A multi-scale energy phase fusion and adaptive difference domain-invariant underwater acoustic feature extraction method

Abstract

The application discloses a multi-scale energy-phase fusion and adaptive difference domain-invariant underwater acoustic feature extraction method, and particularly relates to the design of four core stages of multi-time-frequency resolution spectrum extraction, multi-scale energy-phase fusion, learnable feature alignment and channel splicing, and adaptive time-frequency difference reconstruction. In view of the performance decline problem of the existing method in cross-sea area deployment, the method has the functions of underwater acoustic target robust recognition and generalization application of resisting amplitude fading and suppressing stationary environment artifacts as a light-weight front-end feature extraction module.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic feature extraction technology, specifically involving a domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference. Background Technology

[0002] Underwater acoustic target recognition (UATR) plays a crucial role in applications such as marine environmental monitoring and maritime traffic management. However, unlike aeroacoustic scenarios, underwater channels exhibit highly time- and space-varying characteristics and are strongly affected by multipath propagation. Variations in seawater depth, salinity, seabed characteristics, and wind and wave environmental noise can severely distort ship radiated noise. Therefore, training data collected in specific sea areas is often deeply coupled with background noise and channel distortion in the source domain. When models are deployed in unknown sea areas with entirely different physical conditions, the mismatch between the source and target domain distributions can lead to severe domain shift. To address this issue, existing technologies have introduced domain generalization (DG) methods, which have shown numerous attempts in their development but have also revealed significant limitations in their underlying mechanisms.

[0003] Existing underwater domain generalization methods primarily focus on backend network architecture design and deep learning training strategies. For example, in cross-domain pre-training and distribution alignment, some studies have proposed cross-domain pre-training networks based on FBank fusion features to improve adaptability to unknown underwater data; in extracting domain-invariant features through contrastive learning, some methods have proposed cross-domain contrastive few-shot recognition methods to reduce the distance between similar targets across domains. Compared to traditional algorithms, these backend methods based on Transformers or complex contrastive networks improve recognition performance in specific scenarios.

[0004] However, the aforementioned mainstream domain generalization methods generally suffer from a fundamental problem: they primarily rely on the fitting and alignment capabilities of the back-end network, neglecting the physical generalization capabilities of the front-end acoustic feature extraction. Most features used for underwater target recognition are directly borrowed from the field of speech recognition. For example, applying nonlinear compression to features based on Mel-frequency cepstral coefficients (MFCC) in the low-frequency region may discard low-frequency mechanical line spectra; Short-Time Fourier Transform (STFT), the most widely used time-frequency analysis method, relies on a fixed sliding window; Continuous Wavelet Transform (CWT), while alleviating resolution conflicts to some extent, heavily depends on prior wavelet selection and hyperparameter settings such as center frequency. Because these front-end representations themselves lack cross-domain physical consistency, even if the back-end employs advanced contrastive or adversarial strategies, the input features are often deeply coupled with biases in the source domain environment.

[0005] Compared to the complex iterations of backend networks, existing deep learning methods suffer from problems such as single scale and poor anti-interference ability in the front-end extraction of underwater acoustic features. Especially when faced with strong domain shifts, how to extract domain-invariant features remains a technical bottleneck restricting recognition accuracy. Specifically, existing technologies have the following three core problems: (1) Single-scale time-frequency resolution bottleneck leads to domain overfitting: Existing feature extraction methods (such as STFT, CWT, etc.) are essentially constrained by the Heisenberg uncertainty principle, which imposes an irreconcilable contradiction between time resolution and frequency resolution. The single static analysis window makes the model too sensitive to the energy distribution of the source domain. For example, under high-frequency resolution configuration, the model can accurately capture the narrowband harmonics of the source domain target; however, when transferred to an unknown sea area with severe multipath effects and low signal-to-noise ratio, its performance may deteriorate sharply due to the loss of temporal details.

[0006] (2) The lack of phase information makes the features extremely vulnerable to channel amplitude fading: Most existing time-frequency representation methods only retain amplitude or energy, while discarding phase information. In underwater propagation, energy is easily affected by channel amplitude fading and absorption, while phase retains relative delay information and a more stable intrinsic periodic structure across domains. Although some recent studies have attempted to combine amplitude and phase, the lack of explicit handling of phase wrapping and nonlinear changes has resulted in a lack of effective physical fusion between the two.

[0007] (3) Deep coupling between static features and source domain stationary environment artifacts: The extracted static features inevitably contain stationary background noise specific to the source domain. This static environment bias can lead to severe overfitting in domain generalization. Although existing techniques attempt to introduce dynamic difference features (such as first-order or second-order time difference) to capture transient evolution and suppress stationary background, traditional difference features rely on fixed intervals and hard-coded subtraction, lacking adaptability to different time-frequency scales. This rigid mechanism may lead to over-differentiation, which not only removes background interference but also destroys the target's own weak but persistent low-frequency mechanical spectrum. If the underlying acoustic features are deeply coupled with this source domain environment bias, the backend network will use these biased cues as classification shortcuts, making it difficult to learn the true domain-invariant physical features. Summary of the Invention

[0008] This invention provides a domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive differential, addressing the performance degradation of existing methods when deployed across sea areas.

[0009] This invention is achieved through the following technical solution: A domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference, comprising the following steps: Step 1: Extract the underwater acoustic multi-time-frequency resolution spectrum; Step 2: Perform multi-scale energy phase fusion on the underwater acoustic multi-time-frequency resolution spectrum extracted in Step 1; Step 3: Perform learnable feature alignment and channel concatenation on the multi-branch energy-phase fusion feature matrix extracted in Step 2; Step 4: Perform adaptive time-frequency difference reconstruction on the multi-scale energy-phase fusion representation spliced ​​in Step 3 to achieve domain-invariant underwater acoustic feature extraction based on multi-scale energy-phase fusion and adaptive difference.

[0010] Furthermore, step 1 specifically includes the following steps: Step 1.1: Original signal resampling and time-domain segmentation preprocessing; Step 1.2: Configure the parameters of the multi-resolution adjustable sliding window; Step 1.3: Decoupling parallel time-frequency integral transform from complex spectrum; Step 1.4: Time-frequency feature map dimension interpolation adjustment.

[0011] Furthermore, step 1.1 specifically involves first acquiring the original one-dimensional underwater acoustic time-domain signal. ;Will Resampling is performed uniformly, and the resampling frequency is configured as follows: Obtain the standard one-dimensional time-domain sequence after resampling. Based on the set time span, the long-term audio recording is divided into 3-second independent underwater acoustic samples that do not overlap, which serve as the standard input unit for subsequent time-frequency analysis. Step 1.2 specifically involves configuring a three-way parallel sliding Hamming window operator to construct a time-frequency observation basis at the physical level that can complement transient features and fine line spectrum features; Time-domain window function of each branch Satisfy the following general expression: (1) In the formula, Indicates the first The length of the sliding window for each branch; the number of points for the Fast Fourier Transform. Window length and sliding step size Configure the scaling to be doubled, with the following settings: Enable high frequency resolution / low temporal resolution branching. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Balance the branches with moderate time-frequency resolution. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Enable low frequency resolution / high temporal resolution branching. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Specifically, step 1.3 involves processing the standard one-dimensional underwater acoustic signal obtained in step 1.1. The signals are fed into the three short-time Fourier transform branches configured in step 1.2 in parallel; the one-dimensional time-domain signal is discretized and mapped onto the complex two-dimensional time-frequency space through the time-frequency integral transform operator, and the calculation formula is shown in equation (2): (2) By parallel solving of equation (2), the corresponding single-branch two-dimensional complex time-frequency spectrum matrices are obtained respectively. , and ; Step 1.4 specifically involves obtaining the original matrices of each branch after transformation, with the following feature dimension sizes: high-frequency branch The original shape is Balanced branches The original shape is High-frequency branches The original shape is ; Dimensional interpolation adjustment is performed on the three complex spectrum matrices mentioned above: (3) (4) (5) In equations (3) to (5), The spatial interpolation operator outputs three sets of fully normalized, complementary time-frequency resolution, all-even-order complex time-frequency spectrum matrices. , as the input for the second stage of energy-phase fusion.

[0012] Furthermore, step 2 specifically includes the following steps: Step 2.1: Amplitude-phase decoupling and decibel energy mapping of complex spectrum matrix; Step 2.2: Cosine periodic encoding of phase information and normalization of structural weights; Step 2.3: Amplitude-phase dynamic fusion based on residual multiplication modulation mechanism.

[0013] Furthermore, step 2.1 specifically involves, for each branch of the output from the first stage ( The complex spectrum matrix of ) First, complex decoupling is performed to separate the amplitude spectrum, which reflects the absolute intensity of the acoustic signal. and the phase spectrum reflecting the relative time delay of the physical structure Extract separately and ; For amplitude spectrum Using the global maximum amplitude as a reference, it is mapped to the interval [0, 1] to construct the energy mapping matrix. : (6) Step 2.2 specifically involves processing the phase spectrum after decoupling and cosine coding. After normalization, the phase mapping matrix mapped to the interval [0, 1] is obtained. : (7) Step 2.3 specifically involves dynamically modulating the energy using phase structure weights and calculating the energy-phase fusion feature matrix of a single branch. : (8) In equation (8), Represents the element-wise dot product operation of a matrix; In the high-energy region of the target signal Acting as activation weights, the final output of each branch It will serve as the direct input for the next stage of feature alignment and channel concatenation.

[0014] Furthermore, step 3 specifically includes the following steps: Step 3.1: Learnable extension of the anisotropic dimension; Step 3.2: Local texture smoothing and checkerboard artifact removal; Step 3.3: High-dimensional semantic compression and physical plane reconstruction; Step 3.4: Cross-scale feature channel stitching and panoramic space construction.

[0015] Furthermore, the specific expansion operation of step 3.1 is as follows; against Branch: Employs an asymmetric one-dimensional time-transposed convolution kernel, upsampling only along the time dimension; Output features The calculation formula is: (9) against Branch: Employs an asymmetric one-dimensional frequency transpose convolution kernel, expanding only along the frequency dimension; Output features The calculation formula is: (10) against Branching: Employing symmetrical transposed convolution kernels, performing balanced expansion along both the time and frequency axes; Output features The calculation formula is: (11) Specifically, step 3.2 involves appending a standard layer after the dimensionality expansion layer of each learnable upsampling module. Convolutional layers are cascaded with batch normalization and ReLU activation functions; the formula is as follows: (12) Step 3.3 specifically involves using a The convolution operator forcibly compresses high-dimensional features back into a single channel; the calculation formula is: (13) Step 3.4 specifically involves concatenating and stitching these single-channel feature maps along the feature channel dimension after the three branches are precisely aligned to a uniform size; the output after stitching is the multi-scale energy-phase fusion representation. The calculation formula is as follows: (14).

[0016] Furthermore, step 4 specifically includes the following steps: Step 4.1: Initial channel enhancement and cross-channel fusion; Step 4.2: Intra-channel time decoupling differential; Step 4.3: Interactive sharing of time-dynamic features; Step 4.4: Intra-channel frequency decoupling differential; Step 4.5: Combining channel compression with global residuals.

[0017] Furthermore, step 4.1 specifically involves using a standard two-dimensional convolution operator to stitch together the multi-scale panoramic features obtained. Mapping to a high-dimensional tensor space, the operation is defined as follows: (15) The kernel parameter is set to the kernel size. Step length ,filling Number of groups The output high-dimensional tensor is denoted as ; Specifically, step 4.2 involves... Each independent channel in the process employs an asymmetric one-dimensional convolution kernel that expands along the time dimension. The configuration parameters are First, calculate the standard convolutional response within the local receptive field. : (16) Define a set of learnable time difference weight parameters The forward, center, and backward differential responses are then normalized using the Softmax operator and adaptively fused. (17) The left, center, and right temporal tensor slices are extracted using a space-filling and translation slicing strategy. Constructing time-domain adaptive local background anchor points using normalized weights : (18) Subsequently, the total spatial response intensity of the convolution kernel is calculated. The anchor point is multiplied at the channel level to precisely construct a correction term for eliminating artifacts in stable environments; dynamic time-difference response. Equivalent to standard convolution minus the adaptive background correction term: (19) Introducing weighted parameters Controlled soft residual reconstruction mechanism Calculate and output time dynamic characteristics : (20) Specifically, step 4.3 involves inserting a standard two-dimensional convolution as a central node between the temporal difference and the subsequent frequency difference. (twenty one) The convolution kernel parameters are configured as follows: ; Specifically, step 4.4 involves capturing the frequency domain dynamic evolution using a mechanism dual to step 4.2; and employing an asymmetric one-dimensional frequency-axis depth convolution kernel. The configuration parameters are ; Calculate the standard frequency response ; Similarly, frequency can be defined to learn differential weights. And normalized by Softmax to Extract the top, middle, and bottom tensor slices along the frequency axis to construct frequency background anchor points. : (twenty two) The frequency differential output is obtained by subtracting the local adaptive environment correction term. This is then combined with frequency soft residual weights. Calculate the frequency dynamic features that both filter out static biases and retain the original features. : (twenty three) (twenty four) The above operator ultimately achieves the decoupled extraction of dynamic evolution modes along both time and frequency axes; Step 4.5 specifically involves compressing the high-dimensional two-dimensional difference tensor features back into a low-dimensional space using a standard convolution operator, with the parameters set as follows: ; (25) The compressed differential feature map is compared with the original feature map input to the module. Add: (26) Low-dimensional features of the output This provides a robust underlying physical representation for subsequent networks.

[0018] The beneficial effects of this invention are: This invention is a lightweight front-end feature extraction method with robust underwater acoustic target recognition and generalization functions that resist amplitude fading and suppress artifacts in stable environments.

[0019] This invention can solve the technical defects of traditional feature extraction methods, such as the single-scale time-frequency resolution bottleneck caused by the Heisenberg uncertainty principle, the extreme vulnerability to channel amplitude fading due to the lack of phase information, and the decline in cross-sea generalization performance caused by the deep coupling of static features with artifacts in the source domain stationary environment.

[0020] This invention demonstrates significant performance advantages in recognition tasks under zero-sample cross-domain conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the multi-time-frequency resolution spectrum extraction process of the present invention.

[0022] Figure 2 This is a flowchart of the multi-scale energy phase fusion process of the present invention.

[0023] Figure 3 This is a flowchart of the learnable feature alignment and channel splicing process of the present invention.

[0024] Figure 4 This is a flowchart of the adaptive time-frequency differential reconstruction of the present invention.

[0025] Figure 5 The above are visualization results of T-SNE average vector clustering under four domain offset conditions according to the present invention, where (a) is the visualization result of T-SNE average vector clustering of STFT features, and (b) is the visualization result of T-SNE average vector clustering of the proposed features. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0027] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] The following is in conjunction with the appendix to this application specification. Figure 1-5 The technical solutions in the embodiments of this application are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0031] Implementation Method 1 This embodiment provides a domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference. The underwater acoustic feature extraction method includes the following steps: Step 1: Extract the underwater acoustic multi-time-frequency resolution spectrum; Step 2: Perform multi-scale energy phase fusion on the underwater acoustic multi-time-frequency resolution spectrum extracted in Step 1; Step 3: Perform learnable feature alignment and channel concatenation on the multi-branch energy-phase fusion feature matrix extracted in Step 2; Step 4: Perform adaptive time-frequency difference reconstruction on the multi-scale energy-phase fusion representation spliced ​​in Step 3 to achieve domain-invariant underwater acoustic feature extraction based on multi-scale energy-phase fusion and adaptive difference.

[0032] Furthermore, step 1 specifically includes the following steps: Step 1.1: Original signal resampling and time-domain segmentation preprocessing; Step 1.2: Configure the parameters of the multi-resolution adjustable sliding window; Step 1.3: Decoupling parallel time-frequency integral transform from complex spectrum; Step 1.4: Time-frequency feature map dimension interpolation adjustment.

[0033] Furthermore, step 1.1 specifically involves first acquiring the original one-dimensional underwater acoustic time-domain signal. Given that the radiated noise energy of underwater acoustic targets is mainly concentrated in the low and mid-frequency bands, in order to filter high-frequency random noise and reduce the complexity of subsequent parallel computing, Resampling is performed uniformly, and the resampling frequency is configured as follows: Obtain the standard one-dimensional time-domain sequence after resampling. Subsequently, based on the set time span, the long-term audio recording was divided into 3-second independent underwater acoustic samples that do not overlap, which served as the standard input unit for subsequent time-frequency analysis. Step 1.2 specifically involves configuring a three-way parallel sliding Hamming window operator to construct a time-frequency observation basis at the physical level that can complement transient features and fine line spectrum features; Time-domain window function of each branch Satisfy the following general expression: (1) In the formula, Indicates the first The sliding window length of each branch; the number of points in the Fast Fourier Transform of this invention. Window length and sliding step size Configure the scaling to be doubled, with the following settings: Let the high frequency resolution / low time resolution branch (denoted as) The parameters of ) satisfy: , , The time-frequency resolution unit size corresponding to this branch is , Specifically designed for precise locking of narrow-band mechanical spectra; Let the medium time-frequency resolution balanced branch (denoted as) The parameters of ) satisfy: , , The time-frequency resolution unit size corresponding to this branch is , This is used to balance timing envelope and frequency harmonics; Let the low frequency resolution / high time resolution branch (denoted as) The parameters of ) satisfy: , , The time-frequency resolution unit size corresponding to this branch is , It is specifically designed to accurately capture the rhythmic characteristics of cavitation bursts or high-frequency transient impacts; Specifically, step 1.3 involves processing the standard one-dimensional underwater acoustic signal obtained in step 1.1. The signals are fed into the three short-time Fourier transform branches configured in step 1.2 in parallel; the one-dimensional time-domain signal is discretized and mapped onto the complex two-dimensional time-frequency space through the time-frequency integral transform operator, and the calculation formula is shown in equation (2): (2) By parallel solving of equation (2), the corresponding single-branch two-dimensional complex time-frequency spectrum matrices are obtained respectively. , and Each output complex matrix unit implicitly contains acoustic intensity amplitude information and physical structure time delay phase information, thus breaking the limitation of traditional single analysis window overfitting source domain ocean energy distribution at the source of numerical extraction. Step 1.4 specifically involves obtaining the original matrices of each branch after transformation, with the following feature dimension sizes: high-frequency branch The original shape is Balanced branches The original shape is High-frequency branches The original shape is ; To provide normalized tensor inputs for subsequent cascaded learnable upsampling network layers, the dimensions of the three complex spectral matrices are adjusted by interpolation: (3) (4) (5) In equations (3) to (5), The spatial interpolation operator adjusts the original matrices of each branch to all even dimensions; through interpolation adjustment, it finally outputs three sets of fully normalized, complementary time-frequency resolution, all-even-order complex time-frequency spectrum matrices. , as the input for the second stage of energy-phase fusion.

[0034] Furthermore, step 2 specifically includes the following steps: Step 2.1: Amplitude-phase decoupling and decibel energy mapping of complex spectrum matrix; Step 2.2: Cosine periodic encoding of phase information and normalization of structural weights; Step 2.3: Amplitude-phase dynamic fusion based on residual multiplication modulation mechanism.

[0035] Furthermore, step 2.1 specifically involves, for each branch of the output from the first stage ( The complex spectrum matrix of ) First, complex decoupling is performed to separate the amplitude spectrum, which reflects the absolute intensity of the acoustic signal. and the phase spectrum reflecting the relative time delay of the physical structure In cross-domain underwater acoustic scenarios, pure energy features are extremely fragile due to propagation attenuation in the underwater acoustic channel, while phase structure inherently has stronger resistance to amplitude distortion. To inject this anti-fading characteristic into the underlying representation, this invention extracts... and ; For amplitude spectrum Using the global maximum amplitude as a reference, it is converted to a relative decibel scale and then subjected to forced normalization to map it to the [0, 1] interval, thereby constructing an energy mapping matrix. : (6) Step 2.2 specifically involves processing the phase spectrum after decoupling and cosine coding. Considering the nonlinear jump problem caused by phase wrapping in the original angle space, this invention utilizes the periodic symmetry of cosine encoding to map the original phase space with broken ends to a continuous weight space with closed boundaries; this operation transforms abstract angle information into structural weights that characterize the physical phase correlation. After normalization, the phase mapping matrix mapped to the [0, 1] interval is obtained. : (7) The essence of normalizing the two mapping matrices is to force the decibel energy and cosine coding phase, which have huge differences in physical dimensions and numerical ranges, to be unified into a dimensionless space. This step ensures that the two maintain a comparable numerical scale in the subsequent fusion calculation and avoids the gradient instability problem in the subsequent network training process. Specifically, step 2.3 states that although cosine coding solves the phase boundary wrapping problem, it does not fundamentally eliminate the phase discontinuity in the background noise region. In the background region of the spectrum, the phase is often completely random and oscillates violently. If the energy and phase are directly spliced ​​along the feature channel dimension, these disordered two-dimensional gradient abrupt changes can easily mislead the network. To effectively suppress random phase abrupt changes in the low-energy region while preserving phase information carrying physical coherence, this invention designs a residual multiplication modulation mechanism based on phase mapping. Specifically, it utilizes phase structure weights to dynamically modulate the energy and calculates the energy-phase fusion feature matrix of a single branch. : (8) In equation (8), This represents the element-wise dot product operation of a matrix; at the signal processing level, the above residual multiplicative modulation mechanism is equivalent to amplitude modulation; and from the perspective of deep learning, it constitutes an energy-gated attention mechanism. The operation of equation (8) achieves adaptive physical reconstruction: in the background noise region with extremely low energy, noisy and strongly disordered phase discontinuities are automatically suppressed, fundamentally reducing the possibility of the network learning phase noise; while in the high-energy region of the target signal, Acting as activation weights, this allows the phase skeleton, containing inherent mechanical periodicity and anti-multipath fading modes, to be deeply imprinted into the energy information; this modulation not only endows the fusion characteristics Physical robustness against amplitude fading also strongly encourages subsequent network learning of energy information coherent with the phase structure; the final output of each branch It will serve as the direct input for the next stage of feature alignment and channel concatenation.

[0036] Furthermore, step 3 specifically includes the following steps: Step 3.1: Learnable extension of the anisotropic dimension; Step 3.2: Local texture smoothing and checkerboard artifact removal; Step 3.3: High-dimensional semantic compression and physical plane reconstruction; Step 3.4: Cross-scale feature channel stitching and panoramic space construction.

[0037] Furthermore, step 3.1 specifically involves the energy-phase fusion characteristics of the three branches output from the first and second stages. Having distinct time-frequency physical dimensions, direct alignment using bilinear interpolation or other fixed resampling scaling methods would severely disrupt the inherent topological structure of acoustic signals. To address this, this invention designs a learnable upsampling module corresponding to each branch. Based on the missing dimensional features of each branch, transposed convolutions with different receptive field shapes are used to perform anisotropic dimensional expansion. The specific expansion operation is as follows. For branches with relatively low time resolution ( Branch): Employs an asymmetric one-dimensional time transpose convolution kernel, upsampling only along the time dimension; output features The calculation formula is: (9) For branches with relatively low frequency resolution ( Branching): Employs an asymmetric one-dimensional frequency transpose convolution kernel, expanding only along the frequency dimension; output features The calculation formula is: (10) For medium resolution branches ( Branching): Employing a symmetrical transposed convolution kernel, it performs balanced expansion along both the time and frequency axes; output features The calculation formula is: (11) This heterogeneous, learnable alignment method fundamentally avoids the information loss caused by hard-coded scaling; Specifically, step 3.2 involves the following: after the initial dimensionality expansion, due to the stride operation characteristics of the transposed convolution operator, significant checkerboard artifacts will appear on the feature map; to eliminate these artifacts and restore temporal continuity, a standard checkerboard pattern is appended after the dimensionality expansion layer of each learnable upsampling module. Convolutional layers are cascaded with batch normalization (BN) and ReLU activation functions; this operation aims to smooth local acoustic textures and reconstruct the continuous distribution of the time-frequency plane, as shown in the following formula: (12) Step 3.3 specifically involves, in order to reconstrain the high-dimensional abstract features extracted during the upsampling process to an expression space with actual physical meaning, employing a... The convolution operator forcibly compresses high-dimensional features back into a single channel; this step maps the expanded acoustic semantic information back to a physically meaningful two-dimensional time-frequency plane, calculated as follows: (13) Step 3.4 specifically involves cascading and stitching these single-channel feature maps along the feature channel dimension after the three branches have passed through the corresponding learnable upsampling modules and been precisely aligned to a uniform size; the output after stitching is the multi-scale energy-phase fusion representation. The calculation formula is as follows: (14) This representation effectively constructs a panoramic acoustic observation space, greatly expanding the boundary of the feature manifold and providing a solid data foundation for subsequent cross-domain sharing pattern search and background artifact removal.

[0038] Furthermore, step 4 specifically includes the following steps: Step 4.1: Initial channel enhancement and cross-channel fusion; Step 4.2: Intra-channel time decoupling differential; Step 4.3: Interactive sharing of time-dynamic features; Step 4.4: Intra-channel frequency decoupling differential; Step 4.5: Combining channel compression with global residuals.

[0039] Furthermore, step 4.1 specifically involves using a standard two-dimensional convolution operator to stitch together the multi-scale panoramic features obtained. Mapping to a high-dimensional tensor space enables deep mixing of multi-scale features within the tensor, providing initial fusion information for subsequent dynamic differencing; the operation is defined as follows: (15) The kernel parameter is set to the kernel size. Step length ,filling Number of groups The output high-dimensional tensor is denoted as ; Specifically, step 4.2 involves using a depthwise separable convolution with the same number of channels to extract temporal dynamic features in order to prevent cross-channel interference during the differential process. against Each independent channel in the process employs an asymmetric one-dimensional convolution kernel that expands along the time dimension. The configuration parameters are First, calculate the standard convolutional response within the local receptive field. : (16) Define a set of learnable time difference weight parameters The forward, center, and backward differential responses are then normalized using the Softmax operator and adaptively fused. (17) To avoid the high cost of three independent difference calculations, a space-filling and translation slicing strategy is used to extract the left, center, and right time tensor slices. Constructing time-domain adaptive local background anchor points using normalized weights : (18) Subsequently, the total spatial response intensity of the convolution kernel is calculated. The anchor point is multiplied at the channel level to precisely construct a correction term for eliminating artifacts in stable environments; dynamic time-difference response. Equivalent to standard convolution minus the adaptive background correction term: (19) To prevent excessive differential filtering by pure high-pass filtering from disrupting the target's low-frequency mechanical spectrum, a weighted parameter is introduced. Controlled soft residual reconstruction mechanism ( ), calculate and output time dynamic characteristics. : (20) Step 4.3 specifically involves inserting a standard two-dimensional convolution as a central node between the time difference and the subsequent frequency difference, enabling the temporal dynamic features extracted within independent channels to be shared and interacted between channels; this process prepares sufficiently mixed temporal dynamic information for subsequent frequency domain analysis. (twenty one) The convolution kernel parameters are configured as follows: ; Specifically, step 4.4 involves capturing the frequency domain dynamic evolution using a mechanism dual to step 4.2; and employing an asymmetric one-dimensional frequency-axis depth convolution kernel. The configuration parameters are ; Calculate the standard frequency response ; Similarly, frequency can be defined to learn differential weights. And normalized by Softmax to Extract the top, middle, and bottom tensor slices along the frequency axis to construct frequency background anchor points. : (twenty two) The frequency differential output is obtained by subtracting the local adaptive environment correction term. This is then combined with frequency soft residual weights. Calculate the frequency dynamic features that both filter out static biases and retain the original features. : (twenty three) (twenty four) The above operator ultimately achieves the decoupled extraction of dynamic evolution modes along both time and frequency axes; Step 4.5 specifically involves compressing the high-dimensional two-dimensional difference tensor features back into a low-dimensional space using a standard convolution operator, with the parameters set as follows: This not only completed the final aggregation of time-frequency information, but also strictly controlled the size of the tensors passed to subsequent networks. (25) Finally, considering the risk of original feature degradation that may result from differential processing, a global residual connection mechanism spanning the entire adaptive differential module is used to connect the compressed differential feature map with the original features input to the module. Add: (26) Low-dimensional features of the output It not only adaptively suppresses environmental artifacts, but also better preserves the original physical characteristics of the target, providing a highly robust underlying physical representation for subsequent networks.

[0040] To verify the effectiveness of the proposed method, comprehensive testing and evaluation were conducted on three representative underwater acoustic datasets: DeepShip, ShipsEar, and the WH-HEU dataset. Compared with 16 existing mainstream underwater acoustic feature extraction methods (including classical time-frequency features, cepstral features, military acoustic features, and other fused features), this invention demonstrates a significant performance advantage in recognition tasks under zero-sample cross-domain conditions. Specific effects are as follows: (1) Significantly improves cross-domain recognition accuracy and distortion robustness under extreme sea conditions As shown in Table 1, in the zero-shot cross-domain recognition experiment on the DeepShip dataset, the features constructed in this invention achieved a highest accuracy of 75.42% under pure, offset-free conditions. More importantly, when facing unknown target domains with complex physical distortions, the performance of this invention is significantly superior: Multipath fading resistance: In scenarios with severe frequency-selective fading caused by multipath reverberation, the performance of traditional energy-dependent features (such as MEL and Fbank) deteriorates sharply. However, this invention, with its multi-scale energy-phase fusion mechanism, utilizes phase as the anchor point of the underlying physical structure, achieving a highest accuracy of 65.87%, effectively resisting channel amplitude fading. Stationary environment artifact resistance: Under the background of wind and wave noise with different signal-to-noise ratios (5dB, 0dB, -5dB), this invention achieved accuracies of 71.38%, 67.91%, and 62.44%, respectively, consistently surpassing all baseline methods. The adaptive differential module successfully acts as a multi-element high-pass filter, removing stationary environment artifacts while preserving key mechanical line spectra. High-frequency dissipation resistance: Under high-frequency attenuation (simulating long-distance propagation) conditions, the accuracy of this invention reached 63.49%, ranking first, proving that the panoramic space constructed by the learnable upsampling module can effectively prevent semantic breaks caused by the loss of a single frequency band. Overall, under all domain offset conditions, this invention achieved an average cross-domain accuracy of 61.29%, which is 3.25% higher than the second-ranked STFT feature, and the average performance degradation remained at an extremely low level.

[0041] Table 1 shows the experimental results of zero-shot cross-domain recognition on the DeepShip dataset.

[0042] Note: Rank by numerical value. (2) Possesses excellent backend architecture generalization capabilities and "plug-and-play" universality. The feature extraction framework proposed in this invention does not rely on a specific classifier design, exhibiting strong versatility. As shown in Table 2, in the evaluation of five mainstream architectures, including classic convolutional networks (ResNet18, VGG16, GoogLeNet) and visual Transformers (ViT-B, Mobile-ViT-S), replacing the traditional STFT with the features of this invention consistently improved the clean conditional accuracy and average cross-domain accuracy of all back-ends: on VGG16 and GoogLeNet, the average cross-domain accuracy improved by 2.68% and 1.48%, respectively. On the pure visual Transformer (ViT-B) lacking local inductive bias, this invention, by injecting local texture prior information in advance at the front end, caused its clean accuracy to surge by 6.52% (from 65.31% to 71.83%), and the average cross-domain accuracy to improve by 5.98%, perfectly compensating for the structural weakness of ViT in focusing fine-grained acoustic textures in complex underwater channels.

[0043] Table 2 Evaluation results of STFT and the features of this invention on five mainstream architectures.

[0044] Unlike domain generalization methods that rely on a large number of backend parameters, this invention, as a lightweight and learnable frontend module, achieves extremely low computational overhead. When using ResNet18 as the backend, the feature extraction process of this invention introduces only a tiny increase in the number of parameters of 0.01M (millions). On a single GPU, its average feature reconstruction inference time is only 1.48 milliseconds, providing powerful generalization performance while fully meeting the stringent deployment requirements of edge devices for real-time performance and low power consumption in actual marine monitoring scenarios.

[0045] (3) Strong consistency in distribution across datasets, and clear boundaries of clustering of underlying features. As shown in Table 3, in addition to its excellent performance on the DeepShip dataset, this invention also demonstrates stable superiority in cross-domain generalization evaluations on the ShipsEar and WH-HEU datasets, with average cross-domain accuracy improvements of 7.69% and 3.24%, respectively. This verifies that the method is independent of specific data distributions and is a widely applicable acoustic front-end paradigm. Meanwhile, the T-SNE average vector clustering visualization results (such as...) Figure 5 As shown in the figure, compared with traditional features that are easily affected by environmental disturbances and diverge, the domain-invariant features extracted by this invention can still form more compact intra-class clusters and clearer inter-class boundaries after experiencing different physical distortions, which proves its excellent cross-domain consistency from the underlying logic of the feature manifold.

[0046] Table 3. Experimental results of cross-domain generalization evaluation on different datasets

[0047] The implementation process of the domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference proposed in this invention is as follows: Multi-time-frequency resolution spectrum extraction: First, the input one-dimensional underwater acoustic radiation noise signal is processed, and complementary complex spectra at different scales are extracted in parallel. While fully exploring multi-level physical information such as transient impact and fine line spectrum, the bottleneck of single-scale observation is broken, laying the foundation for subsequent feature alignment and fusion.

[0048] Multi-scale energy-phase fusion: Through amplitude-phase decoupling and mapping transformation, a residual multiplication modulation mechanism is introduced to achieve dynamic physical fusion of energy and phase. The phase is used as the anchor point of the underlying physical structure, so that the features can effectively resist the multipath effect and amplitude fading of complex underwater acoustic channels, and improve the model's adaptability to harsh marine environments.

[0049] Learnable feature alignment and channel stitching: Multi-scale feature manifold alignment is achieved by using anisotropic transposed convolution. After eliminating chessboard artifacts, channel-level cascade stitching is performed to integrate multi-scale information and construct a panoramic acoustic observation space, breaking down the dimensional barriers between features of different time and frequency scales.

[0050] Adaptive time-frequency difference reconstruction: The stitched features are decoupled from the intra-channel time-frequency dual-axis adaptive difference through cross-channel fusion and dynamic subtraction of the source domain stationary background correction term. This not only adaptively suppresses environmental artifacts, but also better preserves the original target physical features, ultimately improving the cross-domain consistency of features.

[0051] The specific process is as follows: Multi-time-frequency resolution spectrum extraction: To capture multi-scale acoustic features ranging from broadband transient impacts to narrowband mechanical line spectra, enhance the adaptability of the underlying representation to the time-frequency resolution of complex underwater acoustic channels, and improve the model's generalization ability to source domain energy distribution, inspired by the successful application of Short Time Fourier Transform (STFT) in signal time-frequency analysis and multi-scale feature extraction, we achieve comprehensive mining of broadband and narrowband acoustic features through parallel configuration of multi-resolution window functions and complex spectrum extraction. Specifically: 1) First, the original underwater acoustic time-domain signal is resampled to 4096 Hz to filter high-frequency random noise, and the long-time audio is divided into 3-second independent samples that do not overlap to adapt to the front-end network; then, three parallel Hamming window operators (high frequency resolution / low time resolution, medium time-frequency resolution, low frequency resolution / high time resolution branches) are configured to build a physically complementary time-frequency observation basis.

[0052] 2) Then, the one-dimensional underwater acoustic signal is fed in parallel into the three configured STFT branches for time-frequency integral transformation to extract three sets of two-dimensional complex time-frequency spectrum matrices; each complex matrix implicitly contains the amplitude information of acoustic intensity and the time delay phase information of physical structure, breaking the limitation of traditional single analysis window overfitting of marine energy distribution from the source.

[0053] 3) Finally, the original complex matrices of each branch obtained after the integral transform are adjusted for even-numbered dimensions by spatial interpolation, and the size normalized complex spectrum matrices that retain complementary time-frequency resolution are output respectively, providing high-quality input with well-structured structure for subsequent energy phase fusion.

[0054] Multi-scale energy-phase fusion: To implicitly inject phase robustness against channel multipath fading into the underlying acoustic representation, enhance the physical coherence between energy and phase, and avoid misleading interference from phase entanglement and low-energy phase noise, inspired by the successful applications of decibel energy mapping in audio processing, cosine coding in periodic signal representation, and gated attention mechanisms in feature activation, we achieve dynamic physical construction of distortion-resistant robust features through amplitude-phase space mapping and residual multiplication modulation mechanisms. Specifically: 1) First, the normalized complex spectrum matrix of the input is decoupled to separate the amplitude spectrum and the phase spectrum. The amplitude spectrum is converted to a relative decibel scale based on the global maximum value and normalized to the interval [0, 1] to form a dimensionless energy mapping matrix. The phase spectrum is mapped to a closed continuous weight space using the periodic symmetry of the cosine function to form a phase structure matrix, so as to explicitly eliminate the nonlinear jump caused by phase winding.

[0055] 2) Then, a residual multiplication modulation mechanism is used for dynamic amplitude-phase fusion, and the energy mapping matrix is ​​modulated element-wise using phase structure weights. In the low-energy background noise region, noisy and highly disordered two-dimensional phase gradient abrupt changes are automatically suppressed, fundamentally reducing the possibility of the network learning phase noise. In the high-energy region of the target signal, the above modulation mechanism is used to imprint a continuous phase skeleton containing inherent mechanical periodicity and anti-multipath fading modes into the energy intensity information, completing the fusion and reconstruction of physical characteristics, outputting single-branch fused features, and significantly enhancing the physical robustness of the features against amplitude fading.

[0056] Learnable Feature Alignment and Channel Stitching: To eliminate the temporal-frequency dimensional differences in energy phase fusion features from different branches, achieve complementary alignment of multi-resolution features, and enhance the model's resistance to broadband semantic breaks, inspired by the successful application of transposed convolution in feature upsampling and panoramic feature stitching in expanding the observation field, we achieve manifold alignment of multi-scale information and panoramic spatial construction through anisotropic dimensional expansion, local smoothing, and channel-level stitching. Specifically: 1) First, for the three fusion features with different physical dimensions, according to their missing dimensions, asymmetric or symmetric learnable transposed convolution kernels (time transposed convolution is applied to the high-frequency branch, frequency transposed convolution is applied to the high-time branch, and biaxial symmetric transposed convolution is applied to the balanced branch) are used to expand the spatial dimension, and they are independently upsampled to a unified common two-dimensional size to avoid hard-coded scaling from destroying the acoustic topology.

[0057] 2) Then, cascade standards after dimensional expansion. Convolutional layers, batch normalization, and ReLU activation functions are used to suppress checkerboard artifacts caused by transposed convolution stride operations and restore the local texture continuity of the time-frequency plane; then, utilizing... The convolution operator forces high-dimensional abstract semantics back to a single-channel physical time-frequency plane.

[0058] 3) Finally, the three single-channel feature maps aligned to a uniform physical size are cascaded and stitched along the feature channel dimension, integrating complementary scale information from each branch to construct a three-channel, multi-scale panoramic acoustic observation spatial feature. This process greatly expands the boundary of the feature manifold, providing a global perspective and multi-scale support for subsequent background artifact removal.

[0059] Adaptive Time-Frequency Differential Reconstruction: To completely remove deeply coupled source-domain stationary background noise (such as wind waves and reverberation artifacts) from panoramic stitching features and eliminate the interference of static environmental biases, inspired by the successful application of depthwise separable convolution in decoupling features, dynamic high-pass filters in background subtraction, and residual connections in preventing feature degradation, we achieve adaptive elimination of source-domain environmental biases and restoration of physical features through time-frequency dual-axis adaptive decoupling differential and soft residual reconstruction. Specifically: 1) First, standard two-dimensional convolution is used to upscale the multi-scale stitched panoramic feature mapping to a 24-dimensional high-dimensional tensor space to achieve deep mixing of multi-scale features within the channel; then, channel depth separable convolution is used to extract the time dynamic evolution features, and local neighborhood slices are adaptively fused by Softmax normalized learnable weights to construct temporal local background anchors and deduct them from the standard response to complete the adaptive decoupling differential filtering in the time dimension.

[0060] 2) Then, standard convolutional interactions are inserted to share cross-channel temporal features, and a dual mechanism is used for intra-channel frequency decoupling differential. Frequency-axis tensor slices are extracted to construct adaptive frequency environment background correction terms, and stationary interference is stripped away through dynamic subtraction to capture transient edges in the frequency domain. At the same time, within this dual-axis differential process, a soft residual reconstruction mechanism controlled by learnable weights is introduced to retain the low-frequency persistent mechanical spectrum of the original target in the best proportion and avoid over-differentiation.

[0061] 3) Finally, the high-dimensional tensor that has undergone dual-domain difference is compressed back into a 3-channel low-dimensional space through standard convolution to control the feature scale transmitted to the back-end classification network; and a global residual connection mechanism that spans the entire difference module is introduced to perform element-wise summation of the pure difference dynamic features with the original panoramic features input to the module; the final output is a domain-invariant physical representation that both filters out static biases and does not lose the original target topology.

[0062] Implementation Method 2 This embodiment provides a domain-invariant underwater acoustic feature extraction system based on multi-scale energy phase fusion and adaptive difference, using the domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference as described in Embodiment 1. The underwater acoustic feature extraction system includes: Extraction module: Extracts underwater acoustic multi-time-frequency resolution spectra; Phase fusion module: performs multi-scale energy phase fusion on the extracted underwater acoustic multi-time-frequency resolution spectrum; Feature alignment and channel stitching module: Performs learnable feature alignment and channel stitching on the extracted multi-branch energy-phase fusion feature matrix; Adaptive Time-Frequency Differential Reconstruction Module: Adaptive time-frequency differential reconstruction is performed on the stitched multi-scale energy-phase fusion representation to realize domain-invariant underwater acoustic feature extraction based on multi-scale energy-phase fusion and adaptive differential.

Claims

1. A domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference, characterized in that, The underwater acoustic feature extraction method includes the following steps: Step 1: Extract the underwater acoustic multi-time-frequency resolution spectrum; Step 2: Perform multi-scale energy phase fusion on the underwater acoustic multi-time-frequency resolution spectrum extracted in Step 1; Step 3: Perform learnable feature alignment and channel splicing on the multi-branch energy-phase fusion feature matrix extracted in Step 2; Step 4: Perform adaptive time-frequency difference reconstruction on the multi-scale energy-phase fusion representation spliced ​​in Step 3 to achieve domain-invariant underwater acoustic feature extraction based on multi-scale energy-phase fusion and adaptive difference.

2. The underwater acoustic feature extraction method according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1: Original signal resampling and time-domain segmentation preprocessing; Step 1.2: Configure the parameters of the multi-resolution adjustable sliding window; Step 1.3: Decoupling parallel time-frequency integral transform from complex spectrum; Step 1.4: Time-frequency feature map dimension interpolation adjustment.

3. The underwater acoustic feature extraction method according to claim 2, characterized in that, Specifically, step 1.1 involves first acquiring the original one-dimensional underwater acoustic time-domain signal. ;Will Resampling is performed uniformly, and the resampling frequency is configured as follows: Obtain the standard one-dimensional time-domain sequence after resampling. Based on the set time span, the long-term audio recording is divided into 3-second independent underwater acoustic samples that do not overlap, which serve as the standard input unit for subsequent time-frequency analysis. Step 1.2 specifically involves configuring a three-way parallel sliding Hamming window operator to construct a time-frequency observation basis at the physical level that can complement transient features and fine line spectrum features; Time-domain window function of each branch Satisfy the following general expression: (1) In the formula, Indicates the first The length of the sliding window for each branch; the number of points for the Fast Fourier Transform. Window length and sliding step size Configure the scaling to be doubled, with the following settings: Enable high frequency resolution / low temporal resolution branching. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Balance the branches with moderate time-frequency resolution. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Enable low frequency resolution / high temporal resolution branching. The parameters are satisfied; , , The time-frequency resolution unit size corresponding to this branch is , ; Specifically, step 1.3 involves processing the standard one-dimensional underwater acoustic signal obtained in step 1.

1. The data is fed into the three short-time Fourier transform branches configured in step 1.2 in parallel. The one-dimensional time-domain signal is discretized and mapped onto the complex two-dimensional time-frequency space using the time-frequency integral transform operator. The calculation formula is shown in equation (2). (2) By parallel solving of equation (2), the corresponding single-branch two-dimensional complex time-frequency spectrum matrices are obtained respectively. , and ; Step 1.4 specifically involves obtaining the original matrices of each branch after transformation, with the following feature dimension sizes: high-frequency branch The original shape is Balanced branches The original shape is High-level branches The original shape is ; Dimensional interpolation adjustment is performed on the three complex spectrum matrices mentioned above: (3) (4) (5) In equations (3) to (5), The spatial interpolation operator outputs three sets of fully normalized, complementary time-frequency resolution, all-even-order complex time-frequency spectrum matrices. , as the input for the second stage of energy-phase fusion.

4. The underwater acoustic feature extraction method according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Amplitude-phase decoupling and decibel energy mapping of complex spectrum matrix; Step 2.2: Cosine periodic encoding of phase information and normalization of structural weights; Step 2.3: Amplitude-phase dynamic fusion based on residual multiplication modulation mechanism.

5. The underwater acoustic feature extraction method according to claim 4, characterized in that, Step 2.1 specifically involves, for each branch of the output from the first stage ( The complex spectrum matrix of ) First, complex decoupling is performed to separate the amplitude spectrum, which reflects the absolute intensity of the acoustic signal. and the phase spectrum reflecting the relative time delay of the physical structure Extract separately and ; For amplitude spectrum Using the global maximum amplitude as a reference, it is mapped to the interval [0, 1] to construct the energy mapping matrix. : (6) Step 2.2 specifically involves processing the phase spectrum after decoupling and cosine coding. After normalization, the phase mapping matrix mapped to the interval [0, 1] is obtained. : (7) Step 2.3 specifically involves dynamically modulating the energy using phase structure weights and calculating the energy-phase fusion feature matrix of a single branch. : (8) In equation (8), Represents the element-wise dot product operation of a matrix; In the high-energy region of the target signal Acting as activation weights, the final output of each branch It will serve as the direct input for the next stage of feature alignment and channel concatenation.

6. The underwater acoustic feature extraction method according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Learnable extension of the anisotropic dimension; Step 3.2: Local texture smoothing and checkerboard artifact removal; Step 3.3: High-dimensional semantic compression and physical plane reconstruction; Step 3.4: Cross-scale feature channel stitching and panoramic space construction.

7. The underwater acoustic feature extraction method according to claim 6, characterized in that, The specific expansion operation in step 3.1 is as follows; against Branch: Employs an asymmetric one-dimensional time transpose convolution kernel, upsampling only along the time dimension; Output features The calculation formula is: (9) against Branch: Employs an asymmetric one-dimensional frequency transpose convolution kernel, expanding only along the frequency dimension; Output features The calculation formula is: (10) against Branching: A symmetrical transposed convolution kernel is used to expand evenly along both the time and frequency axes; Output features The calculation formula is: (11) Specifically, step 3.2 involves appending a standard layer after the dimensionality expansion layer of each learnable upsampling module. Convolutional layers, cascaded with batch normalization and ReLU activation functions; The formula is as follows: (12) Step 3.3 specifically involves using a The convolution operator forcibly compresses high-dimensional features back into a single channel; The calculation formula is: (13) Step 3.4 specifically involves concatenating and stitching these single-channel feature maps along the feature channel dimension after the three branches are precisely aligned to a uniform size; the output after stitching is the multi-scale energy-phase fusion representation. The calculation formula is as follows: (14)。 8. The underwater acoustic feature extraction method according to claim 2, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Initial channel enhancement and cross-channel fusion; Step 4.2: Intra-channel time decoupling differential; Step 4.3: Interactive sharing of time-dynamic features; Step 4.4: Intra-channel frequency decoupling differential; Step 4.5: Combining channel compression with global residuals.

9. The underwater acoustic feature extraction method according to claim 8, characterized in that, Step 4.1 specifically involves using a standard two-dimensional convolution operator to stitch together the multi-scale panoramic features obtained from the stitching process. Mapping to a high-dimensional tensor space, the operation is defined as follows: (15) The kernel parameter is set to the kernel size. Step length ,filling Number of groups The output high-dimensional tensor is denoted as ; Specifically, step 4.2 involves... Each independent channel in the process employs an asymmetric one-dimensional convolution kernel that expands along the time dimension. The configuration parameters are First, calculate the standard convolutional response within the local receptive field. : (16) Define a set of learnable time difference weight parameters The forward, center, and backward differential responses are then normalized using the Softmax operator and adaptively fused. (17) The left, center, and right temporal tensor slices are extracted using a space-filling and translation slicing strategy. Constructing time-domain adaptive local background anchor points using normalized weights : (18) Subsequently, the total spatial response intensity of the convolution kernel is calculated. The anchor point is multiplied at the channel level to precisely construct a correction term for eliminating artifacts in stable environments; dynamic time-difference response. Equivalent to standard convolution minus the adaptive background correction term: (19) Introducing weighted parameters Controlled soft residual reconstruction mechanism Calculate and output time dynamic characteristics : (20) Specifically, step 4.3 involves inserting a standard two-dimensional convolution as a central node between the temporal difference and the subsequent frequency difference. (21) The convolution kernel parameters are configured as follows: ; Specifically, step 4.4 involves capturing the frequency domain dynamic evolution using a mechanism dual to step 4.2; and employing an asymmetric one-dimensional frequency-axis depth convolution kernel. The configuration parameters are ; Calculate the standard frequency response ; Similarly, frequency can be defined to learn differential weights. And normalized by Softmax to Extract the top, middle, and bottom tensor slices along the frequency axis to construct frequency background anchor points. : (22) The frequency differential output is obtained by subtracting the local adaptive environment correction term. This is then combined with frequency soft residual weights. Calculate the frequency dynamic features that both filter out static biases and retain the original features. : (23) (24) The above operator ultimately achieves the decoupled extraction of dynamic evolution modes along both time and frequency axes; Step 4.5 specifically involves compressing the high-dimensional two-dimensional difference tensor features back into a low-dimensional space using a standard convolution operator, with the parameters set as follows: ; (25) The compressed difference feature map is compared with the original feature map input to the module. Add: (26) Low-dimensional features of the output This provides a robust underlying physical representation for subsequent networks.

10. A domain-invariant underwater acoustic feature extraction system based on multi-scale energy phase fusion and adaptive differential calculus, characterized in that, The underwater acoustic feature extraction system uses a domain-invariant underwater acoustic feature extraction method based on multi-scale energy phase fusion and adaptive difference as described in any one of claims 1-9, and the underwater acoustic feature extraction system includes: Extraction module: Extracts underwater acoustic multi-time-frequency resolution spectra; Phase fusion module: performs multi-scale energy phase fusion on the extracted underwater acoustic multi-time-frequency resolution spectrum; Feature alignment and channel stitching module: performs learnable feature alignment and channel stitching on the extracted multi-branch energy-phase fusion feature matrix; Adaptive Time-Frequency Differential Reconstruction Module: Adaptive time-frequency differential reconstruction is performed on the stitched multi-scale energy-phase fusion representation to realize domain-invariant underwater acoustic feature extraction based on multi-scale energy-phase fusion and adaptive differential.

Images