering algorithm is: in, for The adaptive filter weight vector at time 1. This is the step size factor, used to control the convergence speed and steady-state error. In this embodiment, its value is set to a fixed range. for The filtered error signal at time 10:00. for Reference input signal at time (common-mode phase noise reference signal). for The original sound signal at that moment, This is a regularization parameter used to avoid zero denominators; its value is a fixed small constant. The square norm of the reference input signal, This is the transpose of the weight vector.

[0038] This adaptive filtering algorithm filters out common-mode noise components related to the output signal of the reference interferometer from the original acoustic signal, including some platform vibration noise and the system's own photoelectric noise.

[0039] Phase tracking dynamic compensation: Even after adaptive filtering, the acoustic signal still contains some non-stationary residual noise, mainly consisting of remaining flow-induced noise and random noise. Phase tracking technology is employed to dynamically compensate for the phase shift caused by non-stationary residual noise by monitoring the phase change of the filtered signal in real time and establishing a phase change model. Specifically, the phase information of the filtered signal is extracted, compared with an ideal phase model, and the phase deviation is calculated. Then, the signal is phase corrected based on the phase deviation, thereby suppressing the non-stationary residual noise and ultimately obtaining the denoised acoustic signal.

[0040] Feature extraction: The signal processing module performs feature extraction on the denoised acoustic signal and the original magnetic signal respectively to obtain acoustic signal features and magnetic signal features: Feature extraction of the denoised acoustic signal: An improved empirical mode decomposition algorithm incorporating sample entropy evaluation metrics is used to decompose the denoised acoustic signal. First, the parameters of the improved Empirical Mode Decomposition (EMD) algorithm are determined, including the number of decomposition levels and stopping conditions. The decomposition process is then monitored based on the sample entropy evaluation metric. Sample entropy measures the complexity of a signal; the smaller the sample entropy value, the more regular the signal. By using the sample entropy evaluation metric, the mode aliasing problem that may occur in traditional EMD algorithms is avoided, ensuring the reliability of the decomposition results.

[0041] Then, following the steps of the improved empirical mode decomposition algorithm, the denoised acoustic signal is decomposed into multiple intrinsic mode functions. Each intrinsic mode function represents an inherent vibration mode of the signal and contains signal components within a specific frequency range.

[0042] Next, each eigenmode function is analyzed to extract instantaneous frequency, instantaneous amplitude, and energy distribution characteristics: the instantaneous frequency is solved using the Hilbert transform for each eigenmode function. Perform Hilbert transform to obtain Then the instantaneous angular frequency Instantaneous frequency Instantaneous amplitude Energy distribution characteristics are obtained by calculating the energy of each intrinsic mode function. and energy percentage Received, among which The total number of intrinsic mode functions. and To analyze the start and end times of a time period.

[0043] Finally, based on the extracted instantaneous frequency and amplitude information, the acoustic energy flow vector and phase information of the sound field are calculated: acoustic energy flow vector ,in The density of seawater, The speed of sound in seawater. The sound pressure signal is obtained by integrating the particle velocity of the sound signal after noise reduction. The particle velocity signal (the direct component of the noise-reduced acoustic signal); the phase information is the phase angle obtained through Hilbert transform. Comprehensive calculation.

[0044] The instantaneous frequency, instantaneous amplitude, energy distribution characteristics, acoustic energy flow vector, and phase information extracted above together constitute the acoustic signal characteristics.

[0045] Feature extraction of the original magnetic signal: The original magnetic signal contains vector magnetic field components in three directions. Time-domain and frequency-domain analyses are performed on each component to extract characteristic parameters. Time-domain features include statistics such as peak value, mean, variance, peak factor, and kurtosis of the magnetic field components. Frequency-domain features are obtained by performing a Fourier transform on the original magnetic signal, including parameters such as spectral peak value, spectral centroid, and spectral width. Simultaneously, the correlation coefficients between the three magnetic field components are calculated to reflect the spatial distribution characteristics of the magnetic field. These characteristic parameters collectively constitute the magnetic signal features.

[0046] Initial target detection: Before performing asynchronous loop feedback fusion processing, the initial target detection step is executed first. The specific process is as follows: The acoustic and magnetic signal features are normalized to ensure they fall within the same numerical range before being input into a deep neural network. This deep neural network employs a cross-stage partial connection structure as its backbone. This structure divides the input feature map into two parts: one part undergoes multiple convolutional processing, while the other part is directly concatenated with the output feature map of subsequent convolutional layers, effectively improving feature propagation efficiency and network gradient flow performance.

[0047] An attention mechanism module is embedded in the backbone network. This module enhances the focus on target signal features and suppresses interference from background noise features by assigning different weights to different channels and spatial locations of the feature maps. The output features of the attention mechanism module are then fused into multi-scale features through a feature pyramid network and a path aggregation network: the feature pyramid network extracts feature maps from different levels of the backbone network and generates multi-scale feature maps through upsampling and fusion operations; the path aggregation network passes semantic information downward from the top-level feature maps and positional information upward from the bottom-level feature maps, achieving full fusion of multi-scale features.

[0048] The output layer of the deep neural network uses the softmax activation function, outputting the initial presence probability, category, and approximate location region of the target. The initial presence probability represents the confidence level of detecting an underwater target; when this probability is greater than a set threshold, an underwater target is determined to exist. The category output indicates the possible types of the target (such as submarines, unmanned underwater vehicles, mines, etc.). The approximate location region is the approximate spatial range of the target in the sensor coordinate system, represented in the form of a three-dimensional coordinate interval. This approximate location region is used for the initialization of the subsequent first prediction model.

[0049] Asynchronous cyclic feedback fusion processing: When the initial target detection determines that an underwater target exists, the signal processing module initiates asynchronous cyclic feedback fusion processing to obtain the underwater target's state information based on acoustic and magnetic signal characteristics. The specific process is as follows: First position estimation acquisition: The first prediction model uses a Kalman filter to obtain the first position estimate based on magnetic signal characteristics. First, the underwater target is modeled as a magnetic dipole, and the magnetic moment of the magnetic dipole model is... Based on the initial settings of the target's typical dimensions, material, and motion state, its expression is: ,in These are the magnetic moment components of the magnetic dipole in the X, Y, and Z axes, respectively.

[0050] The three components of the vector magnetic field in the characteristics of magnetic signals As the observations of the Kalman filter, the observation equation is: ,in, for The observation vector at time (three components of the vector magnetic field). for The observation matrix at each time step is determined by the spatial distribution characteristics of the magnetic dipole model. for The state vector at any given time (target's position coordinates) and speed of movement ), for The observation noise at any given time follows a Gaussian distribution, and its covariance matrix is... The initial setting is a fixed matrix.

[0051] The center coordinates of the coarse location region from the initial target detection output are used as the state initialization values ​​for the Kalman filter. The state equation is: in, for The state vector at time t, for The state transition matrix at each moment is set based on the target's motion characteristics (assuming the target is moving at a constant velocity in a straight line). (Extended form of the identity matrix) for Time-based control input (There is no active control input in this embodiment, therefore...) ), for The process noise at time step follows a Gaussian distribution, and its covariance matrix is... The initial setting is a fixed matrix.

[0052] The Kalman filter recursion includes a prediction step and an update step, with the prediction step calculated based on the state equation. Prior state estimation at time 1 and prior covariance matrix The update step is calculated based on the observation equation and prior estimates. Posterior state estimation at time 1 and posterior covariance matrix The position coordinate components in the posterior state estimate are the first position estimates.

[0053] Theoretical acoustic feature calculation: The first position estimate is used as the sound source position, combined with the current marine environmental parameters, including water temperature, salinity, depth profile and seabed sediment. These parameters are acquired in real time by environmental sensors carried by the autonomous underwater vehicle, or called from a pre-stored marine environmental database.

[0054] A vector sound field propagation model based on normal mode theory is constructed. The core of this model is to solve the governing equations of the underwater sound field through normal mode expansion, considering the stratification of seawater, the acoustic properties of the seabed, and the multipath propagation effect of the sound signal. The solution process of the vector sound field propagation model is as follows: First, the layered structure of the sound field is determined based on marine environmental parameters, dividing the seawater into several uniform layers. The sound velocity in each layer is calculated from water temperature, salinity, and depth, using an empirical formula. ,in Water temperature Salinity For depth.

[0055] Then, based on the normal mode theory, the eigenvalues ​​and eigenfunctions of the acoustic signal propagating from the sound source position (first position estimation) to the sensor position are calculated. The eigenvalues ​​correspond to the order and propagation constant of the normal mode, and the eigenfunctions correspond to the spatial distribution of each order of normal mode.

[0056] Finally, based on the eigenvalues ​​and eigenfunctions, the theoretical propagation path, time delay, multipath structure, and phase relationship of the acoustic signal are calculated. Based on these calculation results, theoretical acoustic features corresponding to the acoustic signal features are generated. The theoretical acoustic features include theoretical instantaneous frequency, theoretical instantaneous amplitude, theoretical energy distribution, theoretical acoustic energy flow vector, and theoretical phase information, which are consistent with the dimensions of the acoustic signal features.

[0057] Matching residual calculation: Organizing theoretical acoustic features into theoretical feature vectors The actual acoustic features extracted from the noise-reduced sound signal will be organized into actual feature vectors. Matching residuals Calculated using the following formula: in, The 2-norm operation represents a vector and is used to measure the magnitude of the vector. It is a diagonal weight matrix, whose diagonal elements are pre-set according to the signal-to-noise ratio or prior importance of each acoustic feature component. The feature components with high signal-to-noise ratio or high prior importance have larger weight coefficients, which are used to adjust the contribution of different feature dimensions in the residual calculation. This means that the residual vector is obtained by subtracting the corresponding elements of the actual eigenvector from the corresponding elements of the theoretical eigenvector.

[0058] Feedback Correction and Model Optimization: Based on the calculated matching residuals, a feedback correction is generated. Specifically, the matching residuals are converted into the observation noise covariance matrix of the Kalman filter. The adjustment amount is as follows: When the matching residual is large, it indicates that the deviation between the theoretical acoustic features and the actual acoustic features is large, which may be due to inaccurate first position estimation or unreasonable setting of the observation noise covariance matrix. In this case, increase the diagonal elements of the observation noise covariance matrix and reduce the weight of the observations in the Kalman filter update step. When the matching residual is small, it indicates that the deviation between the theoretical acoustic features and the actual acoustic features is small and the first position estimation is relatively accurate. In this case, decrease the diagonal elements of the observation noise covariance matrix and increase the weight of the observations.

[0059] Using the adjusted observation noise covariance matrix Re-execute the update step of the Kalman filter recursion to calculate the new posterior state estimate. and posterior covariance matrix This enables the updating of the first position estimate.

[0060] Iterative convergence and result output: Iteratively execute the steps of driving the acoustic propagation model, calculating the matching residual, and optimizing the first prediction model. In each iteration, the updated first position estimate is used as the new sound source position, the theoretical acoustic characteristics and matching residual are recalculated, and then the Kalman filter parameters are adjusted to update the first position estimate.

[0061] The convergence condition for the iterative process is set as either the matching residual being less than a set threshold, or the number of iterations reaching a set maximum value. When the convergence condition is met, the iteration stops, and the position coordinates in the posterior state estimate output by the Kalman filter are then displayed. Speed ​​of movement The target state information is formed by combining the target category output by the deep neural network with the heading calculated from the motion speed.

[0062] Target status information output and storage: The signal processing module transmits the final target status information to the data storage module for storage. At the same time, it can be transmitted to the shore-based control center or mother ship through the communication module of the autonomous underwater vehicle to provide data support for subsequent target tracking and identification decisions.

[0063] This embodiment achieves synchronous acquisition of raw acoustic and magnetic signals from underwater targets using a fiber optic acousto-magnetic composite sensor. A noise suppression algorithm combining adaptive filtering and phase tracking is employed to effectively reduce interference from platform vibration noise, flow-induced noise, and system-specific noise, thereby improving the signal-to-noise ratio. An improved empirical mode decomposition algorithm is used to extract acoustic signal features, which, combined with magnetic signal features, are then used to perform initial target detection using a deep neural network, providing reliable initialization information for subsequent fusion processing.

[0064] Asynchronous cyclic feedback fusion processing uses a Kalman filter to estimate the position of the magnetic signal, drives the acoustic propagation model to generate theoretical acoustic features, and implements feedback correction of the Kalman filter based on the matching residual. Iteratively optimizes the position estimation results, giving full play to the advantages of long acoustic detection range and strong anti-interference of magnetic detection, and realizing accurate acquisition of underwater target state information.

[0065] The technical solution of this embodiment can effectively adapt to the complex hydrological environment of nearshore waters, improve the accuracy and reliability of underwater target detection, and is suitable for target detection missions of underwater submersible platforms such as submarines and autonomous underwater vehicles.

[0066] Example 2 like Figures 1 to 3 As shown, this embodiment uses a fiber optic acoustomagnetic composite sensor mounted on a submarine as the core detection unit. It is designed for underwater target detection scenarios in complex near-shore waters. The following is a detailed description of the specific implementation.

[0067] The underwater acoustic-magnetic joint sensing system in this embodiment includes a fiber optic acoustic-magnetic composite sensor signal synchronous acquisition module, a signal processing module, a data storage module, and a power supply module. Each module is integrated within the submarine's underwater detection compartment and achieves data exchange and power supply through a dedicated interface.

[0068] Fiber Optic Acoustomagnetic Composite Sensor: The fiber optic acoustomagnetic composite sensor adopts an integrated structural design, with an overall cylindrical shape and an external pressure-resistant sealing layer, adaptable to submarine mounting interfaces. Internally, it integrates a fiber optic vector hydrophone unit and a fiber optic integrated diamond NV color center vector magnetometer unit, with the two units arranged coaxially to ensure consistent detection direction.

[0069] The fiber optic vector hydrophone unit employs a three-branch orthogonally distributed fiber optic Michelson interferometer structure, with each interferometer corresponding to one of the three coordinate axes of a three-dimensional spatial coordinate system. The sensing arm of each interferometer is wound around an elastic body at the center of a mass block made of high-density alloy. The elastic body is made of corrosion-resistant titanium alloy, ensuring both sensitive response to the vibration velocity of sound field particles and adaptability to long-term underwater operation. The core function of this unit is to sense the vibration velocity of particles in a three-dimensional sound field, providing raw data for subsequent acoustic signal feature extraction.

[0070] The core structure of the fiber-integrated diamond NV color center vector magnetometer unit consists of a tapered optical fiber end face integrated with diamond containing nitrogen-vacancy color centers. The tapered optical fiber is fabricated using a fused taper process, with optimized tapered length and tip size to improve fluorescence excitation and collection efficiency. High-purity diamond synthesized under high temperature and pressure is used, and the concentration of nitrogen-vacancy color centers within it is controlled to ensure the ability to detect weak magnetic fields. The core function of this unit is to sense three-dimensional vector magnetic fields and capture magnetic anomaly signals generated by underwater target motion.

[0071] Signal Synchronization Acquisition Module: This module incorporates a high-precision clock synchronization unit, providing a unified sampling clock signal for both the fiber optic vector hydrophone unit and the fiber optic integrated diamond NV color center vector magnetometer unit. This clock synchronization unit allows both units to initiate signal acquisition at the same time, ensuring the time synchronization of the original acoustic and magnetic signals and avoiding fusion errors caused by acquisition time differences. Simultaneously, this module features signal preprocessing capabilities, performing preliminary amplification and filtering on the acquired raw signals to lay the foundation for subsequent processing.

[0072] The specific implementation steps in this embodiment are as follows: Simultaneous acquisition of raw acoustic and magnetic signals: After the submarine sails to the target detection area, the underwater acoustic-magnetic joint sensing system is activated. The clock synchronization unit of the signal synchronization acquisition module sends a synchronization trigger signal, and the fiber optic vector hydrophone unit and the fiber optic integrated diamond NV color center vector magnetometer unit start working simultaneously.

[0073] The fiber optic vector hydrophone unit's three fiber optic Michelson interferometers sense the particle velocities of the sound field from different directions. When the acoustic signal generated by the underwater target acts on the sensor, the particle velocities cause the mass block to move inertially, resulting in opposite deformations of the two sensing arms of the interferometer, which in turn causes a change in the phase difference of the interference light. This phase difference change is demodulated by the optical signal and converted into an electrical signal, which is the original acoustic signal containing particle velocity components from the three directions.

[0074] In the fiber-optic integrated diamond NV center vector magnetometer unit, excitation light emitted from a laser source is transmitted to the diamond via a tapered optical fiber, exciting nitrogen-vacancy centers to produce fluorescence. The fluorescence returns via the same optical fiber and is received by a fluorescence detector, while a microwave drive unit outputs a swept-frequency microwave signal that acts on the diamond. The magnetic anomaly signal generated by the movement of an underwater target causes Zeeman splitting of the electronic energy levels of the nitrogen-vacancy centers, resulting in a specific spectral line formed by the variation of fluorescence intensity with microwave frequency. Through preliminary analysis of this spectral line, the original magnetic signal containing three directional components is obtained.

[0075] The raw acoustic and magnetic signals are transmitted in real time to the signal processing module via the signal synchronization acquisition module, and are also stored in the data storage module for backup.

[0076] Noise suppression processing: The original acoustic signal is subject to noise interference such as the submarine's own vibration and water flow disturbance during the acquisition process, and noise suppression processing is required to improve the signal quality.

[0077] First, a reference interferometer, which shares the same origin as the sensing fiber in the fiber-optic acoustomagnetic composite sensor, is used to monitor common-mode phase noise. This reference interferometer uses the same fiber material and arrangement as the sensing interferometer, but it does not directly contact the sound field; it is only used to capture the system's own phase noise, such as noise caused by environmental disturbances during fiber transmission.

[0078] Secondly, an empirical noise model incorporating platform vibration noise and flow-induced noise was established. The characteristics of platform vibration noise were summarized based on measured data from equipment such as the propulsion system steering gear during submarine navigation, while the characteristics of flow-induced noise were obtained by fitting data from tank simulation tests and actual sea tests. This model can accurately characterize the frequency distribution and amplitude characteristics of the two main types of noise.

[0079] Subsequently, a normalized minimum mean square error adaptive filtering algorithm is adopted, using the common-mode phase noise signal output by the reference interferometer as a reference, to filter out noise components related to the reference signal from the original acoustic signal, mainly including common-mode noise caused by platform vibration and some system electromagnetic interference noise.

[0080] Finally, phase tracking technology is used to dynamically compensate for non-stationary residual noise in the filtered signal. By monitoring the phase change of the filtered signal in real time and comparing it with the ideal phase characteristics, the phase shift caused by non-stationary noise is corrected, and the denoised acoustic signal is finally obtained.

[0081] Feature extraction: Acoustic signal feature extraction: The denoised acoustic signal is processed using an improved empirical mode decomposition algorithm that incorporates sample entropy as an evaluation metric. Sample entropy is used to evaluate the complexity of each modal component after decomposition, avoiding the mode aliasing problem that may occur in traditional empirical mode decomposition algorithms. Through this algorithm, the denoised acoustic signal is decomposed into multiple intrinsic mode functions, each corresponding to signal components in a different frequency range.

[0082] For each intrinsic mode function (EMF), instantaneous frequency, instantaneous amplitude, and energy distribution characteristics are extracted. Instantaneous frequency reflects the frequency variation of the signal at different times, instantaneous amplitude reflects the intensity fluctuation of the signal, and energy distribution characteristics characterize the energy proportion of different frequency components. Simultaneously, combining the instantaneous parameters of each EEMF, the acoustic energy flow vector and phase information of the sound field are calculated. The acoustic energy flow vector reflects the propagation direction and intensity of the sound field energy, while the phase information is used to characterize the phase change pattern of the signal. The extracted parameters collectively constitute the acoustic signal characteristics.

[0083] Magnetic signal feature extraction: Time-domain and frequency-domain analysis is performed on the original magnetic signal to extract statistical features such as peak value, mean, variance, and peak factor in the time domain. These features reflect the intensity variation and distribution pattern of the magnetic signal. Frequency-domain analysis involves spectral transformation of the original magnetic signal to extract features such as spectral peaks and spectral centroid, characterizing the frequency distribution characteristics of the magnetic signal. Simultaneously, the correlation coefficients between the magnetic signal components in the three directions are calculated to reflect the spatial distribution characteristics of the magnetic anomaly signal. These features collectively constitute the magnetic signal characteristics.

[0084] Initial Target Detection: Acoustic and magnetic signal features are input into a deep neural network for initial target detection. The backbone of this deep neural network employs a cross-stage partial connection structure, which improves feature propagation efficiency and network gradient flow performance, avoiding the gradient vanishing problem during deep network training. Simultaneously, an attention mechanism module is embedded in the backbone network, enhancing the network's focus on target signal features, suppressing interference from background noise features, and improving the targeting of feature extraction.

[0085] Based on the multi-scale features output by the backbone network, feature fusion is performed through a feature pyramid network and a path aggregation network. The feature pyramid network extracts features from different levels of the backbone network and achieves multi-scale feature alignment through upsampling; the path aggregation network passes semantic information downward from the top-level features and positional information upward from the bottom-level features, ensuring that the fused features contain both rich semantic information and accurate positional information.

[0086] The output of the deep neural network includes the initial probability category of the target's presence and a rough location region. When the initial probability of presence is greater than a set threshold, it is determined that an underwater target exists in the detection area; the category result indicates the possible type of the target; the rough location region is output in the form of a spatial coordinate interval, which is directly used for the state initialization of the subsequent first prediction model.

[0087] Asynchronous cyclic feedback fusion processing: First position estimation: The first prediction model employs a Kalman filter, and its initial state values ​​are derived from the center coordinates of the approximate position region in the initial target detection output. The target is equivalent to a magnetic dipole model, which is based on the typical material and structural characteristics of underwater targets and can approximately reflect the target's magnetic radiation properties.

[0088] The three components of the vector magnetic field in the magnetic signal are used as observations in a Kalman filter. Through recursive calculations in the prediction and update steps of the Kalman filter, the target's position at the next moment is predicted; this position is the first position estimate. The core function of the Kalman filter is to utilize the stability of the magnetic signal to initially determine the approximate location of the target, providing input for the acoustic propagation model.

[0089] Theoretical acoustic characteristic calculation: The acoustic propagation model adopts a vector sound field propagation model based on normal mode theory. This model has pre-stored the sound field propagation laws corresponding to different marine environmental parameters. The first position estimate is used as the sound source position, and combined with the current marine environmental parameters obtained in real time by the environmental sensors on the submarine, including water temperature, salinity, depth profiles, and seabed sediment, the vector sound field propagation model is driven to run.

[0090] The model generates theoretical acoustic features consistent with the feature dimensions of the acoustic signal by calculating the theoretical propagation path time delay multipath structure and phase relationship of the acoustic signal from the sound source location to the sensor location. These features include parameters such as theoretical instantaneous frequency, theoretical instantaneous amplitude, and theoretical energy distribution.

[0091] Matching residual calculation and feedback optimization: The theoretical acoustic features are organized into theoretical feature vectors, and the actual acoustic features extracted from the denoised acoustic signal are organized into actual feature vectors. By comparing the corresponding feature parameters of the two vectors, the matching residual, reflecting the difference between them, is obtained. The core function of the matching residual is to determine the accuracy of the first position estimation, providing a basis for subsequent model optimization.

[0092] Feedback corrections are generated based on the matching residuals, specifically converting these residuals into adjustments to the observation noise covariance matrix of the Kalman filter. When the matching residuals are large, it indicates a significant difference between the theoretical and actual acoustic characteristics, resulting in insufficient accuracy in the first position estimation. In this case, the correlation elements of the observation noise covariance matrix are increased, reducing the weight of the observations in the filter update. Conversely, when the matching residuals are small, it indicates a smaller difference between the two, resulting in relatively accurate first position estimation. In this case, the correlation elements of the observation noise covariance matrix are decreased, increasing the weight of the observations.

[0093] Using the adjusted observation noise covariance matrix, the Kalman filter recursion is re-executed to update the first position estimate.

[0094] Iterative Convergence and Result Output: The iterative execution process involves calculating the matching residual and optimizing the first prediction model using the driven acoustic propagation model. Each iteration uses the updated first position estimate as the new sound source location, regenerates the theoretical acoustic features, calculates the matching residual, and continuously optimizes the Kalman filter parameters.

[0095] The iteration stops when the matching residual is less than the set convergence threshold or when the number of iterations reaches the preset maximum. At this point, the position coordinates and velocity output by the Kalman filter, combined with the target category output by the deep neural network and the heading calculated based on the velocity, constitute the final target state information. This target state information is output to the submarine's command system via the signal processing module and simultaneously stored in the data storage module.

[0096] This embodiment utilizes a submarine-mounted fiber optic acousto-magnetic composite sensor to achieve synchronous acquisition of acousto-magnetic signals. Noise suppression processing effectively reduces interference from platform vibration and flow-induced noise, improving the signal-to-noise ratio of the acoustic signal. The feature extraction stage comprehensively captures the target's acousto-magnetic signal characteristics through improved empirical mode decomposition algorithms and multi-dimensional statistical analysis. Initial target detection provides a reliable initialization basis for fusion processing. The asynchronous cyclic feedback fusion mechanism fully leverages the advantages of stable magnetic signals and long acoustic signal detection range, continuously improving the accuracy of target state estimation through iterative optimization.

[0097] The technical solution of this embodiment can adapt to the environmental characteristics of complex near-shore waters, improve the reliability and accuracy of underwater target detection, and provide effective technical support for submarine underwater detection missions.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An underwater acoustic-magnetic joint sensing method, characterized in that, The method includes the following steps: Step 1: Simultaneously acquire the raw acoustic and magnetic signals of the underwater target using a fiber optic acousto-magnetic composite sensor; Step 2: Perform noise suppression processing on the original acoustic signal to obtain a noise-reduced acoustic signal. The noise suppression processing adopts an algorithm that combines adaptive filtering and phase tracking. Step 3: Extract features from the denoised acoustic signal and the original magnetic signal respectively to obtain acoustic signal features and magnetic signal features; Step four: Based on the acoustic signal characteristics and the magnetic signal characteristics, perform asynchronous cyclic feedback fusion processing to obtain the state information of the underwater target; The asynchronous cyclic feedback fusion processing includes: obtaining a first position estimate based on the magnetic signal features through a first prediction model; using the first position estimate as input to drive an acoustic propagation model to calculate theoretical acoustic features corresponding to the acoustic signal features; calculating the matching residual between the theoretical acoustic features and the actual acoustic features extracted from the denoised acoustic signal; generating a feedback correction amount based on the matching residual; and using the feedback correction amount to optimize the first prediction model to update the first position estimate; iteratively executing the steps of driving the acoustic propagation model, calculating the matching residual, and optimizing the first prediction model until a preset convergence condition is met, and outputting the final target state information.

2. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, The fiber optic acoustomagnetic composite sensor integrates a fiber optic vector hydrophone unit and a fiber optic integrated diamond NV color center vector magnetometer unit. The fiber optic vector hydrophone unit adopts a three-branch orthogonally distributed fiber optic Michelson interferometer structure to sense the particle velocity of the three-dimensional sound field. The fiber-integrated diamond NV color center vector magnetometer unit adopts a structure in which diamond containing nitrogen vacancy color centers is integrated on the end face of a tapered optical fiber, and is used to sense a three-dimensional vector magnetic field.

3. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, The noise suppression process in step two specifically includes: The common-mode phase noise of the fiber optic acoustomagnetic composite sensor system is monitored using the output signal of a reference interferometer that is of the same origin as the sensing fiber in the fiber optic acoustomagnetic composite sensor. Establish an empirical noise model that includes platform vibration noise and flow-induced noise; The noise components related to the output signal of the reference interferometer are filtered out from the original acoustic signal by a normalized minimum mean square error adaptive filtering algorithm. Phase tracking technology is used to dynamically compensate for non-stationary residual noise in the filtered signal.

4. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, Step three involves feature extraction of the denoised acoustic signal, specifically including: An improved empirical mode decomposition algorithm that incorporates sample entropy evaluation index is used to decompose the denoised acoustic signal into multiple intrinsic mode functions; Extract the instantaneous frequency, instantaneous amplitude, and energy distribution characteristics of each intrinsic mode function; Calculate the acoustic energy flow vector and phase information of the sound field.

5. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, The first prediction model is a Kalman filter. Based on the magnetic signal characteristics, a first position estimate is obtained through the first prediction model, specifically including: The target is equivalent to a magnetic dipole model; The three components of the vector magnetic field in the magnetic signal characteristics are used as the observation values ​​of the Kalman filter; The coarse location information provided by the initial target detection is used as the state initialization value for the Kalman filter; By recursively applying Kalman filtering, the target's position and state at the next moment are predicted, thus obtaining the first position estimate.

6. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, Based on the matching residual, a feedback correction amount is generated, and the first prediction model is optimized using the feedback correction amount, specifically as follows: The matched residuals are converted into adjustments to the observation noise covariance matrix of the Kalman filter; Using the adjusted observation noise covariance matrix, the Kalman filter recursion is re-executed to update the first position estimate.

7. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, The acoustic propagation model is a vector sound field propagation model based on normal mode theory. Driving the acoustic propagation model, theoretical acoustic features corresponding to the acoustic signal characteristics are calculated, specifically including: Using the first location estimate as the sound source location, combined with current marine environmental parameters, including water temperature, salinity, depth profile, and seabed sediment; Run the vector sound field propagation model to calculate the theoretical propagation path, time delay, multipath structure, and phase relationship of the sound signal from the sound source location to the sensor location; Based on the calculation results, the theoretical acoustic features are generated.

8. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, Calculating the matching residual between the theoretical acoustic features and the actual acoustic features extracted from the denoised acoustic signal specifically includes: The theoretical acoustic features are organized into theoretical feature vectors. ; The actual acoustic features extracted from the noise-reduced acoustic signal are organized into an actual feature vector. ; The matching residual Calculated using the following formula: in, Norm operations on vectors A diagonal weight matrix, pre-defined based on the signal-to-noise ratio or prior importance of each acoustic feature component, is used to adjust the contribution of different feature dimensions in the residual calculation. This indicates that the corresponding elements of the actual eigenvector are subtracted from the corresponding elements of the theoretical eigenvector.

9. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, Step four, prior to the asynchronous cyclic feedback fusion process, also includes an initial target detection step: The acoustic signal features and the magnetic signal features are input into a deep neural network; The deep neural network uses a cross-stage partial connection structure as the backbone network and embeds an attention mechanism module. The deep neural network performs multi-scale feature fusion through a feature pyramid network and a path aggregation network. The deep neural network outputs the initial existence probability, category, and approximate location region of the target; The rough location region is used for initializing the first prediction model.

10. The underwater acoustic-magnetic joint sensing method according to claim 1, characterized in that, The method is applied to an underwater submersible platform equipped with the fiber optic acousto-magnetic composite sensor. The underwater submersible platform includes submarines or autonomous underwater vehicles; The final target status information includes the target's three-dimensional coordinates, speed, heading, and identity category.

Images