Underwater sonar target identification system based on multi-domain feature fusion and lightweight modeling

Abstract

The invention discloses an underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling, and the system comprises a Trifusion block, a novel lightweight attention residual network, a long and short time attention LSTM and a Mamba module which are connected in sequence, and achieves target recognition through multi-domain parallel extraction of fusion features, lightweight convolution and attention optimization, long and short time dependence capture and long sequence modeling. The method has the advantages that complex noise is comprehensively represented, the model efficiency and stability are improved, the bottleneck of time sequence processing is broken through, and the method has high performance, high robustness and wide adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater sonar signal processing and target recognition, and in particular to an underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling. Background Art

[0002] In the field of underwater sonar target recognition, the complex characteristics of ship-radiated noise and the inherent challenges of the underwater environment together constitute the core bottleneck restricting technological development. As the primary target of underwater sonar monitoring, ships radiate noise from multiple physical processes such as propulsion systems, mechanical vibrations, and propeller cavitation, resulting in a high degree of complexity in the frequency, amplitude, and temporal evolution of noise signals. Due to differences in propulsion type, hull structure, and navigation status, the noise characteristics of different types of ships exhibit wide-band, multimodal distributions in the frequency domain and non-stationary dynamic changes in the time domain. Furthermore, there is often a significant imbalance in the distribution of the number of samples of different categories, making it difficult to effectively capture the characteristics of minority targets due to sparse samples.

[0003] Traditional underwater sonar target recognition methods face fundamental limitations at the feature extraction level. Early studies mostly relied on single-domain signal analysis. For example, steady-state spectrum analysis based on Mel-Frequency Cepstral Coefficients (MFCCs) can simulate the characteristics of human hearing, but cannot characterize transient high-frequency components such as propeller cavitation. Time-frequency analysis based on Short-Time Fourier Transform (STFT) is limited by a fixed window function, making it difficult to simultaneously consider time and frequency resolution within a wide bandwidth. And simple time-domain statistical features (such as RMS value and crest factor) lack effective representation of frequency structure. The limitations of these single features make it difficult for the model to fully capture the multi-scale characteristics of noise signals, resulting in the loss or redundancy of key information. This incompleteness of feature representation is particularly prominent when processing non-stationary signals.

[0004] Unbalanced data distribution and noise interference further exacerbate the difficulty of model training. In practical applications, the sample size of common ship categories may be far greater than that of rare categories. Traditional classification models (such as support vector machines and hidden Markov models) are prone to bias towards the majority class due to sample imbalance during training, resulting in a significant decrease in the recognition performance of minority class targets. At the same time, underwater sonar signals are inevitably affected by ocean environmental noise (such as wave turbulence and biological vocalizations) and channel effects (such as multipath propagation and signal attenuation). Feature extraction and classification under low signal-to-noise ratio conditions become difficult. Traditional methods lack an effective compensation mechanism for uneven data distribution and are sensitive to noise. They are prone to overfitting or insufficient generalization in complex environments.

[0005] The lack of the ability to model long-term dependencies is another shortcoming of traditional technologies. The time series of sonar signals contains key information such as the target's motion trajectory and the mechanical operation cycle. However, traditional recurrent neural networks (RNNs) and their variants (such as standard LSTMs) suffer from the gradient vanishing problem when processing long sequences, making it difficult to capture long-range dependencies exceeding tens of seconds. For example, the periodic cavitation noise of a ship's propeller may show regular changes in long sequences, but traditional models are unable to effectively model such long-term dependencies, resulting in incomplete feature representation. On the other hand, although pure attention mechanism models can capture global dependencies, their computational complexity increases quadratically with the length of the sequence. When processing long sequences of underwater sonar data in real time, they face hardware resource limitations and cannot meet the timeliness requirements of practical applications.

[0006] The contradiction between lightweight models and high performance is particularly prominent in underwater scenarios. Underwater sonar equipment is typically deployed on resource-constrained platforms such as sensor nodes and unmanned underwater vehicles. While traditional deep neural networks (such as ResNet and DenseNet) have strong feature learning capabilities, they have a large number of parameters and high computational costs, resulting in long training cycles and difficulty in running efficiently on edge devices. Furthermore, traditional methods lack systematic design for multi-domain feature fusion and cannot fully utilize the complementarity of different feature domains, further limiting the model's adaptability to complex noisy environments.

[0007] These challenges have a significant impact on key areas such as national defense security, maritime navigation, and environmental monitoring. Therefore, overcoming the limitations of single features, improving the ability to model long-sequence dependencies, and balancing model lightweighting with high performance requirements have become core issues that need to be addressed in the field of underwater sonar target recognition.

[0008] In response to the above technical bottlenecks, the present invention proposes an innovative solution based on multi-domain feature fusion and lightweight modeling. Through the integration of cross-domain feature collaborative extraction, lightweight network architecture design and long sequence dependency modeling technology, it aims to improve the accuracy, robustness and real-time performance of target recognition in complex underwater environments, and provide a more efficient solution for practical applications. Summary of the Invention

[0009] The purpose of the present invention is to provide an underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling to address the core problems in the field of underwater sonar target recognition, such as the complex characteristics of ship radiation noise, uneven data distribution, incomplete feature extraction of traditional methods, insufficient long sequence dependency modeling capabilities and high model calculation complexity.

[0010] To achieve the above objectives, the present invention adopts the following technical means:

[0011] The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling includes a sequentially connected TriFusionblock multi-domain feature fusion module, a new lightweight attention residual network, a long-short-term attention LSTM module, and a Mamba module. Each module forms a hierarchical processing chain through data flow transmission;

[0012] The TriFusion block multi-domain feature fusion module serves as a front-end processing unit, receiving the original sonar signal, differential signal, and cumulative signal in parallel, and extracting Mel-frequency cepstral coefficient features, constant Q transform features, and Mel-frequency spectrum features respectively. The three-branch features are concatenated into a three-channel feature tensor in the channel dimension according to the order of constant Q transform, Mel-frequency cepstral coefficient, and Mel-frequency spectrum. A weighted fusion formula is used to form a composite feature vector containing full-band information, which serves as the input of a new lightweight attention residual network.

[0013] The new lightweight attention residual network is based on the streamlined ResNet architecture, receives the composite feature tensor output by the TriFusion block, performs dimensionality reduction and spatial feature extraction on the features through 9 layers of lightweight convolutional layers, and embeds the SE attention module after the Conv2 layer. It "squeezes" the global information of the channel through global average pooling, and "stimulates" the generation of channel weights through the fully connected layer to enhance the key features and suppress the interference information. The feature vector after output dimension compression is transmitted to the long-short-term attention LSTM module.

[0014] The Long-Short-Term Attention (LSTM) module, as a temporal feature modeling unit, receives the feature vector sequence output by the novel lightweight attention residual network. It controls the flow of information through a gating mechanism consisting of forget gates, input gates, output gates, and memory units. Combined with the Long-Short-Term Attention (LSTM) mechanism, it captures the long-term and short-term dependencies of sonar signals. The resulting feature representations are then passed to the Mamba module.

[0015] As the final processing unit, the Mamba module first stabilizes the input data distribution through layer normalization to enhance noise resistance. The core state-space model module is based on a discrete-time state-space model, captures long-sequence dependencies with linear computational complexity, and outputs a state vector containing long-range temporal features. Finally, the input features are added to the output features of the state-space model module through residual connections to solve the gradient vanishing problem of deep networks and ultimately output the target classification results.

[0016] A further solution of the present invention is that in the three-branch feature extraction path of the TriFusion block module, the original signal is transformed through short-time Fourier transform, Mel frequency mapping and discrete cosine transform to generate Mel frequency cepstral coefficient features, focusing on the steady-state spectrum structure; the differential signal is transformed through Hilbert transform and constant Q transform to generate constant Q transform features, capturing high-frequency transient components; the cumulative signal is decomposed by wavelet and Mel spectrum feature extraction to enhance the expression of low-frequency trends.

[0017] A further solution of the present invention is that in the TriFusion block module, the original signal is processed by the Mel-frequency cepstral coefficient, first a short-time Fourier transform is performed to obtain X(m,k), the time domain signal is converted to the frequency domain, and then the linear frequency is mapped to the Mel frequency, and the filter bank energy E is calculated. i Finally, the Mel frequency cepstral coefficient feature is obtained through logarithmic transformation and discrete cosine transform. The specific formula is:

[0018]

[0019] Where x(n) is the audio information, w(n) is the Hamming window function, n is the frame index, k is the frequency index, N is the number of FFT points, R is the number of frame shift points, and M is the number of Mel filter banks.

[0020] A further solution of the present invention is that in the TriFusion block module, the differential signal first performs a first-order difference on the original audio signal to obtain v(n), amplifies the instantaneous rate of change, and highlights the high-frequency transient components; then extracts the envelope A(n) through Hilbert transform to enhance the amplitude change information; finally, uses constant Q transform to perform time-frequency analysis to obtain CQT(m,k), and converts the amplitude into dB scale CQT dB (m,k), the specific formula is:

[0021] v(n)=x(n+1)-x(n) (4)

[0022] A(n)=|Hilbert(v(n))| (5)

[0023]

[0024] Among them, h(n) is the analysis window function, f k is a specific frequency point, and ref takes the maximum value of the constant Q transformation amplitude.

[0025] A further solution of the present invention is that in the TriFusion block module, the cumulative signal first accumulates and calculates the original signal to smooth high-frequency fluctuations and highlight low-frequency trends; then the cumulative signal is normalized and wavelet decomposition is performed to obtain d norm(n), take the approximate coefficient to separate the low-frequency component and suppress the high-frequency interference; finally, use a method similar to the Mel frequency cepstral coefficient to extract the Mel spectrum features and convert them into dB scale Fbank dB (m,i) and adjust the shape. The specific formula is:

[0026]

[0027] Among them, J is the number of decomposition layers equal to 5, c j,k is the wavelet coefficient, ψ j,k is the wavelet basis function.

[0028] A further solution of the present invention is that in the TriFusion block module, the weighted fusion formula is:

[0029] Final = αMFCC n +βCQT dB (m,k)+γFbank dB (m,i) (10)

[0030] Among them, α, β, and γ are weight coefficients, satisfying α+β+γ=1.

[0031] A further solution of the present invention is that the novel lightweight attention residual network is simplified to 10 layers based on the ResNet18 architecture, including 9 convolutional layers and 1 fully connected layer, and the SE attention module added after the Conv2 layer generates channel weight coefficients through ReLU and sigmoid activation functions.

[0032] A further solution of the present invention is that in the gating mechanism of the LSTM module, the forget gate helps the model filter out useless information when processing sequence data, preventing the memory unit from being flooded with irrelevant information, so that the model can focus on important information. The specific formula is:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f )(11)

[0034] Among them, f t is the output of the forget gate at time step t, W f is the weight matrix of the forget gate, h t-1 is the hidden state of the previous moment, b f is the bias vector of the forget gate, σ is the activation function of sigmoid;

[0035] The input gate is mainly responsible for determining the current input x tWhich information will be stored in the memory unit, and generate candidate memory units at the same time Used to update the memory unit; the control signal i of the input gate t and candidate memory cells The specific formula is:

[0036] i t =σ(W i ·[h t-1 ,x t ]+b i ) (12)

[0037]

[0038] Among them, W i and W c are the weight matrices of the inputs and candidate memory units, b i and b c are the bias vectors of the input and candidate memory units respectively, and tanh is the hyperbolic tangent activation function;

[0039] The memory unit is responsible for storing and transmitting important information in the sequence. Through the synergy of the forget gate and the input gate, the memory unit can effectively update and retain information between different time steps. The specific formula is:

[0040]

[0041] Among them, c t and c t-1 are the memory unit states at the current moment and the previous moment respectively, is the element-wise multiplication operator;

[0042] The output gate determines the memory cell c t Which information in will be output as the hidden state h at the current moment t ; Hidden state h t As the final output, the long-term and short-term attention is calculated, and then combined with the fused context vector. The intermediate representation is generated through the fully connected layer and then mapped to the final prediction output. The specific formula is:

[0043] o t =σ(W o ·[h t-1 ,x t ]+b0) (15)

[0044]

[0045] Among them, t is the output of the output gate at time step t.

[0046] A further solution of the present invention is that, in the attention mechanism of the long-short-term attention LSTM module, the long-short-term attention mechanism divides attention into short-term attention and long-term attention;

[0047] First calculate the weights and context vectors of long and short attention, as shown in the formula:

[0048]

[0049] Among them, score s and score1 are the scoring functions of short-term attention weight and short-term attention weight respectively. and They are short-term attention weight and short-term attention weight, and They are short-term context vector and long-term context vector respectively;

[0050]

[0051] Among them, γ is the proportion of long-term information in the final fusion with fusion weight ∈ [0,1];

[0052] After fusion, the final context vector is obtained:

[0053]

[0054] The current hidden state h of LSTM t and the fused context vector c t Combine them, generate an intermediate representation through the fully connected layer, and then map it to the final prediction output:

[0055]

[0056] A further solution of the present invention is that in the Mamba module, layer normalization stabilizes data distribution through the formula LN(x), which is specifically:

[0057]

[0058] Among them, μ is the mean of feature dimension, σ 2 is the variance, ε prevents division by zero, γ and β are learnable parameters;

[0059] The core state space model module is based on the discrete time state space model Capturing long sequence dependencies by iterative state updates, where s t is the state vector, u t is the input vector, y t is the output vector, and A, B, C, D are the model parameter matrices.

[0060] Beneficial effects of the present invention:

[0061] 1. Multi-domain feature fusion enables comprehensive characterization of complex noise

[0062] The TriFusion block module processes the original signal, differential signal, and cumulative signal in parallel, extracting Mel-frequency cepstral coefficients, constant Q transform, and Mel-frequency spectrum features, and deeply fuses them to form a multi-scale feature system covering the entire frequency band:

[0063] The Mel-frequency cepstral coefficient feature focuses on the steady-state spectrum structure, simulating the human ear's auditory characteristics to compress redundant frequencies and effectively characterize the basic spectrum envelope of ship noise;

[0064] The constant Q transform feature uses high-resolution characteristics in the high-frequency band and differential signal processing to accurately capture transient high-frequency components such as propeller cavitation and mechanical shock;

[0065] Mel spectrum features enhance the trend expression of low-frequency noise such as ship mechanical vibration by accumulating signal smoothing and extracting low-frequency trends.

[0066] The organic integration of the three types of features breaks through the limitations of traditional single features, enabling the model to fully capture the "steady state-transient-low frequency" multi-dimensional characteristics of the noise signal and reduce information loss and redundancy.

[0067] 2. Lightweight architecture improves model efficiency and stability

[0068] The new lightweight attention residual network achieves a balance between model complexity and performance through a streamlined architecture and innovative attention mechanisms:

[0069] Layer Streamlining: The traditional ResNet18 has 18 layers reduced to 10 (9 convolutional layers + 1 fully connected layer), significantly reducing the number of parameters and computational costs, adapting to the hardware requirements of underwater resource-constrained equipment.

[0070] SE attention mechanism: Through global average pooling and fully connected layers, it automatically learns inter-channel dependencies, strengthening key features and suppressing interfering information. Across datasets of varying sizes, this module improves the targeted nature of feature extraction, ensuring model accuracy and stability. In particular, it significantly enhances the ability to recognize minority samples in scenarios with uneven data distribution.

[0071] 3. Long-sequence dependency modeling breaks through the bottleneck of traditional time series processing

[0072] The cascade design of the Long Short-Term Attention (LSTM) module and the Mamba module builds an efficient temporal feature modeling system:

[0073] Long-Short-Term Attention (LSTM): This solves the vanishing gradient problem of traditional RNNs through a gating mechanism. It combines short-term and long-term attention to adaptively filter the critical time segments in sonar signals for target recognition.

[0074] The Mamba module utilizes a linear computational complexity design based on a state-space model, breaking through the computational bottleneck of traditional self-attention mechanisms and enabling efficient processing of long sequences of sonar data. Furthermore, layer normalization stabilizes data distribution, and residual connections mitigate the vanishing gradient problem in deep networks, ensuring accurate modeling of long-range dependencies.

[0075] 4. The unity of high performance and strong robustness

[0076] Lightweight and real-time: The model has significantly fewer parameters and floating-point operations than classic models such as ResNet, enabling real-time inference on embedded platforms and reducing hardware performance requirements. It is suitable for resource-constrained scenarios such as unmanned underwater vehicles and underwater sensors.

[0077] Interference resistance: The multi-module collaborative noise resistance mechanism enables the model to stably extract features even in low signal-to-noise ratio environments. Combined with the attention mechanism to filter out noise-related components, it significantly improves robustness in complex environments.

[0078] Generalization ability: The model maintains high recognition accuracy under different sample label ratios.

[0079] 5. Wide adaptability of application scenarios

[0080] Through technological integration, this invention provides efficient and reliable solutions for key areas such as national defense security, maritime navigation, and environmental monitoring. Its combination of lightweight design, high robustness, and long-sequence processing capabilities significantly enhances the practical application value of underwater sonar target recognition and promotes the transition of related technologies from the laboratory to engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a structural diagram of the present invention;

[0082] Figure 2 This is a structural diagram of the TriFusion block of the present invention;

[0083] Figure 3 The present invention shows the processing process and feature extraction results of the TriFusion block module on one of the tug categories in the deep ship dataset from different dimensions;

[0084] Figure 4 The NLARN network structure diagram and the downsampled residual unit structure diagram of the present invention;

[0085] Figure 5 This is the network structure diagram of the SE attention mechanism of the present invention;

[0086] Figure 6 This is the structure diagram of the long-short-term attention LSTM module of the present invention;

[0087] Figure 7 This is the Mamba structure diagram of the present invention;

[0088] Figure 8 The confusion matrix and t-SNE graph of different models of the ShipsEar training dataset of the present invention;

[0089] Figure 9 This is the confusion matrix diagram of different modules of the ShipsEar training dataset of the present invention;

[0090] Figure 10 This is the confusion matrix diagram of different modules in the Deep Ship training data set of the present invention;

[0091] Figure 11 Parameters and flops diagrams of each model of the present invention. DETAILED DESCRIPTION

[0092] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0093] Example

[0094] like Figure 1 As shown in the figure, the underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling includes a TriFusion block multi-domain feature fusion module, a new lightweight attention residual network, a long-short-term attention LSTM module and a Mamba module connected in sequence. Each module forms a hierarchical processing link through data flow transmission.

[0095] The TriFusion block is responsible for multi-domain feature extraction. It processes the ship's radiated noise signal in parallel from three perspectives: the original signal, the differential signal, and the cumulative signal, extracting MFCC, CQT, and Fbank features, respectively. MFCC focuses on steady-state spectral structure, CQT captures high-frequency transient changes, and Fbank enhances low-frequency trends. These features are fused through a multi-channel strategy to form a 3×128×216 fused feature tensor, providing a rich and discriminative initial input for the entire network.

[0096] The new lightweight residual attention network is based on an improved ResNet architecture. By streamlining the number of layers, it reduces model complexity and computational costs. It also incorporates the SE attention mechanism to automatically learn inter-channel dependencies, highlight key features, suppress interference, and improve the ability to extract ship radiated noise features, avoid overfitting, and provide high-quality feature representation for subsequent modules.

[0097] The Long-Short-Term Attention (LSTM) module combines gating and attention mechanisms to effectively capture long-term and short-term dependencies in sonar signals, adaptively filter key information, and overcome the vanishing gradient problem of traditional RNNs when processing long sequences. It accurately focuses on the critical time segments or features in sonar signals for target recognition, providing more valuable feature input for subsequent Mamba modules.

[0098] The Mamba module is a key optimization link in the network. Its normalization layer stabilizes data distribution and enhances the model's robustness to noise. The core SSM Block, based on the state-space model, captures long-sequence dependencies with near-linear computational complexity, accurately identifying characteristic information such as the distance and speed of target objects. The residual connection ensures the effective flow of information, solves the vanishing gradient problem in deep networks, and improves the accuracy of sonar target recognition.

[0099] Through the close collaboration between these four modules, the present invention constructs a complete and efficient underwater sonar target recognition network architecture from feature extraction, feature optimization, long sequence dependency capture to anti-interference recognition, significantly improving the accuracy and robustness of ship radiated noise classification, and playing an important role in the field of underwater sonar target recognition.

[0100] TriFusion block multi-domain feature fusion module

[0101] The TriFusion block multi-domain feature fusion module serves as a front-end processing unit, receiving the original sonar signal, differential signal, and cumulative signal in parallel, and extracting Mel-frequency cepstral coefficient (MFCC) features, constant Q transform (CQT) features, and Mel-frequency spectrum (Fbank) features, respectively. In the three-branch feature extraction path of the TriFusion block module, the original signal undergoes short-time Fourier transform, Mel-frequency mapping, and discrete cosine transform to generate Mel-frequency cepstral coefficient (MFCC) features, focusing on the steady-state spectral structure; the differential signal undergoes Hilbert transform and CQT transform to generate CQT features, capturing high-frequency transient components; the cumulative signal undergoes wavelet decomposition and Fbank feature extraction to enhance the expression of low-frequency trends; the three-branch features are spliced ​​into a three-channel feature tensor in the channel dimension in the order of CQT, MFCC, and Fbank, and a composite feature vector containing full-band information is formed through a weighted fusion formula as the input of the NLARN.

[0102] TriFusion block structure

[0103] The TriFusion block module processes the original signal, differential signal and cumulative signal in parallel, extracting Mel-frequency cepstral coefficient (MFCC), constant Q transform (CQT) features and Mel-frequency spectrum (Fbank) features respectively, and finally forms a composite feature vector containing full-band information through feature fusion. Figure 2 As shown in the figure, the three processing branches of the TriFusion block complement each other and explore the characteristics of noise signals from different angles: MFCC focuses on the steady-state spectrum structure, CQT captures high-frequency transient changes, and Fbank strengthens the expression of low-frequency trends. The organic combination of the three provides a more discriminative feature basis for underwater acoustic target recognition.

[0104] TriFusion block multi-dimensional feature collaborative extraction mechanism

[0105] The first part of the TriFusion block processes the original signal using Mel-Frequency Cepstral Coefficients (MFCC). First, a Short-Time Fourier Transform (STFT) is performed to obtain X(m,k), which converts the time domain signal to the frequency domain. Then, the linear frequency is mapped to the Mel-Frequency, and the filter bank energy E is calculated. i Finally, the MFCC features are obtained through logarithmic transformation and discrete cosine transform (DCT). The specific formula is:

[0106]

[0107] Where x(n) is the audio information, w(n) is the Hamming window function, n is the frame index, k is the frequency index, N is the number of FFT points, R is the number of frame shift points, and M is the number of Mel filter banks.

[0108] MFCC captures dynamic changes in features. This processing simulates human hearing, compressing redundant frequencies, highlighting key spectral components, strengthening steady-state features, suppressing noise interference, and using differential calculations to enhance sensitivity to time-domain changes, providing stable and robust basic features for noise.

[0109] The second part focuses on the differential signal. First, the original audio signal is first differentiated to obtain v(n), which amplifies the instantaneous rate of change and highlights the high-frequency transient components. Then, the envelope A(n) is extracted through Hilbert transform to enhance the amplitude change information. Finally, the constant Q transform is used to perform time-frequency analysis to obtain CQT(m,k), and the amplitude is converted to dB scale CQT. dB (m,k), the specific formula is:

[0110] v(n)=x(n+1)-x(n) (4)

[0111] A(n)=|Hilbert(v(n))| (5)

[0112]

[0113] Among them, h(n) is the analysis window function, f k is a specific frequency point, and ref takes the maximum value of the constant Q transformation amplitude.

[0114] The high-resolution characteristics of CQT in the high-frequency band, combined with the first-order difference and Hilbert envelope, can accurately characterize the time-frequency distribution of transient signals such as ship propeller cavitation, making up for the shortcomings of MFCC in representing high-frequency dynamic information.

[0115] The third part processes the cumulative signal. First, the original signal is accumulated and calculated to smooth high-frequency fluctuations and highlight low-frequency trends. Then the cumulative signal is normalized and wavelet decomposition is performed to obtain d norm (n), take the approximate coefficient to separate the low-frequency component and suppress the high-frequency interference; finally, use a method similar to MFCC to extract Fbank features and convert them into dB scale Fbank dB (m,i) and adjust the shape. The specific formula is:

[0116]

[0117] Among them, J is the number of decomposition layers equal to 5, c j,k is the wavelet coefficient, ψ j,k is the wavelet basis function.

[0118] Fbank preserves energy distribution, efficiently characterizes the basic spectral structure, has low computational complexity, and combines with the first two parts to cover the full frequency band information of noise and provide multi-scale feature support.

[0119] In order to more intuitively demonstrate the processing effect and feature extraction process of the TriFusion block module on the ship radiation noise signal, the present invention draws six sets of visual maps, such as Figure 3 As shown, Figure 3Figure a (center) shows the original signal, Figure b shows the MFCC plot of the original signal, Figure c shows the cumulative signal, Figure d shows the Fbank plot of the cumulative signal, Figure e shows the first-order difference signal, and Figure f shows the CQT plot of the first-order difference signal. The original signal plot uses time (seconds) as the horizontal axis and amplitude as the vertical axis, visually presenting the overall trend of the audio time-domain waveform. The first-order difference signal plot and the cumulative signal plot respectively display the signal characteristics after differentiation and integration, with the former highlighting instantaneous changes in the signal and the latter highlighting low-frequency trend components. At the feature spectrum level, the MFCC plot of the original signal displays the signal's steady-state spectral envelope using the Mel-scale frequency and time dimensions. The CQT plot of the first-order difference signal uses logarithmic frequency and time axes to accurately depict the time-frequency distribution of high-frequency transient components. The Fbank plot of the cumulative signal focuses on low-frequency features, presenting the signal's basic spectral structure on the Mel-scale. Figure 3 (af) From time-domain waveforms to frequency-domain features, and from original signals to derived processing results, the module's analysis of noise signals is presented in a multi-dimensional and multi-level manner, providing an intuitive basis for understanding the feature extraction mechanism and verifying the effectiveness of the method.

[0120] TriFusion block feature fusion

[0121] Considering that the MFCC, CQT, and Fbank features extracted by each branch in the TriFusion block initially exist in single-channel form, they can only express the characteristics of ship radiated noise from a single dimension, such as steady-state spectral structure (MFCC), high-frequency transient changes (CQT), and low-frequency trends (Fbank). To present ship radiated noise information more comprehensively and richly, this paper innovatively introduces a multi-channel fusion strategy to expand single-channel features into multi-channel fusion features.

[0122] Specifically, during the fusion process, the CQT, MFCC, and Fbank derived features are arranged in the order of channel-wise order. CQT features are sensitive to high-frequency transient components and can capture transient high-frequency signals generated by phenomena such as propeller cavitation in ship radiated noise. MFCC features simulate the human hearing characteristics and effectively characterize the steady-state spectral characteristics of noise. Fbank features focus on low-frequency trends, highlighting low-frequency noise components such as ship mechanical vibration. This arrangement ultimately forms a fused feature tensor with a shape of 3×128×216. This fused feature integrates the advantages of different features and achieves a coordinated expression of multi-dimensional information in the channel dimension. This provides a more discriminative input for the subsequent NLARN network, helping to improve the accuracy of ship radiated noise classification and underwater sonar target recognition.

[0123] After completing the feature extraction of the three branches, the TriFusion block integrates the different features through the weighted fusion formula. The formula is as follows:

[0124] Final = αMFCC n +βCQT dB (m,k)+γFbank dB (m,i)(10)

[0125] Among them, α, β, and γ are weight coefficients, satisfying α+β+γ=1.

[0126] Novel Lightweight Attention Residual Network (NLARN)

[0127] The new lightweight residual attention network (NLARN) is based on the streamlined ResNet architecture, receives the composite feature tensor output by the TriFusion block, performs dimensionality reduction and spatial feature extraction on the features through 9 layers of lightweight convolutional layers, and embeds the SE attention module after the Conv2 layer. It "squeezes" the global information of the channel through global average pooling, and "stimulates" the generation of channel weights through the fully connected layer to enhance the key features and suppress the interference information. The feature vector after output dimension compression is passed to the long-short-term attention LSTM module.

[0128] NLARN network structure

[0129] like Figure 4 As shown in the figure, NLARN cleverly improves upon the classic ResNet architecture in terms of network structure design. The traditional ResNet18 suffers from parameter redundancy when handling ship radiated noise classification tasks, which not only increases computational costs but also increases the risk of overfitting. NLARN boldly reduces the number of ResNet18 layers, streamlining it to 10, consisting of 9 convolutional layers and 1 fully connected layer. This optimization initiative has achieved remarkable results. Compared to ResNet18, the number of model parameters has been sharply reduced, while the classification accuracy remains at a comparable level. This clearly demonstrates that, for this specific ship radiated noise dataset, the original ResNet18 structure has room for optimization, while NLARN's streamlined structure successfully reduces model complexity while maintaining good classification performance, laying a solid foundation for subsequent efficient training and application.

[0130] To improve model performance, NLARN incorporates the SE (squeeze-excitation) attention mechanism. The SE block is added after the Conv2 layer of the network. This module acts as an intelligent information filter. First, through the "squeezing" operation of global average pooling, the SE block can collect global information from the feature map and condense it into channel statistics. This step is like performing a macro scan of the entire scene to obtain a summary of key information. Subsequently, through two fully connected layers and using the ReLU and Sigmoid activation functions for "excitation" operations, the SE block can adaptively reconstruct the channel statistics. This process is like assigning different weights to different information, making important information more prominent, thereby effectively capturing the dependencies between channels, enabling the model to focus on key features, and improving the ability to extract the characteristics of ship radiated noise.

[0131] SE Attention Module

[0132] The operation process of SE attention mechanism is as follows Figure 5 As shown:

[0133] Starting with a single input image, the model extracts features from that image. The feature map dimensions for the current feature layer are set to [number of channels, height, width]. These three dimensions describe the image from different perspectives. The number of channels reflects the number of different feature types, while the height and width reflect the spatial distribution of the features.

[0134] 2. Perform average or max pooling on the feature map's [height, width] dimensions. After pooling, the feature map's dimensions change from [number of channels, height, width] to [number of channels, 1, 1]. In the resulting [number of channels, 1, 1] feature map, each channel has a unique value. Pooling compresses and integrates information in the spatial dimensions, highlighting the global information of each channel.

[0135] 3. The features in [number of channels, 1, 1] can be understood as weights extracted from each channel. These weights represent the influence of each channel on feature extraction. The vector obtained through global pooling is input into a multilayer perceptron (MLP) network. After computational processing, more accurate and appropriate weights for each channel are obtained. The MLP network can explore the potential relationships between channels, allowing the model to learn channel weights that are more realistic.

[0136] 4. After obtaining the weights for each channel (number of channels, 1, 1), these weights are applied to the original feature map (number of channels, height, width). Specifically, each channel is multiplied by its corresponding weight. When the weight of a channel is large, the value of the channel feature map will increase accordingly, and its influence on the final output will be greater; conversely, when the weight is small, the value of the channel feature map will decrease, and its influence on the final output will also be reduced.

[0137] In this model, the SE attention mechanism plays a crucial role. It enables the model to automatically focus on the importance of different channel features, strengthening important channel features and weakening unimportant channel features. In the ship radiated noise classification task, the SE attention mechanism can more effectively extract key features related to the ship type and suppress irrelevant or interfering features, thereby improving the model's ability to extract ship radiated noise features and classification accuracy, helping the model to more accurately identify ship types in complex ship radiated noise environments.

[0138] Inputting the features processed by the TriFusion block into the NLARN network has significant advantages. From a feature perspective, the MFCC, CQT, and Fbank multi-dimensional features extracted by the TriFusion block respectively simulate the characteristics of human hearing to well characterize steady-state features, excel at capturing high-frequency transient components, and highlight low-frequency trends, comprehensively covering the various characteristics of ship radiated noise, providing rich information for NLARN and helping it learn complex features. From a network structure perspective, NLARN's lightweight design enables it to quickly process input features and reduce computing resource consumption, and its attention mechanism further enhances its learning ability and stability for complex features. In actual ship radiated noise classification applications, this combination performs excellently, effectively improving classification accuracy and achieving accurate classification in complex underwater acoustic environments. It is an efficient and reliable solution for underwater acoustic target recognition. When the subsequent network structure includes long-term and short-term attention LSTM and Mamba, NLARN has more advantages as a front-end network. Its lightweight design meets overall efficiency requirements, enabling rapid front-end feature extraction, reducing subsequent network burdens and avoiding prolonged training times and excessive hardware resource consumption. The SE attention mechanism improves feature extraction quality and model stability, providing more representative and stable feature inputs for LSTM and Mamba, helping them learn long-term dependencies and complex patterns in sequences. Furthermore, NLARN's high classification accuracy in ship radiated noise classification tasks allows it to efficiently process input data while ensuring accuracy, collaborating with LSTM and Mamba to enhance the performance of the entire model.

[0139] Long-short-term attention LSTM module

[0140] The long-short-term attention (LSTM) module, as a temporal feature modeling unit, receives the feature vector sequence output by the NLARN and controls the flow of information through a gating mechanism consisting of a forget gate, input gate, output gate, and memory unit. Combined with the long-short-term attention mechanism, it captures the long-term and short-term dependencies of sonar signals. The filtered feature representations are then passed to the Mamba module.

[0141] The LSTM module is an improved version of the long short-term memory network (LSTM). Figure 6 As shown in Figure 2, it is specifically designed to solve the gradient vanishing and gradient exploding problems of traditional RNN when processing long sequences. At the same time, the introduction of the attention mechanism gives LSTM the ability to dynamically focus on the key parts of the input sequence.

[0142] The main function of the forget gate is to determine the memory unit C at the previous moment t-1 What information in the memory needs to be forgotten? When processing sequence data, some information may become less important over time. The forget gate can help the model filter out this useless information, preventing the memory unit from being flooded with irrelevant information, so that the model can focus on important information. The specific formula is:

[0143] f t =σ(W f ·[h t-1 ,x t ]+b f ) (11)

[0144] Among them, f t is the output of the forget gate at time step t, W f is the weight matrix of the forget gate, h t-1 is the hidden state of the previous moment, b f is the bias vector of the forget gate, σ is the activation function of sigmoid;

[0145] The input gate is mainly responsible for determining the current input x t Which information will be stored in the memory unit, and generate candidate memory units at the same time Used to update the memory unit; the control signal i of the input gate t and candidate memory cells The specific formula is:

[0146] i t =σ(W i ·[h t-1 ,x t ]+b i ) (12)

[0147]

[0148] Among them, Wi and W c are the weight matrices of the inputs and candidate memory units, b i and b c are the bias vectors of the input and candidate memory units respectively, and tanh is the hyperbolic tangent activation function;

[0149] The memory unit is responsible for storing and transmitting important information in the sequence. Through the synergy of the forget gate and the input gate, the memory unit can effectively update and retain information between different time steps. The specific formula is:

[0150]

[0151] Among them, c t and c t-1 are the memory unit states at the current moment and the previous moment respectively, is the element-wise multiplication operator;

[0152] The output gate determines the memory cell c t Which information in will be output as the hidden state h at the current moment t ; Hidden state h t As the final output, the long-term and short-term attention is calculated, and then combined with the fused context vector. The intermediate representation is generated through the fully connected layer and then mapped to the final prediction output. The specific formula is:

[0153] o t =σ(W o ·[h t-1 ,x t ]+b0) (15)

[0154]

[0155] Among them, t is the output of the output gate at time step t.

[0156] The long-short attention mechanism divides attention into short-term attention and long-term attention;

[0157] First calculate the weights and context vectors of long and short attention, as shown in the formula:

[0158]

[0159] Among them, score s and score1 are the scoring functions of short-term attention weight and short-term attention weight respectively. and They are short-term attention weight and short-term attention weight, and They are short-term context vector and long-term context vector respectively;

[0160]

[0161] Among them, γ is the proportion of long-term information in the final fusion with fusion weight ∈ [0,1];

[0162] After fusion, the final context vector is obtained:

[0163]

[0164] The current hidden state h of LSTM t and the fused context vector c t Combine them, generate an intermediate representation through the fully connected layer, and then map it to the final prediction output:

[0165]

[0166] The Long-Short-Term Attention (LSTM) module offers unique advantages in underwater sonar target recognition. The LSTM's gating mechanism enables it to effectively capture long-term and short-term dependencies in sonar signals, overcoming the limitations of traditional RNNs in processing long sequences and ensuring that important historical or future information is not lost. The integration of the attention mechanism further enhances the model's sensitivity and focus on key information in sonar signals. In underwater environments, sonar signals are complex and variable, containing a large amount of redundant and interfering information. The attention mechanism enables the model to automatically focus on the time segments or features most critical for target recognition, such as the specific frequency components or appearance time of the target echo. This precise information extraction and processing capability provides more valuable feature representation for the subsequent Mamba module, significantly improving the accuracy and robustness of underwater sonar target recognition and significantly promoting the development of underwater target recognition technology.

[0167] Mamba Module

[0168] As the final processing unit, the Mamba module first stabilizes the input data distribution through layer normalization to enhance noise resistance. The core state-space model module (SSM Block) is based on a discrete-time state-space model, captures long-sequence dependencies with linear computational complexity, and outputs a state vector containing long-range temporal features. Finally, the input features are added to the output features of the state-space model module through residual connections to solve the gradient vanishing problem of deep networks and ultimately output the target classification results.

[0169] Mamba modules such as Figure 7 As shown, it has a unique structure and principle, which can effectively solve the problem of underwater sonar target recognition.

[0170] In the Mamba module, layer normalization stabilizes data distribution through the formula LN(x), which is:

[0171]

[0172] Among them, μ is the mean of feature dimension, σ 2 is the variance, ε prevents division by zero, γ and β are learnable parameters;

[0173] The core state space model module is based on the discrete time state space model Capturing long sequence dependencies by iterative state updates, where s t is the state vector, u t is the input vector, y t is the output vector, and A, B, C, D are the model parameter matrices. Its computational complexity is close to linear O(n), compared to the O(n) of the traditional self-attention mechanism. 2 ), when processing long sequences of sonar signals, it can accurately capture long-range dependencies at a lower computational cost, effectively identifying target object characteristics such as distance and speed. Finally, the residual connection output = x + h helps address the vanishing gradient problem in deep networks, ensuring efficient information flow and enabling the network to stably transmit information even in complex noisy environments, maintaining model performance and improving sonar target recognition accuracy.

[0174] In the underwater sonar target recognition system, the Mamba module works closely with the TriFusion block, NLARN, and long-short-term attention LSTM module to improve recognition performance.

[0175] The TriFusion block is responsible for multi-domain feature extraction, converting the raw sonar signal into a composite feature vector containing MFCC, CQT, and Fbank features, comprehensively covering the signal's steady-state, high-frequency transient, and low-frequency trend information. These features serve as input to the NLARN, providing it with a rich information foundation.

[0176] NLARN performs in-depth processing on the features output by the TriFusion block. Its lightweight structure rapidly processes input features, reducing computational resource consumption. The SE attention mechanism focuses on key features, improving feature extraction quality. The processed features are then passed to the Long-Short-Term Attention (LSTM) module.

[0177] The LSTM module receives the output of the NLARN and uses a gating mechanism to capture the long-term and short-term dependencies of sonar signals. It combines the LSTM module with the long-term and short-term attention mechanism to dynamically focus on key information, screen out more valuable feature representations, and provide high-quality input for the Mamba module.

[0178] The Mamba module performs final optimization on the features output by the Long-Short-Term Attention (LSTM) module. Its normalization layer stabilizes data distribution and enhances the model's robustness to noise. The SSM Block captures long-sequence dependencies with near-linear computational complexity, accurately identifying key features of target objects. Residual connections ensure efficient information flow, addressing the vanishing gradient problem in deep networks. Through these operations, the Mamba module efficiently integrates and analyzes the information processed by the front-end modules, significantly improving the accuracy and reliability of final target classification and completing the entire underwater sonar target recognition task.

[0179] In the underwater sonar signal processing pipeline, the TriFusion block first fuses the signals, the NLARN extracts features, and the Long-Short-Term Attention (LSTM) further focuses on key information. The Mamba module then plays an indispensable role. As a key component of the entire process, it further optimizes and strengthens feature processing and target recognition capabilities, efficiently integrating and analyzing information processed by the previous modules, significantly improving the accuracy and reliability of final target classification. In summary, the Mamba module, with its outstanding advantages in long-sequence modeling and anti-interference stability, effectively addresses the shortcomings of traditional underwater sonar target recognition methods, bringing new breakthroughs and development opportunities to underwater target recognition technology. It is a key innovative component in solving the current challenges of underwater sonar target recognition.

[0180] Verification and Analysis

[0181] In the challenging research field of underwater sonar target recognition, the performance of the model is directly related to the effectiveness of practical applications. In order to deeply analyze the characteristics of the present invention, explore its potential and clarify its applicability in complex underwater environments, a series of comprehensive and in-depth experiments will be carried out. These experiments cover multiple key dimensions, from the verification of the effectiveness of the feature extraction method to the in-depth exploration of the functions of each module of the model; from detailed comparison with classical models in terms of parameters, computational complexity, etc., to the test of the generalization ability of the model on different data sets. Through a carefully designed experimental scheme and the use of a variety of evaluation indicators, the advantages of the present invention can be fully revealed, providing a solid experimental basis and innovative ideas for the development of underwater sonar target recognition technology.

[0182] Experimental dataset

[0183] DeepShip dataset

[0184] In order to comprehensively and systematically evaluate the model performance, the DeepShip dataset developed by Northwestern Polytechnical University was selected. This dataset focuses on the field of underwater acoustic analysis and provides rich and valuable data support for the research.

[0185] The DeepShip dataset contains recordings from 265 different types of ships, including cargo ships, passenger ships, tankers, tugboats, and other common vessel types. All recordings were made in the waters of the Georgia Strait Delta, at depths between 141 and 147 meters. The study focused on vessels within a 2-kilometer radius of the sonar, ensuring targeted and effective data collection.

[0186] During data preprocessing, all WAV audio files were converted to a 22,050Hz sampling rate, and the underwater acoustic data was segmented into 5-second segments. This series of processing ultimately generated over 30,000 labeled audio samples. To ensure model training stability and accuracy and reduce the risk of overfitting, the dataset was carefully divided into training, validation, and test sets in an 8:1:1 ratio. The detailed divisions are shown in Table 1 below, which provides detailed information on the four categories of the Deepship dataset after preprocessing.

[0187] Table 1

[0188]

[0189] ShipsEar Database

[0190] The ShipsEar database, a benchmark dataset for underwater acoustic target recognition, is widely used in scientific research. Collected off the Atlantic coast of Spain, it includes ship noise and a variety of man-made and natural sounds. The database contains 90 WAV-formatted recordings, categorized into five categories based on various characteristics.

[0191] During data preprocessing, all audio was standardized to a sampling rate of 22,050 Hz and segmented into 5-second segments, ultimately extracting 2,223 labeled audio samples. To facilitate model training and evaluation, these samples were divided into 1,778 training samples and 445 test samples in an 8:2 ratio. Specifically, the content and sample counts of the five categories are shown in Table 2, which provides detailed information about the four categories after preprocessing.

[0192] Table 2

[0193]

[0194] This data set division method enables the model to fully learn the acoustic characteristics of different ship types during training, continuously adjust and optimize model parameters through the validation set, and finally use the test set to accurately evaluate model performance, laying a solid data foundation for related research on underwater acoustic target recognition.

[0195] Experimental setup

[0196] During model training, the Lion optimizer with a momentum of 0.9 was used to effectively reduce noise interference in the samples. The model was trained for 200 epochs, and the initial learning rate of 0.0004 was adjusted using a cosine decay function to achieve optimal learning speed. During training, the batch size was set to 64, and the LMF loss function was used as the primary evaluation metric. In the LMF loss, α = 1, β = 1, and γ = 2 were set.

[0197] Experimental environment and evaluation indicators

[0198] To minimize inconsistencies between experiments, a meticulous research methodology was employed. This included comprehensive training and evaluation of multiple models, with both qualitative and quantitative assessments of their performance. Extensive comparative studies were also conducted to comprehensively assess the effectiveness of the algorithms.

[0199] The computing infrastructure used in this study includes the Windows 11 operating system, an Intel Core i7-12700H processor with 32GB of random access memory (RAM), an NVIDIA GeForce GTX 3070TI graphics processing unit (GPU), and PyTorch version 1.4.0. The following sections will provide a more in-depth discussion of the experimental results.

[0200] The following table lists key performance metrics used to evaluate model accuracy: Precision refers to the proportion of true positive examples accurately identified by the classifier; Recall measures the proportion of correctly predicted positive examples out of all actual positive examples. F1-score is an important metric in classification tasks, calculating the harmonic mean of precision and recall. Support represents the number of examples for each class in the validation dataset.

[0201] In order to evaluate the network's ability to recognize patterns on a given dataset, precision, recall, and F1-score are used as evaluation metrics. Their respective mathematical formulas are as follows:

[0202]

[0203] Here, TP represents the correct answer of the prediction, FP represents the case where the category is incorrectly predicted from other categories, and FN represents the case where the label of the category is predicted as the label of other categories.

[0204] Ablation experiments

[0205] ShipsEar feature ablation experiment

[0206] To demonstrate the ability of the feature extraction method proposed in this paper to characterize raw underwater acoustic signals, Table 3 provides an extensive comparison of various feature extraction methods on the ShipsEar dataset. Table 3 shows the recognition accuracy of the MultiFuseNet-AID model using different features on the ShipsEar dataset. This comparison includes the original 2D features, the corresponding 3D features, and the 3D feature fusion method proposed in this paper.

[0207] Table 3

[0208]

[0209] In order to verify the characterization ability of the feature extraction method proposed in this invention on the original underwater acoustic signals, a variety of feature extraction methods were compared on the ShipsEar dataset, involving original two-dimensional features, three-dimensional features and the TriFusionblock three-dimensional feature fusion method of this invention. Models using different features cover a variety of classic network structures.

[0210] The tabular data shows that different network structures combined with different features produce varying recognition accuracy. For example, in the classic ResNet series, as network depth increases, recognition accuracy decreases when using the same features. For example, when using MFCC features for ResNet18 to ResNet101, the accuracy drops from 91.90 to 89.30. Among the different features, the TriFusion block demonstrates a clear advantage over single features such as MFCC, 3D_MFCC, Fbank, and CQT.

[0211] In various network models, the recognition accuracy of models using TriFusion block features is generally higher than that of models using other single features. Under the ResNet18 model, the recognition accuracy of the TriFusion block feature is 96.43, which is higher than MFCC's 91.90, 3D_MFCC's 92.81, Fbank's 95.97, and CQT's 94.60; under the Efficientnet_b0 model, the accuracy of the TriFusion block feature is 97.10, which is also higher than the corresponding accuracy of other single features. In the comparison of the present invention itself, the recognition accuracy of the TriFusion block feature reached 99.77, far exceeding MFCC's 92.41, 3D_MFCC's 93.97, Fbank's 97.31, and CQT's 96.63.

[0212] This fully demonstrates that the advantage of the TriFusion block is that by processing the original signal, differential signal and cumulative signal in parallel, extracting and fusing MFCC, CQT and Fbank features, it can comprehensively capture the multi-scale characteristics of the ship radiation noise signal, such as the steady-state spectrum structure, high-frequency transient changes and low-frequency trends, effectively overcome the limitations of the single feature extraction method, reduce information loss and redundancy, and provide the model with a more discriminative composite feature vector, significantly improving the model's performance in underwater sonar target recognition tasks, enabling it to achieve higher recognition accuracy under different network structures.

[0213] In the study of underwater sonar target recognition, in order to deeply analyze the performance of the model and compare the differences in feature extraction and classification capabilities of different models, classic models such as RESNET18, Efficientnet_bo, and Densnet121 are often used for reference and comparison, while this invention is the research focus. The t-SNE graph can project high-dimensional data into a two-dimensional coordinate system through nonlinear dimensionality reduction, displaying the data distribution and clustering, and visually observing the different models’ effects on distinguishing underwater acoustic signal features; the confusion matrix can clearly show the classification accuracy of the model on samples of each category, including correct classification and misclassification. Next, we will use Figure 8 The t-SNE plots and confusion matrices of RESNET18, Efficientnet_bo, Densnet121, and the present invention are displayed. From the perspectives of data visualization and classification result statistics, the performance of each model in the underwater sonar target recognition task is deeply analyzed, the advantages and disadvantages of each model are clarified, and the performance improvement effect of the present invention in this field is evaluated.

[0214] Observing the t-SNE graph of ResNet18, we can find that although the data points of each category tend to cluster, there is significant overlap between some categories, such as ClassA and ClassD, and ClassA and ClassB. This means that ResNet18 cannot effectively distinguish the features of different categories when extracting underwater acoustic signal features. The category boundaries are blurred in low-dimensional space, which can easily lead to classification errors. Its feature extraction ability is limited in underwater sonar target recognition tasks. In the graph of Efficientnet_bo, the degree of clustering of data points of each category has improved, but a small number of data points still deviate from the cluster area of ​​their respective categories. For example, some points in ClassB are close to the ClassA area. This shows that it can better extract features than ResNet18, making data points of the same category more concentrated. However, it is not accurate enough in capturing the subtle differences between different categories of signals, and feature confusion still exists. In the t-SNE diagram of Densnet121, the distribution of data points in each category is more compact, and the overlapping areas between categories are reduced. However, at some category boundaries, such as between ClassA and ClassB, there is still a phenomenon of data point mixing, indicating that it has advantages in feature extraction and can make similar data more tightly clustered, but the ability to distinguish the characteristics of similar category signals needs to be improved. In the t-SNE diagram of the present invention, the data points of each category are clearly and compactly clustered, the boundaries between different categories are clear, and there is almost no mixing. This is because its unique TriFusion block module can extract fused features from multiple domains and explore the multi-scale characteristics of underwater acoustic signals. In addition, with the collaboration of other modules, it effectively avoids feature confusion and improves feature discrimination. Compared with the other three models, it has outstanding advantages in feature extraction and data differentiation, can provide high-quality features for accurate classification, and improve the accuracy and reliability of underwater sonar target recognition.

[0215] Comparing the t-SNE plots of the ResNet18, EfficientNet_bo, Densnet121, and MultiFuseNet-AID models reveals differences in their ability to extract underwater acoustic signal features and distinguish between categories. While the first three models each have their own unique characteristics, they all suffer from certain deficiencies in feature differentiation. However, the present invention, with its unique architecture and modular design, excels in feature extraction and data differentiation, demonstrating a strong advantage in underwater sonar target recognition tasks and providing new and effective approaches for model optimization and performance improvement in this field.

[0216] Deep ship feature ablation experiment

[0217] In the field of underwater sonar target recognition, model performance is highly dependent on effective feature extraction and module collaboration. Previous research based on the Shipsear dataset initially validated the contributions of each model module, but the dataset was relatively limited. To further explore the model's performance on larger and more complex data, the Deep Ship dataset was introduced. Feature ablation experiments were conducted on this dataset to systematically investigate the model's performance after removing different key features or modules. Table 4 below presents the results of feature ablation experiments based on the Deep Ship dataset. By analyzing various metrics under different model configurations, we can more accurately assess the specific contribution of each feature and module to the model's target recognition capabilities in complex scenarios, providing key insights for model optimization and improvement. Table 4 shows the recognition accuracy of the MultiFuseNet-AID model on the Deep Ship dataset using different features.

[0218] Table 4

[0219]

[0220] We conducted experiments on the Deep Ship dataset using different features and compared various classic network structures. Overall, as the depth of the ResNet series network increases, the recognition accuracy decreases when using the same features. For example, the accuracy of ResNet18 using MFCC features is 93.90, while the accuracy of ResNet101 using MFCC features drops to 91.30.

[0221] Specifically for different features, compared with single features such as MFCC, 3D_MFCC, Fbank, and CQT, TriFusionblock has obvious advantages. Under the ResNet18 model, the recognition accuracy of the TriFusion block feature is 97.00, which is higher than MFCC's 93.90, 3D_MFCC's 94.20, Fbank's 94.70, and CQT's 96.30; under the Efficientnet_b0 model, the TriFusion block feature accuracy is 97.10, which is higher than the accuracy corresponding to other single features. In the comparison of the present invention itself, the recognition accuracy of the TriFusion block feature reached 98.39, far exceeding MFCC's 93.92, 3D_MFCC's 94.92, Fbank's 96.47, and CQT's 97.58.

[0222] When using various features, the overall recognition accuracy of Densnet121 is relatively low, which may be related to the adaptability of the network structure in processing such underwater acoustic signal characteristics.

[0223] Experiments on the Deep Ship dataset show that different network structures have different feature processing capabilities, and as network depth increases, the performance of some networks tends to decline. Compared with traditional single feature extraction methods, the TriFusionblock feature fusion method in this paper can significantly improve model recognition accuracy, showing advantages across different network structures. This shows that this method can more effectively extract and fuse underwater acoustic signal features, providing more discriminative feature representations for underwater sonar target recognition, and helping to improve model performance in complex data scenarios.

[0224] Experiments using different features were conducted on the Shipsear and Deep Ship datasets, and the proposed method was compared with various classic network structures. The experimental results show that, on the Shipsear dataset, the recognition accuracy of the ResNet series network decreases with increasing depth when using the same features. However, the TriFusion block exhibits significant advantages over single features such as MFCC, 3D_MFCC, Fbank, and CQT. Across all network models, models using the TriFusion block feature generally achieve higher recognition accuracy, with the proposed method showing even greater advantages in comparisons with its own model. The TriFusion block can fully capture the multi-scale characteristics of ship-radiated noise signals, reduce information loss and redundancy, and thus improve model performance. Experiments on the Deep Ship dataset also show that the performance of some network structures decreases with increasing depth, and the overall accuracy of Densnet121 is relatively low. The TriFusion block feature achieves higher recognition accuracy than single features across different network structures, demonstrating a significant advantage in the proposed method. This demonstrates that the TriFusion block can more effectively extract fused features in complex data scenarios, providing the model with more discriminative feature representations. Combining the experimental results of the two datasets, it can be concluded that different network structures have different abilities to process features, and increasing network depth may affect performance. The TriFusion block feature fusion method shows strong advantages on both datasets. Compared with the traditional single feature extraction method, it can better mine the characteristics of underwater acoustic signals and improve the accuracy and reliability of the model in underwater sonar target recognition tasks, providing an important reference and effective approach for model optimization and performance improvement in the field of underwater sonar target recognition.

[0225] ShipsEar module ablation experiment

[0226] In underwater sonar target recognition research, researchers conducted ablation experiments to further explore the impact of various model modules on performance. In the experiment, S1 represents the model without the novel lightweight residual attention network (NLARN), S2 represents the model without the long-short-term attention (LSTM) module, and S3 represents the model without the Mamba module. By successively removing different modules from the model and observing the changes in model performance, the specific role and contribution of each module in the overall model can be clarified, thereby verifying the rationality of the model design and the necessity of each module. Table 5 shows the results of the ablation experiment based on the Shipsear dataset. Table 5 shows the ablation experiment results for different modules in the Shipsear sample set.

[0227] Table 5

[0228]

[0229]

[0230] Based on the experimental results of the Shipsear dataset, the importance of each module can be clearly seen. Taking NLARN as an example, when the module is removed (ie, the S1 model), the overall performance of the model drops significantly. The average precision drops from 0.9967 of the present invention to 0.8941, the recall rate drops from 0.9988 to 0.9087, the F1-Score drops from 0.9977 to 0.8974, and the accuracy drops from 0.9977 to 0.8946. This is because NLARN is based on the improvement of the ResNet architecture, and the streamlined number of layers reduces the model complexity and computational cost. The SE attention mechanism it integrates can automatically learn the dependencies between channels, highlight key features, suppress interference information, and improve the ability to extract ship radiation noise characteristics, avoid overfitting, and provide high-quality feature representation for subsequent modules, which plays a vital role in improving the classification accuracy and stability of the model.

[0231] The importance of the long-short-term attention (LSTM) module cannot be underestimated. Removing this module (the S2 model) reduced model performance, with average precision dropping to 0.9860, recall to 0.9877, F1-Score to 0.9862, and accuracy to 0.9868. This module, combining gating and attention mechanisms, effectively captures long-term and short-term dependencies in sonar signals, overcoming the vanishing gradient problem experienced by traditional RNNs processing long sequences. It also adaptively filters key information, precisely focusing on time segments or features within sonar signals that are crucial for target recognition. This provides more valuable feature input for the subsequent Mamba module, significantly improving target recognition accuracy and reliability.

[0232] The Mamba module is also an indispensable part of the model. After removing the Mamba module (S3 model), the model performance also declined, with the average precision dropping to 0.9699, the recall rate dropping to 0.9553, the F1-Score dropping to 0.9613, and the accuracy dropping to 0.9621. The Mamba module's normalization layer stabilizes data distribution and enhances the model's robustness to noise. The core SSM Block, based on the state-space model, captures long-sequence dependencies with near-linear computational complexity, accurately identifying target object feature information such as distance and speed. The residual connection ensures the effective flow of information, solves the vanishing gradient problem in deep networks, and improves the accuracy of sonar target recognition. It plays a key role in optimizing and strengthening feature processing and target recognition capabilities in the entire model.

[0233] In the study of the underwater sonar target recognition model of this invention, in order to deeply analyze the impact of each module on the model performance, an ablation experiment was carried out based on the Shipsear dataset. By comparing the performance differences between the complete model and the models with different modules removed (S1 without NLARN, S2 without the long-short-term attention LSTM module, S3 without the Mamba module), the importance of each module has been understood from indicators such as precision and recall. The confusion matrix can more intuitively show the classification of the model on samples of different categories, and clearly show the distribution of the model's misjudgments and correct judgments. The following confusion matrix diagram, Figure 9 The classification results of the complete model of the present invention and the S1, S2, and S3 models on the Shipsear dataset will be presented respectively. The impact of the missing modules on the classification accuracy of the model will be analyzed at a more detailed level, and the key role of each module in accurately identifying different types of ship radiated noise targets will be further clarified.

[0234] Deep ship module ablation experiment

[0235] In underwater sonar target recognition research, the Shipsear dataset, previously used, is relatively small. Considering that dataset size may affect experimental results and, in turn, hinder the accurate assessment of the contribution of each model module, we conducted ablation experiments on the larger Deep Ship dataset to more comprehensively and reliably validate the contributions of different model modules. Compared to the Shipsear dataset, the Deep Ship dataset contains a richer and more diverse sample of ship-radiated noise, covering more complex real-world scenarios, and can more thoroughly test the model's performance under different conditions. The experimental results presented in Table 6 allow for a deeper analysis of the contribution of each model module, observing how model performance changes when removing different modules on a larger dataset. This allows for a more precise understanding of the contribution of each module to the overall model performance. Table 6 shows the ablation experiments for different modules on the Deep Ship dataset.

[0236] Table 6

[0237]

[0238] In an ablation experiment based on the Deep Ship sample set, the tabular data shows that each module has a significant impact on model performance. When the novel lightweight residual attention network (NLARN) is removed to become the S1 model, overall performance decreases significantly, with the average precision dropping from 0.9838 of the present invention to 0.9185, the recall rate dropping from 0.9837 to 0.9182, the F1-Score dropping from 0.9837 to 0.9173, and the accuracy dropping from 0.9837 to 0.9178. This demonstrates that NLARN can also reduce computational costs through a streamlined architecture when processing large-scale data. Its SE attention mechanism effectively improves feature extraction capabilities, which is crucial for maintaining the model's high classification accuracy and stability.

[0239] After removing the long- and short-term attention LSTM module to form the S2 model, performance declined somewhat, with average precision dropping to 0.9623, recall dropping to 0.9618, F1-Score dropping to 0.9616, and accuracy dropping to 0.9620. This demonstrates that, in the larger Deepship dataset, this module, combined with gating and attention mechanisms, can still effectively capture long- and short-term dependencies, filter key information, and provide valuable feature input for subsequent modules, playing a significant role in improving object recognition accuracy and reliability.

[0240] The S3 model without the Mamba module also experienced performance degradation, with average precision dropping to 0.9532, recall to 0.9482, F1-Score to 0.9498, and accuracy to 0.9492. This demonstrates that when processing the complex and diverse Deep Ship dataset, the Mamba module's normalization layer stabilizes data distribution and enhances the model's noise immunity. The SSM Block efficiently captures long-term dependencies, and residual connections ensure information flow, optimizing and strengthening feature processing and object recognition capabilities.

[0241] Previous experiments based on the Deep Ship sample set, analyzing metrics such as average precision, recall, F1-score, and accuracy, concluded that the novel lightweight residual attention network (NLARN), the long-short-term attention LSTM module, and the Mamba module all significantly contributed to improved model performance. However, these metrics only provide a holistic view of model performance. To gain a deeper and more comprehensive understanding of the model's classification performance on samples of different categories, particularly the misclassification of different categories when modules are missing, the confusion matrix is ​​a very effective tool. Figure 10The confusion matrices shown here correspond to the classification results of the Deep Ship sample set for the complete model and the S1, S2, and S3 models after removing different modules. By observing these confusion matrices, we can more intuitively identify problems with the model's classification of various sample categories and accurately pinpoint the impact of each module on the recognition of different object categories, providing a strong basis for further model optimization.

[0242] Based on the ablation experiment results of the Shipsear dataset and the Deep Ship sample set, the new lightweight residual attention network (NLARN) can improve feature extraction capabilities, reduce model complexity, and ensure model accuracy and stability in both datasets through architectural optimization and attention mechanisms. The long-short-term attention (LSTM) module can use gating and attention mechanisms to capture long-term and short-term dependencies and filter key information under different data scales, helping to improve target recognition accuracy. In both datasets, the Mamba module relies on normalization layers, SSM blocks, and residual connections to stabilize data, capture long-term sequence dependencies, and ensure information flow, thereby enhancing feature processing and target recognition capabilities.

[0243] This invention uses the TriFusion block module to extract features from multiple domains, including the original signal, differential signal, and cumulative signal, providing rich information for subsequent modules. The NLARN optimizes and processes these features to improve their quality. The LSTM module captures long- and short-term dependencies and screens key features. The Mamba module further optimizes these features to enhance the model's noise immunity and recognition accuracy. These modules work closely together, from feature extraction and optimization to dependency capture and anti-interference recognition, to build a complete and efficient underwater sonar target recognition network architecture. This architecture demonstrates strong performance across datasets of varying sizes, effectively addressing challenges in underwater sonar target recognition and significantly improving the accuracy and robustness of ship radiated noise classification.

[0244] Performance Analysis

[0245] The performance of the present invention will be compared with the existing most advanced target recognition model under the same experimental conditions, and examined from various aspects.

[0246] Parameter analysis

[0247] In underwater sonar target recognition research, the performance of this method depends not only on recognition accuracy, but also on its complexity and computational resource consumption. An efficient model must ensure accurate target recognition while minimizing hardware resource requirements to adapt to practical application scenarios. Figure 11This paper compares the parameters (params) and floating-point operations (FLOPs) of the proposed method with classic models such as the ResNet series. By observing this data, we can intuitively understand the differences in complexity and computational effort between different models, analyze the advantages and limitations of each model in practical deployment, and clarify the proposed method's performance in resource utilization efficiency.

[0248] from Figure 11 It can be clearly seen that the present invention is significantly lower than the Resnet series models in terms of parameter quantity and floating-point operation number. In terms of parameter quantity, Resnet101 is as high as 44.55M, while the present invention is only 18.13M; in terms of floating-point operation number, Resnet101's 7.87G is much higher than the present invention's 0.51G. This shows that the present invention has successfully achieved lightweight design. Its unique architecture design and module combination, such as NLARN, effectively reduce the number of parameters and calculation amount of the model. This lightweight feature is of great significance in practical applications, which means that the model can run efficiently on resource-constrained underwater equipment, reduce the requirements for hardware performance, and at the same time ensure the accuracy of target recognition, providing a more feasible solution for the optimization and actual deployment of underwater sonar target recognition systems.

[0249] Computational bottleneck analysis

[0250] Through in-depth research and testing of the present invention, different models show different performance when faced with diverse sample data. In order to explore the impact of the sample label ratio on the model effect, multiple sets of experiments were conducted on each model. It can be clearly seen from the experimental data that as the sample label percentage gradually increases from 1% to 100%, each model shows a different trend of change. These data intuitively reflect the intrinsic connection between the sample label ratio and model performance, and provide a key basis for the optimization and application of subsequent models. Specific data are shown in Table 7, which shows the label ratios of different training Shipsear data sets and the recognition accuracy of different models.

[0251] Table 7

[0252]

[0253] Judging from the tabular data, on the Shipsear dataset, when the sample label ratio is only 1%, the present invention leads other models with a recognition accuracy of 50.03%. The recognition accuracies of Densnet121, Efficientnet_bo, and RESNET18 are 48.40%, 45.21%, and 42.72%, respectively, indicating that the present invention has better adaptability in the case of small sample labels. As the sample label ratio increases to 10%, the recognition accuracy of each model increases significantly, among which Densnet121 has the largest increase, from 48.40% to 80.44%. When the label ratio reaches 50%, the recognition accuracy of Efficientnet_bo reaches 94.16%, slightly higher than Densnet121's 93.97%, and both are higher than RESNET18's 92.41%. When the sample label ratio reaches 100%, the recognition accuracy of the present invention reaches 99.77%, which is the best performance among all models and has a significant advantage over other models. The recognition accuracy of Densnet121, Efficientnet_bo, and RESNET18 are 97.99%, 97.10%, and 96.43%, respectively. Overall, as the sample label ratio increases, the recognition accuracy of all models gradually improves, but the present invention maintains a high recognition accuracy at different label ratios, showing good stability and generalization ability, and has a prominent advantage in model selection for this dataset. Other models have different performances at different label ratio stages, which also provides more references for model application in specific scenarios.

[0254] Model identification accuracy analysis

[0255] In order to verify whether the present invention is superior to existing mainstream object recognition models, the present invention and other mainstream models were trained and verified. Table 8 shows the comparison results of different models on the ShipsEar dataset.

[0256] Table 8

[0257]

[0258] Table 8 shows experimental results on the ShipsEar dataset. The performance of various mainstream object recognition models varies across metrics such as accuracy, precision, recall, and F1-score. Yamnet's performance was relatively low, with an accuracy of only 0.7872 and a precision of 0.6864, indicating poor overall performance. VGGish improved over Yamnet, reaching an accuracy of 0.8675, but still lagged behind more advanced models. While ADCNN and CRNN9 showed some progress, their overall performance still fell short of leading models. The LSTM-based model's accuracy improved to 0.9477, demonstrating good performance, but still lagging behind leading models.

[0259] The VFR and Mobile_Vit models performed well, with accuracy rates reaching 0.9850, but they are still inferior to the present invention. The Hybrid model performed well in some indicators, but its overall performance was still inferior to the present invention. The present invention achieved extremely high levels in various indicators, with accuracy, precision, recall and F1-Score all reaching 0.9977 or above, significantly better than other mainstream models, and with obvious advantages in recognition accuracy and comprehensiveness. The experimental results fully demonstrate that the performance of the present invention on the ShipsEar dataset surpasses existing mainstream target recognition models, and it has outstanding superiority and strong competitiveness, showing great application potential in the field of target recognition.

[0260] Generalization experiment

[0261] In practical model applications, generalization ability is a key metric for measuring a model's quality. It determines the model's adaptability and reliability under different data distributions and task scenarios. To comprehensively evaluate the generalization performance of each model and verify its stable performance in complex and changing data environments, the DeepShip model was introduced and placed alongside other mainstream models in a rigorous and diverse experimental environment. Table 9 compares the performance of each model under the same conditions through training and validation on multiple datasets with different characteristics, thereby deeply analyzing the differences in generalization ability between DeepShip and other models. This provides solid data support and theoretical basis for model optimization and practical application. Table 9 shows the comparison results of different models on the DeepShip dataset.

[0262] Table 9

[0263]

[0264] In testing on the Deepship dataset, the performance of various models varied significantly. Yamnet and VGGish performed relatively poorly overall, with accuracies of only 0.6953 and 0.6685, respectively. All metrics were relatively low, making them unable to meet high-precision recognition requirements. While CRNN9 and MobilenetV2 showed improvements over the previous two, reaching accuracies of 0.8614 and 0.9018, respectively, they still lagged behind the best models. ADCNN achieved an accuracy of 0.9023, which was acceptable for this group of models, but still had room for improvement in overall performance. Mobile_Vit and CA_MobilenetV2 achieved accuracies of 0.9107 and 0.9350, respectively, showing good performance, but still lagging behind the top models. The VFR model demonstrated high performance, reaching an accuracy of 0.9380 and showing balanced performance across all metrics. The present invention stands out among all models, with accuracy, precision, recall and F1-Score all reaching 0.9837, far exceeding other models in recognition accuracy and stability, fully demonstrating its strong generalization ability and excellent performance advantages on the Deepship dataset. Compared with other models, it is obviously competitive and provides a more reliable solution for target recognition tasks.

[0265] Discussion and Analysis

[0266] Based on the experimental results, the present invention has demonstrated many significant advantages in the field of underwater sonar target recognition. In terms of feature extraction, its unique TriFusion block module can extract and fuse MFCC, CQT, and Fbank features from the original signal, differential signal, and cumulative signal in multiple domains in parallel, comprehensively capturing the multi-scale characteristics of the ship's radiated noise signal, such as the steady-state spectral structure, high-frequency transient changes, and low-frequency trends. This makes the recognition accuracy of the model significantly higher than that of using traditional single features on different data sets (such as ShipsEar and Deep Ship data sets), effectively improving the model's ability to characterize complex underwater acoustic signals and reducing information loss and redundancy.

[0267] From the perspective of model architecture modules, the new lightweight residual attention network (NLARN) reduces model complexity and computational cost by streamlining the number of layers. The integrated SE attention mechanism can automatically learn inter-channel dependencies, highlight key features, and suppress interference information. It plays an important role in improving feature extraction capabilities and ensuring model accuracy and stability in datasets of different sizes. The long-short-term attention LSTM module combines gating and attention mechanisms to effectively capture the long-term and short-term dependencies of sonar signals, overcome the vanishing gradient problem of traditional RNNs when processing long sequences, adaptively filter key information, provide more valuable feature inputs for subsequent modules, and improve target recognition accuracy and reliability. The Mamba module stabilizes data distribution with a normalization layer, enhancing the model's robustness to noise. The core SSM Block captures long-sequence dependencies with near-linear computational complexity, accurately identifying feature information such as the distance and speed of target objects. Residual connections ensure the effective flow of information, solve the vanishing gradient problem of deep networks, and enhance feature processing and target recognition capabilities.

[0268] In terms of performance, the present invention has obvious advantages. Its parameter count and floating-point operations are significantly lower than those of classic models such as the Resnet series, achieving a lightweight design and being able to run efficiently on resource-constrained underwater equipment, reducing hardware performance requirements while ensuring target recognition accuracy. Under different sample label ratios, the model maintains high recognition accuracy, good stability and generalization ability. In comparison with mainstream target recognition models, whether on the ShipsEar dataset or the Deepship dataset, the present invention achieves extremely high levels of accuracy, precision, recall, and F1-Score, far exceeding most mainstream models and demonstrating strong competitiveness.

[0269] The present invention is provided as an example, not as a limitation of the embodiments. Those skilled in the art will appreciate that other variations or modifications may be made based on the above description. It is not necessary and impossible to enumerate all embodiments here, and obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. An underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling, characterized by: It includes the sequentially connected TriFusion block multi-domain feature fusion module, the new lightweight attention residual network, the long-short-term attention LSTM module and the Mamba module. Each module forms a hierarchical processing chain through data flow transmission; The TriFusion block multi-domain feature fusion module serves as a front-end processing unit, receiving the original sonar signal, differential signal, and cumulative signal in parallel, and extracting Mel-frequency cepstral coefficient features, constant Q transform features, and Mel-frequency spectrum features respectively; The three-branch features are concatenated into a three-channel feature tensor in the channel dimension according to the order of constant Q transform, Mel-frequency cepstral coefficient, and Mel-frequency spectrum. A weighted fusion formula is used to form a composite feature vector containing full-band information as the input of the new lightweight attention residual network. The new lightweight attention residual network is based on the streamlined ResNet architecture. It receives the composite feature tensor output by the TriFusion block, performs dimensionality reduction and spatial feature extraction on the features through 9 layers of lightweight convolutional layers, and embeds the SE attention module after the Conv2 layer. It "squeezes" the global information of the channel through global average pooling, and "stimulates" the generation of channel weights through the fully connected layer to enhance the key features and suppress the interference information. The feature vector after output dimension compression is transmitted to the long-short-term attention LSTM module. The Long-Short-Term Attention (LSTM) module, as a temporal feature modeling unit, receives the feature vector sequence output by the novel lightweight attention residual network. It controls the flow of information through a gating mechanism consisting of forget gates, input gates, output gates, and memory units. Combined with the Long-Short-Term Attention (LSTM) mechanism, it captures the long-term and short-term dependencies of sonar signals. The resulting feature representations are then passed to the Mamba module. As the final processing unit, the Mamba module first stabilizes the input data distribution through layer normalization to enhance noise resistance. The core state-space model module is based on a discrete-time state-space model, captures long-sequence dependencies with linear computational complexity, and outputs a state vector containing long-range temporal features. Finally, the input features are added to the output features of the state-space model module through residual connections to solve the gradient vanishing problem of deep networks and ultimately output the target classification results.

2. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 1 is characterized in that: In the three-branch feature extraction path of the TriFusion block module, the original signal is transformed through short-time Fourier transform, Mel frequency mapping and discrete cosine transform to generate Mel frequency cepstral coefficient features, focusing on the steady-state spectral structure; the differential signal is transformed through Hilbert transform and constant Q transform to generate constant Q transform features, capturing high-frequency transient components; the cumulative signal is decomposed by wavelet and Mel spectrum feature extraction to enhance the expression of low-frequency trends.

3. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 2 is characterized in that: In the TriFusion block module, the original signal is processed by Mel-frequency cepstral coefficients. First, a short-time Fourier transform is performed to obtain X(m,k), the time domain signal is converted to the frequency domain, and then the linear frequency is mapped to the Mel frequency to calculate the filter bank energy E i Finally, the Mel frequency cepstral coefficient feature is obtained through logarithmic transformation and discrete cosine transform. The specific formula is: Where x(n) is the audio information, w(n) is the Hamming window function, n is the frame index, k is the frequency index, N is the number of FFT points, R is the number of frame shift points, and M is the number of Mel filter banks.

4. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 3 is characterized in that: In the TriFusion block, the differential signal first performs a first-order difference on the original audio signal to obtain v(n), amplifying the instantaneous rate of change and highlighting the high-frequency transient components; then the envelope A(n) is extracted through the Hilbert transform to enhance the amplitude change information; finally, the constant Q transform is used to perform time-frequency analysis to obtain CQT(m,k), and the amplitude is converted to dB scale CQT dB (m,k), the specific formula is: v(n)=x(n+1)-x(n) (4) A(n)=|Hilbert(v(n))| (5) Among them, h(n) is the analysis window function, f k is a specific frequency point, and ref takes the maximum value of the constant Q transformation amplitude.

5. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 4 is characterized in that: In the TriFusion block module, the cumulative signal is first accumulated and calculated on the original signal to smooth high-frequency fluctuations and highlight low-frequency trends; then the cumulative signal is normalized and wavelet decomposition is performed to obtain d norm (n), take the approximate coefficient to separate the low-frequency component and suppress the high-frequency interference; finally, use a method similar to the Mel frequency cepstral coefficient to extract the Mel spectrum features and convert them into dB scale Fbank dB (m,i) and adjust the shape. The specific formula is: Among them, J is the number of decomposition layers equal to 5, c j,k is the wavelet coefficient, ψ j,k is the wavelet basis function.

6. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 5 is characterized in that: In the TriFusion block module, the weighted fusion formula is: Final=αMFCC n +βCQT dB (m,k)+γFbank dB (m,i) (10) Among them, α, β, and γ are weight coefficients, satisfying α+β+γ=1.

7. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 1 is characterized in that: The novel lightweight attention residual network is streamlined to 10 layers based on the ResNet18 architecture, which includes 9 convolutional layers and 1 fully connected layer. The SE attention module added after the Conv2 layer generates channel weight coefficients through ReLU and sigmoid activation functions.

8. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 1 is characterized in that: In the gating mechanism of the LSTM module, the forget gate helps the model filter out useless information when processing sequence data, preventing the memory unit from being flooded with irrelevant information, and allowing the model to focus on important information. The specific formula is: f t =σ(W f ·[h t-1 ,x t ]+b f ) (11) Among them, f t is the output of the forget gate at time step t, W f is the weight matrix of the forget gate, h t-1 is the hidden state of the previous moment, b f is the bias vector of the forget gate, σ is the activation function of sigmoid; The input gate is mainly responsible for determining the current input x t Which information will be stored in the memory unit, and generate candidate memory units at the same time Used to update the memory unit; the control signal i of the input gate t and candidate memory cells The specific formula is: i t =σ(W i ·[h t-1 ,x t ]+b i ) (12) Among them, W i and W c are the weight matrices of the inputs and candidate memory units, b i and b c are the bias vectors of the input and candidate memory units respectively, and tanh is the hyperbolic tangent activation function; The memory unit is responsible for storing and transmitting important information in the sequence. Through the synergy of the forget gate and the input gate, the memory unit can effectively update and retain information between different time steps. The specific formula is: Among them, c t and c t-1 are the memory unit states at the current moment and the previous moment respectively, is the element-wise multiplication operator; The output gate determines the memory cell c t Which information in will be output as the hidden state h at the current moment t ; Hidden state h t As the final output, the long-term and short-term attention is calculated, and then combined with the fused context vector. The intermediate representation is generated through the fully connected layer and then mapped to the final prediction output. The specific formula is: o t =σ(W o ·[h t-1 ,x t ]+b0) (15) Among them, t is the output of the output gate at time step t.

9. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 1 is characterized in that: In the attention mechanism of the long-short-term attention LSTM module, the long-short-term attention mechanism divides attention into short-term attention and long-term attention; First calculate the weights and context vectors of long and short attention, as shown in the formula: Among them, score s and score1 are the scoring functions of short-term attention weight and short-term attention weight respectively. and They are short-term attention weight and short-term attention weight, and They are short-term context vector and long-term context vector respectively; Among them, γ is the proportion of long-term information in the final fusion with fusion weight ∈ [0,1]; After fusion, the final context vector is obtained: The current hidden state h of LSTM t and the fused context vector c t Combine them, generate an intermediate representation through the fully connected layer, and then map it to the final prediction output:

10. The underwater sonar target recognition system based on multi-domain feature fusion and lightweight modeling according to claim 1 is characterized in that: In the Mamba module, layer normalization stabilizes data distribution through the formula LN(x), which is: Among them, μ is the mean of feature dimension, σ 2 is the variance, ε prevents division by zero, γ and β are learnable parameters; The core state space model module is based on the discrete time state space model Capturing long sequence dependencies by iterative state updates, where s t is the state vector, u t is the input vector, y t is the output vector, and A, B, C, D are the model parameter matrices.

Images