Ship identification method based on multiple acoustic feature fusion

Abstract

The invention discloses a ship identification method based on multiple acoustic feature fusion. The ship identification method comprises the following steps: acquiring and preprocessing underwater noise signals of a ship; carrying out MEL frequency spectrum, CQT frequency spectrum and STFT phase diagram feature transformation on the preprocessed ship underwater noise signals, and carrying out stacking on channel dimensions to obtain three-channel feature images; constructing a recognition neural network; performing feature extraction on the three-channel feature image through an asymmetric convolution feature extraction module to obtain a preliminary feature; strengthening the preliminary features through a feature strengthening module to obtain strengthened features; and performing classification identification on the reinforced features through a classification module to obtain an identification result. Through multiple acoustic feature extraction, deep information in original data is mined and utilized to the greatest extent while the calculation cost is reduced, the noise robustness is excellent, the model structure is simple, and the recognition effect can be effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of ship identification methods, and in particular relates to a ship identification method based on the fusion of multiple acoustic features. Background Art

[0002] Ship radiation noise contains the characteristic information of the ship. Passive sonar systems can remotely detect, locate and track underwater targets based on radiation noise. However, due to the complexity of the underwater environment and the antagonism of the target, underwater acoustic target recognition has always been a challenging task. In the early days, this task was mainly completed by sonar engineers through manual recognition, but the recognition ability was affected by environmental, psychological and physiological factors, and the accuracy of target recognition often had great uncertainty.

[0003] With the evolution of computer and artificial intelligence technology, many deep learning-based recognition technologies have been developed for ship acoustic recognition, such as multi-layer perceptron (MLP), convolutional neural network (CNN) and long short-term memory network (LSTM). These methods have the advantages of fast speed, high efficiency and high accuracy compared with traditional methods. However, the current deep learning technology usually relies on a single method for noise feature extraction, such as MEL feature transform, Mel frequency cepstral coefficient (MFCC) or constant Q transform (CQT), which may lead to partial loss of original feature information. At the same time, traditional recognition algorithms perform poorly in data expression and anti-noise performance, and have gradually failed to meet the processing requirements of the growing amount of data. In addition, conventional neural network models involve multi-layer feature extraction and recognition, so the computational complexity and cost are high, which increases the time and resources required for deployment. Summary of the invention

[0004] The object of the present invention is to provide a ship identification method based on the fusion of multiple acoustic features and to provide at least the advantages to be described below.

[0005] In view of the problems existing in the prior art, the present invention provides a ship identification method based on the fusion of multiple acoustic features. By extracting multiple acoustic features, it aims to reduce the computing cost while maximally mining and utilizing the deep information in the original data. It has excellent noise robustness, a simple model structure, and can effectively improve the recognition effect.

[0006] The technical solution of the present invention is as follows:

[0007] The ship identification method based on multiple acoustic feature fusion includes the following steps:

[0008] Obtain the underwater noise signal of the ship and pre-process it;

[0009] The pre-processed ship underwater noise signal is subjected to MEL spectrum, CQT spectrum and STFT phase Figure 3 The feature transformations are performed and stacked in the channel dimension to obtain a three-channel feature image;

[0010] Constructing a recognition neural network, wherein the recognition neural network includes two asymmetric convolution feature extraction modules, two feature enhancement modules and a classification module;

[0011] Performing feature extraction on the three-channel feature image through the asymmetric convolution feature extraction module to obtain preliminary features;

[0012] Strengthening the preliminary features by the feature enhancement module to obtain enhanced features;

[0013] The enhanced features are classified and identified through a classification module to obtain an identification result.

[0014] Preferably, in the ship identification method based on fusion of multiple acoustic features, the preprocessing includes pre-emphasis and resampling operations.

[0015] Preferably, in the ship identification method based on multiple acoustic feature fusion, the pre-processed ship underwater noise signal is subjected to MEL spectrum, CQT spectrum and STFT phase Figure 3 Feature transformations include:

[0016] The pre-processed underwater noise signal of the ship is divided into overlapping short time windows, and Fourier transform is applied to each window to obtain the spectrum of each time window. The spectrum of all time windows is spliced ​​together to form a time-frequency diagram;

[0017] Calculate the square of the time-frequency graph to obtain the power spectrum density, filter the power spectrum density through a set of Mel filters, each filter corresponds to a Mel frequency band, and take the logarithm of the filtering result to obtain the logarithmic Mel spectrum;

[0018] The preprocessed underwater noise signal of the ship is subjected to wavelet transform of different scales, each scale corresponds to a frequency band, and the wavelet transform result is reconstructed to obtain the CQT spectrum.

[0019] Preferably, in the ship identification method based on multiple acoustic feature fusion, the asymmetric convolution feature extraction module includes:

[0020] Two-dimensional convolutional layer, used to extract features of the input image, so as to capture the local features of the image;

[0021] The batch normalization layer normalizes the input of each mini-batch of data, that is, adjusts the data so that its mean is 0 and its variance is 1;

[0022] ReLU activation function introduces nonlinear characteristics to the network;

[0023] Residual layer, which solves the gradient vanishing problem in deep networks;

[0024] The spatial activation layer emphasizes important features and suppresses unimportant features by adaptively recalibrating feature channels;

[0025] and max pooling layers, which downsample or reduce the spatial size of the input feature map by selecting the maximum value in the covered area.

[0026] Preferably, in the ship identification method based on multiple acoustic feature fusion, the three-channel feature image is subjected to feature extraction by the asymmetric convolution feature extraction module to obtain preliminary features, including:

[0027] Small convolution kernels are used to extract spatial features, and batch normalization and ReLU activation functions are used to regularize the data and introduce nonlinearity.

[0028] Preferably, in the ship identification method based on multiple acoustic feature fusion, the feature enhancement module includes:

[0029] Residual layer;

[0030] The channel attention layer pays attention to the importance of different channels and adjusts the response strength of each channel accordingly;

[0031] and spatial attention modules, which enable the network to focus on key areas in the image.

[0032] Preferably, in the ship identification method based on fusion of multiple acoustic features, the feature enhancement module is used to enhance the preliminary features to obtain enhanced features, including:

[0033] The preliminary features are further integrated by combining high-capacity residual modules with channel and spatial attention modules;

[0034] Adaptive average pooling is used to reduce the feature map to 1x1 size.

[0035] Preferably, in the ship identification method based on multiple acoustic feature fusion, the classification and identification module includes:

[0036] The fully connected layer, at the end of the network, maps the learned ship category features to the final output space;

[0037] And the signal loss layer randomly discards some of the network connections so that the model will not be too dependent on any specific part of the input features during training.

[0038] The present invention has the following beneficial effects:

[0039] By integrating the feature transformations of multiple acoustic signals, including CQT spectrogram, MEL spectrogram and STFT phase map, as the input for feature extraction, this feature fusion method not only retains rich information and multiple features about the target, but also significantly improves the model's robustness to noise.

[0040] The current advanced network components, including convolutional layers, batch normalization layers, channel attention mechanisms, and residual mechanisms, are used to form a four-stage feature extraction and recognition network. This network design is not only simple in structure, easy to understand and deploy, but also can achieve fast convergence due to its small number of parameters and efficient optimization.

[0041] The present invention is a low-cost, high-efficiency ship identification method, which can meet the growing demand for maritime monitoring.

[0042] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of an embodiment of a ship identification method based on multiple acoustic feature fusion provided by the present invention;

[0044] Figure 2 An audio processing flow chart of an embodiment of a ship identification method based on multiple acoustic feature fusion provided by the present invention;

[0045] Figure 3 A neural network structure diagram in one embodiment of a ship identification method based on multiple acoustic feature fusion provided by the present invention;

[0046] Figure 4 This is a diagram showing the result of ship identification using an embodiment of a ship identification method based on multiple acoustic feature fusion provided by the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0048] In existing deep learning methods, feature extraction of noise signals usually focuses on a single feature, which inevitably leads to a certain degree of original data loss during the feature extraction process. Due to this data loss, the effect of target recognition is reduced. Therefore, these traditional feature extractions perform poorly in terms of data expression ability and robustness to different noises. At the same time, traditional feature extraction and recognition modules often contain multi-layer network structures, which not only increases the computational complexity and cost of recognition through the network, but also increases the resources and time cost required to deploy the algorithm.

[0049] like Figure 1 As shown, the present invention proposes a ship identification method based on multiple acoustic feature fusion, which includes the following steps:

[0050] like Figure 2 As shown, the underwater noise signal of the ship is obtained and preprocessed;

[0051] The pre-processed ship underwater noise signal is subjected to MEL spectrum, CQT spectrum and STFT phase Figure 3 The feature transformations are performed and stacked in the channel dimension to obtain a three-channel feature image;

[0052] Construct recognition neural network, such as Figure 3 As shown, the recognition neural network includes two asymmetric convolution feature extraction modules, two feature enhancement modules and a classification module;

[0053] Performing feature extraction on the three-channel feature image through the asymmetric convolution feature extraction module to obtain preliminary features;

[0054] Strengthening the preliminary features by the feature enhancement module to obtain enhanced features;

[0055] The enhanced features are classified and identified through a classification module to obtain an identification result.

[0056] Specifically:

[0057] 1. Feature extraction module:

[0058] Two-dimensional convolution layer (conv2d): used to extract features of the input image, so as to capture local features of the image, such as edges, corners, etc. In this example, asymmetric convolution (3×5) is used to better extract spatial information.

[0059] Batch Normalization Layer (BatchNorm2d): Normalizes the input of each mini-batch of data, that is, adjusts the data so that its mean is 0 and its variance is 1. This makes the input data distribution of the middle layer of the network more stable and reduces the sensitivity of the model to initialization.

[0060] ReLU activation function: introduces nonlinear characteristics to the network. The formula of ReLU (Rectified Linear Unit) function is f(x) = max(0,x). It sets all negative values ​​to 0 and keeps positive values ​​unchanged. Introducing nonlinearity helps the network learn complex patterns.

[0061] Residual Block: It solves the gradient vanishing problem in deep networks, allowing the network to be safely deepened without losing performance.

[0062] Spatial activation layer (SEBlock): Emphasizes important features and suppresses unimportant features by adaptively recalibrating feature channels. This significantly improves the model's ability to focus on useful features and helps improve the overall performance of the model.

[0063] Max pooling layer (Maxpool2d): The max pooling operation downsamples or reduces the spatial size of the input feature map by selecting the maximum value in the coverage area. It reduces the feature dimension and improves computational efficiency while enhancing the model's robustness to input changes.

[0064] 2. Feature enhancement module:

[0065] The residual layer is as described above.

[0066] Channel Attention Layer: Pay attention to the importance of different channels (i.e. different types of features) and adjust the response strength of each channel accordingly. By focusing on information-rich feature channels, the representation ability of the model is further improved.

[0067] Spatial Attention: Allows the network to focus on key areas in the image, thereby improving the performance and efficiency of the model. It effectively suppresses unimportant information in complex backgrounds, allowing the model to focus more on useful signals.

[0068] 3. Classification and recognition module:

[0069] Fully connected layer (Linear): At the end of the network, the learned features are mapped to the final output space (such as the category label in the classification task), in this case the ship category.

[0070] Dropout: Randomly discard (set to zero) a part of the network connections so that the model will not be too dependent on any specific input features during training, thereby enhancing the generalization ability of the model.

[0071] In one embodiment of the ship identification method based on multiple acoustic feature fusion provided by the present invention, the preprocessing includes pre-emphasis and resampling operations.

[0072] In one embodiment of the ship identification method based on multiple acoustic feature fusion provided by the present invention, the pre-processed ship underwater noise signal is subjected to MEL spectrum, CQT spectrum and STFT phase Figure 3 Feature transformations include:

[0073] The pre-processed underwater noise signal of the ship is divided into overlapping short time windows, and Fourier transform is applied to each window to obtain the spectrum of each time window. The spectrum of all time windows is spliced ​​together to form a time-frequency diagram;

[0074] Calculate the square of the time-frequency graph to obtain the power spectrum density, filter the power spectrum density through a set of Mel filters, each filter corresponds to a Mel frequency band, and take the logarithm of the filtering result to obtain the logarithmic Mel spectrum;

[0075] The preprocessed underwater noise signal of the ship is subjected to wavelet transform of different scales, each scale corresponds to a frequency band. The wavelet transform result is reconstructed to obtain the CQT spectrum.

[0076] In one embodiment of the ship identification method based on multiple acoustic feature fusion provided by the present invention, the three-channel feature image is subjected to feature extraction by the asymmetric convolution feature extraction module to obtain preliminary features, including:

[0077] Small convolution kernels are used to extract spatial features, and batch normalization and ReLU activation functions are used to regularize the data and introduce nonlinearity.

[0078] In one embodiment of the ship identification method based on multiple acoustic feature fusion provided by the present invention, the preliminary feature is enhanced by the feature enhancement module to obtain an enhanced feature, including:

[0079] The preliminary features are further integrated by combining high-capacity residual modules with channel and spatial attention modules;

[0080] Adaptive average pooling is used to reduce the feature map to 1x1 size.

[0081] The working principle of the present invention is as follows:

[0082] (1) Noise data preprocessing. First, the underwater noise signal of the ship is obtained. In this example, the specific implementation method uses the public dataset DeepShip, which contains 609 underwater ship noise records, including 240 records of tankers, 69 records of tugboats, 191 records of passenger ships, and 109 records of cargo ships. The duration of each record is about a few minutes, and the total length of the data is about 47 hours and 13 minutes. The sampling frequency of SRN is 32kHz, and all data are mono records. In order to preprocess these audio data, including pre-emphasis and resampling operations, to optimize the effect of subsequent processing. A part of the records is selected, and each record is divided into 3-second segments as independent samples. These segments will be randomly divided into training sets and test sets in a ratio of 7:3.

[0083] (2) Feature image synthesis. In this stage, each independent segment is passed through the audio signal processing module, and three audio transformations (MEL spectrum, CQT spectrum, and STFT phase map) are performed on the processed audio signal, and the converted data is saved in decibel image format. The MEL spectrum map (128×188), CQT spectrum map (84×188), and STFT spectrum map (257×188) are obtained respectively. Subsequently, these three images are appropriately scaled and stacked in the channel dimension to form a three-channel feature image (3, 84, 128) containing rich acoustic information. In this stage, we use images fused with multiple acoustic features as feature image input. Compared with the traditional single feature input, the data expression ability and robustness to different noises are significantly improved.

[0084] (3) Image feature extraction. The synthesized feature image is input into the feature extraction module. This module first uses a small convolution kernel (such as 3x5 size) to extract spatial features, and then normalizes the data and introduces nonlinearity through batch normalization and ReLU operations. The SEBlock and ResidualBlock structures help the model focus on useful features and improve the gradient flow through skip connections. Finally, the maximum pooling technology is used to reduce the amount of calculation. In this way, a data feature with better spatial feature extraction effect is obtained compared with the traditional feature extraction method.

[0085] (4) Image feature enhancement. The extracted preliminary features are then fed into the feature enhancement module. The learned complex features are further integrated by combining high-capacity residual blocks with channel and spatial attention modules. This design enables the model to not only focus on spatial information, but also effectively process time series information in audio signals, greatly improving the recognition ability of acoustic features and the overall performance of the model. In the second stage of feature enhancement, the adaptive average pooling technique is used to reduce the feature map to a 1x1 size in preparation for input to the fully connected layer.

[0086] (5) Network recognition module. In the recognition stage, a simple fully connected network is used for classification. The previously extracted features are input into the network, which consists of three fully connected layers, combined with ReLU activation function and Dropout technology to prevent overfitting and optimize the generalization ability of the model. The last fully connected layer maps the learned features to the final output space (such as the category label in the classification task), which is the ship category in this case.

[0087] The audio conversion parameters are as follows:

[0088] #MELSpectrum

[0089] #Specify the number of Mel filters and frame size

[0090] n_mels (number of mel filters) = 128

[0091] n_fft (FFT window size) = 2048

[0092] hop_length(frame shift)=512

[0093] #CQT Spectrum

[0094] n_bins = 84 # The number of keys of a standard 88-key piano minus 4

[0095] Sr=32000

[0096] hop_length(frame shift)=512

[0097] #STFT spectrum

[0098] Sr=32000

[0099] n_fft=512

[0100] hop_length=512.

[0101] The recognition result is as follows Figure 4 And as shown in Table 1.

[0102] Label Accuracy Recall F1 score cargo ship 98.62％ 86.64％ 92.24％ passenger ship 90.02％ 98.20％ 93.94％ tanker 93.50％ 93.80％ 93.65％ tug 88.98％ 93.45％ 91.16％

[0103] In machine learning classification problems, accuracy, recall, and F-score are commonly used metrics to evaluate model performance, especially when the dataset is imbalanced.

[0104] Precision:

[0105] Definition: Precision is the proportion of items predicted to be positive (such as "cargo ship" in ship type identification) that are actually positive.

[0106] Precision=TP / (TP+FP)

[0107] Recall or sensitivity:

[0108] Definition: Recall is the proportion of items that are actually positive that are correctly predicted as positive by the model.

[0109] Recall=TP / (TP+FN)

[0110] F1 score

[0111] Definition: It is a metric specially designed to comprehensively consider both precision and recall. It is the harmonic average of these two metrics. The F1 score is particularly suitable for situations where both precision and recall are equally important.

[0112] F1 = [(precision × recall) / (precision + recall)] × 2

[0113] The present invention has the following characteristics:

[0114] (1) Audio processing innovation: In this module, the present invention applies three different acoustic transformations (MEL spectrum, CQT spectrum and STFT phase map) to the acquired audio data. These transformed data are stacked in the channel dimension to form a multi-channel feature image. This step allows the subsequent neural network to obtain deeper information from multiple dimensions through a simple feature extraction process, thereby improving the accuracy and efficiency of recognition.

[0115] Mel spectrum transform; is a spectrum analysis method that simulates human auditory perception by converting the frequency information of the audio signal to the Mel scale. The Mel scale is nonlinear and better matches the human ear's perception of frequency. It is more refined in the low frequency part and coarser in the high frequency part. First, perform STFT on the audio signal to obtain a time-frequency diagram. Calculate the square of the STFT result to obtain the power spectrum density. Filter the power spectrum density through a set of Mel filters, each filter corresponding to a Mel frequency band. Take the logarithm of the filtered result to obtain the logarithmic Mel spectrum.

[0116] CQT (Constant-Q Transform): CQT is a transform method that decomposes an audio signal into different frequency components. It is similar to the Fourier transform, but uses wavelets instead of sine waves. The frequency resolution of CQT is higher in the low-frequency part and lower in the high-frequency part, which better matches the nature of music signals. Choose a wavelet that suits the signal characteristics (such as the Morlet wavelet). Perform wavelet transforms of different scales on the audio signal, each scale corresponding to a frequency band. Reconstruct the wavelet transform result to obtain the CQT spectrum.

[0117] Short-time Fourier transform (STFT): STFT is a method to analyze the time-varying spectrum of an audio signal by dividing it into short time windows and applying Fourier transform to each window. It provides a time-frequency representation. Divide the audio signal into overlapping short time windows (such as Hamming window or Hanning window) and apply Fourier transform to each window to obtain the spectrum of each time window. The spectra of all time windows are concatenated to form a time-frequency plot.

[0118] (2) Feature extraction optimization: This module initially uses an asymmetric convolution kernel (e.g., 3x5) to process audio images. This specific convolution design helps to capture key spatial features more effectively. On this basis, padding and batch normalization operations are introduced to optimize the network training process. In addition, a residual block and a SE block (channel attention module) are integrated to strengthen feature expression and improve gradient flow. Finally, a maximum pooling layer is used to reduce computational complexity.

[0119] (3) Feature enhancement structure: At the beginning of this module, a high-capacity residual block is used to further enhance the extracted features, followed by the introduction of a channel attention module and a temporal attention block. This design enables the model to not only focus on spatial information, but also effectively process the time series information in the audio signal, greatly improving the recognition ability of acoustic features and the overall performance of the model.

[0120] Traditional feature image acquisition methods usually rely on a single audio feature transformation method. This singleness often leads to the inability to fully capture all useful information in the audio data during the feature extraction process, resulting in partial loss of the original data, which directly affects the effect of target recognition and leads to insufficient recognition accuracy. The present invention uses multiple feature transformations (such as MEL spectrum, CQT spectrum and STFT phase map) for ship acoustic signals and fuses these different feature images, which can include more extensive information in the same input image. This method significantly improves the ability of neural networks to capture target features in complex acoustic environments, thereby enhancing the accuracy of target recognition and the robustness of the system.

[0121] Traditional neural network models often use symmetric convolution kernels and integrate multi-layer network structures, which not only increases the number of parameters of the model, but also increases the computational complexity and cost. Complex network structures and a large number of parameters require more computing resources, increasing the deployment cost of the algorithm. The present invention effectively improves the ability to capture audio features by using asymmetric convolution kernels (e.g., 3x5 size) for feature extraction, especially when processing non-standardized data. The design of combining the channel attention mechanism and the residual mechanism not only improves the network's attention to important features, but also optimizes the gradient flow, reducing the training difficulty and convergence time. This network design simplifies the model structure, reduces the number of parameters and computing requirements, making the model easy to understand and implement, and can quickly and effectively reach an optimized state.

[0122] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A ship identification method based on the fusion of multiple acoustic features, characterized in that: The following steps are involved: Obtain the underwater noise signal of the ship and pre-process it; The preprocessed ship underwater noise signal is transformed into three features: MEL spectrum, CQT spectrum and STFT phase map, and then stacked in the channel dimension to obtain a three-channel feature image. Constructing a recognition neural network, wherein the recognition neural network includes two asymmetric convolution feature extraction modules, two feature enhancement modules and a classification module; Performing feature extraction on the three-channel feature image through the asymmetric convolution feature extraction module to obtain preliminary features; Strengthening the preliminary features by the feature enhancement module to obtain enhanced features; The enhanced features are classified and identified through a classification module to obtain an identification result.

2. The ship identification method based on multiple acoustic feature fusion according to claim 1, characterized in that: The preprocessing includes pre-emphasis and resampling operations.

3. The ship identification method based on multiple acoustic feature fusion according to claim 2, characterized in that: The pre-processed ship underwater noise signal is subjected to three feature transformations: MEL spectrum, CQT spectrum and STFT phase diagram, including: The pre-processed underwater noise signal of the ship is divided into overlapping short time windows, and Fourier transform is applied to each window to obtain the spectrum of each time window. The spectrum of all time windows is spliced ​​together to form a time-frequency diagram; Calculate the square of the time-frequency graph to obtain the power spectrum density, filter the power spectrum density through a set of Mel filters, each filter corresponds to a Mel frequency band, and take the logarithm of the filtering result to obtain the logarithmic Mel spectrum; The preprocessed underwater noise signal of the ship is subjected to wavelet transform of different scales, each scale corresponds to a frequency band, and the wavelet transform result is reconstructed to obtain the CQT spectrum.

4. The ship identification method based on multiple acoustic feature fusion according to claim 3, characterized in that: The asymmetric convolution feature extraction module includes: Two-dimensional convolutional layer, used to extract features of the input image, so as to capture the local features of the image; The batch normalization layer normalizes the input of each mini-batch of data, that is, adjusts the data so that its mean is 0 and its variance is 1; ReLU activation function introduces nonlinear characteristics to the network; Residual layer, which solves the gradient vanishing problem in deep networks; The spatial activation layer emphasizes important features and suppresses unimportant features by adaptively recalibrating feature channels; and the maximum pooling layer downsamples or spatially reduces the input feature map by selecting the maximum value in the coverage area.

5. The ship identification method based on multiple acoustic feature fusion according to claim 4, characterized in that: The asymmetric convolution feature extraction module is used to extract features from the three-channel feature image to obtain preliminary features, including: Small convolution kernels are used to extract spatial features, and batch normalization and ReLU activation functions are used to regularize the data and introduce nonlinearity.

6. The ship identification method based on multiple acoustic feature fusion according to claim 5, characterized in that: The feature enhancement module comprises: Residual layer; The channel attention layer pays attention to the importance of different channels and adjusts the response strength of each channel accordingly; and spatial attention modules, which enable the network to focus on key areas in the image.

7. The ship identification method based on multiple acoustic feature fusion according to claim 6, characterized in that: The preliminary features are enhanced by the feature enhancement module to obtain enhanced features, including: The preliminary features are further integrated by combining high-capacity residual modules with channel and spatial attention modules; Adaptive average pooling is used to reduce the feature map to 1:1 size.

8. The ship identification method based on multiple acoustic feature fusion according to claim 7, characterized in that: The classification recognition module comprises: The fully connected layer, at the end of the network, maps the learned ship category features to the final output space; And the signal loss layer randomly discards some of the network connections so that the model will not be too dependent on any specific part of the input features during training.

Images