A method for extracting weak underwater acoustic lines based on the coupling of dynamics and attention priors

By using a dynamics and attention prior coupling method, and generating dynamic response sequences through Gammatone filtering and adaptive Duffing oscillator array, combined with 1D-ViT attention calculation, the problems of missed detection and false alarms in weak underwater acoustic spectrum in shallow sea environment are solved, and efficient and reliable weak spectrum extraction is achieved.

CN123065576APending Publication Date: 2026-11-10HARBIN ENGINEERING UNIVERSITY SANYA NANHAI INNOVATION & DEVELOPMENT BASE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611549374.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In shallow sea environments, the stable discrete line spectrum signal of underwater navigation machinery has an extremely low signal-to-noise ratio due to ocean noise interference. Traditional detection methods are difficult to effectively extract weak underwater acoustic line spectra, and existing technical solutions have problems such as high false alarm rate and high false alarm rate.

Method used

A method based on dynamics and attention prior coupling is adopted. Dynamic response sequences are generated by Gammatone filtering and adaptive Duffing oscillator array. Combined with 1D-ViT attention calculation, signal preprocessing, physical prior generation, attention-guided feature enhancement and dynamic consistency verification are performed. The model is trained using binary cross-entropy and Focal Loss to output reliable underwater acoustic weak line spectra.

Benefits of technology

It significantly improves the detection rate of line spectra under low signal-to-noise ratio, reduces the false alarm rate, enhances environmental adaptability, and can efficiently extract weak underwater acoustic line spectra in extremely low signal-to-noise ratio environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123065576A_ABST
    Figure CN123065576A_ABST
Patent Text Reader

Abstract

This invention discloses a method for extracting weak underwater acoustic line spectra based on the coupling of dynamics and attention priors, belonging to the field of underwater acoustic signal processing and target detection technology. The method includes: acquiring dual-component signals of sound pressure and vibration velocity using a vector hydrophone; obtaining sound pressure and vibration velocity features through Gammatone biomimetic filtering; then processing the signals through an adaptive Duffing oscillator array, dynamically updating the damping based on an energy adaptive adjustment mechanism, and obtaining a dynamic response sequence through weighted aggregation of dynamic equations; constructing a 1D-ViT model for self-attention calculation by linear layer dimensional matching; performing physical-guided training using a combination of cross-entropy and Focal Loss loss and an AdamW optimizer to output candidate line spectrum trajectories; finally, verifying a triple consistency index using a sliding window, and determining if two or more criteria are met for a true line spectrum before outputting it. This invention integrates nonlinear dynamic physical priors with deep learning feature extraction capabilities, reducing false alarms and missed detections of weak line spectra in shallow sea environments with extremely low signal-to-noise ratios, and improving extraction robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic signal processing and target detection technology, specifically relating to a method for extracting weak underwater acoustic line spectra based on the coupling of dynamics and attention priors. Background Technology

[0002] Underwater mechanical targets continuously radiate stable discrete line spectrum signals. These line spectra are core features for long-range target identification, trajectory tracking, and target type discrimination. However, shallow sea areas suffer from multiple interferences, including surface wave noise, seabed reverberation, marine biological noise, and multipath propagation loss. The target's radiated acoustic signal attenuates significantly after propagating through long-distance underwater acoustic channels, resulting in an extremely low signal-to-noise ratio for the line spectrum signal acquired at the receiving end. The line spectrum amplitude is completely submerged in the broad-spectrum environmental noise, making effective detection difficult using traditional signal detection methods. Existing technical solutions fall into two independent categories, both with inherent technical limitations: (1) Pure physical mechanism driven detection method Represented by the Duffing chaotic oscillator detection, the Duffing oscillator relies on the critical phase transition characteristics of chaotic systems to have a high sensitivity to detect weak periodic signals. Its physical mechanism is clear and it does not require a large number of experimental samples. However, it has two core shortcomings: The system output is only a binary judgment result (with or without line spectrum), which can only provide discrete detection status and cannot output quantifiable, continuously changing signal dynamic response features. It cannot provide fine-grained time-frequency prior information for the back-end feature extraction network. Ocean noise is time-varying and fluctuates dynamically in intensity. The Duffing oscillator with fixed parameters is prone to deviating from the chaotic critical sensitive state. After the noise increases, the detection sensitivity drops sharply. It has poor adaptability to complex and variable shallow sea environments, and the false alarm rate is high when used alone.

[0003] (2) Pure data-driven deep learning detection method Represented by one-dimensional temporal Transformers and CNN temporal networks, deep learning models rely on massive labeled samples to learn the differences between noise and spectral features. They perform well in large-sample, high signal-to-noise ratio scenarios, but have a core drawback: The network training process is not guided by underwater acoustic physical laws and relies solely on data statistical features to fit the mapping relationship. When there are insufficient training samples, mismatch between the measured sea conditions and the training dataset environment, or extremely low input signal-to-noise ratio, the network is very likely to misjudge high-intensity ocean noise as effective features and cannot autonomously focus on the real weak line spectrum region, resulting in prominent weak line spectrum missed detection problems. The model outputs candidate line spectrum trajectories containing a large number of false trajectories generated by noise, lacking physical criteria for screening true and false trajectories, resulting in a high number of false alarms in the final output, and the reliability cannot meet the standards for use in sonar engineering.

[0004] In summary, neither single physical mechanism models nor purely data-driven deep learning models can independently meet the extraction requirements of high detection rate and low false alarm rate for weak underwater acoustic line spectra in shallow sea environments with extremely low signal-to-noise ratios. The industry urgently needs an integrated extraction framework that deeply couples and integrates prior knowledge of nonlinear dynamic physics with the feature learning capabilities of deep learning, breaking through the performance bottlenecks of using the two existing technical approaches individually. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as discrete features output by a single physical model, low signal-to-noise ratio and easy missed detections in pure deep learning, and shallow coupling in existing fusion architectures, this invention aims to propose a method for extracting weak underwater acoustic lines based on the coupling of dynamics and attention priors. The method involves acquiring signals through a vector hydrophone, filtering them with Gammatone, and then passing them through an adaptive Duffing oscillator array to generate a dynamic response sequence. This method couples the signal to 1D-ViT attention computation to enhance weak line spectrum features; then, it eliminates false trajectories through triple dynamic consistency checks of time, frequency, and amplitude. Training employs a combined loss of binary cross-entropy and FocalLoss, coupled with the AdamW optimizer to alleviate sample imbalance. This invention integrates the advantages of physical mechanisms and deep learning, significantly improving the line spectrum detection rate under low signal-to-noise ratio conditions, greatly reducing false alarms, and enhancing environmental adaptability.

[0006] The technical solution adopted in this invention: A method for extracting weak underwater acoustic line spectra based on the coupling of dynamics and attention priors includes a signal preprocessing stage, a physical prior generation stage, an attention-guided feature enhancement stage, and a dynamic consistency verification stage, as detailed below: Step 1: Signal Preprocessing Stage: A vector hydrophone is used to acquire two components: the time-domain signal of sound pressure and the time-domain signal of particle velocity. A Gammatone cochlear bionic filter is constructed, and the two components are filtered to obtain the sound pressure characteristics. Vibration velocity characteristics ; Step 2: Physics Prior Generation Stage: ... , Parallel input improved parameter adaptive Duffing oscillator array, with dynamic damping update based on sliding window local noise energy adaptive adjustment mechanism. The transient response energy of each oscillator is then obtained by numerically solving the Duffing dynamic equation using the fourth-order Runge-Kutta method. The transient response energies of multiple oscillators are then weighted and aggregated to obtain a global continuous dynamic response sequence. ; Step 3: Attention-Guided Feature Enhancement Stage: Based on Linear Layer Matching The dimensions are used to obtain embedded physical prior features. The self-attention computing mechanism of the 1D-ViT model is constructed to perform attention calculation and enhance the features. Noisy training samples are constructed and input into the 1D-ViT model after performing the operations in steps 1 and 2 in sequence. Physically guided supervised training is completed by using a combination loss of binary cross-entropy and Focal Loss, AdamW optimizer, and cosine annealing strategy. After training, candidate line spectrum trajectories are output. Step 4: Dynamic Consistency Verification Stage: The candidate line spectrum trajectories output by the 1D-ViT model are processed in batches by window, and the triple consistency index verification is completed frame by frame to determine the authenticity of each candidate trajectory. If two or more indicators are met, it is determined to be a real line spectrum trajectory, and the remaining false trajectories are eliminated. After the verification is completed, all line spectrum trajectories determined to be real are summarized as the final underwater acoustic weak line spectrum extraction result.

[0007] Preferably, the specific process of the signal preprocessing stage in step 1 is as follows: Step 1.1: Simultaneously acquire raw underwater acoustic signals radiated by the target in the marine environment using an underwater vector hydrophone array, including sound pressure time-domain signals. Time-domain signal of particle vibration This is a two-component signal; Step 1.2: Construct a Gammatone cochlear biomimetic filter based on the frequency domain decomposition characteristics of the human ear basilar membrane. Its mathematical expression is:

[0008] In the formula, It is the impulse response function. This is the filter gain factor. For continuous time variables, Let the filter order be . For filter bandwidth parameters, The center frequency of the single-channel filter. The initial phase of the filter; Step 1.3: Based on the line spectrum distribution of the LOFAR time-frequency plot, select filter channels whose center frequencies fall within the core frequency band, retaining 4-5 effective core channels and discarding irrelevant frequency band channels; process the acquired sound pressure time-domain signal. Time-domain signal of particle vibration The sound pressure characteristics are obtained by inputting the Gammatone cochlear bionic filter into the sound pressure level sensor and performing the filtering calculations in parallel. and vibration velocity characteristics .

[0009] Preferably, the specific process of the physical prior generation stage in step 2 is as follows: Step 2.1: Construct two sets of parallel improved parameter-adaptive Duffing oscillator arrays and apply sound pressure characteristics and vibration velocity characteristics Inputting into the Duffing oscillator array, the equation expression for the Duffing oscillator array is:

[0010] In the formula, For damping parameters, and For nonlinear restoring force with fixed parameters, The oscillator has a built-in periodic driving force. The signal to be detected is the filtered signal from the oscillator. The signal coupling coefficient, The built-in periodic driving force angular frequency, For the displacement state variables of the oscillator, The first derivative of the oscillator's displacement state variable is the oscillator's velocity. The second derivative of the displacement state variable of the oscillator is the acceleration of the oscillator. Step 2.2 Based on the critical threshold condition of the Melnikov analytical method, a sliding window local noise energy adaptive adjustment mechanism is constructed, with a fixed-length sliding time window set as... The local background noise energy estimate within a window is calculated window by window. Dynamically update damping Its equation is expressed as follows:

[0011] In the formula, A critical damping reference value is preset for an ideal noise-free environment. The damping adaptive adjustment step size factor, This serves as a reference energy level for the marine environment and is used in noise energy normalization calculations. The function is a hyperbolic tangent, which constrains the noise energy adjustment amplitude. Step 2.3 will dynamically update the damping. The Duffing dynamic equations were numerically solved using the fourth-order Runge-Kutta method, and the transient response energy of each oscillator was calculated. The calculation formula is as follows:

[0012] In the formula, It is a time-domain displacement sequence. It is a time-domain velocity sequence; Step 2.4 Let the first digit of the Duffing oscillator array be... The weights of the oscillators are: For all Transient energy of an oscillator Perform exponential statistical weighted aggregation calculations to obtain the global continuous dynamic response sequence. The calculation formula is:

[0013] Preferably, the specific process of the attention-guided feature enhancement stage in step 3 is as follows: Step 3.1 Based on linear layers Complete the dimension mapping by using linear layers. Mapping to the same dimension as the attention score matrix yields embedded physical prior features with the same dimension as the attention score matrix. ; Step 3.2 Construct the self-attention calculation mechanism of the 1D-ViT model, embedding the dimension-aligned physical prior features. As a trainable bias correction term, it is input into the attention score calculation process of the self-attention calculation mechanism, and the attention calculation formula is as follows:

[0014] In the formula, For query, key-value matrix, For normalized calculations, Key matrix Matrix transpose calculation, Key matrix Feature dimensions, These are trainable weight coefficients used to adjust the strength of physical prior influences; Step 3.3 Use the underwater acoustic propagation model to generate a clean line spectrum signal with different frequency bands, time-varying Doppler frequency shift and amplitude fluctuation. Superimpose the clean line spectrum onto the measured ocean background noise to generate noisy training samples. Simultaneously generate a binarized mask matrix as the ground truth label, where the time-frequency unit of the real line spectrum is marked as 1 and the background noise region is marked as 0. The sound pressure features are obtained by performing steps 1 and 2 sequentially on the noisy training samples. Vibration velocity characteristics and the corresponding dynamic response sequence sound pressure characteristics Vibration velocity characteristics and the corresponding dynamic response sequence After training the 1D-ViT model, a weighted combination of binary cross-entropy loss and Focal Loss is used as the total loss function to dynamically increase the weights of difficult-to-classify training samples. Then, the AdamW optimizer is used to update the trainable weight coefficients, and the learning rate is dynamically adjusted in conjunction with a cosine annealing strategy to promote model convergence. Through iterative training, until the line spectrum detection rate and false alarm rate reach the preset convergence threshold, candidate line spectrum trajectories are output, and the optimal weights of the 1D-ViT model are saved.

[0015] Preferably, the specific process of the dynamic consistency verification stage in step 4 is as follows: A fixed-length sliding window is set up to process the candidate line spectrum trajectories output by the 1D-ViT model in batches, and a triple consistency index verification is performed frame by frame to determine the authenticity of each candidate trajectory. The triple consistency index includes time consistency verification, frequency consistency verification, and amplitude consistency verification, as detailed below: Temporal consistency verification: Peak moments of candidate line spectral trajectories and dynamic response sequences Peak time overlap rate > 70%, overlap rate = number of common peak times / total number of times in total trajectory duration; Frequency consistency verification: The deviation between the center frequency of the candidate line spectrum and the characteristic frequency corresponding to the peak value of the dynamic response is less than 2 times the frequency resolution; Amplitude Consistency Verification: Temporal Amplitude and Dynamic Response Sequence of Candidate Line Spectra The time-series amplitude Pearson correlation coefficient is >0.6; If two or more criteria are met, the trajectory is considered a true line spectrum trajectory and is retained for output; if only one criterion is met or none are met, the trajectory is considered a false trajectory generated by noise and is directly removed; after verification, all the line spectrum trajectories determined to be true are summarized as the final underwater acoustic weak line spectrum extraction result.

[0016] Compared with existing technologies, this invention proposes a method for extracting weak underwater acoustic lines based on the coupling of dynamics and attention priors. The advantages of this method are: (1) Breaking through the limitations of single physical Duffing oscillator technology: Traditional Duffing oscillators only output binary discrete detection results. This invention generates a continuous and quantifiable dynamic response sequence by weighted aggregation of multiple oscillators. This transforms the physical sensitivity of nonlinear chaotic detection into fine-grained time-frequency prior information that can be directly integrated into deep learning networks, significantly improving the utilization rate of physical information. Coupled with a damped adaptive update mechanism, it maintains high detection sensitivity under time-varying noise conditions, significantly enhancing environmental adaptability.

[0017] (2) Solving the problem of low signal-to-noise ratio and missed detection in pure deep learning: The dynamic physical prior is embedded as a bias term into the 1D-ViT self-attention calculation process. The network training and inference process is constrained by the laws of underwater acoustic physics. Attention weights are actively allocated to the potential line spectrum region. In the background of extremely low signal-to-noise ratio noise, noise feature interference can be effectively suppressed, and the missed detection rate of weak line spectrum can be greatly reduced; at the same time As trainable parameters, the model can adaptively balance the weights of physical priors and data features without requiring manual adjustment of the prior strength.

[0018] (3) Back-end dynamic consistency verification suppresses false alarms: Based on the objective physical criteria output by the dynamic system, the authenticity of candidate trajectories is verified in three dimensions: time, frequency and amplitude. This is different from the output results of pure deep learning without physical basis. It accurately filters out false line spectrum trajectories induced by noise, significantly reduces the false alarm rate of the extraction results, and the output reliability meets the standards for use in sonar engineering.

[0019] (4) The fusion framework has strong generalization ability and low sample dependence: it relies on physical priors to make up for the performance degradation caused by insufficient training samples and mismatch of sea conditions, and can achieve stable extraction without massive full-sea area labeled samples; the whole dynamics + attention coupling fusion paradigm can be transferred to other weak signal detection fields such as radar weak target spectrum and vibration weak fault characteristics, and has general expansion value.

[0020] (5) High feasibility of the complete project: The overall process is modularly decomposed. Each module, including signal preprocessing, dynamic array calculation, 1D-ViT inference, and consistency verification, can be independently deployed on embedded or GPU parallel computing hardware. The vector hydrophone is an existing mature underwater acoustic acquisition device, which does not require modification of the hardware acquisition end and can be directly adapted to the upgrade and transformation of the existing passive sonar detection system. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall framework of the method of the present invention; Figure 2 This is a flowchart illustrating the physical prior generation stage of the present invention. Figure 3 This is a flowchart illustrating the dynamic consistency verification logic of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be further described clearly and completely below with reference to the accompanying drawings. It should be noted that the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0023] To make the inventive objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings: In order to better understand the above-mentioned objectives, features, and advantages of this invention, the advantages of this invention will be further illustrated below by comparing the embodiments with the accompanying drawings and specific implementation methods.

[0024] This invention proposes a method for extracting weak underwater acoustic lines based on the coupling of dynamics and attention priors, such as... Figure 1 As shown, the method includes a signal preprocessing stage, a physical prior generation stage, an attention-guided feature enhancement stage, and a dynamic consistency verification stage. The process of this method is detailed below: Step 1: Signal Preprocessing Stage: A vector hydrophone is used to acquire two components: the time-domain signal of sound pressure and the time-domain signal of particle velocity. A Gammatone cochlear bionic filter is constructed, and the two components are filtered to obtain the sound pressure characteristics. Vibration velocity characteristics ; Specifically, the signal preprocessing stage in step 1 is as follows: Step 1.1: Simultaneously acquire raw underwater acoustic signals radiated by the target in the marine environment using an underwater vector hydrophone array, including sound pressure time-domain signals. Time-domain signal of particle vibration This is a two-component signal; Step 1.2: Construct a Gammatone cochlear biomimetic filter based on the frequency domain decomposition characteristics of the human ear basilar membrane. The filter's center frequency is uniformly distributed along a logarithmic scale, covering the conventional frequency band of the underwater target line spectrum from 10Hz to 4kHz. The initial total number of channels in the filter is set to 16. Its mathematical expression is:

[0025] In the formula, It is the impulse response function. This is the filter gain factor. For continuous time variables, Let the filter order be . For filter bandwidth parameters, The center frequency of the single-channel filter. The initial phase of the filter; Step 1.3: Based on the line spectrum distribution of the LOFAR time-frequency plot, select filter channels whose center frequencies fall within the core frequency bands, retaining 4-5 effective core channels and eliminating irrelevant frequency band channels to reduce subsequent computational load; process the acquired sound pressure time-domain signal. Time-domain signal of particle vibration The sound pressure characteristics are obtained by inputting the Gammatone cochlear bionic filter into the sound pressure level sensor and performing the filtering calculations in parallel. and vibration velocity characteristics Furthermore, a high-frequency adaptive gain is applied to the output of the vibration velocity filter. The gain value decreases linearly from 0.5 to 0.3 as the frequency increases, compensating for the weakening of vibration velocity characteristics caused by high-frequency attenuation in the underwater acoustic channel.

[0026] Step 2: Physics Prior Generation Stage: ... , Parallel input improved parameter adaptive Duffing oscillator array, with dynamic damping update based on sliding window local noise energy adaptive adjustment mechanism. The transient response energy of each oscillator is then obtained by numerically solving the Duffing dynamic equation using the fourth-order Runge-Kutta method. The transient response energies of multiple oscillators are then weighted and aggregated to obtain a global continuous dynamic response sequence. ; Specifically, in step 2, such as Figure 2 As shown, the specific process of the physical prior generation stage is as follows: Step 2.1: Construct two sets of parallel improved parameter-adaptive Duffing oscillator arrays and apply sound pressure characteristics and vibration velocity characteristics The input is fed into a Duffing oscillator array. Each array contains N independent Duffing oscillators, and the angular frequency of the built-in periodic driving force of each oscillator has a slight difference, completely covering the entire measured line spectrum frequency band of the target. The equation expression for the Duffing oscillator array is as follows:

[0027] In the formula, These are damping parameters (which can be updated adaptively over time). and The nonlinear restoring force is a fixed parameter, initialized as a constant. The oscillator has a built-in periodic driving force. The signal to be detected is the filtered signal from the oscillator. The signal coupling coefficient controls the coupling strength between the input signal and the nonlinear system. The built-in periodic driving force angular frequency, For the displacement state variables of the oscillator, The first derivative of the oscillator's displacement state variable is the oscillator's velocity. The second derivative of the displacement state variable of the oscillator is the acceleration of the oscillator. Step 2.2 Ocean noise changes dynamically in real time. Fixed damping can cause the oscillator to escape the chaotic critical sensitive state. Based on the critical threshold condition of Melnikov's analytical method, a sliding window local noise energy adaptive adjustment mechanism is constructed, with a fixed-length sliding time window set as... The local background noise energy estimate within a window is calculated window by window. Dynamically update damping To ensure that each Duffing oscillator remains in a chaotic critical state highly sensitive to weak line spectra under time-varying and extremely low signal-to-noise ratio shallow sea background noise, the equation expression is as follows:

[0028] In the formula, A critical damping reference value is preset for an ideal noise-free environment. The damping adaptive adjustment step size factor, This serves as a reference energy level for the marine environment and is used in noise energy normalization calculations. The function is a hyperbolic tangent, which constrains the noise energy adjustment amplitude. When local noise energy When it rises, Automatically lowered to ensure the oscillator remains in a critical chaotic state; Step 2.3 will dynamically update the damping. The Duffing dynamic equations were numerically solved using the fourth-order Runge-Kutta method, and the transient response energy of each oscillator was calculated. This characterizes the response intensity of the oscillator to the input signal and monitors the phase trajectory change of each oscillator. The calculation formula is as follows:

[0029] In the formula, It is a time-domain displacement sequence. It is a time-domain velocity sequence; Step 2.4 Let the first digit of the Duffing oscillator array be... The weights of the oscillators are: The weight value is determined by the degree of matching between the oscillator's built-in center frequency and the target line spectrum frequency band; the higher the matching degree, the greater the weight. Transient energy of an oscillator Perform exponential statistical weighted aggregation calculations to obtain the global continuous dynamic response sequence. The calculation formula is:

[0030] This globally continuous dynamic response sequence is continuously quantized data in the entire time domain, capable of reflecting the response intensity of a nonlinear system to an input signal in real time and quantitatively. When the input signal contains a weak line spectrum close to the built-in frequency of the oscillator, significant synchronization or state transitions will occur, and... Peaks are formed at corresponding positions in the sequence, thus accurately indicating the time-frequency positions of potential weak line spectra.

[0031] Step 3: Attention-Guided Feature Enhancement Stage: Based on Linear Layer Matching The dimensions are used to obtain embedded physical prior features. The self-attention computing mechanism of the 1D-ViT model is constructed to perform attention calculation and enhance the features. Noisy training samples are constructed and input into the 1D-ViT model after performing the operations in steps 1 and 2 in sequence. Physically guided supervised training is completed by using a combination loss of binary cross-entropy and Focal Loss, AdamW optimizer, and cosine annealing strategy. After training, candidate line spectrum trajectories are output. Specifically, the attention-guided feature enhancement stage in step 3 is as follows: Step 3.1 Based on linear layers Complete the dimension mapping by using linear layers. ( (Since it is a one-dimensional scalar time series, it cannot be directly superimposed onto the attention score matrix.) Mapping it to the same dimension as the attention score matrix yields embedded physical prior features with the same dimension as the attention score matrix. This ensures bidirectional alignment of temporal length and feature dimension. Step 3.2 Constructs a self-attention calculation mechanism for the 1D-ViT model. Unlike the standard 1D-ViT model, which relies solely on temporal master input features to solve feature associations without constraints, this mechanism embeds dimension-aligned physical prior features. As a trainable bias correction term, it is input into the attention score calculation process of the self-attention calculation mechanism, and the attention calculation formula is as follows:

[0032] In the formula, For query, key-value matrix, To normalize the calculation, output the attention weights at each temporal position; Key matrix Matrix transpose calculation, Key matrix Feature dimensions, Training initialization settings are used to adjust the trainable weights that influence the strength of prior physical knowledge. It updates automatically with backpropagation; Embedded physical prior features The higher the value of the temporal position, the higher the attention score, and the more computing resources will be automatically allocated to focus on the potential line spectral region, so as to achieve targeted enhancement of weak line spectral features; Step 3.3 Generate a clean line spectrum signal with different frequency bands, time-varying Doppler frequency shift and amplitude fluctuation using the underwater acoustic propagation model. Superimpose the clean line spectrum onto the measured ocean background noise to generate a noisy training sample that closely matches the real shallow sea environment. Simultaneously generate a binary mask matrix that matches the time-frequency dimension of the noisy training sample as the ground truth label, where the time-frequency unit of the real line spectrum is marked as 1 and the background noise region is marked as 0. The sound pressure features are obtained by performing steps 1 and 2 sequentially on the noisy training samples. Vibration velocity characteristics and the corresponding dynamic response sequence sound pressure characteristics Vibration velocity characteristics and the corresponding dynamic response sequence After training the 1D-ViT model, to address the extreme sparseness of weak line spectra across the entire time-frequency space and the severe imbalance between positive and negative samples, the conventional single mean square error loss is abandoned. Instead, a weighted combination of binary cross-entropy loss and Focal Loss loss is used as the total loss function. This dynamically increases the weight of difficult-to-classify training samples, enhances the training weight of weak line spectra, and suppresses the interference of a large number of background noise and simple negative samples on gradient updates. Next, the AdamW optimizer is used to update the trainable weight coefficients, and a cosine annealing strategy is used to dynamically adjust the learning rate to promote model convergence. Through iterative training, until the line spectrum detection rate and false alarm rate reach the preset convergence threshold, candidate line spectrum trajectories (including real line spectra and false trajectories generated by noise) are output, and the optimal weights of the 1D-ViT model are saved.

[0033] Step 4: Dynamic Consistency Verification Stage: The candidate line spectrum trajectories output by the 1D-ViT model are processed in batches by window, and the triple consistency index verification is completed frame by frame to determine the authenticity of each candidate trajectory. If two or more indicators are met, it is determined to be a real line spectrum trajectory, and the remaining false trajectories are removed. After the verification is completed, all line spectrum trajectories determined to be real are summarized as the final underwater acoustic weak line spectrum extraction result. Specifically, in step 4, such as Figure 3 As shown, the specific process of the dynamic consistency verification stage is as follows: A fixed-length sliding window is set up to process the candidate line spectrum trajectories output by the 1D-ViT model in batches, and a triple consistency index verification is performed frame by frame to determine the authenticity of each candidate trajectory. The triple consistency index includes time consistency verification, frequency consistency verification, and amplitude consistency verification, as detailed below: Temporal consistency verification: Peak moments of candidate line spectral trajectories and dynamic response sequences Peak time overlap rate > 70%, overlap rate = number of common peak times / total number of times in total trajectory duration; Frequency consistency verification: The deviation between the center frequency of the candidate line spectrum and the characteristic frequency corresponding to the peak value of the dynamic response is less than 2 times the frequency resolution; Amplitude Consistency Verification: Temporal Amplitude and Dynamic Response Sequence of Candidate Line Spectra The time-series amplitude Pearson correlation coefficient is >0.6; If two or more criteria are met (number of criteria ≥ 2), it is determined to be a true line spectrum trajectory (i.e., a true and effective weak underwater acoustic line spectrum trajectory) and the output is retained; if only one criterion is met or none of the criteria are met, it is determined to be a false trajectory generated by noise and is directly removed; after the verification is completed, all line spectrum trajectories determined to be true are summarized as the final weak underwater acoustic line spectrum extraction result, which can be directly used for underwater target recognition and tracking backend modules.

[0034] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0035] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for extracting weak underwater acoustic line spectra based on the coupling of dynamics and attention priors, characterized in that, It includes a signal preprocessing stage, a physical prior generation stage, an attention-guided feature enhancement stage, and a dynamic consistency verification stage, as detailed below: Step 1: Signal Preprocessing Stage: A vector hydrophone is used to acquire two components: the time-domain signal of sound pressure and the time-domain signal of particle velocity. A Gammatone cochlear bionic filter is constructed, and the two components are filtered to obtain the sound pressure characteristics. Vibration velocity characteristics ; Step 2: Physics Prior Generation Stage: ... , Parallel input improved parameter adaptive Duffing oscillator array, with dynamic damping update based on sliding window local noise energy adaptive adjustment mechanism. The transient response energy of each oscillator is then obtained by numerically solving the Duffing dynamic equation using the fourth-order Runge-Kutta method. The transient response energies of multiple oscillators are then weighted and aggregated to obtain a global continuous dynamic response sequence. ; Step 3: Attention-Guided Feature Enhancement Stage: Based on Linear Layer Matching The dimensions are used to obtain embedded physical prior features, and the self-attention computing mechanism of the 1D-ViT model is used to perform attention calculation to enhance the features. Noisy training samples are constructed and then input into the 1D-ViT model for training after performing steps 1 and 2 in sequence. Physically guided supervised training is completed using a combination of binary cross-entropy and Focal Loss, AdamW optimizer, and cosine annealing strategy. After training, candidate line spectral trajectories are output. Step 4: Dynamic Consistency Verification Stage: The candidate line spectrum trajectories output by the 1D-ViT model are processed in batches by window, and the triple consistency index verification is completed frame by frame to determine the authenticity of each candidate trajectory. If two or more indicators are met, it is determined to be a real line spectrum trajectory, and the remaining false trajectories are eliminated. After the verification is completed, all line spectrum trajectories determined to be real are summarized as the final underwater acoustic weak line spectrum extraction result.

2. The underwater acoustic weak line spectrum extraction method based on the coupling of dynamics and attention priors as described in claim 1, characterized in that, In step 1, the specific process of the signal preprocessing stage is as follows: Step 1.1: Simultaneously acquire raw underwater acoustic signals radiated by the target in the marine environment using an underwater vector hydrophone array, including sound pressure time-domain signals. Time-domain signal of particle vibration This is a two-component signal; Step 1.2: Construct a Gammatone cochlear biomimetic filter based on the frequency domain decomposition characteristics of the human ear basilar membrane. Its mathematical expression is: , In the formula, It is the impulse response function. This is the filter gain factor. For continuous time variables, Let the filter order be . For filter bandwidth parameters, The center frequency of the single-channel filter. The initial phase of the filter; Step 1.3: Based on the line spectrum distribution of the LOFAR time-frequency plot, select filter channels whose center frequencies fall within the core frequency band, retaining 4-5 effective core channels and discarding irrelevant frequency band channels; process the acquired sound pressure time-domain signal. Time-domain signal of particle vibration The sound pressure characteristics are obtained by inputting the Gammatone cochlear bionic filter into the sound pressure level sensor and performing the filtering calculations in parallel. and vibration velocity characteristics .

3. The underwater acoustic weak line spectrum extraction method based on the coupling of dynamics and attention priors as described in claim 1, characterized in that, In step 2, the specific process of the physical prior generation stage is as follows: Step 2.1: Construct two sets of parallel improved parameter-adaptive Duffing oscillator arrays and apply sound pressure characteristics and vibration velocity characteristics Inputting into the Duffing oscillator array, the equation expression for the Duffing oscillator array is: , In the formula, For damping parameters, and For nonlinear restoring force with fixed parameters, The oscillator has a built-in periodic driving force. The signal to be detected is the filtered signal from the oscillator. The signal coupling coefficient, The built-in periodic driving force angular frequency, For the displacement state variables of the oscillator, The first derivative of the oscillator's displacement state variable is the oscillator's velocity. The second derivative of the displacement state variable of the oscillator is the acceleration of the oscillator. Step 2.2 Based on the critical threshold condition of the Melnikov analytical method, a sliding window local noise energy adaptive adjustment mechanism is constructed, with a fixed-length sliding time window set as... The local background noise energy estimate within a window is calculated window by window. Dynamically update damping Its equation is expressed as follows: , In the formula, A critical damping reference value is preset for an ideal noise-free environment. The damping adaptive adjustment step size factor, This serves as a reference energy level for the marine environment and is used in noise energy normalization calculations. The function is a hyperbolic tangent, which constrains the noise energy adjustment amplitude. Step 2.3 will dynamically update the damping. The Duffing dynamic equations were numerically solved using the fourth-order Runge-Kutta method, and the transient response energy of each oscillator was calculated. The calculation formula is as follows: , In the formula, It is a time-domain displacement sequence. It is a time-domain velocity sequence; Step 2.4 Let the first digit of the Duffing oscillator array be... The weights of the oscillators are: For all Transient energy of an oscillator Perform exponential statistical weighted aggregation calculations to obtain the global continuous dynamic response sequence. The calculation formula is: 。 4. The underwater acoustic weak line spectrum extraction method based on the coupling of dynamics and attention priors as described in claim 1, characterized in that, In step 3, the specific process of the attention-guided feature enhancement stage is as follows: Step 3.1 Based on linear layers Complete the dimension mapping by using linear layers. Mapping to the same dimension as the attention score matrix yields embedded physical prior features with the same dimension as the attention score matrix. ; Step 3.2 Construct the self-attention calculation mechanism of the 1D-ViT model, embedding dimension-aligned physical prior features. As a trainable bias correction term, it is input into the attention score calculation process of the self-attention calculation mechanism, and the attention calculation formula is as follows: , In the formula, For querying key-value matrices, For normalized calculations, Key matrix Matrix transpose calculation, Key matrix Feature dimensions, These are trainable weight coefficients used to adjust the strength of physical prior influences; Step 3.3 Use the underwater acoustic propagation model to generate a clean line spectrum signal with different frequency bands, time-varying Doppler frequency shift and amplitude fluctuation. Superimpose the clean line spectrum onto the measured ocean background noise to generate noisy training samples. Simultaneously generate a binarized mask matrix as the ground truth label, where the time-frequency unit of the real line spectrum is marked as 1 and the background noise region is marked as 0. The sound pressure features are obtained by performing steps 1 and 2 sequentially on the noisy training samples. Vibration velocity characteristics and the corresponding dynamic response sequence sound pressure characteristics Vibration velocity characteristics and the corresponding dynamic response sequence After training the 1D-ViT model, a weighted combination of binary cross-entropy loss and Focal Loss is used as the total loss function to dynamically increase the weights of difficult-to-classify training samples. Then, the AdamW optimizer is used to update the trainable weight coefficients, and the learning rate is dynamically adjusted in conjunction with a cosine annealing strategy to promote model convergence. Through iterative training, until the line spectrum detection rate and false alarm rate reach the preset convergence threshold, candidate line spectrum trajectories are output, and the optimal weights of the 1D-ViT model are saved.

5. The underwater acoustic weak line spectrum extraction method based on the coupling of dynamics and attention priors as described in claim 1, characterized in that, In step 4, the specific process of the dynamic consistency verification stage is as follows: A fixed-length sliding window is set up to process the candidate line spectrum trajectories output by the 1D-ViT model in batches, and a triple consistency index verification is performed frame by frame to determine the authenticity of each candidate trajectory. The triple consistency index includes time consistency verification, frequency consistency verification, and amplitude consistency verification, as detailed below: Temporal consistency verification: Peak moments of candidate line spectral trajectories and dynamic response sequences Peak time overlap rate > 70%, overlap rate = number of common peak times / total number of times in total trajectory duration; Frequency consistency verification: The deviation between the center frequency of the candidate line spectrum and the characteristic frequency corresponding to the peak value of the dynamic response is less than 2 times the frequency resolution; Amplitude Consistency Verification: Temporal Amplitude and Dynamic Response Sequence of Candidate Line Spectra The time-series amplitude Pearson correlation coefficient is >0.6; If two or more criteria are met, it is determined to be a true line spectrum trajectory and the output is retained; if only one criterion is met or none of the criteria are met, it is determined to be a false trajectory generated by noise and is directly rejected. After verification, all line spectrum trajectories determined to be true are summarized as the final underwater acoustic weak line spectrum extraction result.