A method for underwater acoustic target recognition based on vector linear arrays based on deep learning

Abstract

The present invention relates to a method for underwater acoustic target recognition based on vector linear arrays based on deep learning. The recognition method uses data received by vector linear arrays to establish a database. The vector linear arrays can not only receive sound pressure information, but also obtain omnidirectional vibration velocity information at the same position. The method includes the following steps: establishing a database; data preprocessing; and deep learning model construction. The recognition method described in the present invention uses a vector array signal feature vector with a strong anti-interference ability, which is more conducive to the feature vector extraction of low-noise underwater acoustic targets. For underwater acoustic target recognition, its recognition success rate is higher than that of sound pressure signals under the same conditions. At the same time, in the database establishment stage, an improved recursive neural network is used to predict the ocean background noise, eliminate part of the interference noise in the target channel, improve the data sample quality, and facilitate the accurate and rapid learning and establishment of the model.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic equipment design and manufacturing, and in particular to a vector linear array underwater acoustic target recognition method based on deep learning. Background Art

[0002] Hydroacoustic target recognition has always been the focus of hydroacoustic target research. Hydroacoustic target feature vector is the main means of hydroacoustic target recognition. The hydroacoustic target feature vector is a kind of information that can represent the target attributes and status obtained by signal processing of the original hydroacoustic data received by the acoustic array. Hydroacoustic targets mainly include characteristic information such as noise, motion, wake, and geometric structure.

[0003] Traditional underwater acoustic target recognition requires target feature extraction. This process requires the extraction of time-frequency spectrum features or other types of features through manual analysis. The target recognition process has a strong human-computer interaction characteristic, which requires researchers to have strong data analysis capabilities, prior knowledge and professional qualities. The signal-to-noise ratio of weak targets is very low, and the difficulty of feature extraction increases accordingly, making target recognition more difficult. Summary of the invention

[0004] To this end, the technical problem to be solved by the present invention is to overcome the problem in the prior art that underwater acoustic target recognition relies on the participation of human factors and is difficult to identify, thereby providing a vector linear array underwater acoustic target recognition method based on deep learning.

[0005] In order to solve the above technical problems, the present invention provides a vector linear array underwater acoustic target recognition method based on deep learning. The recognition method uses the data received by the vector linear array to establish a database. The vector linear array can not only receive sound pressure information, but also obtain omnidirectional vibration velocity information at the same position. The method includes the following steps:

[0006] Step S1: Use the data received by the vector linear array to establish a database. The vector sensor can simultaneously measure the sound pressure p and two mutually orthogonal vibration velocity components ν in the sound field at the same position x , ν y , under the plane wave condition:

[0007]

[0008] In the above formula, x(t) is the sound pressure waveform, θ s is the horizontal azimuth angle of sound wave propagation;

[0009] Step S2: Using the weighted linear combination of the two velocity components of the sound field, the beam can be rotated in two-dimensional space. Let:

[0010] ν c (t) = ν x(t)cosψ+ν y (t)sinψ

[0011] ν s (t)=-ν x (t)sinψ+ν y (t)cosψ

[0012] In the above formula, Ψ is the guiding direction;

[0013] Step S3: The data preprocessing model uses 10 LSTM multicells connected in series. The output of each multicell is a linear regression, and the cost function is selected as the mean square error loss function;

[0014] Step S4: Select the logistic regression algorithm to build a learning model. The logistic regression model represents:

[0015] h θ (x) = g(θ T X)

[0016] Where: X represents the feature vector;

[0017] g represents the logic function,

[0018] but

[0019] h θ The role of (x) is to calculate the probability of the output variable being equal to 1 for a given input variable based on the selected parameters, thereby classifying the input variable;

[0020] Step S5: Fitting the logistic regression model, that is, fitting the parameters θ in the logistic regression model, requires defining the optimization objective or cost function used to fit the parameters:

[0021]

[0022] in, Substitute it into the derivative and simplify it to find the min θ When J(θ)

[0023] In one embodiment of the present invention, in step S2, c The directivity maximum of (t) is at θ s =ψ, and ν s The directivity zero point of (t) is also at θ s =ψ. As long as the value of ψ is changed, the directivity of the vibration velocity can be electronically rotated to form a zero point in any direction, thereby suppressing interference and being more conducive to the extraction of feature vectors of low-noise underwater acoustic targets.

[0024] In one embodiment of the present invention, the data preprocessing in step S3 is performed at the stage of establishing a database, and an improved recursive neural network is used to predict the ocean background noise.

[0025] In one embodiment of the present invention, an overfitting problem may occur in step S5, which may deteriorate the prediction effect of the model. Introducing regularization can retain all features of the sample and solve the overfitting problem. The regularized logistic regression model is:

[0026]

[0027] Among them, λ is called the regularization parameter. If the value of λ is too large, all parameters will be minimized, that is, underfitting will occur.

[0028] The above technical solution of the present invention has the following advantages over the prior art: the recognition method described in the present invention adopts a vector array signal feature vector with strong anti-interference ability, which is more conducive to the feature vector extraction of low-noise underwater acoustic targets. For underwater acoustic target recognition, its recognition success rate is higher than that of the sound pressure signal under the same conditions.

[0029] At the same time, during the database establishment phase, an improved recursive neural network was used to predict ocean background noise, eliminate some interference noise in the target channel, improve the quality of data samples, and facilitate accurate and rapid learning and establishment of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.

[0031] Figure 1 It is a time domain waveform diagram after normalization of data samples of the vector linear array underwater acoustic target recognition method based on deep learning of the present invention;

[0032] Figure 2 It is a schematic diagram of a model of the identification method of the present invention;

[0033] Figure 3 is a flow chart for learning and using the identification method of the present invention;

[0034] Figure 4 is a linear regression graph of the simulation data learning result of the recognition method of the present invention;

[0035] Figure 5 It is a linear regression diagram of the actual data learning result of the recognition method described in the present invention. DETAILED DESCRIPTION

[0036] This embodiment provides a vector linear array underwater acoustic target recognition method based on deep learning. The recognition method uses the data received by the vector linear array to establish a database. The vector linear array can not only receive sound pressure information, but also obtain omnidirectional vibration velocity information at the same position. The method includes the following steps:

[0037] Step S1: Use the data received by the vector linear array to establish a database. The vector sensor can simultaneously measure the sound pressure p and two mutually orthogonal vibration velocity components ν in the sound field at the same position x , ν y , under the plane wave condition:

[0038]

[0039] In the above formula, x(t) is the sound pressure waveform, θ s is the horizontal azimuth angle of sound wave propagation;

[0040] Step S2: Using the weighted linear combination of the two velocity components of the sound field, the beam can be rotated in two-dimensional space. Let:

[0041]

[0042] In the above formula, Ψ is the guiding direction;

[0043] Step S3: The data preprocessing model uses 10 LSTM multicells connected in series. The output of each multicell is a linear regression, and the cost function is selected as the mean square error loss function;

[0044] Step S4: Select the logistic regression algorithm to build a learning model. The logistic regression model represents:

[0045] h θ (x) = g(θ T X)

[0046] Where: X represents the feature vector;

[0047] g represents the logic function,

[0048] but

[0049] h θ The role of (x) is to calculate the possibility of the output variable being equal to 1 for a given input variable based on the selected parameters, thereby classifying the input variable, such as Figure 2 As shown;

[0050] Step S5: Fitting the logistic regression model, that is, fitting the parameters θ in the logistic regression model, requires defining the optimization objective or cost function used to fit the parameters:

[0051]

[0052] in, Substitute it into the derivative and simplify it to find the min θ When J(θ)

[0053] It can be seen from this that the velocity output of the vector sensor has a dipole directivity. c The directivity maximum of (t) is at θ s =ψ, and ν s The directivity zero point of (t) is also at θ s =ψ. As long as the value of ψ is changed, the directivity of the vibration velocity can be electronically rotated to form a zero point in any direction, thereby suppressing interference and being more conducive to the extraction of feature vectors of low-noise underwater acoustic targets.

[0054] In the data preprocessing step S3, during the database establishment phase, an improved recursive neural network is used to predict the ocean background noise.

[0055] The learning results eliminate some interference noise in the target channel, improve the quality of data samples, facilitate accurate and rapid learning and establishment of the model, and improve the efficiency of model establishment.

[0056] In order to make the deep learning process more convergent, the input data samples are normalized and the most classic scaling and data centering method is used, that is, the mean value is subtracted from each data and scaled according to the maximum value. The processed results are as follows: Figure 1 shown.

[0057] In step S5, there will be an overfitting problem, which will make the prediction effect of the model worse. Introducing regularization can retain all the features of the sample and solve the overfitting problem at the same time. The regularized logistic regression model is:

[0058]

[0059] Among them, λ is called the regularization parameter. If the value of λ is too large, all parameters will be minimized, that is, underfitting will occur.

[0060] The present invention uses the gradient descent algorithm to find the minimum parameter of the cost function. The samples in the database are divided into two parts, one as training samples and the other as cross-validation samples. The estimation function is obtained by training the training samples, and the cross-validation samples are calculated by the estimation function to obtain the corresponding estimation value. The estimation value is compared with the actual category to verify the reliability of the algorithm, so that we can use this relatively reliable estimation function to predict the category of new data. The process is as follows: Figure 3 shown.

[0061] A set of data was simulated using an underwater acoustic signal simulator and classified using a logistic regression algorithm. The learning accuracy was 89%, meeting the classification requirements, such as Figure 4 shown.

[0062] Using historical data with clear target attributes, we extracted the target’s spectral features (two of which are more obvious line spectra) for fishing boats and large transport ships as feature data for machine learning. A total of 16 samples were used in this experiment.

[0063] Learning results such as Figure 5 As shown in the figure, the learning accuracy is 88.24%, which meets the classification requirements of this time. Due to the small number of data samples and insufficient fitting, the prediction accuracy of new data may be low, and the number of data samples needs to be increased to further optimize the training model.

[0064] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A vector linear array underwater acoustic target recognition method based on deep learning, wherein the recognition method uses the data received by the vector linear array to establish a database, and the vector linear array can not only receive sound pressure information, but also obtain omnidirectional vibration velocity information at the same position, characterized in that: The steps include: Step S1: Use the data received by the vector linear array to establish a database. The vector sensor can simultaneously measure the sound pressure p and two mutually orthogonal vibration velocity components ν in the sound field at the same position x , ν y , under the plane wave condition: In the above formula, x(t) is the sound pressure waveform, θ s is the horizontal azimuth angle of sound wave propagation; Step S2: Using the weighted linear combination of the two velocity components of the sound field, the beam can be rotated in two-dimensional space. Let: n c (t)=n x (t)cosψ+ν y (t)sinψ n s (t)=-ν x (t)sinψ+ν y (t)cosψ In the above formula, Ψ is the guiding direction; Step S3: The data preprocessing model uses 10 LSTM multicells connected in series. The output of each multicell is a linear regression, and the cost function is selected as the mean square error loss function; Step S4: Select the logistic regression algorithm to build a learning model. The logistic regression model represents: h θ (x)=g(θ T X) Where: X represents the feature vector; g represents the logic function, but h θ The role of (x) is to calculate the probability of the output variable being equal to 1 for a given input variable based on the selected parameters, thereby classifying the input variable; Step S5: Fitting the logistic regression model, that is, fitting the parameters θ in the logistic regression model, requires defining the optimization objective or cost function used to fit the parameters: in, Substitute it into the derivative and simplify it to find the min θ When J(θ) 2. The method for underwater acoustic target recognition based on vector linear array based on deep learning according to claim 1, characterized in that: In step S2, c The directivity maximum of (t) is at θ s =ψ, and ν s The directivity zero point of (t) is also at θ s =ψ; as long as the value of ψ is changed, the directivity of the vibration velocity can be rotated electronically to form a zero point in any direction, thereby suppressing interference and being more conducive to the extraction of characteristic vectors of low-noise underwater acoustic targets.

3. The method for underwater acoustic target recognition based on vector linear array based on deep learning according to claim 1, characterized in that: In the data preprocessing step S3, during the database establishment phase, an improved recursive neural network is used to predict the ocean background noise.

4. The method for underwater acoustic target recognition based on vector linear array based on deep learning according to claim 1, characterized in that: In step S5, there will be an overfitting problem, which will make the prediction effect of the model worse. Introducing regularization can retain all the features of the sample and solve the overfitting problem at the same time. The regularized logistic regression model is: Among them, λ is called the regularization parameter. If the value of λ is too large, all parameters will be minimized, that is, underfitting will occur.

Images