Method and apparatus for constructing an underwater acoustic target recognition model

Abstract

An embodiment of the present invention discloses a method and apparatus for constructing an underwater acoustic target recognition model. The method for constructing the underwater acoustic target recognition model includes: selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples; training a densely connected network with the training samples to obtain a trained densely connected network; and inputting the test samples into the trained densely connected network to obtain a converged densely connected network. Through the present invention, the problem of insufficient underwater acoustic samples in the related art is solved, the complex and heavy feature engineering is avoided, and the feature reuse is greatly strengthened, and the technical effect of alleviating the insufficient underwater acoustic samples to a certain extent is achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology applications, and particularly to a method and device for constructing an underwater acoustic target recognition model. Background Technique

[0002] Underwater acoustic target recognition is a technology for classifying targets using target radiated noise signals. The classification and recognition method based on traditional statistical models mainly consists of three steps: preprocessing, feature extraction and selection, and classifier. Common features include power spectrum, auditory spectrum, Detection of Envelope Modulation On Noise (DEMON) spectrum, Low Frequency Analysis Recording (LOFAR) spectrum, wavelet features, loudness features, Mel-Frequency Cepstral Coefficients (MFCC) features, Perceptual Linear Predictive (PLP) features, etc. Whether reliable features can be extracted will directly affect the recognition rate of underwater acoustic targets. In the context of the big data era, the structure and algorithm of artificial neural networks have been continuously optimized and have achieved great success in the field of computer image recognition. At the same time, with the rapid development of computer hardware technology, the computing speed of neural networks has also been greatly improved, thus further promoting the development of neural networks. How to use deep neural networks to complete the underwater acoustic target recognition task has also received more and more attention. Deep neural networks can directly extract features from the original waveform signal, avoiding complex feature extraction and selection work, and higher recognition rates can also be obtained using neural networks.

[0003] In practical engineering applications, it is difficult to obtain underwater acoustic data and the confidentiality is relatively strong, and real underwater acoustic target data is very scarce, and the high-quality underwater acoustic samples for training the target recognition model are further reduced.

[0004] Aiming at the problem of insufficient underwater acoustic samples in the current related technologies, no effective solution has been obtained yet. Summary of the Invention

[0005] Embodiments of the present invention provide a method and device for constructing an underwater acoustic target recognition model to at least solve the problem of insufficient underwater acoustic samples in the related technologies.

[0006] According to one aspect of an embodiment of the present invention, a method for constructing an underwater acoustic target recognition model is provided, including: selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples; training a densely connected network with the training samples to obtain a trained densely connected network; inputting the test samples into the trained densely connected network to obtain a converged densely connected network.

[0007] Optionally, selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples includes: selecting at least three types of underwater acoustic targets as underwater acoustic data samples; classifying the underwater acoustic data samples according to a preset ratio to obtain training samples and test samples.

[0008] Optionally, the method further includes: constructing a densely connected network, where the densely connected network includes: a Steam module, a dense connection module, a Transition Layer module, and a classification module.

[0009] Furthermore, optionally, the method further includes: determining hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, an iteration number, and a batch size; the loss function includes: a cross-entropy loss function.

[0010] Optionally, training a densely connected network with the training samples to obtain a trained densely connected network includes: training the densely connected network by inputting the training samples to obtain labels corresponding to the training samples and the trained densely connected network.

[0011] Furthermore, optionally, inputting the test samples into the trained densely connected network to obtain a converged densely connected network includes: inputting the test samples into the trained densely connected network to identify the data in the test samples to obtain an identification result; determining whether the trained densely connected network converges according to the identification result; and obtaining a converged densely connected network when the determination result is yes.

[0012] According to another aspect of an embodiment of the present invention, a device for constructing an underwater acoustic target recognition model is provided, including: a selection module for selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples; a training module for training a densely connected network with the training samples to obtain a trained densely connected network; and an identification module for inputting the test samples into the trained densely connected network to obtain a converged densely connected network.

[0013] Optionally, the selection module includes: a selection unit for selecting at least three types of underwater acoustic targets as underwater acoustic data samples; and a classification unit for classifying the underwater acoustic data samples according to a preset ratio to obtain training samples and test samples.

[0014] Optionally, the device further includes: a construction module configured to construct a densely connected network, where the densely connected network includes: a Steam module, a densely connected module, a Transition Layer module, and a classification module.

[0015] Further optionally, the device further includes: a parameter determination module configured to determine hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, the number of iterations, and a batch size; the loss function includes: a cross-entropy loss function.

[0016] In an embodiment of the present invention, based on selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples; training a densely connected network with the training samples to obtain a trained densely connected network; inputting the test samples into the trained densely connected network to obtain a converged densely connected network. That is to say, the embodiment of the present invention can solve the problem of insufficient underwater acoustic samples, thereby avoiding complex and heavy feature engineering, and greatly strengthening feature reuse, and alleviating the technical effect of insufficient underwater acoustic samples to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0018] Figure 1 It is a schematic flowchart of a method for underwater acoustic target recognition provided by an embodiment of the present invention;

[0019] Figure 2 It is a schematic diagram of a Dense Block structure in a method for constructing an underwater acoustic target recognition model provided by an embodiment of the present invention;

[0020] Figure 3 It is a schematic diagram of a device for constructing an underwater acoustic target recognition model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the drawings are used to distinguish different objects, rather than to limit a specific order.

[0023] According to one aspect of the embodiments of the present invention, a method for constructing an underwater acoustic target recognition model is provided. Figure 1 It is a schematic flowchart of a method for constructing an underwater acoustic target recognition model provided by the embodiments of the present invention. As Figure 1 shown, the method for constructing an underwater acoustic target recognition model provided by the embodiments of the present application includes:

[0024] Step S102, select underwater acoustic data samples and divide the underwater acoustic data samples into training samples and test samples;

[0025] Optionally, in step S102, selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples includes: selecting at least three types of underwater acoustic targets as underwater acoustic data samples; classifying the underwater acoustic data samples according to a preset ratio to obtain training samples and test samples.

[0026] Specifically, the dataset used in the method for constructing an underwater acoustic target recognition model provided by the embodiments of the present application is a lake trial dataset, which is collected on the Danjiangkou Lake. The lake trial dataset is divided into 4 categories, which are the radiated noises of 4 different types of ships, namely iron-hulled ships, Shuhanghao, Guotaihao, and Xinshijihao. Two 8-element linear arrays are used for all-weather data collection. Take 15 sound samples of each type of target, each sample with a duration of 5s. After frame segmentation of the data, randomly select 80% of them as the training sample set, and the remaining 20% as the test sample set.

[0027] Optionally, the method for constructing an underwater acoustic target recognition model provided by the embodiments of the present application further includes: constructing a dense connection network, where the dense connection network includes: a Steam module, a dense connection module, a Transition Layer module, and a classification module.

[0028] Specifically, Figure 2 It is a schematic diagram of the DenseBlock structure in a method for constructing an underwater acoustic target recognition model provided by the embodiments of the present invention. As Figure 2 shown, build a DenseNet network model. The DenseNet network specifically includes a Steam module, a dense connection module (Dense Block), a Transition Layer module, and a classification module (ClassificationLayer):

[0029] (1) Determine the structure of the Steam module. The Steam module is placed at the very front of the entire network and consists of a convolutional layer and a max-pooling layer. The convolutional layer has a convolutional kernel size of 3*1 and a stride of 2. The size of the pooling in the max-pooling layer is 3*1 and the stride is 2. Through this Steam module, the feature map is reduced to 1 / 4 of the original.

[0030] (2) Determine the structure of the Dense Block module. The Dense Block module is the core part of the network of the present invention. The key point is to adopt dense connections, that is, the feature maps of all previous layers are concatenated in the Channel dimension and used as the input of this layer. Through this dense connection method, this network can greatly enhance feature reuse and complete the underwater acoustic target recognition task in the case of insufficient underwater acoustic samples. The basic unit block adopted in the present invention is composed of a convolutional layer with a convolution kernel size of 1*1 and a convolutional layer with a convolution kernel size of 3*1 connected in series. Then, an unequal number of basic unit blocks are connected in a dense connection manner to form a Dense Block module. The 1st, 2nd, 3rd, and 4th Dense Block modules respectively contain 6, 12, 24, and 16 basic unit blocks;

[0031] (3) Determine the structure of the Transition Layer module. The Transition Layer module consists of a convolutional layer and an average pooling layer. The convolution kernel size of the convolutional layer is 1*1, the stride is 1, and the size of the pooling of the average pooling layer is 2*1, and the stride is 2. Through this Transition Layer module, the size of the feature map is reduced to half of the original, which can accelerate the model training speed;

[0032] (4) Connect the 4 Dense Block modules and 3 Transition Layer modules built above alternately;

[0033] (5) Determine the structure of the Classification Layer module. The Classification Layer module consists of an average pooling layer and a fully connected layer. The size of the pooling of the average pooling layer is 7*1, the stride is 21, and then the softmax function is used to realize underwater acoustic target classification.

[0034] Further, optionally, the method for constructing an underwater acoustic target recognition model provided by the embodiments of the present application further includes: determining the hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, an iteration number, and a batch size; the loss function includes: a cross-entropy loss function.

[0035] Specifically, in the method for constructing an underwater acoustic target recognition model provided by the embodiments of the present application, the cross-entropy loss function is selected as the loss function, the learning rate is set to 0.0001, the iteration number is 100, and the batch size is 128.

[0036] Step S104, train the dense connection network with training samples to obtain the trained dense connection network;

[0037] Optionally, training the dense connection network with the training samples in step S104 to obtain the trained dense connection network includes: training the dense connection network by inputting the training samples to obtain the labels corresponding to the training samples and the trained dense connection network.

[0038] Specifically, use the training sample set divided in step S102 as the input of the DenseNet network, and the corresponding labels as the expected outputs to complete the training of the network model.

[0039] Step S106: Input the test samples into the trained dense connection network to obtain a converged dense connection network.

[0040] Optionally, inputting the test samples into the trained dense connection network in step S106 to obtain a converged dense connection network includes: inputting the test samples into the trained dense connection network to identify the data in the test samples to obtain the identification results; determining whether the trained dense connection network converges according to the identification results; and obtaining a converged dense connection network when the determination result is yes.

[0041] Specifically, input the test sample set divided in step S102 into the trained DenseNet network model to test the robustness of the model. The recognition rate of the final model on the test sample set reached 0.9352, exceeding 0.8523 of the convolutional neural network.

[0042] In summary, combining steps S102 to S106, in practical applications, it is difficult to obtain underwater acoustic data. For confidentiality reasons, few people publicly publish the obtained underwater acoustic data sets on the Internet. Therefore, underwater acoustic target data is relatively scarce. The model proposed in the method for constructing an underwater acoustic target recognition model provided by this application embodiment uses a dense connection structure, which greatly strengthens feature reuse and effectively solves the problem of few underwater acoustic samples. In addition, in traditional statistical models, it is often necessary to manually select what features to extract. Therefore, whether high-quality features can be extracted directly affects the final recognition results. The method for constructing an underwater acoustic target recognition model provided by this application embodiment uses a neural network, getting rid of the heavy feature engineering. The model can actively learn the corresponding features according to the input waveform to complete the underwater acoustic target recognition task.

[0043] In the embodiments of the present invention, based on selecting underwater acoustic data samples, and dividing the underwater acoustic data samples into training samples and test samples; training a dense connection network with the training samples to obtain a trained dense connection network; inputting the test samples into the trained dense connection network to obtain a convergent dense connection network. That is to say, the embodiments of the present invention can solve the problem of insufficient underwater acoustic samples, thereby avoiding complex and heavy feature engineering, and greatly strengthening feature reuse, and to a certain extent alleviating the technical effect of insufficient underwater acoustic samples.

[0044] According to another aspect of the embodiments of the present invention, there is provided an apparatus for constructing an underwater acoustic target recognition model. Figure 3 As shown in the schematic diagram of an apparatus for constructing an underwater acoustic target recognition model provided by the embodiments of the present invention, Figure 3 As shown, the apparatus for constructing an underwater acoustic target recognition model provided by the embodiments of the present application includes: a selection module 32, configured to select underwater acoustic data samples and divide the underwater acoustic data samples into training samples and test samples; a training module 34, configured to train a dense connection network with the training samples to obtain a trained dense connection network; an identification module 36, configured to input the test samples into the trained dense connection network to obtain a convergent dense connection network.

[0045] Optionally, the selection module 32 includes: a selection unit, configured to select at least three types of underwater acoustic targets as underwater acoustic data samples; a classification unit, configured to classify the underwater acoustic data samples according to a preset ratio to obtain training samples and test samples.

[0046] Optionally, the apparatus for constructing an underwater acoustic target recognition model provided by the embodiments of the present application further includes: a construction module, configured to construct a dense connection network, where the dense connection network includes: a Steam module, a dense connection module, a TransitionLayer module, and a classification module.

[0047] Further, optionally, the apparatus for constructing an underwater acoustic target recognition model provided by the embodiments of the present application further includes: a parameter determination module, configured to determine hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, an iteration number, and a batch size; the loss function includes: a cross-entropy loss function.

[0048] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A method for constructing an underwater acoustic target recognition model, characterized in that, Including: Select underwater acoustic data samples and divide the underwater acoustic data samples into training samples and test samples; Train a densely connected network with the training samples to obtain the trained densely connected network; Input the test samples into the trained densely connected network to obtain the converged densely connected network, including: Input the test samples into the trained densely connected network to identify the data in the test samples and obtain an identification result; Determine whether the trained densely connected network converges according to the identification result; When the determination result is yes, obtain the converged densely connected network; The densely connected network includes a Steam module, a Dense Block module, a Transition Layer module, and a classification module, and the determination method is as follows: Determine the structure of the Steam module. The Steam module is placed at the front of the entire network and consists of a convolutional layer and a max pooling layer. The convolutional layer has a convolutional kernel size of 3*1 and a stride of 2. The size of the pooling in the max pooling layer is 3*1 and the stride is 2. Through this Steam module, the feature map is reduced to 1 / 4 of the original; Determine the structure of the Dense Block module. The basic unit block used consists of a convolutional layer with a convolutional kernel size of 1*1 and a convolutional layer with a convolutional kernel size of 3*1 connected in series. Then, an unequal number of basic unit blocks are connected in a densely connected manner to form a Dense Block module. The 1st, 2nd, 3rd, and 4th Dense Block modules contain 6, 12, 24, and 16 basic unit blocks respectively; Determine the structure of the Transition Layer module. The Transition Layer module consists of a convolutional layer and an average pooling layer. The convolutional layer has a convolutional kernel size of 1*1 and a stride of 1. The size of the pooling in the average pooling layer is 2*1 and the stride is 2. Through this Transition Layer module, the size of the feature map is reduced to half of the original, which can accelerate the model training speed; Alternately connect the 4 Dense Block modules and 3 Transition Layer modules built above; Determine the structure of the Classification Layer module. The Classification Layer module consists of an average pooling layer and a fully connected layer. The size of the pooling in the average pooling layer is 7*1 and the stride is 21. Then, the softmax function is used to achieve underwater acoustic target classification.

2. The method according to claim 1, wherein The selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples includes: Select at least three types of underwater acoustic targets as the underwater acoustic data samples; Classify the underwater acoustic data samples according to a preset ratio to obtain the training samples and the test samples.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Construct the densely connected network, where the densely connected network includes a Steam module, a densely connected module, a Transition Layer module, and a classification module.

4. The method according to claim 3, wherein The method further includes: Determine the hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, the number of iterations, and the batch size; the loss function includes: a cross-entropy loss function.

5. The method according to claim 4, characterized in that, The training of the densely connected network with the training samples to obtain the trained densely connected network includes: Training the densely connected network by inputting the training samples to obtain the labels corresponding to the training samples and the trained densely connected network.

6. An apparatus for implementing the method of constructing an underwater acoustic target recognition model according to claim 1, characterized in that, Includes: A selection module for selecting underwater acoustic data samples and dividing the underwater acoustic data samples into training samples and test samples; A training module for training a densely connected network with the training samples to obtain the trained densely connected network; An identification module for inputting the test samples into the trained densely connected network to obtain the converged densely connected network.

7. The device according to claim 6, characterized in that, The selection module includes: A selection unit for selecting at least three types of underwater acoustic targets as the underwater acoustic data samples; A classification unit for classifying the underwater acoustic data samples according to a preset ratio to obtain the training samples and the test samples.

8. The device according to claim 6 or 7, characterized in that, The apparatus further includes: A construction module for constructing the densely connected network, where the densely connected network includes: a Steam module, a densely connected module, a Transition Layer module, and a classification module.

9. The device according to claim 8, characterized in that, The apparatus further includes: A parameter determination module for determining the hyperparameters of the network, where the hyperparameters include a loss function, a learning rate, the number of iterations, and the batch size; the loss function includes: a cross-entropy loss function.

Images