An underwater acoustic target positioning method based on self-supervised learning

Abstract

The application discloses a kind of underwater acoustic target positioning methods based on self-supervised learning, first to the collected underwater acoustic original data is preprocessed, according to whether with position label data set is divided into labeled data and unlabeled data, labeled data is further divided into independent training set and test set, again randomly destroy the sampling covariance matrix of unlabeled data sample, to reconstruct the damaged part as target, train a self-supervised module based on Transformer model, the self-supervised module is a self-encoder structure based on Transformer model, the role is to reconstruct the damaged sampling covariance matrix, then after the training of self-encoder is completed, the parameters of encoder are used as the initialization parameters of the locator model, finally the training set in labeled data set is used to fine-tune the parameters of the locator module, so as to significantly improve the model performance, generalization ability, reduce the role of label samples required for training in related tasks.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross technical field of digital signal processing and oceanography, and particularly relates to a method for underwater acoustic target positioning based on self-supervised learning. BACKGROUND

[0002] Sound waves can travel long distances in water with little loss, and underwater targets can be identified through sound wave signals, sound source positions can be located, and underwater communication can be performed. Sonar systems have many applications in many fields, such as civilian applications, sonar systems can provide convenience for tracking marine life and exploring marine resources.

[0003] Passive sonar positioning uses a sensor array to receive sonar signals emitted by a target sound source, and processes and analyzes the received signals to obtain the position of the target sound source. Traditional sound source positioning methods use mathematical modeling, mostly modeling the physical environment, and finding the most likely target position by searching the target area. This method relies heavily on physical modeling of the real environment and has poor generalization ability, and the search method has low precision and efficiency. With the rapid development of deep learning technology, the use of deep neural networks for underwater acoustic source positioning has achieved good results. Compared with traditional methods, deep learning methods have lower estimation error and stronger generalization ability.

[0004] In underwater sound source positioning, it is often difficult to obtain enough labeled training samples, and insufficient training samples can cause overfitting of the deep learning model, that is, the model has not learned enough generalization feature expressions, so when applied to actual scenarios, it cannot achieve the effect on the training samples. Therefore, how to improve the generalization ability of the deep learning model on small-scale data sets and reduce the impact of overfitting has become a bottleneck for related research. SUMMARY

[0005] The present invention is aimed at the problem that the model in the prior art has not learned sufficiently generalized feature expressions and the effect is poor in actual application. It provides an underwater acoustic target positioning method based on self-supervised learning. First, the collected underwater acoustic raw data is preprocessed, and the data set is divided into labeled data and unlabeled data according to whether it has a position label. The labeled data is further divided into a training set and a test set that are independent of each other. Then, the sampling covariance matrix of the unlabeled data samples is randomly destroyed. With the goal of reconstructing the destroyed part, a self-supervised module based on the Transformer model is trained. The self-supervised module is an autoencoder structure based on the Transformer model. Its function is to reconstruct the destroyed sampling covariance matrix. After the autoencoder training is completed, the encoder parameters are used as the initialization parameters of the locator model. Finally, the training set in the labeled data set is used to fine-tune the parameters of the locator module, thereby significantly improving the model performance and generalization ability in related tasks and reducing the role of labeled samples required for training.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a method for underwater acoustic target localization based on self-supervised learning, comprising the following steps:

[0007] S1: Collect the raw data of underwater acoustic signals and preprocess the raw data. The data set is divided into labeled data and unlabeled data according to whether it has position labels. The labeled data is further divided into independent training sets and test sets; all data samples are represented by a standard normalized sampling covariance matrix;

[0008] S2: Randomly destroy the sampling covariance matrix of the unlabeled data samples, with the goal of reconstructing the destroyed part, and train a self-supervised module based on the Transformer model. The self-supervised module is an autoencoder structure based on the Transformer model, which is used to reconstruct the destroyed sampling covariance matrix. The parameters of the encoder are used as the initialization parameters of the locator model in the subsequent step S3; the goal of the model is to minimize the mean square error between the reconstructed input and the masked part of the original input:

[0009]

[0010] Among them, w is the parameter of the self-supervision module f, w * is the final parameter of the model, x m represents the masked part of the sampling covariance, and N is the number of samples;

[0011] S3: After the autoencoder training is completed, the encoder parameters are used to initialize the locator module parameters. The goal of the locator module is to minimize the error between the true value and the predicted value, that is, to minimize the following two objective functions respectively:

[0012]

[0013]

[0014] Among them, L r ,L d They are distance loss function and depth loss function respectively. is the true distance and depth of the sample, is the model's predicted distance and depth for the sample;

[0015] S4: Fine-tune the parameters of the locator module using the training set in the labeled dataset in step S1.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) The present invention uses a self-supervised learning method to pre-train the model. This pre-training method does not require labeled samples and can learn feature expressions with stronger generalization capabilities. Samples are randomly destroyed during the self-supervised training process. This method makes the training task more complex and allows the model to learn deeper feature expressions.

[0018] (2) Compared with direct training without pre-training, the training of downstream tasks in the present invention can significantly reduce the positioning error of the model and improve positioning accuracy. During the training process, when the same effect as direct training is achieved, only half the training time is spent; and after the model converges, its positioning error is significantly smaller than the error of direct training. In addition, in the case of environmental mismatch (sensor array offset of 0.6°), the error using the pre-trained model is also significantly smaller than the error of direct training.

[0019] (3) When the training set is large enough, using the pre-training module yields only slightly better results than direct training. However, as the training set becomes smaller, the effect of not using the pre-training module decreases dramatically, while the effect of using the pre-training module decreases within an acceptable range. Experiments show that using the pre-training module can save about half of the labeled training samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the overall framework of the method of the present invention;

[0021] Figure 2 A flow chart for training the method of the present invention;

[0022] Figure 3 Schematic diagram of the model structure of the self-supervision module of the present invention;

[0023] Figure 4 Schematic diagram of the model structure of the locator module of the present invention;

[0024] Figure 5 This is a test example of the present invention, showing the changing trend of the test set accuracy as the number of iterations increases under the two training strategies;

[0025] Figure 6 Schematic diagram of the sound source localization results of two training strategies in an ideal environment in the test example of the present invention;

[0026] Figure 7 This is a schematic diagram of the sound source localization results of two training strategies in a mismatched environment (sensor array offset 0.6°) in the test example of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0028] Example 1

[0029] A method for underwater acoustic target localization based on self-supervised learning. Figure 1 is the overall framework diagram of the method of the present invention, wherein the ovals represent data or outputs and the rectangles represent frameworks or modules; Figure 2 This is the training flow chart of this framework, namely: first, the data preprocessing module is used to preprocess the unlabeled data and labeled data; after the data processing is completed, the self-supervision module is trained using the unlabeled data; then, the parameters of the self-supervision module are used to initialize the locator module, and finally the locator module is fine-tuned using the labeled data.

[0030] A method for underwater acoustic target localization based on self-supervised learning specifically includes the following steps:

[0031] S1: Collect the raw data of underwater acoustic signals and preprocess them. The data set is divided into labeled data and unlabeled data according to whether it has position labels. The labeled data is further divided into independent training and test sets.

[0032] The preprocessing operation is divided into three steps: data normalization, calculation of sampling covariance matrix, and standard normalization. All samples described in this embodiment are represented by the standard normalized sampling covariance matrix.

[0033] In order to make the processing independent of the complex source spectrum, the received array sound field data is converted into a normalized sampling covariance matrix. By performing discrete Fourier transform on the input data of the L receivers in the hydrophone array, the sound field data at frequency f can be expressed as p(f) = [p1(f),…,p L (f)] T The sound field is modeled as:

[0034] p(f) = S(f)g(f, r) + ε

[0035] Here ε is the noise, S(f) is the source, and g is the Green's function. To reduce the effect of the magnitude of the sound field and |S(f)|, the composite sound field is normalized by:

[0036]

[0037] Then the sample covariance matrix is calculated:

[0038]

[0039] where H represents the conjugate transpose operation, and C(f) = C H (f), N s represents the number of snapshots formed. After the sample covariance is calculated, all samples are normalized by the standardization.

[0040] The preprocessing of the label uses 0-1 normalization, and the depth and position of the target position are normalized to 0-1 respectively:

[0041]

[0042]

[0043] The preprocessed data is a 2x20x20 matrix. The preprocessed data is divided into three independent data sets, 80% of which are unlabeled samples, and the rest are labeled samples divided into a training set and a test set. All unlabeled samples are used to train the self-supervised module, and the labeled training set is used to train the locator module. The performance of the final model is evaluated on the locator module using the labeled test set. In this embodiment, the unlabeled data set contains 400,000 unlabeled samples; the number of labeled data is 100,000, which is divided into independent training and test sets.

[0044] S2: Randomly destroy the sample covariance matrix of the unlabeled data sample, and train a self-supervised module based on the Transformer model to reconstruct the destroyed part as the target.

[0045] The pre-trained self-supervised module is a Transformer-based autoencoder structure, and the input is the destroyed sample covariance matrix. The function of the autoencoder is to reconstruct the destroyed sample covariance matrix. Through this "destroy-reconstruct" method, the self-supervised pre-training model can learn features with stronger generalization ability. The basic structure of the self-supervised module is as follows: Figure 3As shown in Figure 1, it mainly consists of four parts: patch embedding, position embedding, Transformer encoder and Transformer decoder. Table 1 below shows the structural information of each part of the Transformer-based autoencoder.

[0046] Table 1 Transformer-based autoencoder structure

[0047]

[0048] Patch embedding splits the original input into sub-blocks of equal size and then maps each sub-block to a new feature space. In this embodiment, two-dimensional convolution is used to implement this operation. The size of the convolution kernel and the padding are both set to the size of the sub-block, which is set to 4 in this embodiment. Position embedding adds position information to the feature vector of each sub-block. In this method, sincos position encoding is used. The Transformer encoder and Transformer decoder have the same structure, both consisting of a multi-head attention mechanism and MLP, as shown in Tables 2 and 3:

[0049] Table 2 Transformer encoder structure

[0050] TransformerEncoder Layer Type Number of layers Number of neurons LayerNorm Norm / / MutiHeadAttention Fc 1 2304 ResidualAdd Add 2 / LayerNorm Norm / / MLP Fc 2 3072、768 ResidualAdd Fc 2 /

[0051] Table 3 Transformer decoder structure

[0052] TransformerDecoder Layer Type Number of layers Number of neurons LayerNorm Norm / / MutiHeadAttention Fc 1 1536 ResidualAdd Add 2 / LayerNorm Norm / / MLP Fc 2 2048、512 ResidualAdd Fc 2 /

[0053] First, the sample is divided into mutually exclusive sub-blocks of equal size. A certain proportion of the sub-blocks are randomly selected and masked. The self-supervisory module uses the visible sub-blocks as input and reconstructs the original sample. After feature extraction by the Transformer encoder, the features of the visible sub-blocks are obtained. Due to the characteristics of the Transformer encoder / decoder, the first dimension of its input and the first dimension of its output are usually the same size. The first dimension of the input of the autoencoder model is the number of visible sub-blocks, and the first dimension of its output is the number of all sub-blocks. Therefore, to reconstruct the entire original sample, the Transformer decoder cannot only use the features of the visible sub-blocks as input; some features must also be added to represent the invisible sub-blocks. These invisible sub-block features are randomly generated, shared across all invisible sub-blocks, and not updated. After adding the features of the invisible sub-blocks, the features are decoded into the original sample through the Transformer decoder.

[0054] In this example, the width of the encoder is 768, while the width of the decoder is 512. In addition, the attention in both the encoder and decoder is eight-head attention. The forward propagation of the input data in this module is as follows:

[0055] 1. The size of the initial sample is 2×20×20, which is divided into 25 equal-sized sub-blocks of 2×4×4;

[0056] 2. Each 2×4×4 sub-block is represented as a 768-dimensional feature vector after the embedding layer. The entire input is then mapped into a 25×768-dimensional feature matrix, where each row represents the feature vector of a sub-block.

[0057] 3. Add fixed-weight sincos position encoding to the 25×768-dimensional feature matrix to ensure that the original position information of each eigenvector is contained in it;

[0058] 4. Randomly select a portion of the eigenvectors and mask them. In the present invention, 40% of the eigenvectors are masked. At this time, it can be seen that the eigenmatrix becomes a matrix of size 15×768;

[0059] 5. It can be seen that the size of the feature matrix obtained after the feature matrix passes through the 12-layer Transformer encoder remains unchanged, which is still 15×768;

[0060] 6. Generate a random and shared feature vector for each discarded sub-block. In this invention, 10 sub-blocks are masked, so a total of 10 weight-shared feature vectors are generated. These weight-shared intermediate codes and the feature vectors of the visible sub-blocks generated by the encoder form a new feature vector of dimension 25×768;

[0061] 7. After the embedding operation, the eigenvector is reduced to a 25×512 two-dimensional matrix. Similarly, each row of the matrix is ​​the eigenvector of a sub-block;

[0062] 8. Add sincos position encoding to all feature vectors again to ensure that the feature vectors of invisible and visible sub-blocks contain position information;

[0063] 9. After being decoded by the 8-layer Transformer decoder, the feature matrix retains its dimension. After passing through the fully connected layer, it is mapped into a 25×2×4×4 matrix and reconstructed back to the original sampling covariance matrix of 2×20×20 after rearrangement.

[0064] During training, the goal of the model is to minimize the mean square error between the reconstructed input and the masked portion of the original input:

[0065]

[0066] Among them, w is the parameter of the self-supervision module f, w * is the final parameter of the model, x m represents the masked part of the sampling covariance, and N is the number of samples.

[0067] S3: After the autoencoder training is completed, the encoder parameters are used to initialize the locator module parameters.

[0068] The model structure of the locator module is as follows Figure 4 It consists of two main parts. The first part is a Transformer encoder, which has the same structure as the Transformer encoder in the self-supervised module. After the self-supervised module is trained, the parameters of the self-supervised module are used for initialization.

[0069] The second part consists of several fully connected layers that map the features extracted by the encoder into distance and depth. In this embodiment, depth and distance are mapped to a value between 0 and 1. The goal of the locator is to minimize the error between the true value and the predicted value, that is, to minimize the following two objective functions:

[0070]

[0071]

[0072] Using the joint training method, the final loss function is as follows:

[0073] L=(1-α)L r +αL d

[0074] α is used to balance the weights of distance and depth. In this embodiment, α is set to 0.5.

[0075] S4: Use the training set in the labeled dataset to fine-tune the parameters of the locator module, significantly reducing the positioning error of the model and improving the positioning accuracy.

[0076] Test Case

[0077] During the experiment, the performance of the locator module was compared with that with the pre-training module under different training set sizes. First, the model training steps using the self-supervised module are as follows:

[0078] 1. Use 400,000 unlabeled samples to train the self-supervised module;

[0079] 2. Use the encoder parameters of the self-supervisory module to initialize the encoder parameters of the locator module;

[0080] 3. Train the locator using the labeled training set.

[0081] Model training without the self-supervision module directly uses the labeled training set to train the locator. Figure 5 The following figure shows how the accuracy of the model changes with the number of training times when the training set accounts for 5% of the total data volume (25,000 labeled samples) when the locator is trained in two different ways. It can be clearly seen that when the number of iterations is the same, the accuracy of using the pre-training module is significantly higher than that of not using the pre-training module. When the accuracy of using the self-supervised module reaches the highest accuracy of not using the self-supervised module, only about 50% of the training time is used. At the same time, after the model converges, the accuracy without using the pre-training module is only 61.63%, while the accuracy with the pre-training module can reach 89.14%.

[0082] Tables 5 and 6 below show the errors in depth and distance prediction by the two models in an ideal environment and a mismatched environment (sensor array offset 0.6°).

[0083] Table 5 Errors under ideal conditions when using 5% (25,000 labeled samples) of data for training

[0084]

[0085] Table 6 Errors in mismatched environments when using 5% (25,000 labeled samples) of data for training

[0086]

[0087] As can be seen from the table, the various errors using the pre-training module are significantly smaller than the errors without the pre-training module. Figure 6 、 Figure 7 The model's localization results for an ideal environment and a mismatched environment (with a sensor array offset of 0.6°) are shown. As can be seen, without the pre-trained module, significant errors were observed in the sound source estimation at a distance of approximately 5,500 meters from the sensor array. This error was eliminated when the pre-trained module was used. When the environment was mismatched, the errors for both methods increased significantly, but the localization error using the pre-trained module was significantly lower, resulting in significantly fewer mispredictions.

[0088] We also conducted experiments on the two models using training sets of varying sizes. These experiments show that, with any amount of training data, the accuracy achieved using the pre-trained module is superior to that achieved without it. Although localization accuracy decreases with the reduction of training data, even with the pre-trained module, the decrease is significantly less than without it. Using the pre-trained module with 5% and 2.5% training sets, respectively, yielded results close to those achieved without it (10% and 5%). This means that using the pre-trained module can save approximately 50% of labeled training data.

[0089] The present invention introduces self-supervised learning into underwater sound source positioning. In self-supervised learning, a model is usually pre-trained using samples without labels, and the model parameters are used as the initialization model parameters for specific downstream tasks. When training downstream tasks, labeled data is used to perform fine-tuning on the basis of the parameters. Because the pre-trained model is trained using unlabeled samples and does not require any annotation information, the cost of obtaining labeled training data on a large scale is greatly reduced. At the same time, it learns the generalization features of a large number of samples during the training process. When fine-tuning specific tasks, it can effectively alleviate the overfitting problem of the model, improve the generalization ability of the model, and reduce the number of labeled samples required during the training process.

[0090] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for underwater acoustic target localization based on self-supervised learning, characterized in that: The steps include: S1: Collect the raw data of underwater acoustic signals and preprocess the raw data. The data set is divided into labeled data and unlabeled data according to whether it has position labels. The labeled data is further divided into independent training sets and test sets; all data samples are represented by a standard normalized sampling covariance matrix; S2: Randomly destroy the sampling covariance matrix of the unlabeled data samples, with the goal of reconstructing the destroyed part, and train a self-supervised module based on the Transformer model. The self-supervised module is an autoencoder structure based on the Transformer model, which is used to reconstruct the destroyed sampling covariance matrix. The parameters of the encoder are used as the initialization parameters of the locator model in the subsequent step S3; the goal of the model is to minimize the mean square error between the reconstructed input and the masked part of the original input: Among them, w is the parameter of the self-supervision module f; w * is the final parameter of the model; x m Represents the masked part of the sampling covariance; N is the number of samples; S3: After the autoencoder training is completed, the encoder parameters are used to initialize the locator module parameters. The goal of the locator module is to minimize the error between the true value and the predicted value, that is, the objective function of the minimum distance loss function and the depth loss function respectively: Among them, L r ,L d They are distance loss function and depth loss function respectively. is the true distance and depth of the sample; is the model's predicted distance and depth for the sample; S4: Fine-tune the parameters of the locator module using the training set in the labeled dataset in step S1.

2. The underwater acoustic target localization method based on self-supervised learning according to claim 1, characterized in that: The preprocessing steps of step S1 include: transforming the time domain signal into the frequency domain through Fourier transform, calculating the sampling covariance matrix of the received signal in each frequency band, and normalizing the calculated sampling covariance matrix.

3. The underwater acoustic target positioning method based on self-supervised learning according to claim 2, characterized in that: In step 1, 80% are unlabeled samples. All unlabeled samples are used to train the self-supervisory module, and the labeled training set is used to train the locator module. The performance of the final model is tested on the locator module using the labeled test set.

4. The underwater acoustic target localization method based on self-supervised learning according to claim 2 or 3, characterized in that: The self-supervision module of the Transformer model in step S2 includes: The first submodule is Patch Embedding, which consists of convolutional layers that split the original input into sub-blocks of equal size, and then maps each sub-block into a new feature space, stretching the input block into a vector of the same dimension. The second submodule is Position embedding, which has fixed parameters and is used to add position information to the feature; The third submodule Transformer encoder and the fourth submodule Transformer decoder have the same structure, consisting of a multi-head attention mechanism and MLP. The encoder maps the original input to the feature space, and the decoder decodes the features back to the original input.

5. The underwater acoustic target positioning method based on self-supervised learning according to claim 4, characterized in that: The working principle of each sub-module in the self-supervision module in step S2 is as follows: divide the data sample into sub-blocks of the same size and mutually exclusive, randomly select a certain proportion of sub-blocks and mask them, the self-supervision module takes the visible part of the sample as input, and obtains the features of the visible sub-blocks after feature extraction by the third sub-module Transformer encoder; according to the characteristics of the third sub-module Transformer encoder or the fourth sub-module Transformer decoder, the first dimension of its input and the first dimension of its output are the same size, so the first dimension of the input of the self-supervision module is the number of visible sub-blocks, and the first dimension of the output is the number of all sub-blocks. The fourth sub-module Transformer decoder needs to add some features to represent the invisible sub-blocks. After decoding by the fourth module Transformer decoder, the features are decoded into the original samples.

6. The underwater acoustic target positioning method based on self-supervised learning according to claim 5, characterized in that: The model structure of the locator module in step S3 includes at least a Transformer encoder submodule and a fully connected layer. The Transformer encoder submodule has the same structure as the third submodule Transformer encoder in the self-supervisory module; the fully connected layer maps the features extracted by the Transformer encoder submodule into distance and depth.

7. The underwater acoustic target positioning method based on self-supervised learning according to claim 6, characterized in that: This method adopts a joint training method, and the final loss function is as follows: L=(1-α)L r +αL d Among them, α is a weight parameter that balances distance and depth.

Images