A multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention

Abstract

The application belongs to the technical field of marine observation, and specifically discloses a multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention. The method discloses a multi-dimensional feature fusion method based on a gamma filter to realize feature fusion of obtained three-dimensional logarithmic gamma spectrum and three-dimensional gamma cepstrum coefficients; the gamma filter has stronger robustness to noise, so that the method for extracting multi-dimensional fusion features can more comprehensively capture useful acoustic information. In addition, in the multi-dimensional feature fusion process, the feature data is denoised through a self-encoding network, noise interference is reduced, feature quality is enhanced, and the underwater acoustic target recognition rate is improved. In addition, the application also discloses a U-Net network combined with a channel attention mechanism to enhance the representation ability of key features, and the underwater acoustic target is classified through a classification head. The method has good robustness in a noisy environment, and has high underwater acoustic target recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the field of marine observation and relates to a method for underwater acoustic target recognition based on multidimensional fusion features of channel attention. Background Technology

[0002] With the increasing number of underwater acoustic signal datasets and the development of underwater acoustic target recognition technology, it can be widely applied to many fields such as marine biological monitoring, seabed resource exploration, and marine noise pollution monitoring. Currently, most research on underwater acoustic target recognition is based on one-dimensional features fused from multiple acoustic features or multi-dimensional single acoustic features. These mostly rely on the original underwater acoustic signal or data features such as gamma filter banks and Mel-frequency cepstral coefficients. Research on underwater acoustic target recognition based on the fusion of multiple acoustic features in multiple dimensions is relatively rare. Wang Xingmei et al. from Harbin Engineering University proposed an acoustic-embedding memory unit modified space autoencoder (ASAE) for underwater acoustic target recognition. This method merges the Mel filter bank (FBank) and the gamma-frequency filter bank (GBank) into the SAE spectrogram to obtain high acoustic discrimination capability and noise robustness. However, the accuracy of the acoustic-embedding memory unit modified space autoencoder method is insufficient in low signal-to-noise ratio or more complex noise environments.

[0003] Building upon spatial autoencoders modified with acoustic embedded memory units, Wang Xingmei et al. from Harbin Engineering University proposed a dual-channel self-attention acoustic encoder (DASE) network to extract features from the Mel filter bank and gamma filter bank, respectively. Through a local self-attention (LSA) mechanism combined with Swin-Transformer blocks, features are extracted, enhancing local information and noise resistance. Compared to spatial autoencoders modified with acoustic embedded memory units, the dual-channel self-attention acoustic encoder method enhances local information and noise resistance through local self-attention. However, the dual-channel self-attention acoustic encoder method requires processing a large number of positive and negative sample pairs and more iterations to converge, resulting in lower training efficiency and longer training time.

[0004] Furthermore, Dr. Wang Yan and colleagues from Ocean University of China proposed a fast ship radiated noise identification method based on 3D Mel-spectrograms and an additive attention-based Transformer (ADDTr). This method enhances feature representation by combining Mel-spectrograms, first-order difference, and second-order difference features, thereby capturing the temporal variations of the target signal and environmental noise. Simultaneously, the additive attention mechanism improves the model's focus on informative acoustic features, suppressing the influence of noise or distortion components. However, while this fast ship radiated noise identification method based on 3D Mel-spectrograms and an additive attention-based Transformer captures more feature information by combining multidimensional features, it cannot efficiently capture useful acoustic information when dealing with complex non-stationary signals.

[0005] In summary, while the aforementioned related works can achieve relatively accurate underwater acoustic target recognition, the extracted and recognized features are mostly features resulting from the fusion of multiple acoustic features in a single dimension or single acoustic features in multiple dimensions. Therefore, it is difficult to capture the comprehensive dynamic changes of the signal and the robustness to noise is not strong. Thus, there is an urgent need to propose a multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention to effectively solve the above-mentioned technical problems in the existing technology. Summary of the Invention

[0006] To address the problem of underwater acoustic target recognition in complex noise environments, this invention proposes a multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention. By fusing three-dimensional gamma spectral features and three-dimensional gamma cepstral coefficients, the method captures comprehensive dynamic changes in the signal and enhances robustness to noise, thereby enabling accurate recognition of underwater acoustic targets.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for underwater acoustic target recognition based on multidimensional fusion features using channel attention includes the following steps:

[0009] Step 1. Extract the underwater acoustic signal dataset and confirm the corresponding label of the underwater acoustic signal dataset. Extract underwater acoustic audio according to the underwater acoustic target recognition task, and select underwater acoustic audio of various categories from the extracted underwater acoustic audio dataset.

[0010] Step 2. Preprocess the extracted underwater acoustic audio signal to obtain the speech signal;

[0011] Step 3. Input the preprocessed speech signal into the Gammatone filter, then take the logarithm of the filter output to obtain the logarithmic gamma spectrum, and finally denoise the obtained logarithmic gamma spectrum using an autoencoder.

[0012] Step 4. Perform discrete cosine transform on the logarithmic gamma spectrum before processing by the autoencoder in step 3 to obtain gamma cepstral coefficients, and then perform noise reduction processing on the obtained gamma cepstral coefficients by the autoencoder.

[0013] Step 5. Based on the denoised log-gamma spectrum obtained in Step 3 and the denoised gamma cepstral coefficients obtained in Step 4, obtain the three-dimensional log-gamma spectrum and the three-dimensional gamma cepstral coefficients respectively, and perform feature fusion on the two.

[0014] Step 6. Construct a multi-dimensional fusion feature underwater acoustic target recognition model for the underwater acoustic target recognition task;

[0015] The model uses a Unet network with a channel attention mechanism for feature extraction. By assigning different weights to the channels of multi-dimensional features, it enhances the representation ability of key features and integrates a classification head to achieve underwater acoustic target classification.

[0016] The multidimensional fusion features obtained in step 5 are input into the target recognition model, and the classification prediction results are obtained through forward propagation.

[0017] The loss between the predicted result and the true label is then calculated, and the model parameters are optimized through the backpropagation algorithm to complete the model training. Finally, the trained model is used to realize the underwater acoustic target recognition based on multi-dimensional fusion features of channel attention.

[0018] The present invention has the following advantages:

[0019] As described above, this invention relates to a multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention. This method addresses the problem that current underwater acoustic target recognition tasks often extract underwater acoustic signal features that are mostly single-dimensional acoustic features or multi-dimensional fusion features of single acoustic features, making it difficult to fully capture all useful acoustic information and exhibiting limited robustness to noise. Therefore, this invention proposes a multi-dimensional feature fusion method based on a gamma filter. The preprocessed speech signal is input into a Gammatone filter, and the logarithm of the filter output is taken to obtain the logarithmic gamma spectrum. The obtained logarithmic gamma spectrum is then denoised using an autoencoder. A discrete cosine transform is performed on the logarithmic gamma spectrum before autoencoder processing to obtain gamma cepstral coefficients, which are then denoised using an autoencoder. Using the denoised logarithmic gamma spectrum and gamma cepstral coefficients, three-dimensional logarithmic gamma spectrum and three-dimensional gamma cepstral coefficients are obtained, and feature fusion is performed on the three-dimensional logarithmic gamma spectrum and three-dimensional gamma cepstral coefficients. The gamma filter is more robust to noise, thus the method of extracting multi-dimensional fused features captures more comprehensive and useful acoustic information. Furthermore, in the multi-dimensional feature fusion process, this invention uses an autoencoder network to denoise the feature data, reducing noise interference, enhancing feature quality, and improving the final underwater acoustic target recognition rate. In addition, this invention proposes a U-Net network combined with a channel attention mechanism. This U-Net network extracts features from the multi-dimensional fused features and classifies underwater acoustic targets using a classification head. This invention enhances the representation of key features and effectively improves recognition accuracy by assigning different weight ratios to the channels of multidimensional features. In summary, the method of this invention exhibits good robustness in noisy environments and effectively extracts the spectral features of signals for target recognition tasks with a high recognition rate, providing important guidance for marine scientific research. Attached Figure Description

[0020] Figure 1 This is a flowchart of the multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention in an embodiment of the present invention;

[0021] Figure 2 This is a structural diagram of the self-encoder in an embodiment of the present invention;

[0022] Figure 3 This is a flowchart illustrating the generation of three-dimensional fusion features in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the C-Attention U-net neural network model in an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of the channel attention mechanism in an embodiment of the present invention;

[0025] Figure 6 This is a confusion matrix diagram of the recognition results of the multi-dimensional fusion feature + Campplus model in the example of this invention;

[0026] Figure 7 Confusion matrix diagram of the recognition results of multi-dimensional fusion features + ERes2net model in the example of this invention;

[0027] Figure 8 Confusion matrix diagram of the recognition results of the multi-dimensional fusion feature + C-Attention U-Net model in the example of this invention;

[0028] Figure 9 The confusion matrix diagram of the recognition results of the three-dimensional Mel spectrum + C-Attention U-Net model in the example of this invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0030] like Figure 1 As shown, this embodiment addresses the problem of underwater acoustic target recognition in complex noise environments by describing a multi-dimensional fusion feature-based underwater acoustic target recognition method based on channel attention. This method includes the following steps:

[0031] Step 1. Extract the underwater acoustic signal dataset and confirm the corresponding label of the underwater acoustic signal dataset. Extract underwater acoustic audio according to the underwater acoustic target recognition task, and select underwater acoustic audio of various categories from the extracted underwater acoustic audio dataset.

[0032] This embodiment selects 12 underwater audio categories, including fishing boats, trawlers, mussel boats, tugboats, dredgers, motorboats, pilot boats, sailboats, passenger ferries, ocean liners, roll-on / roll-off ships, and background noise, which serve as reference data, i.e., labels.

[0033] Step 2. Perform pre-emphasis, framing, and windowing preprocessing on the extracted underwater acoustic audio signal to obtain the speech signal.

[0034] The pre-emphasis processing is implemented using a simple high-pass filter, and its basic formula is as follows:

[0035] ;

[0036] in, It is the signal after pre-emphasis. It is the original signal. It is the pre-weighting coefficient. These are sampling points.

[0037] During the extraction process, the audio signal is amplified by a pre-emphasis filter, which plays a crucial role in equalizing the signal spectrum. This amplification is particularly beneficial because high-frequency components tend to exhibit lower noise levels compared to low-frequency components. By mitigating the natural attenuation of high frequencies, the pre-emphasis filter effectively restores the balance of the spectrum. Therefore, signal clarity is enhanced, noise prominence is reduced, and the overall quality of the original data is improved.

[0038] Framing typically divides a signal into fixed-length frames. By framing, the original signal is broken down into multiple short, stationary segments, each of which can be processed independently. Local frequency information can be extracted within each frame, allowing analysis of the signal's instantaneous characteristics. This effectively preserves the signal's time-domain properties, facilitating subsequent spectral analysis and feature extraction.

[0039] After dividing the signal into frames, each frame is multiplied by a window function to increase the continuity at the left and right ends of the frame, counteracting the FFT assumption (that the data is infinite) and reducing spectral leakage. The mathematical formula for windowing is:

[0040] Let the original signal be The window function is Then the signal after windowing for:

[0041] ;

[0042] in It is a window function that weights each frame of the signal.

[0043] Step 3. Obtain the logarithmic gamma spectrum and perform noise reduction using an autoencoder.

[0044] The preprocessed speech signal from step 2 is input into a Gammatone filter. The logarithm of the filter output is then taken to obtain the logarithmic gamma spectrum. Finally, the obtained logarithmic gamma spectrum is denoised using an autoencoder.

[0045] The frequency response function of the Gammatone filter is as follows:

[0046] ;

[0047] in The frequency response of the filter describes the gain of the filter for signals of different frequencies.

[0048] This is an amplitude constant that determines the overall gain of the filter. It can usually be adjusted according to design requirements.

[0049] Frequency, measured in Hz, is the input to the filter's response to the signal frequency.

[0050] The order is usually taken as . = 1 or = 2, which controls the shape of the filter's response and affects the steepness of the in-band gain. Generally, the higher the order, the steeper the frequency response.

[0051] The center frequency is the frequency at which the filter's center frequency (i.e., the frequency at which the filter has the strongest response in the frequency domain) is determined, and its unit is Hz. This is the bandwidth parameter, which controls the bandwidth of the filter, and is expressed in Hz.

[0052] bandwidth It is related to the frequency range of the filter, and is usually related to the center frequency in the design. Related.

[0053] The obtained log-gamma spectral feature set for:

[0054] ;

[0055] in For the first Log-gamma spectral features, log-gamma spectral features total indivual, .

[0056] No. The data dimension of the log-gamma spectral features is .

[0057] in, For time step, This represents the number of channels in the Gammatone filter bank.

[0058] To increase the number of samples to meet training requirements, the signal is cut into 1-second time intervals. The processed signal is then input into a gamma filter bank, where fs is the original sampling frequency of the audio signal, the lowest frequency is set to 50Hz, and the analysis window duration is 0.025s. To ensure that the output data can be downsampled multiple times in the model, the number of Gammatone channels is set to 32, the target time step is set to 32, and the output time step is set to 32.

[0059] Frame shift duration It needs to be calculated, and the calculation formula is as follows:

[0060] ;

[0061] The duration of the input audio. To analyze window duration, The target time step.

[0062] To enhance the low-energy components and avoid numerical overflow, the output of the gamma filter bank is logarithmically processed to obtain the log-gamma spectrum. Therefore, the resulting feature set is: .

[0063] in This represents the i-th logarithmic gamma spectral feature, of which there are 11147.

[0064] The data dimension of the i-th log-gamma spectral feature is .

[0065] The 32 in the first dimension represents the time step, and the 32 in the second dimension represents the number of channels in the Gammatone filter bank.

[0066] The obtained log-gamma spectrum is then denoised using an autoencoder. The network structure of the autoencoder is as follows: Figure 2 As shown, in this embodiment, the autoencoder consists of two symmetrical parts and a hidden layer representation.

[0067] The first part is the encoder, which maps the input data to a low-dimensional hidden layer representation. Its formula is:

[0068] ;

[0069] in For input data, For encoding the weight matrix, It's a bias. It is a non-linear activation function.

[0070] The decoder part represents the hidden layer. The formula for reconstructing the input data by mapping it back to the original data space is as follows:

[0071] ;

[0072] in For the reconstructed input data, yes transpose, yes transpose, yes The transpose of .

[0073] The data dimension of the log-gamma spectral features obtained after processing by the autoencoder is the same as the data dimension before denoising.

[0074] To reduce noise interference, enhance feature quality, and thus improve the final accuracy of underwater acoustic target recognition, this embodiment uses an autoencoder to denoise the extracted log-gamma spectrum. By compressing the dimension of the encoder, the network is forced to learn the most discriminative patterns in the gamma spectrum features, while ignoring minor fluctuations that may be noise. When the decoder reconstructs the signal from the compressed representation, it prioritizes the recovery of the main feature patterns, thereby achieving the denoising function.

[0075] Step 4. Calculate the gamma cepstral coefficients and perform denoising using an autoencoder.

[0076] The logarithmic gamma spectrum before processing by the autoencoder in step 3 is subjected to a discrete cosine transform (DCT) to obtain gamma cepstral coefficients. The obtained gamma cepstral coefficients are then denoised using an autoencoder.

[0077] The Discrete Cosine Transform (DCT) is a process of applying a discrete cosine transform to the logarithmic gamma spectrum to remove high-frequency components, compress data, and enhance feature representation. Its formula is:

[0078] ;

[0079] in, It is the first One gamma cepstral coefficient, It is the number of Gammatone filters. The log-gamma spectrum represents the first... The count value on each channel, This refers to the final number of gamma cepstral coefficients obtained; the first 32 gamma cepstral coefficients are selected. , .

[0080] The gamma cepstral coefficient feature set obtained through the above operations is as follows:

[0081] ;in Indicates the first The gamma cepstral coefficient features of each audio audio, the total number of gamma cepstral coefficient features In this embodiment For example, there are 11,147. .

[0082] No. The data dimension of each audio gamma cepstral coefficient feature is .

[0083] in The time step, for example, is 32. This represents the number of characteristics of the gamma cepstral coefficients, for example, 32.

[0084] Next, the obtained gamma cepstral coefficients are denoised using an autoencoder. The method of denoising using the autoencoder is the same as that used in step 3 above, and will not be repeated here.

[0085] Similarly, in order to reduce noise interference, enhance feature quality, and thus improve the final underwater acoustic target recognition accuracy, this embodiment uses an autoencoder to denoise the extracted gamma cepstral coefficients. By compressing the dimension of the encoder, the network is forced to learn the most discriminative pattern in the gamma cepstral coefficient features, while ignoring minor fluctuations that may be noise. When the decoder reconstructs the signal from the compressed representation, it prioritizes the recovery of the main feature patterns, thereby achieving the denoising function.

[0086] The data dimension of the gamma cepstral coefficient features obtained after processing by the autoencoder is the same as that of the data before denoising.

[0087] Step 5. Based on the denoised log-gamma spectrum obtained in Step 3 and the denoised gamma cepstral coefficients obtained in Step 4, obtain the three-dimensional log-gamma spectrum and the three-dimensional gamma cepstral coefficients respectively, and perform feature fusion on the two.

[0088] like Figure 3 The three-dimensional fusion feature process of three-dimensional log-gamma spectrum and three-dimensional gamma cepstral coefficients is shown.

[0089] First, the log-gamma spectrum obtained in step 3 is differentially divided twice to obtain the delta-log-gamma spectrum and the delta-delta-log-gamma spectrum features. Then, the three-dimensional log-gamma spectrum is obtained by splicing the features according to the third dimension, i.e. the channel dimension.

[0090] Next, the gamma cepstral coefficients obtained in step 4 are differentially divided twice to obtain delta-gamma cepstral coefficients and delta-delta-gamma cepstral coefficient features. Then, they are spliced ​​according to the third dimension, i.e. the channel dimension, to obtain the three-dimensional gamma cepstral coefficients.

[0091] Finally, the three-dimensional logarithmic gamma spectrum and the three-dimensional gamma cepstral coefficients are fused. In this embodiment, feature fusion is achieved by feature splicing. By fusing the two features, a multi-dimensional fused feature is obtained.

[0092] The multidimensional fusion features here specifically refer to three-dimensional fusion features. The resulting set of multidimensional fusion features is as follows:

[0093] .

[0094] in Indicates the first The fusion features of each audio element, total One, that is .

[0095] No. The data dimension of each audio fusion feature is .

[0096] in Represents the time step. The number of features representing the gamma cepstral coefficients and the number of channels in the Gammatone filter bank.

[0097] Step 6. Construct a multi-dimensional fusion feature underwater acoustic target recognition model (classification head attention U-Net, C-Attention U-Net) for the underwater acoustic target recognition task. Its network structure is as follows: Figure 4 As shown.

[0098] The C-Attention U-Net model in this embodiment is constructed based on the characteristics of the target recognition task.

[0099] This model is based on the standard U-Net architecture and introduces an attention mechanism to enhance the ability to focus on key features, thereby improving recognition accuracy. In addition, a classification head is also included to achieve underwater acoustic target recognition.

[0100] Specifically, the multidimensional fusion feature underwater acoustic target recognition model includes the Unet network and a classification head.

[0101] The Unet network consists of an encoder and a decoder; skip connections are set between corresponding layers of the encoder and decoder, and a channel attention mechanism is introduced at each skip connection. The channel attention mechanism adopts the SE attention mechanism.

[0102] After inputting multidimensional fused features, the encoder extracts high-level features step by step through multiple convolutional layers and max pooling layers, while reducing the spatial resolution of the features; each convolutional layer contains two convolutional operations followed by a ReLU activation function.

[0103] The decoder gradually restores the image resolution through upsampling, and at the same time merges the feature maps of the corresponding layers in the encoder and the corresponding layers in the decoder through skip connections to retain more feature information.

[0104] Before the skip connections are made between the feature maps of the corresponding layers of the encoder and decoder, the output features of the corresponding layer of the encoder are first processed by the channel attention mechanism before being skip connected with the features of the corresponding layer of the decoder.

[0105] At each jump connection, the process is handled via a channel attention mechanism, and the process is roughly as follows:

[0106] First, global average pooling is performed on the two-dimensional feature map of each channel.

[0107] It compresses the information of all pixels in each channel into a single value. The vector is then fed into a small two-layer fully connected neural network. The first fully connected layer compresses the vector to a lower dimension and applies a ReLU activation function to capture the non-linear dependencies between channels. The second fully connected layer restores this low-dimensional vector to its original dimension and applies a Sigmoid activation function. Finally, the weight vector of this dimension is multiplied channel by channel with the original input feature map to obtain a calibrated feature map in which features of unimportant channels are suppressed and features of important channels are amplified, thereby further enhancing the ability to capture key features.

[0108] Specifically, such as Figure 5 As shown, the SE attention mechanism (Squeeze-and-Excitation Networks) adds an attention mechanism to the channel dimension, with the key operations being squeeze and excitation.

[0109] This part uses an automatic learning approach, employing a new neural network to determine the importance of each channel in the feature map. This importance is then used to assign a weight to each feature, allowing the neural network to focus on certain feature channels. This process boosts the channels of the feature map that are useful for the current task and suppresses those that are less useful.

[0110] First is Squeeze ( In the Squeeze part, global average pooling is used to compress the two-dimensional features of each channel of the feature map into a single real number, resulting in a vector of length C equal to the number of channels. The formula is expressed as:

[0111] ;

[0112] in It is the first Feature map of each channel These are the corresponding global descriptors, where H and W represent the height and width, respectively. It is the first The feature map of each channel in spatial location Eigenvalue representation on;

[0113] The excitation section uses two fully connected layers (FC) to form a bottleneck structure.

[0114] The first FC layer compresses the number of channels to , To achieve the desired compression ratio, ReLU activation is used. The second fully connected (FC) layer restores the number of channels to C and uses Sigmoid activation, outputting the weight of each channel. The formula is as follows:

[0115] ;

[0116] in It is the weight vector output by Excitation. , For ReLU, It is Sigmoid; It is the vector output by Squeeze. It is a component of it;

[0117] The Scale part applies the channel weights obtained from the excitation to the original feature map, i.e., the feature map input to the SE attention mechanism, to weight the channels and achieve feature recalibration.

[0118] Specifically, the feature map of each channel is multiplied by its corresponding weight value, using the following formula:

[0119] ;

[0120] in It is the first The weight of each channel, all composition , It is the original feature map. This is the recalibrated feature map.

[0121] The Squeeze operation masks spatial distribution information, allowing subsequent channel weight adjustments to be based on overall channel information rather than local spatial features. The Excitation operation learns the non-linear dependencies between channels to generate weights for each channel, which are used to adaptively recalibrate channel features. The Scale part multiplies the channel weights with the original features to achieve channel-level feature recalibration, enhancing useful features and suppressing ineffective features.

[0122] This invention uses Attention U-Net as a feature extractor to extract features from 3D fused features. After the feature extractor, the model converts the feature map into a classification result through a simple classification head. The classification head consists of a global average pooling layer and a flattening layer. The global average pooling layer compresses the spatial dimension of the feature map to 1×1, while the flattening layer converts it into a shape of (batch size, num) The output vector of classes; batch size represents the batch size. The label represents the type of underwater acoustic signal. During the forward propagation of the model, the multidimensional fused features are first extracted by the Unet network, which combines channel attention mechanism, i.e., the feature extractor, and then the final underwater acoustic target classification result is obtained through the classification head.

[0123] After building the above model, the model is trained. The training process is roughly as follows: input the multi-dimensional fusion features obtained in step 5 into the target recognition model, and obtain the classification prediction result through forward propagation; then calculate the loss between the prediction result and the real label, and optimize the model parameters through the backpropagation algorithm to complete the training of the model.

[0124] During model training, the input fused feature parameters are resized to fit the dimensions of the Attention U-Net neural network. Samples with the same numerical value are randomly extracted from each label sample (e.g., 100 random samples) for training testing, and the training and test sets are split in a 4:1 ratio. The input 3D feature parameters undergo downsampling via the encoder, which includes double convolutional blocks and pooling operations, reducing the dimension of the input feature parameters from... Transform into ;in Indicates the time step. This indicates the number of channels in the gamma filter and the number of gamma cepstral coefficients. This represents the number of channels; after the encoder output, the data is imported into the decoder for upsampling. The upsampling operation includes transposed convolution, channel attention processing, feature concatenation, and double convolutional block operations, restoring the dimension of the input feature parameters to [the desired value]. Finally, the decoder output is fed into the classification head, and after global average pooling and flattening operations, the final output is... ,in The type of label indicating underwater acoustic signals.

[0125] Specifically, the input to the C-Attention U-net model is the multidimensional fusion feature obtained in step 5.

[0126] This embodiment uses a 32×32×6 matrix as an example, where the last dimension represents the number of channels. The input first passes through two 3×3 convolutional layers to extract subtle features from the multi-dimensional fusion features while maintaining the same spatial size. Then, it sequentially passes through a first downsampling and convolutional module, a second downsampling and convolutional module, a third downsampling and convolutional module, and a fourth downsampling and convolutional module to extract high-level features from the multi-dimensional fusion features. Within each downsampling and convolutional module, the features are first downsampled using a max pooling layer to halve the spatial size, and then passed through two 3×3 convolutional layers to extract even more subtle high-level features. The output features after the fourth downsampling and convolutional module are progressively processed through four upsampling and convolutional modules to gradually restore the original scale and information of the input features. In the first upsampling and convolutional module, the features are first upsampled through the upper convolutional layer, then connected to the output of the fourth downsampling and convolutional module via a channel attention mechanism, and then processed by a convolutional layer before being input into the second upsampling and convolutional module. In the second upsampling and convolutional module, the features are first upsampled through the upper convolutional layer, then connected to the output of the third downsampling and convolutional module via a channel attention mechanism, and then processed by a convolutional layer before being input into the third upsampling and convolutional module. In the third upsampling and convolutional module, the features are first upsampled through the upper convolutional layer, then connected to the output of the second downsampling and convolutional module via a channel attention mechanism, and then processed by a convolutional layer before being input into the third upsampling and convolutional module. The output of the module is processed by a channel attention mechanism and then skipped. After passing through a convolutional layer, it is input into the fourth upsampling and convolutional module. In the fourth upsampling and convolutional module, the features are first upsampled through an upper convolutional layer, then skipped with the output of the first downsampling and convolutional module after processing by a channel attention mechanism. This connection is then passed through a convolutional layer and input into the second upsampling and convolutional module. After passing through two 3×3 convolutional layers, the input is fed into a 1×1 convolutional layer. Finally, the output from the 1×1 convolutional layer is fed into the classification head for classification. In the classification head, the data is compressed to 1×1 spatial dimension by a global average pooling layer, and the final output is flattened. ,in There are 12 types of tags representing underwater acoustic signals.

[0127] During model validation, the recognition rates of each round are compared, and the model with the highest recognition rate is retained to complete the target recognition task.

[0128] Among them, the recognition rate The calculation formula is:

[0129] ;

[0130] in For the sample size, For real labels, For predicting labels, This is an indicator function (1 for correct prediction, 0 for incorrect prediction). If the recognition rate in the new round is better than the previous round, the model is updated and the process proceeds to the next round of training and validation. If the recognition rate in the new round is not better than the previous round, the model does not need to be updated and the process proceeds directly to the next round of training and validation.

[0131] Finally, the trained model was used to achieve underwater acoustic target recognition based on multidimensional fusion features of channel attention.

[0132] The method of this invention can extract and fuse features of underwater acoustic audio with significant noise impact, and accurately identify targets from the fused features, effectively solving the problem of low recognition rate of underwater acoustic signals affected by strong noise.

[0133] Furthermore, to verify the superiority of the C-Attention Unet model proposed in this invention, existing models Campplus and ERes2net are selected below, and comparative experiments are conducted using multi-dimensional fusion features. Figure 6 The confusion matrix is ​​the result of multi-dimensional fusion features + Campplus model recognition. Figure 7 This is the confusion matrix of the recognition results of the multidimensional fusion feature + ERes2net model. Figure 8 This is the confusion matrix of the recognition results from the multi-dimensional fusion feature + C-Attention U-Net model. (See attached...) Figures 6 to 8 A precise comparison shows that the C-Attention Unet model proposed in this invention outperforms the Campplus and ERes2net models in underwater acoustic target recognition tasks. The accuracy comparison of different models is shown in Table 1.

[0134] Table 1. Accuracy Comparison of Different Models

[0135]

[0136] The results show that the Unet model with multidimensional fusion features and C-Attention achieves a better recognition rate of 94.33%.

[0137] In addition, to verify the superiority of the multidimensional fusion feature of the present invention, a three-dimensional Mel spectrum was selected as the input feature and compared with the multidimensional fusion feature in a comparative experiment. Figure 8 The confusion matrix of the recognition results of the multi-dimensional fusion feature + C-Attention U-Net model. Figure 9 This is the confusion matrix of the recognition results from the 3D Mel-spectrum + C-Attention U-Net model. (From the appendix...) Figures 8 to 9The comparison demonstrates that the multi-dimensional fusion feature proposed in this invention can capture more effective underwater acoustic feature information compared to traditional three-dimensional Mel spectrum features, thus achieving a higher accuracy in underwater acoustic target recognition. Table 2 shows a comparison of the recognition accuracy for different input features.

[0138] Table 2 Comparison of accuracy for different features

[0139]

[0140] As can be seen from the comparison of the results in Table 2, the multi-dimensional fusion feature adopted in this invention can achieve a higher accuracy. The final result shows that the multi-dimensional fusion feature + C-Attention Unet model achieves the best recognition rate, which is 94.33%.

[0141] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for underwater acoustic target recognition based on multidimensional fusion features using channel attention, characterized in that, Includes the following steps: Step 1. Extract the underwater acoustic signal dataset and confirm the corresponding label of the underwater acoustic signal dataset. Extract underwater acoustic audio according to the underwater acoustic target recognition task, and select underwater acoustic audio of various categories from the extracted underwater acoustic audio dataset. Step 2. Preprocess the extracted underwater acoustic audio signal to obtain the speech signal; Step 3. Input the preprocessed speech signal into the Gammatone filter, then take the logarithm of the filter output to obtain the logarithmic gamma spectrum, and finally denoise the obtained logarithmic gamma spectrum using an autoencoder. Step 4. Perform discrete cosine transform on the logarithmic gamma spectrum before processing by the autoencoder in step 3 to obtain gamma cepstral coefficients, and then perform noise reduction processing on the obtained gamma cepstral coefficients by the autoencoder. Step 5. Based on the denoised log-gamma spectrum obtained in Step 3 and the denoised gamma cepstral coefficients obtained in Step 4, obtain the three-dimensional log-gamma spectrum and the three-dimensional gamma cepstral coefficients respectively, and perform feature fusion on the two. First, the log-gamma spectrum obtained in step 3 is differentially divided twice to obtain the delta-log-gamma spectrum and the delta-delta-log-gamma spectrum features. Then, they are stitched together according to the channel dimension to obtain the three-dimensional log-gamma spectrum. Next, the gamma cepstral coefficients obtained in step 4 are differentially divided twice to obtain delta-gamma cepstral coefficients and delta-delta-gamma cepstral coefficient features, which are then stitched together according to the channel dimension to obtain three-dimensional gamma cepstral coefficients. Finally, the three-dimensional log-gamma spectrum and the three-dimensional gamma cepstral coefficients are fused to obtain multi-dimensional fused features. Step 6. Construct a multi-dimensional fusion feature underwater acoustic target recognition model for the underwater acoustic target recognition task; The model uses a Unet network with a channel attention mechanism for feature extraction. By assigning different weights to the channels of multi-dimensional features, it enhances the representation ability of key features and integrates a classification head to achieve underwater acoustic target classification. The multidimensional fusion features obtained in step 5 are input into the target recognition model, and the classification prediction results are obtained through forward propagation. The loss between the predicted result and the true label is then calculated, and the model parameters are optimized through the backpropagation algorithm to complete the model training. Finally, the trained model is used to realize the multi-dimensional fusion feature underwater acoustic target recognition based on channel attention.

2. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 1, 12 types of underwater audio are selected, including fishing boats, trawlers, mussel boats, tugboats, dredgers, motorboats, pilot boats, sailboats, passenger ferries, ocean liners, roll-on / roll-off ships, and background noise.

3. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 2, the extracted underwater acoustic audio signal is preprocessed by pre-emphasis, framing, and windowing.

4. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 3, the log-gamma spectral feature set is obtained. for: ； in For the first Log-gamma spectral features, log-gamma spectral features total indivual, ; No. The data dimension of the log-gamma spectral features is ; in For time step, The number of channels in the Gammatone filter bank; The data dimension of the log-gamma spectral features obtained after processing by the autoencoder is the same as the data dimension before denoising.

5. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 4, the process of obtaining the gamma cepstral coefficients and performing denoising is as follows: The Discrete Cosine Transform (DCT) is the discrete cosine transform applied to the logarithmic gamma spectrum. Its formula is: ； in, It is the first One gamma cepstral coefficient, It is the number of Gammatone filters. The log-gamma spectrum represents the first... The count value on each channel It is the number of gamma cepstral coefficients obtained in the end. ; The gamma cepstral coefficient feature set obtained through the above operations is as follows: ； in Indicates the first The gamma cepstral coefficient features of each audio audio, the total number of gamma cepstral coefficient features indivual, ; No. The data dimension of each audio gamma cepstral coefficient feature is ; in For time step, The characteristic number of gamma cepstral coefficients; The data dimension of the gamma cepstral coefficient features obtained after processing by the autoencoder is the same as that of the data before denoising.

6. The multi-dimensional fusion feature underwater acoustic target recognition method based on channel attention according to claim 4 or 5, characterized in that, The autoencoder consists of two symmetrical parts and a hidden layer representation; The first part is the encoder, which maps the input data to a low-dimensional hidden representation. Its formula is: ； in For input data, For encoding the weight matrix, It's a bias. It is a non-linear activation function; The decoder section is used to represent the hidden layer. The formula for reconstructing the input data by mapping it back to the original data space is as follows: ； in For the reconstructed input data, yes transpose, yes transpose, yes The transpose of .

7. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 6, the multi-dimensional fusion feature underwater acoustic target recognition model includes a Unet network and a classification head; The Unet network includes an encoder and a decoder; skip connections are set between corresponding layers of the encoder and decoder, and a channel attention mechanism is introduced at each skip connection. The channel attention mechanism adopts the SE attention mechanism. After inputting multidimensional fused features, the encoder extracts high-level features step by step through multiple convolutional layers and max pooling layers, while reducing the spatial resolution of the features; Each convolutional layer contains two convolution operations followed by a ReLU activation function. The decoder gradually restores the image resolution through upsampling, and at the same time merges the feature maps of the corresponding layers in the encoder and the corresponding layers in the decoder through skip connections to retain more feature information. Before the skip connections are made between the feature maps of the corresponding layers of the encoder and decoder, the output features of the corresponding layer of the encoder are first processed by the channel attention mechanism and then skip connected with the features of the corresponding layer of the decoder. During the forward propagation of the model, the multidimensional fusion features are first extracted by the Unet network, which combines channel attention mechanism, i.e., the feature extractor, and then the final underwater acoustic target classification result is obtained through the classification head.

8. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 7, characterized in that, The SE attention mechanism includes Squeeze, Excitation, and Scale components; The Squeeze part uses global average pooling to compress the two-dimensional features of each channel of the feature map into a single real number, resulting in a vector of length C equal to the number of channels. The formula is as follows: ； in It is the first Feature map of each channel These are the corresponding global descriptors, where H and W represent the height and width, respectively. It is the first The feature map of each channel in spatial location Eigenvalue representation on; The Excitation section uses two fully connected layers (FC) to form a bottleneck structure. The first FC layer compresses the number of channels to , To achieve the desired compression ratio, ReLU activation is used. The second fully connected (FC) layer restores the number of channels to C and uses Sigmoid activation, outputting the weight of each channel. The formula is as follows: ； in It is the weight vector output by Excitation. , For ReLU, It is Sigmoid; It is the vector output by Squeeze. yes Components; The Scale part applies the channel weights obtained from the excitation to the original feature map, i.e., the feature map input to the SE attention mechanism, to weight the channels and achieve feature recalibration. Specifically, the feature map of each channel is multiplied by its corresponding weight value, using the following formula: ； in It is the first The weight of each channel, all composition , It is the original feature map. This is the recalibrated feature map.

9. The underwater acoustic target recognition method based on channel attention using multidimensional fusion features according to claim 1, characterized in that, In step 6, the classification head consists of a global average pooling layer and a flattening layer; Global average pooling layers compress the spatial dimension of feature maps to 1×1, while flattening layers transform them into shapes of (batchsize, num). The output vector of classes; batch size represents the batch size. The type of label indicating underwater acoustic signals.

Images