Underwater sound target identification method based on comparative learning semi-supervised neural network

Through a semi-supervised neural network based on comparison learning, combined with supervised and unsupervised networks, the classification characteristics of water acoustic signals are optimized, and the problems of insufficient samples and imbalance in water target recognition are solved, and efficient water acoustic target recognition is achieved.

CN118964983APending Publication Date: 2024-11-15XIAN CHANGFENG ELECTROMECHANICAL RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411011230.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing in-water target recognition methods mainly rely on supervised learning, requiring a large amount of labeled data, and there is a problem of sample imbalance in small sample data, which makes the labeling process time-consuming and labor-intensive, and the number of unlabeled samples is limited, hindering the development of deep learning.

Method used

Using a semi-supervised neural network based on contrast learning, by constructing a contrast learning database and training data set, combining a supervised network and an unsupervised contrast learning network, the contrast loss function and cross entropy function are used to optimize, extract the classification characteristics of the water acoustic signals, and improve the Identify performance.

Benefits of technology

In the case where the number of labeled samples and unlabeled samples is insufficient and the class is uneven, the effective identification of the water acoustic target category is improved, and the identification performance of the network is overcome, and the problems of insufficient samples and uneven categories are overcome.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964983A_ABST
    Figure CN118964983A_ABST
Patent Text Reader

Abstract

The invention provides an underwater acoustic target recognition method based on a contrast learning semi-supervised neural network. Comprising the following steps: processing an underwater sound sample, respectively constructing a contrast learning database, a training data set and a test data set, then constructing a semi-supervised neural network based on contrast learning, carrying out network training, and finally inputting the test data set into the trained network to obtain a classification result. The semi-supervised neural network based on comparative learning constructed by the invention can effectively identify the type of the underwater acoustic target under the condition that the number of labeled samples and the number of unlabeled samples are insufficient.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of underwater acoustic signal target recognition, and in particular relates to an underwater acoustic target recognition method based on contrastive learning semi-supervised neural network. Background Art

[0002] Early deep learning methods for underwater target recognition mainly focused on supervised learning, which required a large number of accurately labeled underwater targets for training. If real-time and accurate labeling of target categories can be achieved in any sea area, the supervised network can meet the accuracy requirements of underwater acoustic recognition. However, the time-consuming and labor-intensive labeling process hinders the applicability of supervised learning in underwater target recognition. Many deep learning methods in underwater target recognition use supervised learning methods. However, supervised networks require a large amount of data as a driving factor, which is a challenge for small sample underwater target recognition. Compared with supervised networks, semi-supervised networks require much less driving data, so they are widely used in various fields. The difficulty in collecting underwater targets is the main reason for the small samples of underwater acoustics, so the limited number of unlabeled samples is also an important factor restricting the development of target recognition methods.

[0003] Further research shows that there may be sample imbalance in both labeled and unlabeled samples in small sample data. In the field of language recognition, self-supervised contrastive learning methods are used to learn in limited samples and achieve better optimization recognition results. The focus of contrastive learning is to learn common features between similar samples, distinguish differences between non-similar samples, and learn to distinguish data at the abstract level of feature space, rather than focusing on a single task feature. Contrastive learning increases the edge of the decision boundary by forcing positive sample pairs to be close to each other in the latent space and negative sample pairs to be far away from each other. Due to the regularization operation of contrastive learning, the feature representation space should make similar samples as close as possible, while dissimilar samples are evenly distributed. Therefore, the feature representation space of contrastive learning is linearly separable and can be used as a linear classifier sufficient for classification. However, the construction of positive samples in contrastive learning is by applying perturbations, keeping the original data intact, and tending to produce similar features, which will result in a large number of samples containing similar background noise in the samples of underwater acoustic targets, which is not conducive to the back-propagation optimization method of deep learning. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a method for underwater acoustic target recognition based on contrastive learning semi-supervised neural network. The method includes processing underwater acoustic samples and constructing a contrastive learning database, a training data set and a test data set respectively, then constructing a semi-supervised neural network based on contrastive learning, and performing network training, and finally inputting the test data set into the trained network to obtain a classification result. The semi-supervised neural network based on contrastive learning constructed by the present invention can effectively identify the category of underwater acoustic targets when the number of labeled samples and unlabeled samples is insufficient.

[0005] A method for underwater acoustic target recognition based on contrastive learning semi-supervised neural network, characterized by the following steps:

[0006] Step 1: The original underwater acoustic signal is downsampled and windowed, and then normalized. For the unlabeled samples after processing, samples of different working conditions belonging to the same type of target are positive sample pairs, and samples of different types of targets are negative samples to form a comparative learning database; for the labeled samples after processing, 30% of the samples are divided into the training data set, and the rest of the samples are divided into the test data set;

[0007] Step 2: construct a semi-supervised neural network based on contrastive learning, and train it using the sample data obtained in step 1; the semi-supervised neural network based on contrastive learning is composed of a supervised network and an unsupervised contrastive learning network, the labeled samples in the training data set are input into the supervised network, and the output of the supervised network is the classification result of the sample; the positive and negative sample pairs in the contrastive learning database are input into the unsupervised contrastive learning network, and the classification effect of the supervised network is optimized through contrastive learning;

[0008] Step 3: Input the labeled samples in the test data set into the trained semi-supervised neural network, and the output of the supervised network is the classification result of the corresponding sample.

[0009] Furthermore, the supervised network in the semi-supervised neural network based on contrastive learning adopts the SE_ResNet model, and the unsupervised contrastive learning network is composed of 8 SE_ResNet models connected with 8 DNN models respectively.

[0010] Furthermore, the contrast loss function is used as the loss function of the unsupervised contrastive learning network, and the expression is as follows:

[0011]

[0012] Among them, Loss_con represents the contrast loss value of the network, N represents the number of samples, Y represents the label corresponding to the sample pair in the input data set, when Y = 1, it means that the two samples in the sample pair are of the same type, when Y = 0, it means that the two samples in the sample pair are of different types, d represents the Euclidean distance between the two input samples, and m is a setting parameter, set m = 0.5;

[0013] The cross entropy function is used as the loss function of the SE_ResNet model, and the expression is as follows:

[0014]

[0015] Among them, Loss_cross represents the cross entropy loss value of the network, N′ represents the number of samples in the current batch, Z represents the total number of categories of the identified target, y represents the true label, and y′ represents the predicted label;

[0016] When training the semi-supervised neural network based on contrastive learning, first, the SE_ResNet model and the DNN model in the unsupervised contrastive learning network are optimized using the function Loss_con, and the optimization rule is the Euclidean distance between positive and negative samples. Then, the function Loss_cross is used to guide the SE_ResNet model to be optimized again, and the optimization rule is the marking rule when marking samples, and its optimization result is reversely guided to the Loss_con function.

[0017] The beneficial effects of the present invention are: in view of the fact that both the labeled samples and the unlabeled samples of the underwater acoustic signal are insufficient, and the insufficient samples have the problem of imbalance between classes, the present invention constructs a semi-supervised neural network based on contrastive learning, and combines the contrastive unsupervised method with the supervised network to effectively solve the problem of insufficient samples and imbalance between sample categories, improve the recognition performance of the network, and can effectively identify the category of underwater acoustic targets when the number of labeled samples and the unlabeled samples are insufficient and there is imbalance between classes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of a semi-supervised neural network based on contrastive learning of the present invention;

[0019] Figure 2 It is a schematic diagram of the unsupervised contrastive learning network of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.

[0021] In view of the recognition problem when the number of labeled and unlabeled samples is insufficient, the present invention provides an underwater acoustic target recognition method based on a contrastive learning semi-supervised neural network. The unlabeled samples carry the acoustic characteristics of underwater targets. By extracting the geometric distance between different signals and comparing the size of the distance, the same type of targets and different types of targets are adaptively distinguished. The labeling process of the labeled samples is based on a specific task as a standard, so the learning process of the supervised network will lose some potential effective features in the sample information, and it is impossible to explore the relationship pattern within the entire data set from a global perspective. Unlabeled samples can make up for this shortcoming of supervised learning. The present invention constructs a semi-supervised neural network based on contrastive learning, which can extract both the classification features of labeled samples and the scalability feature information of unlabeled samples. The contrastive learning method between unlabeled samples is used to extract the classification feature information of the signal about the Euclidean distance, and then the network is optimized using labeled samples. This method can learn the deterministic classification features of the labeled samples and the abstract classification features of the unlabeled samples. The contrastive learning model learns the category representation of the target by maximizing the contrast loss of different category information in the latent space and minimizing the data with the same consistency.

[0022] The present invention provides an underwater acoustic target recognition method based on contrastive learning semi-supervised neural network, and its specific implementation process is as follows:

[0023] Step 1: First, the original underwater acoustic signal is downsampled and processed with a Hamming window of 600, and then normalized. For the processed unlabeled samples, samples of different working conditions belonging to the same class of targets constitute positive sample pairs, and samples of different classes of targets are used as their negative samples to constitute negative sample pairs, forming a comparative learning database. For the processed labeled samples, the samples are divided into training data sets (30%) and test data sets (70%), where the training data and test data are not repeated.

[0024] Step 2: First, a semi-supervised neural network based on contrastive learning is constructed.

[0025] like Figure 1 As shown, the semi-supervised neural network based on contrastive learning constructed by the present invention includes two parts, one part is a supervised network, which uses the SE_ResNet model, inputs the labeled samples into the SE_ResNet model, and outputs the reduced-dimensional data representing the classification results. By comparing these data with the labels, the classification effect of the SE_ResNet model can be optimized. When training the network to optimize the network, this part adopts supervised learning, that is, the cross entropy function between the network output and the label corresponding to the labeled sample is used to update the network. The cross entropy can enable the network to learn the recognition parameters directly from the sample label. The mathematical expression is:

[0026]

[0027] Among them, Loss_cross represents the cross entropy loss value of the network, N′ represents the number of samples in the current batch, Z represents the total number of categories of the identified targets, y represents the true label, and y′ represents the predicted label.

[0028] In underwater target recognition, unlabeled samples contain a large amount of effective recognition information, which can continuously supplement the learning features of the supervised model. In order to better realize the classification of similar samples and different samples, the present invention adopts a semi-supervised model based on contrastive learning, and uses contrastive learning to optimize the network model without labeled samples. The task of the network training is to minimize the difference between the label and the predicted value, and at the same time, minimize the difference between the same category and maximize the difference between different categories. Therefore, another part of the semi-supervised neural network based on contrastive learning constructed by the present invention is an unsupervised contrastive learning network composed of a SE_ResNet model and a DNN model. A positive sample pair is input, and the unsupervised contrastive learning network outputs the predicted probability of the sample. The optimization process is to reduce the loss function. A negative sample pair is input, and the unsupervised network outputs the predicted probability of the sample. The optimization process is to increase the loss function. Among them, DNN is the backbone architecture of projection mapping, which represents the space mapped to the contrast loss. Using the DNN network to reduce the dimensionality of the features can provide scalable features without clear targets. Such as Figure 2 As shown, after the output features of the DNN network, the positive sample pair ( and k represents the number of target categories, i and j represent different sample representations in the positive sample pair) have similar Euclidean distances, and the negative sample pair (x k and x c (c≠k)) have different distance boundaries. The contrast prediction task can cluster similar samples in the same boundary area by minimizing the distance between positive sample pairs and maximizing the distance between negative sample pairs, while samples of different categories have other boundary areas. The contrast loss function is defined as follows:

[0029]

[0030] Among them, Loss_con represents the contrast loss value of the network, N represents the number of samples, Y represents the label corresponding to the sample pair in the input data set, when Y = 1, it means that the two samples in the sample pair are of the same type, when Y = 0, it means that the two samples in the sample pair are of different types, d represents the Euclidean distance between the two input samples, and m is a setting parameter, which is set to 0.5.

[0031] The labeled samples follow specific standards during the labeling process, which makes it impossible to fully define the acoustic feature categories of the samples, and some potential classification features will be lost. In contrast, unlabeled samples do not have specific labeling standards as constraints, so they are highly scalable for the learning characteristics of the network. The features learned from unlabeled samples are global features, including features of various scales, while the features learned from labeled samples are features at specific scales. The inconsistency between the features learned from the two will affect the recognition results. The semi-supervised neural network based on contrastive learning constructed by the present invention does not process labeled samples or unlabeled samples separately, but establishes a unified strategy between unlabeled samples and labeled samples. In each contrastive learning, the labeled samples are first used to guide the update of the feature extraction network in the contrastive network, that is, the SE_ResNet model and the DNN model in the unsupervised contrastive learning network are optimized using the function Loss_con, and the optimization rule is the Euclidean distance between positive and negative samples. Then, the updated model of contrastive learning is used to optimize the classification features of the labeled samples, that is, the function Loss_cross is used to guide the SE_ResNet model to be optimized again, and the optimization rule is the labeling rule when labeling samples, and its optimization result is reversely guided to the Loss_con function.

[0032] Step 3: Input the test data in the labeled samples into the trained semi-supervised neural network, and output the classification results after passing through the supervised network SE_ResNet model.

[0033] In order to verify the effectiveness of the method of the present invention, experiments were conducted on a measured hydroacoustic sample data set. In this data, the unlabeled samples contain a total of four types of target ships, and each type of target contains two different driving states, namely straight and turning. First, the data is downsampled using a sampling frequency of 4800Hz, and then a Hamming window with a window length of 1200 is performed, and the samples are normalized. Then, for unlabeled samples, two different motion states in the same category are taken as a pair of positive sample pairs, and different categories are taken as negative sample pairs. For labeled samples, training sets and test sets are divided. 30% of the labeled samples are training sets, and 70% are test sets. The training set and the test set are not repeated. Train the network model, and stop training when the recognition rate error obtained is less than 0.005.

[0034] In order to make the one-dimensional underwater acoustic data suitable for the input characteristics of the SE_ResNet model with multi-dimensional input, the data is preprocessed, and the convolution operation is selected with a convolution kernel size of 1×64 and a step size of 1. The number of input data channels is changed from 1 to 16, and the convolution output is nonlinearly mapped using the ReLU activation function. Then, in order to obtain the feature data of all convolution channels, the maximum pooling method is used with a pooling size of 1×2 and a step size of 2. In the same number of channel sizes, pooled data of half the size are obtained, and the obtained data is used as the original data of the SE_ResNet model.

[0035] Table 1 shows the recognition rates of different network models under different sample numbers. It can be seen that the DAE model, RNN model and DBN model are three relatively basic network models with simple structures and different basic theories of the network, but they are all driven by large-scale training. When the number of samples is insufficient, the recognition effect is not good; when there are only 600 training samples, the recognition effect of the U_Net model and the GAN model is better than that of several basic networks, but due to the complex background noise and the high similarity between target categories, the network recognition performance is affected; the SE_ResNet model and the camResNet model have significantly improved the recognition effect due to the addition of the attention mechanism, and are currently more effective supervised recognition methods; the semi-supervised network Ladder network model and the SimCLR model are used as comparison models of the network of the present invention. When the training samples include 200 labeled samples and 400 labeled samples, the recognition effect of the Ladder network is lower than that of the supervised network model, and the semi-supervised neural network based on contrastive learning of the present invention also has a better recognition effect than all the comparison networks when the labeled samples are 200 and the unlabeled samples are 400.

[0036] Table 1

[0037] method Correct recognition rate [%] Number of labeled samples The number of unlabeled samples Total sample size DAE 71.25 600 0 600 RNN 82.56 600 0 600 DBN 72.12 600 0 600 U_Net 81.33 600 0 600 Densnet 81.62 600 0 600 GAN 84.49 600 0 600 ABNN 82.43 600 0 600 SE_ResNet 86.78 600 0 600 camResNet 88.39 600 0 600 Ladder 79.12 200 400 600 SimCLR 91.26 200 400 600 CBSN (the present invention) 98.06 200 400 600

Claims

1. A method for underwater acoustic target recognition based on contrastive learning semi-supervised neural network, characterized in that Here are the steps: Step 1: The original underwater acoustic signal is downsampled and windowed, and then normalized. For the unlabeled samples after processing, samples of different working conditions belonging to the same type of target are positive sample pairs, and samples of different types of targets are negative samples to form a comparative learning database; for the labeled samples after processing, 30% of the samples are divided into the training data set, and the rest of the samples are divided into the test data set; Step 2: construct a semi-supervised neural network based on contrastive learning, and train it using the sample data obtained in step 1; the semi-supervised neural network based on contrastive learning is composed of a supervised network and an unsupervised contrastive learning network, the labeled samples in the training data set are input into the supervised network, and the output of the supervised network is the classification result of the sample; the positive and negative sample pairs in the contrastive learning database are input into the unsupervised contrastive learning network, and the classification effect of the supervised network is optimized through contrastive learning; Step 3: Input the labeled samples in the test data set into the trained semi-supervised neural network, and the output of the supervised network is the classification result of the corresponding sample.

2. The underwater acoustic target recognition method based on contrastive learning semi-supervised neural network according to claim 1, characterized in that: The supervised network in the semi-supervised neural network based on contrastive learning adopts the SE_ResNet model, and the unsupervised contrastive learning network is composed of 8 SE_ResNet models connected with 8 DNN models respectively.

3. The underwater acoustic target recognition method based on contrastive learning semi-supervised neural network as claimed in claim 2, characterized in that: The contrast loss function is used as the loss function of the unsupervised contrast learning network, and the expression is as follows: Among them, Loss_con represents the contrast loss value of the network, N represents the number of samples, Y represents the label corresponding to the sample pair in the input data set, when Y = 1, it means that the two samples in the sample pair are of the same type, when Y = 0, it means that the two samples in the sample pair are of different types, d represents the Euclidean distance between the two input samples, and m is a setting parameter, set m = 0.5; The cross entropy function is used as the loss function of the SE_ResNet model, and the expression is as follows: Among them, Loss_cross represents the cross entropy loss value of the network, N′ represents the number of samples in the current batch, Z represents the total number of categories of the identified target, y represents the true label, and y′ represents the predicted label; When training the semi-supervised neural network based on contrastive learning, first, the SE_ResNet model and the DNN model in the unsupervised contrastive learning network are optimized using the function Loss_con, and the optimization rule is the Euclidean distance between positive and negative samples. Then, the function Loss_cross is used to guide the SE_ResNet model to be optimized again, and the optimization rule is the marking rule when marking samples, and its optimization result is reversely guided to the Loss_con function.