An underwater ship radiated noise target recognition method

Abstract

The application provides an underwater ship radiation noise target recognition method, and belongs to the technical field of underwater acoustic signal processing. In view of the problems of mode fusion contradiction, poor robustness and weak physical interpretability of the existing method in the complex underwater acoustic environment of shallow sea, the application proposes a physical anchoring decoupling-recoupling double-track parallel network, constructs a three-branch feature extraction architecture of time domain, frequency domain and time-frequency domain, embeds an attention module with normal wave prior constraint to enhance the core feature, realizes the physical alignment and bidirectional interaction of modes through cyclic stationary guidance, completes the no-reference adaptive weighted fusion based on Bayesian inference, and finally outputs the recognition result. The application effectively alleviates the underlying contradiction between mode independence and cross-modal interaction, has excellent robustness in the low SNR and multi-path distortion scene, and the recognition accuracy is significantly better than that of the existing method. After lightening, it can be deployed to the underwater edge platform, has strong engineering practicability, and is suitable for various shallow underwater target recognition scenes.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic signal processing technology, and more specifically, to a method for identifying underwater ship radiated noise targets, applicable to the accurate identification of underwater targets in scenarios such as underwater security, shipping supervision, and marine environmental monitoring in complex shallow sea underwater acoustic environments. Background Technology

[0002] Sound waves are the only effective carrier for long-distance underwater information transmission. Underwater acoustic target identification technology based on ship radiated noise is a core supporting technology for underwater detection, marine security, and shipping supervision. Shallow coastal areas are the core application scenario, and the underwater acoustic environment has characteristics such as non-stationarity, multipath propagation, and strong noise interference. Ship radiated noise exhibits discrete line spectrum, continuous spectrum, and non-stationary pulse composite characteristics, which pose a great challenge to target identification.

[0003] Existing methods for identifying radiated noise from underwater ships can be mainly categorized into three types: traditional statistical signal processing methods, single-domain deep learning methods, and multimodal fusion methods. These methods face numerous technical bottlenecks in practical applications.

[0004] 1. Traditional methods rely on manual feature design and can only extract shallow statistical features. They have poor feature robustness in low signal-to-noise ratio and multipath distortion scenarios, lack fine-grained classification ability, and cannot adapt to complex shallow sea environments.

[0005] 2. Single-domain deep learning methods have obvious limitations in representation. They rely on only time-domain, frequency-domain, or time-frequency-domain features and cannot fully cover the complex physical characteristics of ship radiated noise. Their recognition performance degrades sharply in extreme scenarios. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing a method for identifying underwater ship radiated noise targets, solving the problems of low identification accuracy, poor robustness, and weak engineering adaptability in complex shallow sea environments.

[0007] This invention provides a method for identifying underwater ship radiated noise targets, comprising:

[0008] An underwater ship radiated noise target recognition network is constructed. The target recognition network includes a multi-domain feature mapping module, a three-branch multi-domain feature extraction module, a channel space joint attention module PCS-JAM with normal mode prior constraints, a physical feature track, an interactive enhancement track, a modal confidence dynamic adaptive fusion module MCDE-AFM based on Bayesian inference, and a classification head module.

[0009] The multi-domain feature mapping module performs multi-domain feature mapping on the original time-domain underwater acoustic signal, mapping the original time-domain underwater acoustic signal to the time-domain space, frequency-domain space, and time-frequency-domain space respectively.

[0010] The three-branch multi-domain feature extraction module extracts three single-mode clean features from the original time-domain underwater acoustic signal. The three single-mode clean features include time-domain clean features, frequency-domain clean features, and time-frequency-domain clean features.

[0011] The PCS-JAM performs directional enhancement on each single-modal pure feature in both the channel dimension and the dimension to obtain single-modal directional enhanced features, which serve as single-modal features of the physical feature track.

[0012] The physical feature track, based on each single-modal directional enhancement feature, performs a pre-confidence score evaluation on each single modality to obtain a pre-confidence score for each single modality;

[0013] The interactive enhancement track completes cross-modal bidirectional interaction of unimodal directional enhancement features based on the pre-confidence score of each unimodal mode, resulting in each unimodal interactive enhancement feature;

[0014] MCDE-AFM calculates the final confidence score and fusion weight of each single modality based on Bayesian inference, based on each single modality directional enhancement feature, and dynamically fuses the three single modality interactive enhancement features based on the fusion weight to obtain global discriminative features.

[0015] The classification head module identifies the target based on the global discriminative features.

[0016] This invention provides an underwater vessel radiated noise target recognition method. Addressing the problems of existing methods in complex shallow-sea underwater acoustic environments, such as modal fusion contradictions, poor robustness, and weak physical interpretability, this invention proposes a physically anchored decoupling-recoupling dual-track parallel network, MC-ACIFNet. It constructs a three-branch feature extraction architecture in the time domain, frequency domain, and time-frequency domain, embedding an attention module with normal mode prior constraints to enhance core features. Through cyclical smooth guidance, it achieves modal physical alignment and bidirectional interaction. Based on Bayesian inference, it completes referenceless adaptive weighted fusion, ultimately outputting the recognition result. This invention effectively alleviates the underlying contradiction between modal independence and cross-modal interaction, exhibits excellent robustness in low signal-to-noise ratio and multipath distortion scenarios, and significantly outperforms existing methods in recognition accuracy. Furthermore, its lightweight design allows for deployment on underwater edge platforms, demonstrating strong engineering practicality and applicability to various shallow-sea underwater target recognition scenarios. Attached Figure Description

[0017] Figure 1 A flowchart of a target recognition method based on a ship radiated noise target recognition network provided by the present invention;

[0018] Figure 2 Here is a flowchart of the temporal feature extraction process;

[0019] Figure 3 Here is a flowchart of the frequency domain feature extraction process;

[0020] Figure 4 Here is a flowchart of the time-frequency domain feature extraction process;

[0021] Figure 5 Flowchart for directional enhancement of single-modal features;

[0022] Figure 6 This is a flowchart illustrating the entire working process of the PCS-JAM module.

[0023] Figure 7 Flowchart of the complete calculation process for channel attention branches;

[0024] Figure 8 Workflow diagram for spatial attention branch;

[0025] Figure 9 Workflow diagram for the cross-modal bidirectional cross-attention interaction module CBC-IFM;

[0026] Figure 10 A flowchart for three-modal physical alignment;

[0027] Figure 11 Here is a flowchart of the bidirectional cross-attention calculation process;

[0028] Figure 12 This is a complete forward flow diagram of the three-modal interaction;

[0029] Figure 13 This is a flowchart of the no-reference modality confidence assessment process within the Bayesian framework. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined with each other to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0031] To achieve target identification of underwater vessel radiated noise, this invention designs a physically anchored, decoupled-recoupled, dual-track parallel adaptive cross-modal interactive fusion network (MC-ACIFNet). The overall architecture of the network, the design ideas of each core module, mathematical derivations, and parameter settings are described in detail below, forming an end-to-end underwater vessel radiated noise target identification network. (See also...) Figure 1 The target recognition network includes a multi-domain feature mapping module, a three-branch multi-domain feature extraction module, a channel space joint attention module PCS-JAM with normal mode prior constraints, a physical feature track, an interaction enhancement track, a modal confidence dynamic adaptive fusion module MCDE-AFM based on Bayesian inference, and a classification head module.

[0032] In the process of target identification using the underwater ship radiated noise target identification network, the parameter settings of each module of the MC-ACIFNet network are shown in Table 1 below.

[0033] Table 1 Parameter Settings for Each Module of the MC-ACIFNet Network

[0034]

[0035] See the flowchart for the entire target identification process using the underwater vessel radiated noise target identification network. Figure 1 The underwater ship radiated noise target identification method includes the following steps:

[0036] Step 1: The multi-domain feature mapping module performs multi-domain feature mapping on the original time-domain underwater acoustic signal, mapping the original time-domain underwater acoustic signal to the time domain space, frequency domain space, and time-frequency domain space respectively.

[0037] Understandably, the input is a normalized time-domain underwater acoustic signal with a length of 5 seconds. The sampling frequency is 16 kHz, and the input dimension is... ,in The batch size is specified. The original time-domain signal is preprocessed in the time domain, and MFCC+GFCC frequency domain features are extracted and STFT time-spectrum transform is performed to obtain the input features in the time domain, frequency domain, and time-frequency domain. The three modal features remain independent throughout the process.

[0038] Step 2: The three-branch multi-domain feature extraction module extracts three single-mode clean features from the original time-domain underwater acoustic signal. The three single-mode clean features include time-domain clean features, frequency-domain clean features, and time-frequency-domain clean features.

[0039] Understandably, in response to the bottleneck of incomplete multi-domain feature coverage in existing methods, this invention designs a three-branch multi-domain feature extraction architecture based on the three core physical components of ship radiated noise, namely time domain, frequency domain, and time-frequency domain. It maintains the physical independence of each mode throughout the process, comprehensively covers the three core physical components of ship radiated noise: time-series pulse, discrete line spectrum, and time-varying cavitation. It fully combines the noise resistance and discriminative advantages of different modal features, making up for the limitations of single-domain feature representation.

[0040] For details, see the temporal feature extraction branch. Figure 2 The time-domain waveform directly contains the non-stationary pulse components and time-series dependent characteristics of ship radiated noise, which are important feature sources for ship target identification. This invention uses one-dimensional convolution + two-layer bidirectional gated recurrent units (BiGRU) to construct the time-domain feature extraction branch, while capturing the local pulse features and long-distance time-series dependent characteristics of the time-series signal, adapting to the non-stationary characteristics of the time-domain signal.

[0041] like Figure 2 As shown, the complete forward propagation process of the time-domain branch is as follows:

[0042] (1) Input layer: Input a normalized time-domain waveform with a length of 5 seconds, with dimension . .

[0043] (2) Primary Convolutional Dimensionality Reduction Layer: Shallow feature extraction and sequence dimensionality reduction are performed through a one-dimensional convolutional layer. The kernel size is 7, the stride is 522, and the number of output channels is 64. Batch normalization and ReLU activation are used to reduce the sequence length from 80,000 to 153, aligning it with the frame length of the frequency domain branch to ensure temporal dimension matching for subsequent cross-modal interactions. The output dimension is... .

[0044] (3) PCS-JAM Physical Enhancement Layer: The primary convolutional features are input into the PCS-JAM module, and the temporal pulse peak prior is embedded to enhance the pulse feature region and suppress background noise, so as to obtain the temporal clean features after physical enhancement.

[0045] (4) Dimension Transformation Layer: The channel dimension and sequence dimension of the physically enhanced temporal clean features are interchanged to obtain a dimension of The time-series feature sequences are adapted to the input format requirements of BiGRU.

[0046] (5) Two-layer BiGRU feature extraction layer: The temporal feature sequence is input into a two-layer BiGRU network. The dimension of each hidden layer is set to 256. After bidirectional concatenation, the output dimension is 512. The dropout rate is set to 0.3 to avoid overfitting. The final output dimension is... The time-domain feature sequence completely preserves the temporal information of the signal.

[0047] (6) Global pooling branch: Perform global average pooling on the temporal feature sequence to obtain the dimension of The temporal global purity features are used for subsequent modal confidence assessment.

[0048] Frequency domain features, containing discrete line spectrum components of ship radiated noise, are the most reliable discriminative features in scenarios with low signal-to-noise ratios and multipath distortion, and are also a core guarantee of model robustness. This invention employs an improved one-dimensional ResNet18 architecture to construct the frequency domain feature extraction branch, adapting and adjusting the original ResNet18 for one-dimensional frequency domain features. Simultaneously, physical prior constraints are embedded in the primary feature layer, focusing on enhancing the 0-2 kHz low-frequency line spectrum features. The frequency domain feature extraction branch can be found in [link to documentation]. Figure 3 The complete forward propagation process of the frequency domain branch is as follows:

[0049] (1) Frequency domain feature extraction: 12-dimensional MFCC static coefficients, 12-dimensional GFCC static coefficients, and their corresponding first-order and second-order difference coefficients are extracted from the original time-domain signal to obtain 36-dimensional frequency domain features. The input dimension is... , where 36 is the feature dimension and 153 is the time frame length.

[0050] (2) Primary Convolutional Layer: Shallow feature extraction is performed through a one-dimensional convolutional layer with a kernel size of 7×1, a stride of 1, 64 output channels, batch normalization + ReLU activation, and an output dimension of .

[0051] (3) PCS-JAM physical enhancement layer: The primary convolutional features are input into the PCS-JAM module, and the low-frequency line spectrum low kurtosis prior and normal mode propagation prior are embedded. Adaptive physical constraints are applied to the low-frequency channels corresponding to 0-2 kHz, focusing on the core line spectrum features, suppressing high-frequency noise, and obtaining the frequency domain pure features after physical enhancement.

[0052] (4) Improved 1D-ResNet18 deep feature extraction: The enhanced frequency domain features are input into the improved one-dimensional ResNet18 architecture, retaining the original 8 residual blocks of ResNet18, divided into 4 stages. By adjusting the stride and padding of the convolutional layers, the length of the feature time frame sequence is always kept at 153 to avoid loss of temporal information. The final output dimension is The frequency domain feature sequence.

[0053] (5) Global pooling branch: Perform global average pooling on the frequency domain feature sequence to obtain the dimension of The frequency domain global purity features are used for subsequent modal confidence assessment.

[0054] The time-frequency spectrogram simultaneously contains the joint time-domain and frequency-domain distribution information of ship radiated noise, effectively characterizing time-varying cavitation components and serving as a core feature source for fine-grained classification in high signal-to-noise ratio scenarios. This invention employs an improved Swin-Tiny architecture to construct a time-frequency domain feature extraction branch, adapting it to the characteristics of the time-frequency spectrogram while embedding physical prior constraints to enhance the peak intensity features of cavitation pulses. The time-frequency domain feature extraction branch can be found in [link to relevant documentation]. Figure 4 The complete forward propagation process of the time-frequency domain branch is as follows:

[0055] (1) Time-domain spectrum generation: Perform STFT transformation on the original time-domain signal to obtain the spectrum with dimension 1. The time-frequency spectrogram was adjusted to a size of 224×224 using bilinear interpolation and converted into a 3-channel RGB image with an input dimension of [missing value]. .

[0056] (2) Primary patch embedding layer: Through the patch embedding layer, using a 4×4 patch size, the time-spectrum graph is divided into 56×56 patches, with the embedding dimension set to 96 and the output dimension being... The patch sequence.

[0057] (3) PCS-JAM physical enhancement layer: Dimensional reconstruction is performed on the embedded features of each patch, input into the PCS-JAM module, cavitation pulse peak prior is embedded, the cavitation feature region above 2 kHz is enhanced, and meaningless background noise region is suppressed to obtain the time-frequency domain pure features after physical enhancement of each patch.

[0058] (4) Improved Swin-Tiny deep feature extraction: The enhanced patch sequence is input into the improved Swin-Tiny architecture, retaining the original Swin-Tiny's 4-stage structure. Multi-scale features are constructed through patch merging operations, and global feature extraction is achieved using window self-attention and shift window mechanisms. The final output feature map size is 7×7, which flattens into a sequence of 49 patches, with an output dimension of... The time-frequency domain feature sequence.

[0059] (5) Temporal alignment interpolation: Based on the ship's cyclic stationary fundamental frequency, cubic spline interpolation is performed on the time-frequency domain feature sequence to align the sequence length from 49 to 153, matching the temporal dimension of the time and frequency domain branches, ensuring the physical consistency of subsequent cross-modal interactions.

[0060] (6) Global pooling branch: Perform global average pooling on the aligned time-frequency domain feature sequence to obtain the dimension of The time-frequency domain global purity features are used for subsequent modal confidence assessment.

[0061] The three branches feature clear functional divisions and strong complementarity, comprehensively covering the three core physical components of ship radiated noise and achieving complementary advantages in different scenarios. The time-domain branch focuses on capturing non-stationary impulse components, offering balanced reliability in medium signal-to-noise ratio scenarios and supplementing the temporal details lost by the frequency-domain and time-frequency-domain branches. The frequency-domain branch focuses on capturing discrete line spectrum components, exhibiting the strongest noise resistance in low signal-to-noise ratio and multipath distortion scenarios, serving as the core guarantee for model robustness. The time-frequency-domain branch focuses on capturing time-varying cavitation components, demonstrating the strongest discriminative power in high signal-to-noise ratio scenarios, and serving as the core support for the model's fine-grained classification capabilities.

[0062] Step 3: PCS-JAM performs directional enhancement on each monomodal clean feature in both channel and dimension to obtain monomodal directional enhanced features, which serve as monomodal features of the physical feature track.

[0063] Understandably, to address the bottlenecks of insufficient physical prior embedding and strong black-box characteristics of existing methods, this invention designs a channel-space joint attention module (PCS-JAM) with normal mode prior constraints. It simultaneously embeds statistical priors of ship noise and shallow sea channel propagation priors at the primary feature layer. It replaces the traditional pooling method with unbiased sliding window kurtosis pooling with degree-of-freedom correction, and combines it with gated residual connections to achieve targeted enhancement of core discriminative features. This avoids the cancellation of enhancement effects by traditional residual structures and improves the physical interpretability and generalization ability of the model.

[0064] Combination Figure 5 and Figure 6 As shown, the entire working process of the PCS-JAM module is described in detail.

[0065] Traditional average pooling smooths out peak features of the line spectrum, while max pooling is susceptible to random noise spikes and cannot accurately distinguish between effective ship features and noise. Furthermore, traditional kurtosis estimation suffers from severe statistical bias under small sample windows. This invention replaces traditional pooling methods with unbiased sliding window kurtosis pooling with degree-of-freedom correction, accurately capturing the spatial distribution of two core discriminative features based on kurtosis statistical characteristics.

[0066] The principle of unbiased sliding window kurtosis pooling is that, for the input primary feature map... ,in The number of feature channels, Given the feature sequence length, the formula for calculating the unbiased kurtosis pooling of the sliding window is:

[0067] ;

[0068] ;

[0069] in, For the first Channel, First Subsequences within a sliding window The window length is set to 15 in this invention (to match the STFT frame length and ensure the validity of physical kurtosis statistics), and the step size is 1. This represents the mean of the subsequences within the window. The preceding term is the degrees of freedom correction coefficient, used to correct for kurtosis estimation bias under small sample conditions, ensuring the accuracy of the statistical results.

[0070] This invention uses unbiased sliding window kurtosis pooling to obtain an absolute value kurtosis feature map for each single-modal pure feature extracted by the three-branch multi-domain feature extraction module. With signed kurtosis feature map The absolute value kurtosis feature map is used to characterize the non-Gaussianity of the feature, providing a physical basis for channel attention; the signed kurtosis feature map is used to distinguish between low-kurtosis line spectra and high-kurtosis impulses, providing a physical basis for spatial attention. Additionally, a kurtosis confidence interval check is added; only when the kurtosis value exceeds the 95% confidence interval is it considered a valid ship feature, avoiding misjudgments caused by noise spikes.

[0071] Based on absolute value kurtosis feature map Channel dimension enhancement is performed on unimodal pure features based on signed kurtosis feature maps. Spatial dimension enhancement is applied to the clean features of a single mode. Based on this, the present invention designs channel attention branches and spatial attention branches. Differentiated channel constraints are designed for different feature branches, namely, soft constraints based on channel reliability priors and adaptive signal-to-noise ratio (SNR) according to the normal mode propagation law, to avoid misjudgment by hard-coded bias in low SNR and multipath distortion scenarios.

[0072] This section first presents the prior calibration parameters and procedures for channel reliability, and then provides the complete calculation process for this branch.

[0073] 1. Channel reliability prior calibration parameters and procedures:

[0074] (1) Calibrate the physical parameters of the scene.

[0075] Average water depth m, the sea surface is an absolutely soft boundary, and the seabed is composed of silt (density). g / cm³, speed of sound m / s, attenuation coefficient The sound velocity profile is a typical summer negative gradient sound velocity profile (sea surface sound velocity 1520 m / s, seabed sound velocity 1505 m / s), the sound source depth is 10 m, and the hydrophone depth is 5 m.

[0076] (2) Solving for normal modes and calculating attenuation coefficient: Based on the Kraken normal mode model, the propagation attenuation coefficient of each frequency point in the 0-8 kHz frequency band is solved, with the frequency interval set to 10 Hz, to obtain the propagation attenuation of each frequency point at a propagation distance of 10 km. .

[0077] (3) Channel reliability weight mapping: The attenuation coefficient is normalized by min-max and mapped to the interval [0.1, 1.0] to obtain the prior weight of channel reliability:

[0078] ;

[0079] in, This represents the maximum attenuation at a frequency of 8 kHz. This represents the minimum attenuation at a frequency of 100 Hz. The resulting weighting curve satisfies the following: the weight of the low-frequency channel (0-2 kHz) increases monotonically as the frequency decreases, while the weight of the high-frequency channel (above 2 kHz) decreases monotonically as the frequency increases. This conforms to the physical law that low-frequency sound wave propagation attenuation is small and anti-multipath capability is strong in shallow sea waveguides.

[0080] (4) Online adaptation logic: During the model deployment phase, the sound velocity profile and waveguide parameters can be updated through measured CTD data, the normal mode attenuation coefficient can be re-solved, and the channel reliability weight can be updated online without retraining the model.

[0081] 2. For the complete calculation process of channel attention branches, please refer to [link / reference]. Figure 7 Its working process is as follows:

[0082] (1) Absolute value kurtosis feature map Perform global average pooling and global max pooling respectively to obtain feature vectors with two channel dimensions, each having dimension 1. ;

[0083] (2) Two feature vectors are input into a shared multilayer perceptron (MLP). The MLP contains two fully connected layers, with the intermediate layer having a dimension of . (Compression ratio 16), ReLU activation, to obtain channel attention feature vectors. , dimension ;

[0084] (3) Constructing the channel reliability weight curve of normal mode prior: Based on the normal mode model, the propagation attenuation coefficient of different frequency channels in a typical shallow sea waveguide is calibrated in advance to obtain the channel reliability prior weight. The weights of the 0-2 kHz low-frequency channels monotonically increase as the frequency decreases, while the weights of the high-frequency channels above 2 kHz monotonically decrease as the frequency increases. The weight range is constrained within... ;

[0085] (4) Calculate the signal-to-noise ratio confidence factor The spectral flatness of the input signal is calculated in real time, with a value range of [value range missing]. At high signal-to-noise ratios, it approaches 1; at low signal-to-noise ratios, it automatically decreases. The formula is:

[0086] ;

[0087] in This represents the spectral flatness of the input signal. The average spectral flatness of environmental noise. This is the Sigmoid activation function.

[0088] (5) Channel attention feature vector By applying normal mode prior constraints, the final channel attention weights are obtained. Channel dimension enhancement is performed on the original input feature map to obtain the channel-enhanced feature map. :

[0089] ;

[0090] ;

[0091] in It is the Sigmoid activation function. This is for the Hadamard product operation.

[0092] The spatial attention branch is based on signed kurtosis feature maps, simultaneously focusing on low-kurtosis line spectrum regions and high-kurtosis pulse regions, while suppressing noise regions where the absolute kurtosis value is close to 0. For example... Figure 8 As shown, the workflow of the spatial attention branch is as follows:

[0093] Feature map after channel enhancement Global average pooling and global max pooling are performed along the channel dimension to obtain feature maps in two spatial dimensions, both of which are... Combined with the absolute value of the signed kurtosis feature map The three feature maps are concatenated along the channel dimension to obtain a dimension of The splicing feature map; through A one-dimensional convolutional layer maps the concatenated features to a single-channel spatial attention feature map, and the final spatial attention weights are obtained through a sigmoid activation function. , dimension Feature maps after channel enhancement Spatial dimension enhancement is performed to obtain spatially enhanced feature maps. :

[0094] ;

[0095] To avoid the cancellation of physical enhancement effects by traditional residual connections, this invention designs a physically-driven gated residual fusion, which avoids gradient vanishing without destroying the physical enhancement effect. This enhances the features of the channel. and spatially enhanced features Perform gated residual fusion:

[0096] ;

[0097] in The learnable gating coefficients are driven in real time by the spectral significance of the input signal, under high signal-to-noise ratio conditions. Enhance physical enhancement effect; under low signal-to-noise ratio Automatically reduces robustness while preserving original features.

[0098] Layer normalization is performed on the final output to stabilize the training process:

[0099] ;

[0100] in, For layer normalization, the output This is known as targeted enhancement feature.

[0101] Step 4: Physical Feature Track: Based on each single-modal directional augmentation feature, perform a pre-confidence assessment for each single modality to obtain a pre-confidence score for each single modality; Interactive Augmentation Track: Based on the pre-confidence score of each single modality, complete the cross-modal bidirectional interaction of the single-modal directional augmentation features to obtain each single-modal interactive augmentation feature.

[0102] Understandably, to avoid the core problems of existing cross-modal interaction methods, such as the lack of physical anchors for modal alignment, the destruction of single-modal independence by full cross-fusion, and the contamination of high-reliability features by low-confidence modal noise diffusion, this invention designs a Cyclostationary-guided Bidirectional Cross-attention Interaction Fusion Module (CBC-IFM). This module uses the inherent cyclostationary characteristics of ship radiated noise as an anchor to achieve physical alignment between modalities. It employs a pre-confidence-guided complementary information injection mechanism to complete the interaction of deep complementary information between modalities while maximizing the preservation of single-modal physical independence, effectively mitigating the noise diffusion and feature homogenization problems of traditional cross-attention mechanisms. The overall structure of the module is as follows: Figure 9 As shown.

[0103] This invention designs a dual-track branch: a physical feature track and an interactive enhancement track. The physical feature track is mainly based on the directional enhancement features of each single mode, and performs a pre-confidence assessment on each single mode to obtain the pre-confidence of each single mode.

[0104] The interactive enhancement track physically aligns multiple modal features, and then performs cross-modal bidirectional interaction of the single-modal directional enhancement features based on the pre-confidence of each single modality, resulting in each single-modal interactive enhancement feature.

[0105] Existing cross-modal alignment methods mostly employ purely data-driven approaches such as linear interpolation and dimension mapping, achieving only numerical dimension matching without considering the physical periodicity of underwater acoustic signals. In multipath distortion scenarios, the feature periods of different modes shift, and linear interpolation completely destroys the physical correspondence of features, leading to complete failure of cross-modal interaction.

[0106] This invention, based on the cyclic stationary characteristics of ship radiated noise, constructs a two-stage joint alignment mechanism for fundamental frequency anchoring, achieving precise physical alignment between modes, such as... Figure 10 As shown, the complete process is as follows:

[0107] (1) Fundamental frequency synchronization detection: Based on the cyclic autocorrelation function defined in Chapter 2, the cyclic frequency of the primary feature sequence of the three modes is detected to extract the fundamental frequency cyclic frequency of the ship propeller. Using the high-reliability fundamental frequency detected by frequency domain modes as a unified anchor point, the frequency domain shift of time-domain and time-frequency-domain modes is corrected, thus solving the line spectrum frequency shift problem caused by multipath propagation.

[0108] (2) Periodic segment alignment: based on the period corresponding to the anchor fundamental frequency Based on this, the feature sequences of the three modalities are segmented at the periodic level. Linear interpolation is used to align the length of the feature segments within each period, rather than performing fixed-length indiscriminate interpolation on the entire sequence. This method ensures that the physical meaning of the features within each period is completely consistent, avoids false information caused by meaningless interpolation, and preserves the temporal periodic structure of the features.

[0109] (3) Temporal dimension unification: The length of the three modal feature sequences after alignment is unified to 153, which is consistent with the original frame length of the temporal and frequency domain branches, providing dimension matching input for subsequent cross-modal attention calculation, while ensuring that the temporal physical meaning of the features is not destroyed.

[0110] The algorithm for detecting the fundamental frequency of ship propellers is as follows.

[0111] (1) Calculation of cyclic autocorrelation function: For the input time-domain signal Calculate its cyclic autocorrelation function:

[0112] ;

[0113] in, The cycle frequency, For time delay, the FFT fast algorithm is used to calculate the cyclic autocorrelation spectrum for discrete signals. The time delay range is set to [-1024, 1024], and the cyclic frequency search range is set to 5-120 Hz, covering the typical range of the fundamental frequency of civilian ship propellers.

[0114] (2) Low signal-to-noise ratio robust processing.

[0115] (3) The cyclic autocorrelation spectrum is smoothed by a 3-point moving average to suppress false peaks caused by random noise;

[0116] (4) The CFAR (Constant False Alarm Rate) detection algorithm is adopted, and the false alarm rate is set. Only when the peak value exceeds the detection threshold is it determined to be a valid cycle frequency peak value, thus avoiding false alarms under low signal-to-noise ratio.

[0117] (5) Perform line spectrum detection on the power spectrum of the frequency domain branch to obtain the harmonic frequency sequence. Perform harmonic matching verification on the peak value of the cyclic frequency. Only when the integer harmonic of the cyclic frequency matches the line spectrum of the power spectrum is it determined to be the effective fundamental frequency.

[0118] (6) Multipath frequency shift correction: Using the high-reliability fundamental frequency detected by the frequency domain branch as a unified anchor point, the fundamental frequencies detected by the time domain and time-frequency domain branches are corrected. The correction formula is as follows:

[0119] ;

[0120] in, These are the fundamental frequencies detected in the time domain, frequency domain, and time-frequency domain branches, respectively. The harmonic matching degree is scored, and the fundamental frequency with the highest matching degree is finally selected as the global anchoring fundamental frequency to alleviate the line spectrum frequency shift problem caused by multipath effect.

[0121] (7) Outlier fallback handling: When no effective fundamental frequency is detected in any of the three branches (extremely low signal-to-noise ratio scenario), the a priori mean of the ship's fundamental frequency of 15 Hz is used as the fallback anchoring fundamental frequency to ensure that the modal alignment process is not interrupted. At the same time, the gating coefficient of cross-modal interaction in this scenario is reduced to avoid feature pollution caused by incorrect alignment.

[0122] Traditional full cross-attention mechanisms indiscriminately interact with all features of two modalities, which can easily lead to noise from the low-confidence modality spreading to the high-confidence modality and undermining the physical independence of individual modalities. This invention, based on modality pre-confidence scoring, designs a pairwise bidirectional cross-attention mechanism with directional injection of complementary information. This mechanism only allows the high-confidence modality to inject complementary information into the low-confidence modality, and only interacts with residual complementary features, preserving the core structure of the original modality.

[0123] For two single-modal targeted augmentation feature sequences to be interacted with , ,in For sequence length, For the feature embedding dimension, the corresponding pre-confidence score is: , ,like Figure 11 As shown, the improved bidirectional cross-attention calculation process is as follows:

[0124] Complementary feature extraction: Self-attention encoding is performed on the feature sequences of the two modalities respectively to obtain self-enhancing features. , Calculate the complementary characteristics of the residuals:

[0125] ;

[0126] in , It is a residual complementary feature of two modalities, which only contains the incremental information after the current modality is self-enhanced, thus avoiding the destruction of the original feature structure caused by full feature interaction.

[0127] Confidence-guided interaction mask: An interaction mask is set based on a pre-confidence score, allowing complementary information injection only when the source modality's confidence is higher than the target modality's.

[0128] ;

[0129] in This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. This masking mechanism ensures that information is transmitted only from the high-reliability mode to the low-reliability mode, thus avoiding noise propagation.

[0130] Bidirectional cross-attention calculation: Based on scaled dot product attention, cross-modal interaction is performed only on complementary features. The calculation process is as follows:

[0131] ;

[0132] ;

[0133] in, , and The weight matrix represents the learnable linear mapping of mode A. , and The weight matrix represents the learnable linear mapping for mode B. For the embedded dimension.

[0134] Feature Enhancement Update: Based on interactive masks, the original features are incrementally updated to ensure that the core structure of the original features is not destroyed.

[0135] ;

[0136] ;

[0137] This mechanism only incrementally injects complementary features instead of replacing all features, thus maximizing the physical independence of each modality. At the same time, it avoids noise propagation through confidence masking, solving the core defects of traditional cross-attention mechanisms.

[0138] The complete forward flow of trimodal interaction is as follows: Figure 12 As shown, this section constructs a pairwise bidirectional trimodal interaction architecture for the three modalities of time domain, frequency domain, and time-frequency domain. The complete forward propagation process is as follows:

[0139] (1) Input preprocessing: Input physical feature track output three single-mode directional enhancement feature sequences , , and pre-confidence score .

[0140] (2) Modal physical alignment: Based on the cyclic stationary fundamental frequency anchoring mechanism, the feature sequences of the three modes are physically aligned at the periodic level to ensure the physical meaning matching of the time dimension.

[0141] (3) Feature embedding mapping: Through a shared linear mapping layer, the features of the three modalities are uniformly mapped to a 128-dimensional embedding space, reducing computational complexity while ensuring the consistency of feature spaces among modalities.

[0142] (4) Two-way cross interaction: Construct three sets of two-way cross attention units to realize the complementary information direction injection of time domain-frequency domain, frequency domain-time-frequency domain, and time domain-time-frequency domain respectively. Each set of units adopts the cross attention mechanism guided by the confidence level mentioned above. The number of attention heads is set to 4 and the dropout rate is set to 0.2 to avoid overfitting.

[0143] (5) Feature aggregation and output: Gated aggregation is performed on the original features of each modality and the enhanced features obtained from the three sets of interactions to obtain the final single-modality interaction enhanced features. , , The three features maintain independent outputs with dimensions completely consistent with the inputs, without any cross-modal splicing or mixing, providing independent single-modal interactive enhancement features for subsequent adaptive fusion.

[0144] Step 5, MCDE-AFM: Based on each single-modal directional enhancement feature, calculate the final confidence score and fusion weight of each single-modal feature based on Bayesian inference. Based on the fusion weight, dynamically fuse the three single-modal interactive enhancement features to obtain the global discriminative features.

[0145] Understandably, such as Figure 13 As shown, this invention designs a no-reference modality confidence assessment under a Bayesian framework. Based on the Bayesian inference framework, physical statistical features are used as prior distributions and deep feature quality indices are used as likelihood functions to output the posterior probability distribution of modality confidence, replacing the traditional black-box MLP regression, so that the confidence assessment has clear probabilistic statistical meaning and physical interpretability throughout the process.

[0146] For the Modality The complete confidence assessment process is as follows:

[0147] (1) Prior distribution construction: Four types of no-reference physical statistical features were selected to construct the prior distribution, namely the significance of line spectral clusters. Periodic stability Normalized excess kurtosis Spectral flatness All four feature classes were normalized to the [0,1] interval using min-max normalization, and the prior probabilities were fitted using a Beta distribution. The two shape parameters of the Beta distribution are obtained by mapping the standardized values ​​of the physical characteristics:

[0148]

[0149]

[0150] in The shape parameter of the Beta distribution is given. The prior distribution is constructed entirely based on physical features, has no training parameters, and has a clear physical meaning.

[0151] (2) Likelihood function modeling: Four types of deep feature quality indicators were selected as observation values, namely intra-class compactness Inter-class separation Feature signal-to-noise ratio Feature entropy Construct the Gaussian likelihood function Characterizes at a given confidence level The probability of observing the corresponding deep feature quality is as follows:

[0152] ;

[0153] in The standardized mean values ​​of the four types of deep feature quality indicators are: The mean, To fix the variance, we need to ensure a positive correlation between the likelihood function and the confidence level.

[0154] Likelihood function with constant variance The calibration is based on the statistical results of deep feature quality indicators of the entire test set. The calibration process is as follows: (a) For all samples in the ShipsEar test set, calculate four types of deep feature quality indicators (intra-class compactness, inter-class separation, feature signal-to-noise ratio, and feature entropy), and perform min-max normalization to the [0,1] interval to obtain the standardized mean. (b) Statistics The Pearson correlation coefficient between the accuracy of single-modal recognition and the accuracy of single-modal recognition showed a strong positive correlation of 0.92, which meets the modeling premise of the likelihood function. (c) Statistics The standard deviation on the entire test set is 0.097. We take an approximate value of 0.1 as the fixed variance of the likelihood function to ensure that the probability distribution of the likelihood function matches the statistical characteristics of the measured data perfectly. This avoids the problems of excessively large variance leading to an unconstrained likelihood function and excessively small variance leading to oversensitivity of the posterior estimate.

[0155] (3) Posterior confidence inference: Based on Bayes' theorem, the posterior probability distribution of modal confidence is calculated, and the posterior expectation is taken as the final confidence score:

[0156] ;

[0157] ;

[0158] in The confidence score is the standardized mean of the deep feature quality index. A higher value indicates a stronger reliability of the mode.

[0159] This evaluation mechanism does not require external signal-to-noise ratio reference information. It can complete the confidence assessment based solely on the physical and deep features of the single mode itself, making it suitable for referenceless recognition scenarios in real marine environments. At the same time, it has clear physical and probabilistic statistical significance throughout the process, solving the black box problem of traditional methods.

[0160] Based on the final confidence score of each modality, the fusion weight of each modality is adaptively calculated. However, the traditional Softmax weighted fusion uses a fixed temperature coefficient, which cannot adapt to the dynamic changes in modality confidence under different environments: when the confidence difference is large, the fixed temperature can easily lead to an overly balanced weight distribution, which cannot strengthen the contribution of highly reliable modes; when the confidence difference is small, the fixed temperature can easily lead to an overly extreme weight distribution, which can lose the complementary information between modes.

[0161] This invention designs an adaptive temperature coefficient based on confidence dispersion, and combines it with weighted hard constraints to achieve dynamic optimization of the fusion weights. The complete process is as follows:

[0162] Adaptive temperature coefficient calculation: Based on the coefficient of variation (dispersion) of three modal confidence levels, dynamically adjust the temperature coefficient of Softmax. :

[0163] ;

[0164] ;

[0165] in, As the reference temperature coefficient, This is the scaling factor. This is the coefficient of variation for the confidence level. When the confidence level dispersion is large, Increase, temperature coefficient Decreasing the confidence level makes the weight distribution of the Softmax output steeper, strengthening the contribution of high-confidence modes; when the confidence level dispersion is small, Decrease, temperature coefficient The weight distribution is smoother, and the complementary information of each mode is preserved in a balanced manner.

[0166] Fusion weight calculation: Initial fusion weights are calculated using the Softmax function based on an adaptive temperature coefficient.

[0167] ;

[0168] Hard weight constraint: To prevent the weight of a certain modality from approaching 0 in extreme scenarios and losing complementary information, a hard weight constraint is set to limit the weight of each modality to a certain value. Within the range, the present invention sets , The constrained weights are then renormalized to ensure that the sum of the weights is 1.

[0169] ;

[0170] Adaptive weighted fusion: Using the fusion weights calculated from the physical feature tracks, the single-modal enhancement features output by the interactive enhancement tracks are weighted and fused to obtain global discriminative features.

[0171] ;

[0172] in, The single-modal interaction enhancement features output by the CBC-IFM module, after global average pooling, have all dimensions. The fusion of global feature dimensions is maintained. , input the subsequent classification header module.

[0173] This fusion mechanism strictly follows the logic of physical feature track confidence assessment and interactive enhancement track feature weighting, ensuring that the physical anchor point of the fusion weight is traceable throughout the process. At the same time, through adaptive temperature coefficient and weight hard constraint, it achieves optimized fusion in dynamic underwater acoustic environment, solving the core defects of traditional static fusion strategy.

[0174] Step 6: The classification head module identifies the target based on the global discriminative features.

[0175] Understandably, global discriminative features are used. Input a classification header, output the category probability distribution, and complete end-to-end target recognition.

[0176] For underwater vessel radiated noise target recognition networks, this invention addresses common problems in underwater acoustic target recognition datasets such as imbalanced samples, large intra-class variability, overlapping inter-class features, and lack of modal decoupling constraints. It designs a multi-task joint loss function comprising four parts: cross-entropy loss, center loss, contrastive loss, and mutual information regularization. This function improves feature discriminativity while maintaining classification accuracy and constrains the independence of single-modal features. The complete loss function definition is as follows:

[0177] ;

[0178] in The weighting coefficients were set to 0.005, 0.1, and 0.01 respectively after cross-validation to balance the contributions of each loss term. The validation process and the logic for selecting the optimal value are as follows:

[0179] (1) Parameter search space: Based on general experience in the field of deep learning, the parameter search range is set as follows: , , A grid search method is used to traverse all parameter combinations.

[0180] (2) Validation and evaluation indicators: OA in the 0 dB low signal-to-noise ratio scenario is the core optimization indicator, while constraining the OA loss in the normal scenario to not exceed 0.5 percentage points, giving priority to ensuring the robustness of the model in complex scenarios and meeting the needs of engineering applications.

[0181] (3) Determination of the optimal value: Cross-validation results show that when , , When the model achieves an optimal OA of 95.25% in the 0dB scenario, it maintains an OA of 98.85% in the normal scenario, thus achieving the optimal balance between classification accuracy, feature discriminativeness, and modal decoupling constraints.

[0182] In conclusion, Balance the center loss and cross-entropy loss to avoid the center loss causing model training instability; To ensure that the contrastive loss optimizes the separation of features between classes without compromising the physical distribution characteristics of the features; The constraint effect of achieving modal decoupling is achieved, ensuring that the mutual information of the three modal pure features is below the independence threshold of 0.1, while not affecting the discriminativeness of the single modal features.

[0183] The specific definitions of each loss term in the multi-task joint loss function are as follows:

[0184] Weighted cross-entropy loss: To address the problem of imbalanced datasets, a class-weighted cross-entropy loss is used, which assigns higher weights to the class with fewer samples to mitigate the classification bias caused by the imbalance. It is defined as follows:

[0185] ;

[0186] in, For training batch size, For the number of target categories, For the first The weight of a class is inversely proportional to the size of the sample in that class. For the true label of the sample, The output of the target recognition network is the first Class prediction probability;

[0187] Center loss: To reduce intra-class feature differences and improve feature compactness, center loss is introduced. It learns the center of each class in the feature space and penalizes the distance between a sample and the center of its corresponding class. It is defined as follows:

[0188] ;

[0189] in, For the first Global discriminative features of each sample For the first The feature centers of the class are dynamically updated during training.

[0190] Contrast Loss: To increase the difference in features between classes and improve the discriminative power of features, intra-batch contrast loss is introduced. This loss brings similar samples closer together and widens the distance between dissimilar samples in the feature space. It is defined as follows:

[0191] ;

[0192] in In order to be with the first Each sample is a sample of the same type. The cosine similarity function is used. For the global discriminative features of the j-th sample, To compare the temperature coefficient of loss.

[0193] Mutual information regularization term: To ensure the single-modal independence of physical feature tracks and achieve the core goal of modal decoupling, a mutual information regularization term is introduced to constrain the mutual information between pure features of different modes, avoiding information redundancy between modes. It is defined as follows:

[0194] ;

[0195] in For mutual information operations, This is a directional enhancement feature for two different modalities output by the physical feature track. The regularization term constrains the mutual information of the pure features of different modalities to be as small as possible, ensuring the physical independence of each modality and providing a pure physical anchor point for subsequent confidence assessment.

[0196] Here, the Kraskov k-nearest neighbor mutual information estimation algorithm is used to calculate the mutual information between two modal pure features. This algorithm is suitable for estimating the mutual information of high-dimensional continuous random variables, avoids the binning bias problem of histogram estimation, and has better estimation accuracy and robustness. The specific implementation is as follows:

[0197] Algorithm core definition: For two d-dimensional continuous random variables and The formula for their mutual information based on k-nearest neighbor estimation is:

[0198] ;

[0199] in, For the digamma function, The nearest neighbor number, The total number of samples, For sample points in Spatial distance less than or equal to The number of samples, For sample points in Spatial distance less than or equal to The number of samples, For the first Each sample point in the joint space The first in Proximity distance.

[0200] Parameter settings: For the 512-dimensional single-modal pure feature of this invention, the nearest neighbor number is set. This value was determined through cross-validation, balancing estimation bias and variance under high-dimensional features; sample batch size The size should be consistent with the training batch size.

[0201] The regularization term implementation logic is as follows: In each training batch, the mutual information of three pairs of modal clean features—time domain-frequency domain, time domain-time frequency domain, and frequency domain-time frequency domain—is calculated separately. The sum of these mutual information is then added to the total loss function as a mutual information regularization term to constrain the mutual information of different modal clean features to be as small as possible, thus ensuring the physical independence of each modality.

[0202] Numerical stability processing: The mutual information estimate is truncated and its numerical range is limited to the interval [0,5] to avoid extreme values ​​that may cause model training instability; at the same time, a moving average is used to smooth the mutual information estimate, further improving the stability of the training process.

[0203] This invention provides a method for underwater vessel radiated noise target recognition. Addressing the problems of existing methods in complex shallow-sea underwater acoustic environments, such as modal fusion contradictions, poor robustness, and weak physical interpretability, this invention proposes a physically anchored decoupling-recoupling dual-track parallel network, MC-ACIFNet. It constructs a three-branch feature extraction architecture in the time domain, frequency domain, and time-frequency domain, embedding an attention module with normal mode prior constraints to enhance core features. Modal physical alignment and bidirectional interaction are achieved through cyclic smooth guidance, and referenceless adaptive weighted fusion is completed based on Bayesian inference, ultimately outputting the recognition result. This invention effectively alleviates the underlying contradiction between modal independence and cross-modal interaction, exhibits excellent robustness in low signal-to-noise ratio and multipath distortion scenarios, and significantly outperforms existing methods in recognition accuracy. Furthermore, its lightweight design allows for deployment on underwater edge platforms, demonstrating strong engineering practicality and applicability to various shallow-sea underwater target recognition scenarios.

[0204] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0205] Although preferred embodiments of the invention 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 both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0206] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying underwater ship radiated noise targets, characterized in that, include: An underwater ship radiated noise target recognition network is constructed. The target recognition network includes a multi-domain feature mapping module, a three-branch multi-domain feature extraction module, a channel space joint attention module PCS-JAM with normal mode prior constraints, a physical feature track, an interactive enhancement track, a modal confidence dynamic adaptive fusion module MCDE-AFM based on Bayesian inference, and a classification head module. The multi-domain feature mapping module performs multi-domain feature mapping on the original time-domain underwater acoustic signal, mapping the original time-domain underwater acoustic signal to the time-domain space, frequency-domain space, and time-frequency-domain space respectively. The three-branch multi-domain feature extraction module extracts three single-mode clean features from the original time-domain underwater acoustic signal. The three single-mode clean features include time-domain clean features, frequency-domain clean features, and time-frequency-domain clean features. The PCS-JAM performs directional enhancement on each single-modal pure feature in both the channel and spatial dimensions to obtain single-modal directional enhanced features, which serve as single-modal features of the physical feature track. The physical feature track, based on each single-modal directional enhancement feature, performs a pre-confidence score evaluation on each single modality to obtain a pre-confidence score for each single modality; The interactive enhancement track completes cross-modal bidirectional interaction of unimodal directional enhancement features based on the pre-confidence score of each unimodal mode, resulting in each unimodal interactive enhancement feature; MCDE-AFM calculates the final confidence score and fusion weight of each single modality based on Bayesian inference, based on each single modality directional enhancement feature, and dynamically fuses the three single modality interactive enhancement features based on the fusion weight to obtain global discriminative features. The classification head module identifies the target based on the global discriminative features; The PCS-JAM performs targeted enhancement on each monomodal clean feature in both the channel and spatial dimensions to obtain monomodal targeted enhanced features. These monomodal features, serving as physical feature tracks, include: For each unimodal pure feature, perform sliding window unbiased kurtosis pooling to obtain the absolute value kurtosis feature map. With signed kurtosis feature map X represents the single-modal pure feature; Based on absolute value kurtosis feature map Channel dimension enhancement is performed on the single-modal pure features to obtain single-modal channel-enhanced features; Based on signed kurtosis feature map Spatial dimension enhancement is performed on the single-modal channel enhancement features to obtain single-modal spatial enhancement features; Gated residual fusion is performed on single-modal channel enhancement features and single-modal spatial enhancement features to obtain single-modal directional enhancement features; The absolute value kurtosis feature map Channel dimension enhancement is performed on the unimodal pure features to obtain unimodal channel-enhanced features, including: For the absolute value kurtosis feature map Perform global average pooling and global max pooling respectively to obtain feature vectors with two channel dimensions, each having dimension 1. C is the number of channels; The feature vectors of the two channel dimensions are input into a shared multilayer perceptron (MLP) to obtain the channel attention feature vectors. , dimension ; Based on the normal mode wave model, the propagation attenuation coefficients of different frequency channels in shallow sea waveguides are pre-calibrated to obtain the prior weights for channel reliability. , This represents the prior weight of the channel reliability for the c-th channel; The signal-to-noise ratio confidence factor of the input signal is calculated based on the spectral flatness of the input signal and the average spectral flatness of the underwater environmental noise. ; Based on channel reliability prior weights and signal-to-noise ratio confidence factor Channel attention feature vectors By applying normal mode prior constraints, the final channel attention weights are obtained. Channel dimension enhancement is performed on the unimodal pure features to obtain unimodal channel-enhanced features. ; The signed kurtosis feature map Spatial dimension enhancement is performed on the single-modal channel enhancement features to obtain single-modal spatial enhancement features, including: Enhancement features for single-mode channels Global average pooling and global max pooling are performed in the spatial dimension to obtain feature maps in two spatial dimensions, each with a dimension of 1. ; Combining the absolute value of the signed kurtosis feature map The three feature maps are concatenated along the channel dimension to obtain a dimension of splicing feature map; The concatenated feature map is mapped to a single-channel spatial attention feature map through a one-dimensional convolutional layer, and the spatial attention weights are obtained through a sigmoid activation function. , dimension ; Enhancement features for single-mode channels Spatial dimension enhancement is performed to obtain single-modal spatial enhancement features. Specifically, it includes: ； in, This is for the Hadamard product operation.

2. The underwater ship radiated noise target identification method according to claim 1, characterized in that, Based on the normal mode wave model, the propagation attenuation coefficients of different frequency channels in shallow sea waveguides are pre-calibrated to obtain the prior weights for channel reliability. ,include: Based on the Kraken normal mode model, the propagation attenuation at each frequency point within the frequency band underwater is calculated. ; The propagation attenuation at each frequency point is normalized using the min-max method and mapped to the interval [0.1, 1.0] to obtain the prior weights for channel reliability. ； in, This represents the maximum propagation attenuation within the frequency band. This is the minimum attenuation within the frequency band; The signal-to-noise ratio confidence factor of the input signal is calculated based on the spectral flatness of the input signal and the average spectral flatness of the underwater environmental noise. ,include: ； in, This represents the spectral flatness of the input signal. The average spectral flatness of environmental noise. Use the Sigmoid activation function; The channel attention feature vector By applying normal mode prior constraints, the final channel attention weights are obtained. Channel dimension enhancement is performed on the unimodal pure features to obtain unimodal channel-enhanced features. ,include: ； ； in It is the Sigmoid activation function. This is for the Hadamard product operation.

3. The underwater ship radiated noise target identification method according to claim 1, characterized in that, The gated residual fusion of the single-modal channel enhancement features and the single-modal spatial enhancement features to obtain the single-modal directional enhancement features includes: ； in, For learnable gating coefficients, This is a single-modal directional enhancement feature. For layer normalization.

4. The underwater ship radiated noise target identification method according to claim 1, characterized in that, The interactive enhancement track, based on the pre-confidence score of each unimodal pre-confidence score, completes cross-modal bidirectional interaction of unimodal directional enhancement features, resulting in each unimodal interactive enhancement feature, including: Input three single-modal targeted augmentation features , and and pre-confidence score and ,in, , and These represent the directional enhancement features of the time-domain mode, frequency-domain mode, and time-frequency-domain mode, respectively. Directed enhancement features for three single modes , and Perform periodic physical alignment; By sharing a linear mapping layer, the mapped single-modal directional enhancement features are... , and A unified mapping is performed to the same dimension of the mapping embedding space; Three sets of bidirectional cross-attention units are constructed to realize the directional injection of complementary information in the time domain-frequency domain, frequency domain-time-frequency domain, and time domain-time-frequency domain, respectively, resulting in three sets of interactive cross-modal complementary features; Gated aggregation is performed on three unimodal directional enhancement features and three sets of cross-modal complementary features obtained from interactions to obtain unimodal interactive enhancement features. , and .

5. The underwater ship radiated noise target identification method according to claim 4, characterized in that, Each set of bidirectional cross-attention units enables the targeted injection of complementary information, including: Directed augmentation features for two single-modal interactions and Self-attention encoding is performed separately to obtain self-enhancing features. , Calculate the complementary characteristics of the residuals: ； in, , These are the residual complementary features of the two modes; Based on the pre-confidence scores of the two unimodals to be interacted with, set an interaction mask, including: ； in, This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise. and These are the pre-confidence scores for the two single modes, respectively. Cross-modal interaction of residual complementary features yields cross-modal complementary features, including: ； ； in, Inject cross-modal complementary features of mode B into mode A. Inject cross-modal complementary features of mode A into mode B. , and The weight matrix represents the learnable linear mapping of mode A. , and The weight matrix represents the learnable linear mapping for mode B. For the embedded dimension; Targeted Enhancement Features for Single-Modal Based on Interactive Masking and Perform incremental updates for each unimodal interaction enhancement feature, including: ； ； in, For the single-modal interaction enhancement features of mode B, This is a single-modal interaction enhancement feature for mode A.

6. The underwater ship radiated noise target identification method according to claim 1, characterized in that, The MCDE-AFM, based on each unimodal directed enhancement feature, calculates the final confidence score and fusion weight for each unimodal using Bayesian inference. Based on the fusion weight, it dynamically fuses the three unimodal interactive enhancement features to obtain global discriminative features, including: For the Modality t represents the time-domain mode, f represents the frequency-domain mode, and tf represents the time-frequency-domain mode. Four classes of no-reference physical statistical features are extracted from the single-mode directional enhancement features of the m-th mode, including the saliency of line spectral clusters. Periodic stability Normalized excess kurtosis Spectral flatness ; Given a pre-set confidence score Below, based on the salience of the spectral clusters Periodic stability Normalized excess kurtosis Spectral flatness The prior probability of the m-th mode is fitted using a Beta distribution. ; Four deep feature quality indices were extracted from the single-modal directional enhancement features of the m-th modality as observations, namely, intra-class compactness. Inter-class separation Feature signal-to-noise ratio and feature entropy Construct the Gaussian likelihood function , The standardized mean of a deep feature quality index represents the performance of a feature at a given confidence level. The probability of observing the corresponding deep feature quality index; Based on Bayes' theorem, the pre-confidence score of the m-th mode is calculated. The posterior probability distribution is used, and the posterior expectation is taken as the final confidence score for the m-th mode. ; The adaptive temperature coefficient is calculated based on the coefficient of variation of the final confidence scores of the three modes. ; Based on adaptive temperature coefficient Calculate the fusion weights for each modality; Based on the fusion weights of each modality, the single-modal interaction enhancement features of the three modalities are weighted and fused to obtain the global discriminative features. .

7. The underwater ship radiated noise target identification method according to claim 1, characterized in that, The joint loss function of the underwater ship radiated noise target identification network is: ； in, These are weighting coefficients. For weighted cross-entropy loss, Loss at the center To compare the losses, For mutual information regularization; ； in, For training batch size, For the number of target categories, For the first The weight of a class is inversely proportional to the size of the sample in that class. For the true label of the sample, The output of the target recognition network is the first Class prediction probability; ； in, For the first Global discriminative features of each sample For the first The feature centers of the class are dynamically updated during training; ； in, In order to be with the first Each sample is a sample of the same type. For the global discriminative features of the j-th sample, The cosine similarity function is used. To compare the temperature coefficient of loss; ； in, For mutual information operations, These are two different modalities of directional enhancement features output from the physical feature track.

Images