A method, device and equipment for underwater acoustic target recognition

Abstract

The present invention provides a method, device and equipment for underwater acoustic target recognition. The method comprises: obtaining underwater acoustic audio information; determining hidden feature vectors and time-frequency feature vectors respectively according to the underwater acoustic audio information; splicing the hidden feature vectors and time-frequency feature vectors, and performing classification prediction according to the spliced ​​feature vectors. The underwater acoustic target recognition method provided by the present invention extracts features from underwater acoustic audio information to extract hidden feature vectors and time-frequency feature vectors respectively, thereby enriching the extracted feature vectors, and then performing prediction according to the feature vectors spliced ​​from the hidden feature vectors and the time-frequency feature vectors, thereby improving the accuracy of the prediction results.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to an underwater acoustic target recognition method, device and equipment. Background Art

[0002] As the strategic position of the ocean becomes increasingly prominent, the development and utilization of marine resources and space has become an important topic. Sound waves are currently the only form of energy that can propagate over long distances in the ocean. Hydroacoustic target recognition is of great significance to marine development and national defense security, and has now become one of the research hotspots in the field of hydroacoustics. In response to the needs of underwater detection in modern warfare, underwater target classification and recognition based on sound signals has also become a research hotspot in the field of underwater detection. Due to the advantages of strong concealment and long transmission distance of passive sonar, underwater target radiation noise collected through passive sonar is the main source of information for hydroacoustic target recognition. Passive sonar systems have good concealment and flexibility. Therefore, how to use ship radiation noise obtained by passive sonar to detect and identify underwater targets has become an urgent problem to be solved. However, due to the application of various ship stealth technologies and the complexity and diversity of the marine environment, the underwater target recognition task based on ship radiation noise faces huge challenges.

[0003] In recent years, deep learning has been developing continuously, and many models with strong end-to-end feature extraction capabilities have emerged. These models are widely used in various fields, and more and more people are applying them to underwater acoustic target recognition. The deep learning method is based on data-driven adaptive learning, which requires a large amount of data to support its learning of deep features from rich samples. The size of the data set is directly related to the accuracy of the final recognition and the generalization ability of the model. A data set that is too small may lead to problems such as overfitting, affecting the accuracy of the final verification set. However, given the application scenarios of underwater target recognition technology and the limitations of the complex underwater environment, it is not realistic to obtain real data in large quantities. The complex underwater environment will also greatly affect the results of signal acquisition, so our experimental data cannot fully cover all the characteristics of the target.

[0004] At present, in the process of underwater acoustic target recognition, it is necessary to convert the original underwater acoustic audio information into an underwater acoustic feature vector, and then perform underwater acoustic recognition based on the underwater acoustic feature vector. However, on the one hand, due to the influence of the objective environment, it is impossible to obtain a large amount of underwater acoustic audio information; on the other hand, the form of the currently extracted underwater acoustic feature vector is relatively simple. Based on these two reasons, the current underwater acoustic recognition results are not accurate enough. Summary of the invention

[0005] The purpose of the present invention is to solve the defects of the above-mentioned prior art and provide a method, device and equipment for underwater acoustic target recognition with more accurate prediction results.

[0006] A method for underwater acoustic target recognition, comprising:

[0007] Obtaining water sound audio information;

[0008] Determine the hidden feature vector and the time-frequency feature vector respectively according to the underwater acoustic audio information;

[0009] The hidden feature vector and the time-frequency feature vector are concatenated, and classification prediction is performed based on the concatenated feature vector.

[0010] Furthermore, in the above-mentioned underwater acoustic target recognition method, the step of determining the hidden feature vector according to the underwater acoustic audio information comprises:

[0011] Performing feature extraction on the underwater acoustic audio information to obtain a feature vector with a compressed length;

[0012] The hidden feature vector is determined according to the feature vector whose length is compressed.

[0013] Furthermore, in the above-mentioned underwater acoustic target recognition method, the step of extracting features from the underwater acoustic audio information to obtain a length-compressed feature vector includes:

[0014] The underwater acoustic audio information is divided into two paths, one path is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature; the other path is used to extract features from the underwater acoustic audio information and extract features from the extracted features again to obtain a second down-sampling feature;

[0015] The first down-sampled feature and the second down-sampled feature are added to obtain the feature vector with the length compressed.

[0016] Furthermore, in the above-mentioned method for underwater acoustic target recognition, the step of determining the hidden feature vector according to the feature vector whose length is compressed comprises:

[0017] Inputting the length-compressed feature vector into a trained waveform coding model for feature extraction to obtain the hidden feature vector;

[0018] The training process is:

[0019] Obtain unlabeled human voice audio information and water sound audio information;

[0020] Mixing the unlabeled human voice audio information and water sound audio information to obtain mixed audio information;

[0021] The mixed audio information is input into the original waveform coding model for training to obtain a trained waveform coding model.

[0022] Furthermore, in the above-mentioned underwater acoustic target recognition method, determining the time-frequency feature vector according to the underwater acoustic audio information comprises:

[0023] Determine a Mel-spectrogram of the underwater acoustic audio information according to the underwater acoustic audio information;

[0024] The time-frequency feature vector is determined according to the Mel-spectrogram of the underwater acoustic audio information.

[0025] An underwater acoustic target recognition device, comprising:

[0026] An acquisition unit, used for acquiring target underwater acoustic audio information;

[0027] A trained waveform coding model for determining a hidden feature vector based on the underwater acoustic audio information;

[0028] A mel-spectrogram coding model, used to determine a time-frequency feature vector according to the underwater acoustic audio information;

[0029] A concatenation unit, used for concatenating the hidden feature vector and the time-frequency feature vector;

[0030] The prediction unit is used to perform classification prediction on the concatenated feature vectors.

[0031] Furthermore, the above-mentioned underwater acoustic target recognition device further includes a feature extraction network;

[0032] The feature extraction network is arranged between the acquisition unit and the trained waveform coding model, and is used to downsample the underwater acoustic audio information to output a feature vector with a compressed length;

[0033] The length-compressed feature vector is used as input to the trained waveform coding model.

[0034] Further, in the above-mentioned underwater acoustic target recognition device, the feature extraction network includes: a first CNN module, a second CNN module, and a third CNN module;

[0035] The third CNN module is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature;

[0036] The first CNN module is used to perform a feature extraction network on the underwater acoustic audio information to obtain a third down-sampling feature; the second CNN module is used to perform a feature extraction on the third down-sampling feature to obtain a second down-sampling feature;

[0037] A merging module is used to add the first down-sampled feature and the second down-sampled feature to obtain the feature vector with compressed length.

[0038] Furthermore, in the above-mentioned underwater acoustic target recognition device, the trained waveform coding model is a model obtained by training the original waveform coding model;

[0039] The training process is:

[0040] Obtain unlabeled human voice audio information and water sound audio information;

[0041] Mixing the unlabeled human voice audio information and water sound audio information to obtain mixed audio information;

[0042] The mixed audio information is input into the original waveform coding model for training to obtain a trained waveform coding model.

[0043] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned underwater acoustic target recognition method when executing the program.

[0044] Beneficial effects:

[0045] The underwater acoustic target recognition method provided by the present invention performs feature extraction on underwater acoustic audio information to respectively extract hidden feature vectors and time-frequency feature vectors, thereby enriching the extracted feature vectors, and then performs prediction based on the feature vector obtained by concatenating the hidden feature vectors and the time-frequency feature vectors, thereby improving the accuracy of the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is one of the flow charts of the underwater acoustic target recognition method of the present invention;

[0047] Figure 2 This is the second flow chart of the underwater acoustic target recognition method of the present invention;

[0048] Figure 3 A schematic diagram of the structure of the underwater acoustic target recognition device provided by the present invention;

[0049] Figure 4 This is a schematic diagram of the feature extraction network structure of the present invention;

[0050] Figure 5 It is a structural schematic diagram of the transformer encoder provided by the present invention;

[0051] Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The existing deep learning models are mainly based on convolutional neural networks (CNNs), which have the advantages of sparse interaction, parameter sharing, and equivariant representation, and are good at processing two-dimensional image data. The local feature extraction is completed through the convolution kernel, and then the feature mapping is completed through pooling and downsampling. The layer-by-layer abstract extraction and mapping of the features enable the model to obtain more abstract and stable features than shallow learning. This method makes the model pay more attention to the local features of the sample, but for the time-frequency features of ship noise, the equally important global features are less concerned. The current deep learning model input usually has two types: waveform graphs and time-frequency images. Both have their advantages and can provide important feature information for the model, but the current methods are basically single-modal inputs, that is, the input is one of the two, which makes the final prediction result not accurate enough.

[0054] Figure 1 This is one of the flow charts of the underwater acoustic target recognition method of the present invention. Figure 2 This is the second flow chart of the underwater acoustic target recognition method of the present invention, as shown in Figure 1 , Figure 2 As shown, the method includes:

[0055] Step 101: Acquire water sound audio information.

[0056] Step 102: Determine the hidden feature vector and the time-frequency feature vector respectively according to the underwater acoustic audio information.

[0057] Step 103: concatenate the hidden feature vector and the time-frequency feature vector, and perform classification prediction based on the concatenated feature vector.

[0058] Specifically, on the one hand, the underwater acoustic audio information is first compressed through a feature extraction network composed of CNN, that is, f:x->e, and a vector e compressed to 1 / 4 of its original length is output. Then, the vector e is sent to a pre-trained waveform coding model to extract the hidden feature vector r of the underwater acoustic audio information. wave ;

[0059] r wave =f wave (e; W wave )

[0060] On the other hand, the Mel spectrogram of the underwater audio information is first extracted, that is, m, and then the Mel spectrogram m is sent to the Mel spectrogram encoding model to extract the time-frequency feature vector r from the time-frequency signal. mel ;

[0061] r mel =f mel (m; W mel )

[0062] Finally, the two are concatenated and sent to the classifier for classification prediction.

[0063] y pred =f cla (r wave ,r mel ; W cla )

[0064] The underwater acoustic target recognition method provided by the present invention performs feature extraction on underwater acoustic audio information to respectively extract hidden feature vectors and time-frequency feature vectors, thereby enriching the extracted feature vectors, and then performs prediction based on the feature vector obtained by concatenating the hidden feature vectors and the time-frequency feature vectors, thereby improving the accuracy of the prediction result.

[0065] Furthermore, the method of determining the hidden feature vector based on the underwater acoustic audio information includes: performing feature extraction on the underwater acoustic audio information to obtain a feature vector with a compressed length; and determining the hidden feature vector based on the feature vector with a compressed length.

[0066] Specifically, in order to reduce the amount of computation of the pre-trained waveform coding model, the present invention first extracts features from the underwater acoustic audio information, automatically downsamples the input underwater acoustic audio information by means of feature extraction, and retains all necessary information as much as possible to obtain a length-compressed feature vector. At the same time, feature extraction can also remove redundant information in the original audio to reduce the amount of computation of the model. Then, the length-compressed feature vector is used as the input of the pre-trained waveform coding model, and a hidden feature vector can be efficiently obtained.

[0067] The underwater acoustic target recognition method provided by the present invention performs feature extraction before sending the underwater acoustic audio information into a pre-trained waveform coding model for feature extraction, thereby improving the efficiency of the waveform coding model for feature extraction.

[0068] The following is a detailed description of how to extract features based on underwater acoustic audio information to obtain a feature vector with a compressed length:

[0069] The underwater acoustic audio information is divided into two paths, one path is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature; the other path is used to extract features from the underwater acoustic audio information and then extract features from the extracted features again to obtain a second down-sampling feature; the first down-sampling feature and the second down-sampling feature are added together to obtain the length-compressed feature vector.

[0070] Specifically, the present invention divides the underwater acoustic audio information into two down-sampling paths for feature extraction. One path is down-sampled through a single-core CNN to extract feature information; the other path is down-sampled through a multi-core CNN to extract some feature information that would otherwise be lost, and then the two are added together to obtain the feature vector with the compressed length.

[0071] The underwater acoustic target recognition method provided by the present invention can, on the one hand, retain necessary information as much as possible by dividing the underwater acoustic audio information into two paths for downsampling, so that the extracted features are richer, thereby improving the accuracy of the final underwater acoustic recognition. On the other hand, through two-path downsampling, redundant information in the original audio can be removed, reducing the amount of calculation of the trained waveform coding model, thereby improving the efficiency of the final underwater acoustic recognition.

[0072] The following is an explanation of how to determine the hidden feature vector based on the feature vector whose length is compressed:

[0073] Input the length-compressed feature vector into a trained waveform coding model for feature extraction to obtain a hidden feature vector. The trained waveform coding model is obtained by training the original waveform coding model. The training process is: obtaining unlabeled human voice audio information and water sound audio information; mixing the unlabeled human voice audio information and water sound audio information to obtain mixed audio information; inputting the mixed audio information into the original waveform coding model for training to obtain a trained waveform coding model.

[0074] Specifically, since the original waveform coding model is used to extract features from human voices, the present invention needs to train the original waveform coding model in order to enable the waveform coding model to also extract features from underwater acoustic audio information. The training process is: unlabeled human voice audio information and underwater acoustic audio information are mixed and input into the original waveform coding model for training, so that the trained waveform coding model can effectively identify underwater acoustic audio information. Thus, the migration of the original waveform coding model is achieved, and through training, the waveform coding model that recognizes human voice audio information is migrated to the waveform coding model that recognizes underwater acoustic audio information.

[0075] The underwater acoustic target recognition method provided by the present invention realizes the migration of the waveform coding model by inputting the length-compressed feature vector into a trained waveform coding model for feature extraction, that is, migrating from a waveform coding model for recognizing human voice audio information to a waveform coding model that can effectively recognize underwater acoustic audio information, thereby improving the efficiency and accuracy of underwater acoustic audio information recognition.

[0076] The present invention also provides a hydroacoustic target recognition device. Figure 3 The schematic diagram of the structure of the underwater acoustic target recognition device provided by the present invention is as follows: Figure 3 As shown, the device comprises:

[0077] An acquisition unit, used for acquiring target underwater acoustic audio information;

[0078] A feature extraction network is used to downsample the underwater acoustic audio information to output a length-compressed feature vector; and the length-compressed feature vector is used as input of a trained waveform coding model.

[0079] A trained waveform coding model for determining a hidden feature vector based on the underwater acoustic audio information;

[0080] A mel-spectrogram coding model, used to determine a time-frequency feature vector according to the underwater acoustic audio information;

[0081] A concatenation unit, used for concatenating the hidden feature vector and the time-frequency feature vector;

[0082] The prediction unit is used to perform classification prediction on the concatenated feature vectors.

[0083] Further, Figure 4 This is a schematic diagram of the feature extraction network structure of the present invention, as shown in Figure 4 As shown, the feature extraction network includes: a first CNN module, a second CNN module, and a third CNN module;

[0084] The third CNN module is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature;

[0085] The first CNN module is used to perform feature extraction network on the underwater acoustic audio information to obtain the third down-sampling feature; the second CNN module is used to perform feature extraction on the third down-sampling feature to obtain the second down-sampling feature;

[0086] A merging module is used to add the first down-sampled feature and the second down-sampled feature to obtain the feature vector with compressed length.

[0087] Specifically, the present invention divides the underwater acoustic audio information into two paths and performs downsampling operations on each path, and adds the features of the two downsampled paths to obtain a length-compressed feature vector. Among them, one path of the underwater acoustic audio information is downsampled by the third CNN module to extract features and obtain the first downsampled features. The third CNN module is a single-core neural network. One path is downsampled by the first CNN module and the second CNN module in turn, and multiple feature extractions are performed, so that while downsampling, some downsampled features that would otherwise be lost are extracted, thereby retaining as much necessary information as possible.

[0088] Furthermore, the output, output, convolution kernel, step size, and padding of the first CNN module are: 1, 64, 10, 2, 4 respectively; the output, output, convolution kernel, step size, and padding of the second CNN module are: 64, 1, 10, 2, 4 respectively; the output, output, convolution kernel, step size, and padding of the third CNN module are: 1, 1, 10, 4, 4 respectively.

[0089] Furthermore, based on the above-mentioned underwater acoustic target recognition device, the trained waveform coding model is a model obtained by training the original waveform coding model; the training process is: obtaining unlabeled human voice audio information and underwater acoustic audio information; mixing the unlabeled human voice audio information and underwater acoustic audio information to obtain mixed audio information; inputting the mixed audio information into the original waveform coding model for training to obtain a trained waveform coding model.

[0090] The following is a detailed description of the underwater acoustic target recognition device provided by the present invention:

[0091] The device provided by the present invention mainly consists of three parts: feature extraction network, waveform coding model and mel-spectrogram coding model. On the one hand, the underwater acoustic audio information is first compressed through a feature extraction network composed of CNN, that is, f:x->e. Then it is sent to the pre-trained model waveform coding model to extract the hidden feature vector r of the underwater acoustic audio information. wave .

[0092] r wave =f wave (e; W wave )

[0093] On the other hand, the Mel spectrogram of the underwater audio information is first extracted, that is, m, and then it is sent to the Mel spectrogram encoding model to extract the time-frequency feature vector r from the time-frequency signal mel .

[0094] r mel =f mel (m; W mel )

[0095] Finally, the two are concatenated and sent to the classifier for classification prediction.

[0096] y pred =f cla (r wave ,r mel ; W cla )

[0097] The following is an explanation of the feature extraction network:

[0098] For underwater targets, the main noise comes from mechanical noise, which is concentrated in the low-frequency band and is different from the frequency band of human voice. Since the pre-trained model is trained on human voice data, if the underwater acoustic audio is directly input, there will be a gap between the pre-training and downstream training data sets. If the underwater acoustic audio is directly downsampled and input, some necessary information may be lost. Therefore, we designed a compression network to automatically downsample the input audio while retaining all necessary information as much as possible. At the same time, it can remove redundant information in the original audio and reduce the amount of computation of the model. The network consists of a multi-layer convolutional neural network, which inputs the original audio x and outputs a vector e compressed to 1 / 4 of its original length.

[0099] Waveform Coding Model

[0100] The waveform coding model is a self-supervised model based on the transformer structure to extract feature vectors from the original waveform. The model consists of three parts: feature coding module, context feature extraction module, and quantization module. First, the original waveform is input into the feature coding module composed of convolution blocks, normalization layers, and GELU activation functions for simple feature coding. The output is then sent to the context feature extraction module after masking to learn context representation, and to the quantization module to learn discrete features. Finally, the context representation and discrete features are compared and learned so that the model can learn the feature vector of the original waveform.

[0101] The feature encoding module is a model composed of several layers of one-dimensional convolutional neural networks, which is used to extract features from underwater acoustic audio information and convert the one-dimensional original audio into a two-dimensional feature vector. A seven-layer convolutional neural network is used, with a step size of (5, 2, 2, 2, 2, 2, and a convolution kernel width of (10, 3, 3, 3, 3, 2, 2), which is equivalent to generating a 512-dimensional feature vector every 20 milliseconds.

[0102] The context feature extraction module is a context feature network based on a transformer structure, which is used to extract the context representation of the feature vector output in the feature encoding module.

[0103] Specifically, the output in the feature encoding module is randomly masked and then fed into the context feature extraction module, where the context representation of the masked feature vector is used to predict the discretized mapping of the vector.

[0104] The quantization module is used to discretize the output of the feature encoding module to achieve discretization of the feature space. Because it is difficult to predict continuous values, it is easier to predict discrete targets. The discretization method is to first perform a gumbel softmax operation, and then find the content of the codebook corresponding to the maximum value in the vector as the output.

[0105] The loss function in the waveform coding model provided by the present invention is introduced below:

[0106] Loss function: The loss function is divided into two parts, namely contrast loss L m and discrete loss L d .

[0107] L=L m +αL d

[0108] The contrast loss is mainly used to take the vector of the mask position t as the center, and calculate ct after masking the position vector t. At the same time, K are selected from all masked frames of the same speech as negative examples, and the positive example is qt. The similarity between them and ct is calculated respectively, so that the positive examples are as similar as possible and the negative examples are as dissimilar as possible.

[0109]

[0110] Discrete loss L d It is mainly used to make the probability of occurrence of each entry in each G as even as possible during the quantization process, and to encourage the most even use of the contents of the codebook.

[0111]

[0112] The following is an introduction to the implementation process of how to obtain the trained waveform coding model:

[0113] The present invention first uses a large amount of unlabeled human voice data and part of the underwater acoustic data to pre-train the waveform coding model, then removes the quantization module of the waveform coding model, and connects a fully connected layer as the trained waveform coding model. The waveform coding model is used to extract the hidden feature vector of the compressed vector e.

[0114] Specifically, the pre-trained waveform coding model includes: a feature encoder, a context feature extraction module, and a fully connected layer. The compressed vector e is subjected to feature extraction by the feature encoding module to obtain a two-dimensional feature vector; the two-dimensional feature vector is respectively input into the context feature extraction module and the quantization module, and the context representation of the two-dimensional feature vector is extracted by the context feature extraction module; the two-dimensional feature vector is subjected to row discretization operation by the quantization module to obtain a discretized feature vector; the context representation of the two-dimensional feature vector is compared with the discretized feature vector to calculate the loss, and the parameters of the context feature extraction module are adjusted by the loss until the loss is within a preset range, and the adjusted context feature extraction module is used as the context feature extraction module of the trained waveform coding model, and then the quantization module in the original waveform coding model is removed, and a fully connected layer is connected after the context feature extraction module to obtain a trained waveform coding model.

[0115] The following is an introduction to the Mel spectrogram encoding model:

[0116] When using the human voice pre-trained coding model to realize underwater acoustic target recognition, although the time sequence information can be better captured, the introduction of the feature extraction network will inevitably affect the frequency domain information. Considering the effectiveness of Mel spectrum in time-frequency analysis, the present invention extracts features from Mel spectrum by introducing a transformer encoder.

[0117] In summary, the waveform coding model and the mel-spectrogram coding model in the model provided by the present invention are both composed of a transformer encoder. Figure 5 is a schematic diagram of the structure of the transformer encoder provided by the present invention, such as Figure 5 As shown in the figure, Transformer is mainly composed of three modules: position encoding module, multi-head attention module and forward feedback module.

[0118] Position encoding module: In order to make the model understand the position information of the sequence, we need to add relative position encoding, that is, position encoding,

[0119]

[0120]

[0121] Among them, t represents the position in the time dimension, and f represents the position in the frequency dimension. This encoding can reflect their relative position relationship, making it easy for the model to learn relative position information.

[0122] Multi-head attention module: Intuitively, the attention mechanism is similar to the human attention allocation process. In the process of information processing, different information is assigned different attention weights. The calculation is mainly divided into three steps. The first step is to calculate the similarity between Q and K to get the weight. The second step is to divide it by the 1 / 2 power of the dimension of K and normalize the weight using the softmax function. The last step is to perform a weighted summation of the weight and the corresponding V.

[0123]

[0124] Multi-head attention means observing the characteristics of information from multiple different angles and assigning different attention weights, so that information can be extracted in different feature subspaces.

[0125] MultiHead(Q,K,V)=concat(head1,...,head h )W O

[0126] where head i =Attention(QW i Q ,KW i K ,VW i V )

[0127] Feedforward Module: Except for the attention layer, each layer in the encoding layer contains a fully connected feedforward, which consists of two linear transformation layers with a ReLU activation in between.

[0128] FFN(x)=max(0,xW1+b1)W2+b2

[0129] The method proposed in the present invention has a higher accuracy rate than other models when detecting underwater targets, and can also achieve good results in the case of small samples.

[0130] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communications interface 620 and the memory 630 communicate with each other via the communication bus 640. The processor 610 may call the logic instructions in the memory 830 to execute the underwater acoustic target recognition method, which includes:

[0131] Obtaining water sound audio information;

[0132] Determine the hidden feature vector and the time-frequency feature vector respectively according to the underwater acoustic audio information;

[0133] The hidden feature vector and the time-frequency feature vector are concatenated, and classification prediction is performed based on the concatenated feature vector.

[0134] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0135] Experimental example:

[0136] The experimental data set was selected from the measured ship radiation noise and collected in the Danjiangkou Lake test. There are three different types of ships, sailing on a variety of routes and speeds, and an eight-channel hydrophone is arranged in the north and south to passively collect data, with a sampling frequency of 48kHz. 9 minutes of data are selected for each ship as the data set, and each data is 2s audio of a single channel, with a total of 3*8*2*9*60 / 2=12960 data. Among them, according to the time division, the first 80% is used as the training set, and the last 20% is used as the test set, that is, the total number of training set samples is 10368, and the total number of test set samples is 2592.

[0137] Hyperparameter setting: We select a batch size of 32, a learning rate of 0.0001, a cross entropy loss, an optimizer of AdamW, and a linear increase in the learning rate in the first iteration. We train for 60 iterations in total, where the learning rate is reduced to 0.2 times the original value every 20 iterations.

[0138] We first conducted an experiment to compare different models and different input formats, mainly selecting two models, VGG and Transformer, and waveforms, short-time Fourier transforms, and Mel spectra. Figure 3Different input formats are used to compare the prediction effect of the proposed model. The experimental results show that the prediction accuracy of our model is 96.99%, which has obvious advantages over other models, as shown in Table 1.

[0139] Table 1

[0140]

[0141] Then, in order to verify the effectiveness of the proposed compression network, we further conducted comparative experiments. The experiment found that when only the waveform encoding model was used without using pre-trained weights, the accuracy was low, which shows that less underwater acoustic data is not enough to support the huge model training. At the same time, the waveform encoder with a sampling rate of 4k is also lower than the feature extraction network + waveform encoder with a sampling rate of 16k. These two show that the better effect of our proposed model is not due to the large model, but the role of transfer learning. It also shows the effectiveness of the feature extraction network (compression network) compared to direct downsampling, see Table 2.

[0142] Table 2

[0143]

[0144] In order to verify the effect of the proposed model under small sample data, the present invention changed the number of data sets and re-experimented. Only one channel in the hydrophone was selected as the data set, that is, the data volume was 810, and 20%, 40%, 60%, and 80% were selected as training sets for training. The transformer model with mel-spectrogram as input and the vgg model with waveform graph as input were selected for comparison and experiment. The experimental results showed that when 20% of the data set was selected for training, each model was difficult to fit due to the small amount of data. However, when the data set is slightly increased, especially after selecting 40%*810=324 samples and more samples as training sets, the model we proposed can have a higher accuracy rate than other models. This shows that the device provided by the present invention can have a more powerful prediction ability in the case of small samples.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for underwater acoustic target recognition, characterized in that: include: Obtaining water sound audio information; Determine the hidden feature vector and the time-frequency feature vector respectively according to the underwater acoustic audio information; Concatenating the hidden feature vector and the time-frequency feature vector, and performing classification prediction based on the concatenated feature vector; Determining the hidden feature vector according to the underwater acoustic audio information comprises: Performing feature extraction on the underwater acoustic audio information to obtain a feature vector with a compressed length; Determining the hidden feature vector according to the feature vector whose length is compressed; The extracting features of the underwater acoustic audio information to obtain a length-compressed feature vector comprises: The underwater acoustic audio information is divided into two paths, one path is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature; the other path is used to extract features from the underwater acoustic audio information and extract features from the extracted features again to obtain a second down-sampling feature; Adding the first down-sampled feature and the second down-sampled feature to obtain the feature vector with compressed length; Determining the hidden feature vector according to the feature vector whose length is compressed comprises: Inputting the length-compressed feature vector into a trained waveform coding model for feature extraction to obtain the hidden feature vector; The training process is: Obtain unlabeled human voice audio information and water sound audio information; Mixing the unlabeled human voice audio information and water sound audio information to obtain mixed audio information; The mixed audio information is input into the original waveform coding model for training to obtain a trained waveform coding model.

2. The underwater acoustic target recognition method according to claim 1, characterized in that: Determining the time-frequency feature vector according to the underwater acoustic audio information comprises: Determine a Mel-spectrogram of the underwater acoustic audio information according to the underwater acoustic audio information; The time-frequency feature vector is determined according to the Mel-spectrogram of the underwater acoustic audio information.

3. An underwater acoustic target recognition device, the characteristics of which include: An acquisition unit, used for acquiring target underwater acoustic audio information; A trained waveform coding model for determining a hidden feature vector based on the underwater acoustic audio information; A mel-spectrogram coding model, used to determine a time-frequency feature vector according to the underwater acoustic audio information; A concatenation unit, used for concatenating the hidden feature vector and the time-frequency feature vector; A prediction unit, used for classifying and predicting the concatenated feature vectors; It also includes feature extraction networks; The feature extraction network is arranged between the acquisition unit and the trained waveform coding model, and is used to downsample the underwater acoustic audio information to output a feature vector with a compressed length; Using the length-compressed feature vector as input to the trained waveform coding model; The feature extraction network includes: a first CNN module, a second CNN module, and a third CNN module; The third CNN module is used to extract features from the underwater acoustic audio information to obtain a first down-sampling feature; The first CNN module is used to perform a feature extraction network on the underwater acoustic audio information to obtain a third down-sampling feature; the second CNN module is used to perform a feature extraction on the third down-sampling feature to obtain a second down-sampling feature; a merging module, configured to add the first down-sampled feature and the second down-sampled feature to obtain the feature vector with the length compressed; The trained waveform coding model is a model obtained by training the original waveform coding model; The training process is: Obtain unlabeled human voice audio information and water sound audio information; Mixing the unlabeled human voice audio information and water sound audio information to obtain mixed audio information; The mixed audio information is input into the original waveform coding model for training to obtain a trained waveform coding model.

4. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the underwater acoustic target recognition method as described in any one of claims 1 to 2 is implemented.

Images