Underwater target classification method, device and equipment and storage medium

Abstract

The invention provides an underwater target classification method, device and equipment and a storage medium, and relates to the technical field of ocean engineering. The method comprises the following steps of: generating a time-frequency characteristic data set by performing wavelet transform on an actually measured noise signal; the adversarial network is generated by using an auxiliary classifier to enhance the data set so as to overcome sample insufficiency; extracting multi-scale features through parallel multi-scale expansion convolution; self-adaptive dynamic weighting is performed on the features in combination with a channel time-frequency attention mechanism, and key information is highlighted; and finally, a bidirectional time sequence dependency relationship is extracted by adopting a bidirectional long-short-term memory network, and accurate classification is realized. Through effective fusion of time-frequency analysis and deep learning, the accuracy and robustness of underwater target classification are significantly improved under the condition of a small number of samples.

Description

Technical Field

[0001] This application relates to the field of marine engineering technology, and in particular to an underwater target classification method, apparatus, equipment and storage medium. Background Technology

[0002] Underwater target identification is of great significance in fields such as marine monitoring and national defense security. Traditional methods mainly rely on manual listening or conventional signal processing techniques, which suffer from problems such as low efficiency, strong subjectivity, and poor generalization ability.

[0003] In recent years, machine learning-based methods have been increasingly applied in this field, but challenges remain, including difficulties in acquiring underwater samples, insufficient data volume, and complex noisy environments. Existing technologies often use raw signals directly for modeling, which can easily lead to overfitting when samples are scarce, resulting in limited recognition accuracy. Therefore, how to achieve stable and high-accuracy automatic classification of underwater targets with limited samples remains a pressing technical challenge. Summary of the Invention

[0004] The main objective of this application is to provide an underwater target classification method, apparatus, device, and storage medium, aiming to solve the technical problem of how to effectively improve the accuracy and stability of underwater target classification under the condition of scarce samples.

[0005] To achieve the above objectives, this application provides an underwater target classification method, the steps of which include: Wavelet transform is performed on the underwater target noise signal obtained from the actual measurement to form the corresponding initial time-frequency feature dataset; The initial time-frequency feature dataset is expanded based on the auxiliary classifier generative adversarial network model to obtain the corresponding target time-frequency feature dataset; Based on a parallel multi-scale dilated convolution model, multi-scale feature extraction is performed on the target time-frequency feature dataset to obtain the corresponding target time-frequency features. Based on the channel-time-frequency attention model, attention weights are assigned to the target time-frequency features in both the channel and time-frequency dimensions to obtain the corresponding weighted target time-frequency features. Based on the bidirectional long short-term memory network model, the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features is extracted, and the underwater target classification result is output based on the bidirectional temporal feature mapping relationship.

[0006] In one embodiment, the step of performing wavelet transform on the measured underwater target noise signal to form a corresponding initial time-frequency feature dataset includes: Perform continuous wavelet transform on the underwater target noise signal to obtain the time-frequency distribution spectrum of the underwater target noise signal after scaling and translation operations; Based on the time-frequency distribution map, the initial time-frequency feature dataset is obtained.

[0007] In one embodiment, before the step of expanding the initial time-frequency feature dataset based on the auxiliary classifier generative adversarial network model to obtain the corresponding target time-frequency feature dataset, the method further includes: An initial generative adversarial network model is constructed based on a generator and a discriminator; The random noise signal and category label are input into the generator to obtain the corresponding synthesized time-frequency feature data; The discriminator is used to distinguish and predict the categories of the synthetic time-frequency feature data and the historical real time-frequency feature data to obtain the discrimination results and the category prediction results. Based on the discrimination result and the category prediction result obtained each time, the generation parameters of the generator and the discrimination parameters of the discriminator are iteratively optimized alternately to obtain the auxiliary classifier generative adversarial network model.

[0008] In one embodiment, before the step of performing multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features, the method further includes: Construct several dilated convolution branches with different dilation rates and kernels; By setting each of the dilated convolution branches in parallel, the parallel multi-scale dilated convolution model is obtained.

[0009] In one embodiment, the channel time-frequency attention model includes a channel attention submodule and a time-frequency attention submodule. The step of allocating attention weights based on the channel time-frequency attention model to the target time-frequency features in terms of channel dimension and time-frequency dimension to obtain the corresponding weighted target time-frequency features includes: The channel attention weights of the target time-frequency features in the channel dimension are obtained through the channel attention submodule. The time-frequency attention weights of the target time-frequency features in the time-frequency dimension are obtained through the time-frequency attention submodule. The weighted target time-frequency features are obtained based on the time-frequency attention weights, the channel attention weights, and the target time-frequency features.

[0010] In one embodiment, the step of obtaining the time-frequency attention weights of the target time-frequency features in the time-frequency dimension through the time-frequency attention submodule includes: Dilated convolution is performed on the target time-frequency features, and global average pooling is performed along the time dimension and frequency dimension respectively to obtain the time dimension feature vector and the frequency dimension feature vector. The time-frequency attention weights are obtained based on the time-dimensional feature vector and the frequency-dimensional feature vector.

[0011] In one embodiment, the step of extracting the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and outputting the underwater target classification result based on the bidirectional temporal feature mapping relationship includes: Based on the bidirectional long short-term memory network model, the forward hidden state sequence and the backward hidden state sequence of the weighted target time-frequency features are obtained respectively. Based on the forward hidden state sequence and the backward hidden state sequence, the bidirectional temporal feature mapping relationship is formed in chronological order; The underwater target classification result is output based on the bidirectional temporal feature mapping relationship.

[0012] Furthermore, to achieve the above objectives, this application also provides an underwater target classification device, the underwater target classification device comprising: The data preprocessing module is used to perform wavelet transform on the underwater target noise signal obtained from the actual measurement to form the corresponding initial time-frequency feature dataset; The dataset expansion module is used to generate an adversarial network model based on an auxiliary classifier, and to expand the initial time-frequency feature dataset to obtain the corresponding target time-frequency feature dataset. The feature extraction module is used to perform multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features. The weight allocation module is used to allocate attention weights in the channel dimension and time-frequency dimension to the target time-frequency features based on the channel time-frequency attention model, and obtain the corresponding weighted target time-frequency features; The target classification module is used to extract the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and output the underwater target classification result based on the bidirectional temporal feature mapping relationship.

[0013] In addition, to achieve the above objectives, this application also provides an underwater target classification device, which includes: a memory, a processor, and an underwater target classification program stored in the memory and executable on the processor, wherein the underwater target classification program is configured to implement the steps of the underwater target classification method as described above.

[0014] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium, on which an underwater target classification program is stored, and when the underwater target classification program is executed by a processor, it implements the steps of the underwater target classification method as described above.

[0015] This application provides an underwater target classification method, apparatus, device, and storage medium. The underwater target classification method includes the following steps: performing wavelet transform on the measured underwater target noise signal to form a corresponding initial time-frequency feature dataset; expanding the initial time-frequency feature dataset based on an auxiliary classifier generative adversarial network model to obtain a corresponding target time-frequency feature dataset; performing multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain corresponding target time-frequency features; assigning attention weights to the target time-frequency features based on a channel time-frequency attention model to obtain corresponding weighted target time-frequency features; and extracting the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on a bidirectional long short-term memory network model, and outputting the underwater target classification result based on the bidirectional temporal feature mapping relationship.

[0016] A time-frequency feature dataset is generated by performing wavelet transform on measured noise signals. A generative adversarial network (GAN) is used to enhance the dataset to overcome insufficient samples. Multi-scale features are extracted through parallel multi-scale dilated convolution. A channel-based time-frequency attention mechanism is combined to adaptively and dynamically weight the features, highlighting key information. Finally, a bidirectional long short-term memory (LSTM) network is used to extract bidirectional temporal dependencies, achieving accurate classification. By effectively integrating time-frequency analysis and deep learning, the accuracy and robustness of underwater target classification are significantly improved under limited sample conditions. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the underwater target classification method of this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the underwater target classification method of this application. Figure 3 A schematic diagram of the structure of a Generative Adversarial Network (GAN) model for an auxiliary classifier; Figure 4 Training flowchart for generating adversarial network models for auxiliary classifiers; Figure 5 This is a schematic diagram of a parallel multi-scale dilated convolution model. Figure 6 This is a flowchart illustrating Embodiment 3 of the underwater target classification method of this application. Figure 7 This is a schematic diagram of the overall structure of the channel time-frequency attention model; Figure 8 This is a partial structural diagram of the channel time-frequency attention model; Figure 9 This is another schematic diagram of a local structure of the channel time-frequency attention model; Figure 10 This is a schematic diagram of the structure of a unidirectional long short-term memory network model; Figure 11 This is a schematic diagram of the structure of a bidirectional long short-term memory network model; Figure 12 This is a schematic diagram of the module structure of the underwater target classification device according to an embodiment of this application; Figure 13 This is a schematic diagram of the underwater target classification device according to an embodiment of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] This application presents a first embodiment of an underwater target classification method; please refer to [link / reference]. Figure 1 The underwater target classification method includes steps S10-S50: Step S10: Perform wavelet transform on the measured underwater target noise signal to form the corresponding initial time-frequency feature dataset; It should be understood that, in this embodiment, the executing entity can be an underwater target classification device or underwater target classification equipment installed on a ship or other carrier that can move in water.

[0024] It should be noted that underwater target noise signals can be understood as noise signals generated by various underwater targets obtained through actual testing. In this embodiment, due to the nonlinear and non-stationary characteristics of ship noise, traditional analysis methods may not be able to fully reveal its features. To capture these time-varying characteristics, time-frequency analysis techniques are needed to obtain the time-frequency distribution characteristics reflecting the dynamic characteristics of ship radiated noise. Corresponding to the above requirements, this embodiment uses wavelet transform to process the actually collected underwater target noise signals. Wavelet transform is a time-frequency analysis method that can achieve fine analysis at different time and frequency scales by scaling and shifting the signal, automatically adapting to the needs of time-frequency signal analysis.

[0025] As is easily understood, the initial time-frequency feature dataset is a two-dimensional matrix dataset, where rows represent time points and columns represent frequency points, with each element representing the energy or amplitude of the (signal) at that time-frequency point. In this embodiment, the underwater target noise signal can be processed using wavelet transform to form an initial time-frequency feature dataset that can more comprehensively characterize the time-frequency properties of the signal.

[0026] Step S20: Generate an adversarial network model based on the auxiliary classifier to expand the initial time-frequency feature dataset and obtain the corresponding target time-frequency feature dataset; It should be noted that in this embodiment, the Auxiliary Classifier Generative Adversarial Network (ACGAN) is a variant of generative adversarial networks, which introduces class labels (class information) into both the generator and the discriminator to generate synthetic samples with specified classes.

[0027] It is easy to understand that underwater target data acquisition is costly and samples are scarce, making direct training of deep learning models prone to overfitting. In this embodiment, to address these issues, an ACGAN model is used for adversarial training between the generator and discriminator. This allows the model to learn the distribution of real data and generate realistic synthetic samples. Introducing category labels (category information) ensures the semantic consistency of the generated samples, thereby effectively expanding the dataset and improving the model's generalization ability and robustness with limited real data (i.e., the initial time-frequency feature dataset mentioned above).

[0028] Step S30: Based on the parallel multi-scale dilated convolution model, perform multi-scale feature extraction on the target time-frequency feature dataset to obtain the corresponding target time-frequency features; It should be understood that, in this embodiment, Parallel Multi-scale Dilated Convolution (PMDC) is a convolutional neural network structure that can extract features under different receptive fields through multiple parallel dilated convolution branches.

[0029] It should be noted that, in this embodiment, the target time-frequency features are high-dimensional feature representations extracted by the PMDC model, which contain semantic information at different scales.

[0030] It is easy to understand that traditional convolutional neural networks have a fixed receptive field, making it difficult to capture both local details and global context simultaneously. In this embodiment, to address this issue, a PMDC model is used to perform multi-scale feature fusion and feature extraction on the target time-frequency feature dataset, ultimately obtaining the required target time-frequency features. This expands the receptive field without increasing parameters or computational load, enabling multi-scale feature fusion, which helps perceive the target's local texture and overall structure, improving the richness and discriminative power of feature representation.

[0031] Step S40: Based on the channel time-frequency attention model, the target time-frequency features are assigned attention weights in the channel dimension and the time-frequency dimension to obtain the corresponding weighted target time-frequency features; It should be understood that, in this embodiment, Channel Time-Frequency Attention (CTFA) is an attention mechanism used to adaptively weight features in both the channel and time-frequency dimensions.

[0032] It should be noted that, in this embodiment, the weighted target time-frequency features are features that have been dynamically weighted by attention, wherein important features are enhanced and minor or noisy features are suppressed.

[0033] It is easy to understand that in this embodiment, the CTFA model can learn the importance of different channels through channel attention and the importance of different positions in the time-frequency graph through time-frequency attention, thereby obtaining the corresponding weighted target time-domain features. This adaptive weighting mechanism enables the model to focus on key regions and feature channels related to the target category, suppressing background noise and irrelevant information, thus improving the accuracy and interpretability of classification.

[0034] Step S50: Based on the bidirectional long short-term memory network model, extract the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features, and output the underwater target classification result based on the bidirectional temporal feature mapping relationship.

[0035] It should be understood that, in this embodiment, the Bidirectional Long Short-Term Memory (Bi-LSTM) network is a recurrent neural network that can capture the temporal dependencies of sequence data from both forward and backward directions simultaneously.

[0036] It should be noted that, in this embodiment, the bidirectional temporal feature mapping relationship refers to the sequential dependency relationship of features in the time dimension, and the input of the fully connected layer used for classification is an abstract representation of this relationship.

[0037] It is easy to understand that underwater acoustic signals have strong temporal correlations, and the evolution patterns of target features over time contain important category information. In this embodiment, a Bi-LSTM model can be used to simultaneously extract the forward and backward information corresponding to the weighted target time-frequency features, thereby enabling a more comprehensive understanding of the temporal dynamics of the weighted target time-frequency features and extracting the corresponding bidirectional temporal feature mapping relationship. Finally, underwater target classification can be achieved through the bidirectional temporal feature mapping relationship, thus obtaining the underwater target classification results corresponding to the measured underwater target noise signals. Through the above method, it is possible to distinguish the change patterns of different categories of targets on the time-frequency map, further improving classification accuracy, especially in scenarios involving target movement or state changes.

[0038] This application provides an underwater target classification method. This method generates a time-frequency feature dataset by performing wavelet transform on measured noise signals; it then uses a generative adversarial network (GAN) to enhance the dataset to overcome insufficient samples; it extracts multi-scale features through parallel multi-scale dilated convolution; it adaptively and dynamically weights the features using a channel time-frequency attention mechanism to highlight key information; and finally, it employs a bidirectional long short-term memory (LSTM) network to extract bidirectional temporal dependencies, achieving accurate classification. By effectively integrating time-frequency analysis and deep learning, this method significantly improves the accuracy and robustness of underwater target classification even with limited samples.

[0039] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 3 , Figure 4 as well as Figure 5 The step of performing wavelet transform on the measured underwater target noise signal to form the corresponding initial time-frequency feature dataset includes: Step S11: Perform continuous wavelet transform on the underwater target noise signal to obtain the time-frequency distribution spectrum of the underwater target noise signal after scaling and translation operations; Step S12: Based on the time-frequency distribution map, obtain the initial time-frequency feature dataset.

[0040] It should be noted that, in this embodiment, the underwater target noise signal can be inner-producted with a set of scaled and translated wavelet basis functions using continuous wavelet transform, i.e., scaling and translating the underwater target noise signal to obtain the corresponding time-frequency distribution map. The obtained time-domain distribution map clearly shows the energy distribution of the underwater target noise signal in the time-frequency plane. This distribution can be represented by a two-dimensional matrix dataset, which is the aforementioned initial time-frequency feature dataset.

[0041] It is easy to understand that, in this embodiment, for square-integrable underwater target noise signals In this regard, the corresponding wavelet transform is: ; in, This is the scaling parameter (used to control the scaling of the wavelet). These are translation parameters (used to control the translation of the wavelet). It is a time variable (i.e., an integral variable). The complex conjugate of the mother wavelet, The mother wavelet function (i.e., an oscillating waveform that satisfies specific conditions and decays rapidly). It is a two-dimensional matrix after wavelet transform (used to represent the amplitude or intensity of a signal at a specific frequency and time).

[0042] Furthermore, in this embodiment, before the step of expanding the initial time-frequency feature dataset based on the auxiliary classifier generative adversarial network model to obtain the corresponding target time-frequency feature dataset, the method further includes: Step S201: Construct an initial generative adversarial network model based on the generator and discriminator; Step S202: Input the random noise signal and category label into the generator to obtain the corresponding synthesized time-frequency feature data; Step S203: The discriminator is used to distinguish and predict the category of the synthesized time-frequency feature data and the historical real time-frequency feature data to obtain the discrimination result and the category prediction result. Step S204: Based on the discrimination result and the category prediction result obtained each time, the generation parameters of the generator and the discrimination parameters of the discriminator are iteratively optimized alternately to obtain the auxiliary classifier generative adversarial network model.

[0043] It should be understood that in actual scientific research, due to limitations imposed by environmental, time, and cost factors, acquiring underwater sample data is extremely difficult, and the insufficient total sample size leads to poor training performance of deep learning models. Generative adversarial networks (GANs), as a novel artificial intelligence technology, have wide applications in data augmentation, image generation, and other fields. However, traditional GAN ​​models are not suitable for underwater echo samples and cannot be directly used for sample data generation. Therefore, to address the problem of insufficient underwater target data, this embodiment proposes an ACGAN model to expand the underwater target dataset. The ACGAN model combines label information and reconstructs label information at the end of the discriminant module, improving the quality of generated data by combining these two features.

[0044] It should be noted that, in this embodiment, the historical real time-frequency feature data can be a kind of time-frequency feature data actually collected in the past, while the above-mentioned initial time-frequency feature dataset can be understood as a collection of newly collected time-frequency feature data.

[0045] It is easy to understand, please combine Figure 3 To understand, Figure 3 This describes the structure of the ACGAN model. In this embodiment, the generator in the ACGAN model receives class labels in addition to random noise signals. These class labels can be numerical labels 0, 1, 2, etc., representing the class of the input data. The discriminator's input remains real data (i.e., the aforementioned historical real time-frequency feature data or initial time-frequency feature dataset) and fake data generated by the generator (i.e., the aforementioned synthetic time-frequency feature data), but the output becomes two-fold. One type is the discrimination result, used to characterize the output that distinguishes between real and fake data; the other type is the class prediction result, used to classify the input real and fake data based on the corresponding class labels. Based on this architecture, label information (i.e., class labels) can be reconstructed while performing real / fake discrimination, greatly improving the quality of the generated samples.

[0046] It should be noted that, please refer to... Figure 4 To understand, Figure 4 This describes the training process of the ACGAN model. In this embodiment, the generator can be used... This means that the discriminator can be used This means that during the training process of the ACGAN model, label information (category labels) is added to the input of the generator, and label information (category labels) is reconstructed in the output of the discriminator.

[0047] The training objective of the generator is to make the generator's loss function... Minimize: ; ; ; in, The loss for determining the truth value of the generator. The classification loss of the generator, For category labels, Indicates "true". It means "fake".

[0048] The training objective of the discriminator is to make the discriminator's loss function... maximize: ; ; ; in, The loss caused by the discriminator's determination of true and false values. The classification loss of the discriminator, For category labels, Indicates "true". It means "fake".

[0049] It should be noted that, in this embodiment, as can be seen from the above formula, the loss function of the discriminator... It can be divided into two parts: the loss of the discriminator in determining whether something is true or false. Classification loss of the discriminator Specifically, this can help determine whether a loss is real or fake. Classification loss of the discriminator The sum. Correspondingly, the generator's loss function. It can be divided into two parts: the generator's loss for determining whether something is true or false. Classification loss of generator Specifically, this can be the loss for determining the truth value of the generator. Classification loss of the generator The sum of these two parameters is crucial. Therefore, the discriminator aims to both distinguish the authenticity of the input data and correctly classify it; while the generator aims to correctly classify the input data, even if the discriminator cannot determine its authenticity. Through this adversarial process, the adjustable parameters of the generator (i.e., the generation parameters) and the adjustable parameters of the discriminator (i.e., the discrimination parameters) are iteratively updated and optimized alternately based on the results of each adversarial interaction until the aforementioned effect is achieved, thus obtaining the auxiliary classifier generative adversarial network model proposed in this embodiment.

[0050] It is worth noting that in this embodiment, the label information (category label) Instead of being directly input into the discriminator, the label information serves as a reference standard, allowing the discriminator's classification results to gradually approach this standard through continuous training. In the ACGAN model, the class labels are only input into the generator during training, while the label information is not input into the network during discriminator training; it is only used as a standard to determine the accuracy of the discriminator's label reconstruction.

[0051] In a specific case, the function for determining the true / false loss in the ACGAN model can be the binary cross-entropy (BCE), while the classification loss function uses the categorical cross-entropy (CCE) function from the Keras library. The formulas are as follows: ; in, The number of samples for each training session. The number of categories in the sample set. For the true label of the j-th category, This represents the actual output value. The BCE mentioned above is actually a special type of multi-class loss function, specifically for true and false classifications. Here, CCE is mainly used as the loss value for the discriminator's reconstruction of the class labels. As training continues, the discriminator's reconstruction of the class labels becomes more and more accurate, and this loss value will decrease.

[0052] Furthermore, in this embodiment, before the step of performing multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features, the method further includes: Step S301: Construct several dilated convolutional branches with different dilation rates and kernels; Step S302: Set each of the dilated convolution branches in parallel to obtain the parallel multi-scale dilated convolution model.

[0053] It should be noted that, please refer to... Figure 5 To understand, Figure 5This document demonstrates the structure of the parallel multi-scale dilated convolution model used in this embodiment. Specifically, the constructed dilated convolution branches can be three, meaning the parallel multi-scale dilated convolution model includes three parallel dilated convolution channels. Each channel can have a convolution kernel with a different dilation rate. By setting dilated convolution blocks with different dilation rates, multi-scale information can be obtained while avoiding grid effects. Specifically, a 1×1 convolution can be used in the first part of each channel to reduce the input feature dimensionality and model computational complexity. In the second part, dilated convolutions with different dilation rates can be used for feature extraction, allowing the three parallel dilated convolution branches of the model to obtain features with different receptive fields, thereby capturing multi-scale information components on the same feature. This design, in turn, can help the model better analyze local and global information in the data. Subsequently, the three channels can be fused by concatenation. In the third part, a 1×1 convolution is used to capture inter-channel information, learn the dependencies between channels, reduce the number of parameters when acquiring channel attention information, and finally generate the output target time-frequency features.

[0054] It is easy to understand that in this embodiment, multi-scale features are captured by parallel dilated convolution of the PMDC model, which maintains feature resolution without increasing computational cost and reduces overfitting.

[0055] Based on the first and / or second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to that in embodiments one and two above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 as well as Figure 11 The channel time-frequency attention model includes a channel attention submodule and a time-frequency attention submodule. The step of allocating attention weights in the channel dimension and time-frequency dimension to the target time-frequency features based on the channel time-frequency attention model to obtain the corresponding weighted target time-frequency features includes: Step S41: Obtain the channel attention weights of the target time-frequency features in the channel dimension through the channel attention submodule; Step S42: Obtain the time-frequency attention weights of the target time-frequency features in the time-frequency dimension through the time-frequency attention submodule; Step S43: Based on the time-frequency attention weight, the channel attention weight, and the target time-frequency feature, obtain the weighted target time-frequency feature.

[0056] It should be understood that as more data is input into the neural network, the number of parameters in the segmentation network also increases, leading to a need for more parameters to represent and extract features. However, too much redundant feature information can cause the neural network to ignore the most important parameters.

[0057] It should be noted that, please refer to... Figure 7 To address the aforementioned issues, this embodiment employs a fusion of a channel attention submodule and a time-frequency attention submodule to construct a channel time-frequency attention model. The channel attention submodule can be understood as a Squeeze-and-Excitation Network (SeNet), which multiplies the learned weights back to the corresponding channels of the original feature map, achieving adaptive recalibration for different channel features. This allows the network to focus more on information-rich feature channels and suppress useless channels. Specifically, a channel attention map is generated using the channel attention mechanism to capture inter-channel relationships and identify effective content in the target time-frequency features. Subsequently, the spatial attention (time-frequency attention) mechanism of the time-frequency attention submodule is used to calculate the output features of the feature map (target time-frequency features), thereby supplementing the channel attention.

[0058] The above design enhances the ability to extract information from different dimensions of feature maps, forming feature mappings in both time and frequency dimensions. By fusing these two approaches, the model can extract weights for each dimension across different channels, thereby achieving more comprehensive extraction of key feature information and significantly improving the correlation between various features in the channel and spatial dimensions. This is more conducive to the extraction of effective features from the target area and enhances the network's sensitivity to information.

[0059] Furthermore, in this embodiment, the step of obtaining the time-frequency attention weights of the target time-frequency features in the time-frequency dimension through the time-frequency attention submodule includes: Step S421: Perform dilated convolution on the target time-frequency features, and perform global average pooling along the time dimension and frequency dimension respectively to obtain the time dimension feature vector and the frequency dimension feature vector. Step S422: Obtain the time-frequency attention weight based on the time dimension feature vector and the frequency dimension feature vector.

[0060] It is worth noting that, for ease of understanding, the construction structures of the channel attention submodule and the time-frequency attention submodule in this embodiment will be described below.

[0061] Please combine Figure 8 To understand, Figure 8This demonstrates the structure of the channel attention submodule. The channel attention submodule achieves this by fusing the mean and maximum values ​​extracted from each region. First, local pooling is used to extract the mean and maximum values ​​within each region. Then, feature fusion is performed based on the number of channels, and global pooling aggregates the spatial information between regions in the entire feature map. Next, a 1×1 convolutional layer is used to compress the spatial dimension of the input feature map, and finally, element-wise summation and merging operations are performed. This attention module, combining local and global pooling techniques, can accurately extract the maximum value information from important temporal features, avoid the influence of outliers on the scoring results, and capture important features from key regions of underwater acoustic signal data.

[0062] Please combine Figure 9 To understand, Figure 9 This demonstrates the structure of the time-frequency attention submodule. The time-frequency attention submodule enriches its feature representation by reconstructing the structural information of the feature map using a larger receptive field. Subsequently, global average pooling is applied to perform pooling operations on the time and frequency dimensions to extract key information from each channel dimension and generate corresponding feature vectors. The dimensional features of these feature vectors can be calculated using the following formula: Time dimension features for: ; Frequency dimension features for: ; in, Indicates the height of the feature map. Indicates the width of the feature map. , These represent the number of channels, respectively. The feature map reconstructed after dilated convolution is denoted as... as well as .

[0063] By concatenating the obtained dimensional features, the final fused features are obtained. That is, feature map: ; in, It is a convolution function. This is a concatenation function.

[0064] Feature map Then, segmentation is performed according to different dimensions. After a 1×1 convolution operation with an activation function, the attention weights for each dimension are obtained. Subsequently, the features from each dimension are weighted and fused to finally output the feature map. That is, the aforementioned weighted target time-frequency characteristics: ; in, Represents matrix multiplication. and This represents the activation function with a 1×1 convolutional layer.

[0065] Furthermore, in this embodiment, the step of extracting the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and outputting the underwater target classification result based on the bidirectional temporal feature mapping relationship includes: Step S51: Based on the bidirectional long short-term memory network model, obtain the forward hidden state sequence and the backward hidden state sequence of the weighted target time-frequency features respectively; Step S52: Based on the forward hidden state sequence and the backward hidden state sequence, form the bidirectional temporal feature mapping relationship in chronological order; Step S53: Output underwater target classification results based on the bidirectional temporal feature mapping relationship.

[0066] It should be understood that, please combine Figure 10 To understand this, the unidirectional LSTM model has four interacting layers, namely the main layer... and input gate Forgotten Gate and output gate Its main function is to synthesize the feature vectors extracted by the model. Specifically, the main layer... Analyze the current input and the previous hidden state Function; Input gate Controlling the input of new information; forget gate Controls the retention of new information and the forgetting of old information; output gates Control information output.

[0067] The expression for the above relationship is as follows: ; ; ; ; ; ; in, for The input to the time-matrix model, for The output of the time-mapping model for The memory state of the time-matter model for The hidden state (vector) of the model at each time step. This represents the weight matrix between the layer and each gate; Indicates the corresponding bias term; Represents the tangent hyperbolic function; The activation function is sigmoid.

[0068] It should be noted that in this embodiment, a Bi-LSTM model is used to learn the inherent correlations and dependencies between the weighted target time-frequency features, thereby improving classification accuracy. This is primarily to address the technical problem that traditional unidirectional LSTM models, when processing time series data, only consider information inference from the beginning to the end, ignoring the influence of later information, thus leading to unreliable accuracy in time series prediction results.

[0069] It is easy to understand, please combine Figure 11 To understand, Figure 11 This embodiment demonstrates the structure of the Bi-LSTM model used in this example. The Bi-LSTM model proposed in this embodiment can compute input vectors containing past and future information within a specific time range (i.e., the forward hidden state sequence and backward hidden state sequence mentioned above). In the Bi-LSTM model, the forward state and forward state neurons constitute a bidirectional LSTM, with one part responsible for the positive time direction (forward computation) and the other part responsible for the opposite time direction (backward computation). Essentially, the Bi-LSTM model can be understood as a bidirectional architecture composed of two independent LSTM model networks, which computes the hidden vector in the positive time direction during solution processing. Generate new hidden vectors and the hidden vector in the opposite time direction Generate new hidden vectors The above relationship is the bidirectional temporal feature mapping relationship described above. Their outputs are combined to form the final prediction result. The calculation formula is as follows: ; ; ; in, These are the weight coefficients that connect each layer to the previous hidden state. It is the bias value.

[0070] The above final prediction results This can be understood as the underwater target classification result mentioned above.

[0071] This application also provides an underwater target classification device. Please refer to... Figure 12The underwater target classification device includes: Data preprocessing module 10 is used to perform wavelet transform on the measured underwater target noise signal to form the corresponding initial time-frequency feature dataset; The dataset expansion module 20 is used to generate an adversarial network model based on an auxiliary classifier to expand the initial time-frequency feature dataset and obtain the corresponding target time-frequency feature dataset. Feature extraction module 30 is used to perform multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features; The weight allocation module 40 is used to allocate attention weights in the channel dimension and the time-frequency dimension to the target time-frequency features based on the channel time-frequency attention model, and obtain the corresponding weighted target time-frequency features. The target classification module 50 is used to extract the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and output the underwater target classification result based on the bidirectional temporal feature mapping relationship.

[0072] The underwater target classification device provided in this application embodiment, employing the underwater target classification method described above, can solve the technical problem of effectively improving the accuracy and stability of underwater target classification when samples are scarce. Compared with the prior art, the beneficial effects of the underwater target classification device provided in this application embodiment are the same as those of the underwater target classification method described above, and other technical features in the underwater target classification device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0073] This application provides an underwater target classification device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the underwater target classification method in Embodiment 1 above.

[0074] The following is for reference. Figure 13 The diagram illustrates a structural schematic suitable for implementing an underwater target classification device according to embodiments of this application. The underwater target classification device in these embodiments may include, but is not limited to, fixed terminals such as vehicle-mounted terminals. Figure 13 The underwater target classification device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0075] like Figure 13As shown, the underwater target classification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the underwater target classification device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the underwater target classification device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows an underwater target classification device with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0076] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0077] The underwater target classification device provided in this application, employing the underwater target classification method described in the above embodiments, can solve the technical problem of effectively improving the accuracy and stability of underwater target classification when samples are scarce. Compared with the prior art, the beneficial effects of the underwater target classification device provided in this application are the same as those of the underwater target classification method provided in the above embodiments, and other technical features of this underwater target classification device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0078] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0080] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the underwater target classification method in the above embodiments.

[0081] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0082] The aforementioned computer-readable storage medium may be included in the underwater target classification device; or it may exist independently and not assembled into the underwater target classification device.

[0083] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the underwater target classification device, the underwater target classification device performs the following actions: performs wavelet transform on the measured underwater target noise signal to form a corresponding initial time-frequency feature dataset; expands the initial time-frequency feature dataset based on an auxiliary classifier generative adversarial network model to obtain a corresponding target time-frequency feature dataset; performs multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain corresponding target time-frequency features; assigns attention weights to the target time-frequency features based on a channel time-frequency attention model to obtain corresponding weighted target time-frequency features; and extracts the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on a bidirectional long short-term memory network model, and outputs the underwater target classification result based on the bidirectional temporal feature mapping relationship.

[0084] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0085] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0086] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0087] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described underwater target classification method. This addresses the technical problem of effectively improving the accuracy and stability of underwater target classification when samples are scarce. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the underwater target classification method provided in the above embodiments, and will not be elaborated upon here.

[0088] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A method for classifying underwater targets, characterized in that, The underwater target classification method includes the following steps: Wavelet transform is performed on the underwater target noise signal obtained from the actual measurement to form the corresponding initial time-frequency feature dataset; The initial time-frequency feature dataset is expanded based on the auxiliary classifier generative adversarial network model to obtain the corresponding target time-frequency feature dataset; Based on a parallel multi-scale dilated convolution model, multi-scale feature extraction is performed on the target time-frequency feature dataset to obtain the corresponding target time-frequency features. Based on the channel-time-frequency attention model, attention weights are assigned to the target time-frequency features in both the channel and time-frequency dimensions to obtain the corresponding weighted target time-frequency features. Based on the bidirectional long short-term memory network model, the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features is extracted, and the underwater target classification result is output based on the bidirectional temporal feature mapping relationship.

2. The underwater target classification method as described in claim 1, characterized in that, The step of performing wavelet transform on the measured underwater target noise signal to form the corresponding initial time-frequency feature dataset includes: Perform continuous wavelet transform on the underwater target noise signal to obtain the time-frequency distribution spectrum of the underwater target noise signal after scaling and translation operations; Based on the time-frequency distribution map, the initial time-frequency feature dataset is obtained.

3. The underwater target classification method as described in claim 1, characterized in that, Before the step of expanding the initial time-frequency feature dataset to obtain the corresponding target time-frequency feature dataset using the auxiliary classifier-based generative adversarial network model, the method further includes: An initial generative adversarial network model is constructed based on a generator and a discriminator; The random noise signal and category label are input into the generator to obtain the corresponding synthesized time-frequency feature data; The discriminator is used to distinguish and predict the categories of the synthetic time-frequency feature data and the historical real time-frequency feature data to obtain the discrimination results and the category prediction results. Based on the discrimination result and the category prediction result obtained each time, the generation parameters of the generator and the discrimination parameters of the discriminator are iteratively optimized alternately to obtain the auxiliary classifier generative adversarial network model.

4. The underwater target classification method as described in claim 1, characterized in that, Before the step of extracting multi-scale features from the target time-frequency feature dataset based on the parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features, the method further includes: Construct several dilated convolution branches with different dilation rates and kernels; By setting each of the dilated convolution branches in parallel, the parallel multi-scale dilated convolution model is obtained.

5. The underwater target classification method as described in claim 1, characterized in that, The channel time-frequency attention model includes a channel attention submodule and a time-frequency attention submodule. The step of allocating attention weights based on the channel time-frequency attention model to the target time-frequency features in terms of channel dimension and time-frequency dimension to obtain the corresponding weighted target time-frequency features includes: The channel attention weights of the target time-frequency features in the channel dimension are obtained through the channel attention submodule. The time-frequency attention weights of the target time-frequency features in the time-frequency dimension are obtained through the time-frequency attention submodule. The weighted target time-frequency features are obtained based on the time-frequency attention weights, the channel attention weights, and the target time-frequency features.

6. The underwater target classification method as described in claim 5, characterized in that, The step of obtaining the time-frequency attention weights of the target time-frequency features in the time-frequency dimension through the time-frequency attention submodule includes: Dilated convolution is performed on the target time-frequency features, and global average pooling is performed along the time dimension and frequency dimension respectively to obtain the time dimension feature vector and the frequency dimension feature vector. The time-frequency attention weights are obtained based on the time-dimensional feature vector and the frequency-dimensional feature vector.

7. The underwater target classification method as described in claim 1, characterized in that, The steps of extracting the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and outputting the underwater target classification result based on the bidirectional temporal feature mapping relationship include: Based on the bidirectional long short-term memory network model, the forward hidden state sequence and the backward hidden state sequence of the weighted target time-frequency features are obtained respectively. Based on the forward hidden state sequence and the backward hidden state sequence, the bidirectional temporal feature mapping relationship is formed in chronological order; The underwater target classification result is output based on the bidirectional temporal feature mapping relationship.

8. An underwater target classification device, characterized in that, The underwater target classification device includes: The data preprocessing module is used to perform wavelet transform on the underwater target noise signal obtained from the actual measurement to form the corresponding initial time-frequency feature dataset; The dataset expansion module is used to generate an adversarial network model based on an auxiliary classifier, and to expand the initial time-frequency feature dataset to obtain the corresponding target time-frequency feature dataset. The feature extraction module is used to perform multi-scale feature extraction on the target time-frequency feature dataset based on a parallel multi-scale dilated convolution model to obtain the corresponding target time-frequency features. The weight allocation module is used to allocate attention weights in the channel dimension and time-frequency dimension to the target time-frequency features based on the channel time-frequency attention model, and obtain the corresponding weighted target time-frequency features; The target classification module is used to extract the bidirectional temporal feature mapping relationship corresponding to the weighted target time-frequency features based on the bidirectional long short-term memory network model, and output the underwater target classification result based on the bidirectional temporal feature mapping relationship.

9. An underwater target classification device, characterized in that, The underwater target classification device includes: a memory, a processor, and an underwater target classification program stored in the memory and executable on the processor, the underwater target classification program being configured to implement the steps of the underwater target classification method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and the computer-readable storage medium stores an underwater target classification program, which, when executed by a processor, implements the steps of the underwater target classification method as described in any one of claims 1 to 7.

Images