A method for underwater acoustic target classification and orientation based on array element correlation characteristic deep convolutional neural network

Abstract

The application provides a method for underwater acoustic target classification and orientation based on an array element correlation feature deep convolutional neural network, comprising: obtaining an underwater acoustic target radiation noise signal x(t), generating a data sample set X of the underwater acoustic target radiation noise signal x(t), and obtaining a spectrum Y of the underwater acoustic target radiation noise signal x(t) through Fourier transform i ; the spectrum Y i ; calculating a covariance along a spectrum frequency dimension to obtain a covariance matrix Rx i ; the covariance matrix Rx i ; performing complex component dimension expansion to obtain a dimension-expanded covariance matrix Rx i '; inputting the covariance matrix Rx i ' into a deep convolutional neural network model based on an array element correlation feature, associating different target types in each branch of the deep convolutional neural network model, and outputting a predicted direction of the underwater acoustic target The application uses a covariance matrix of a spectrum of an underwater acoustic target radiation noise signal as network input, extracts phase-frequency difference features between underwater acoustic target radiation noise signals through a network model, constructs a multi-branch network output structure, and realizes classification and orientation tasks of underwater acoustic targets.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater acoustic target information application, and particularly relates to an underwater acoustic target classification and orientation method based on an array element correlation feature deep convolutional neural network. BACKGROUND

[0002] By utilizing underwater acoustic target radiation noise signals for target classification or orientation, the acquisition of underwater acoustic target type or bearing information can be realized, which has important application value in the fields of ocean exploration, underwater navigation, anti-submarine warfare and the like. However, due to the complexity of the marine environment, the sonar array system inevitably receives radiation noise signals of multiple underwater acoustic targets in the target detection process. The existing method can only realize the analysis of the bearing of the underwater acoustic target or the identification of the target type, and cannot simultaneously associate and correspond the bearing of each target with the underwater acoustic target type, which is not conducive to the subsequent execution of the underwater acoustic target detection task.

[0003] With the continuous development of deep learning technology, its self-adaptive capability enables the model to mine representative features from a large amount of data, and it has been widely applied in many fields such as target detection and identification. The underwater acoustic target radiation noise signals received by different array elements of the sonar array system have phase delays, and the phase delays are jointly determined by the signal coming direction, the array element position and the signal center frequency. Therefore, by analyzing the phase delay information between the signals received by each array element, the classification and orientation tasks of the underwater acoustic target can be realized. In addition, the correlation feature between the array elements of the sonar array system describes the correlation between the underwater acoustic target radiation noise signals received by different array elements, and the correlation feature is usually embodied by the covariance matrix of the target radiation noise signal. The covariance matrix of the target radiation noise signal accurately captures the correlation feature between the signals received by each array element, and contains the power difference and phase difference of the signal. Therefore, the correlation feature between each component in the frequency spectrum covariance matrix of the underwater acoustic target radiation noise signal implies the phase-frequency characteristics of different underwater acoustic target radiation noise signals, which provides a basis for the realization of the underwater acoustic target classification and orientation task. SUMMARY

[0004] The application aims at solving the problem in the prior art that the two tasks cannot be associated with each other by utilizing underwater acoustic target radiation noise signals, and provides an underwater acoustic target classification and orientation method based on an array element correlation feature deep convolutional neural network. The frequency spectrum covariance matrix of the underwater acoustic target radiation noise signal is used as the network input, the phase-frequency difference feature between the underwater acoustic target radiation noise signals is extracted through a phase-frequency feature mining module, a multi-branch network output structure is constructed, and the classification and orientation tasks of the underwater acoustic target are realized.

[0005] To achieve the above-mentioned purpose, the technical solution provided by the application is:

[0006] A method for underwater acoustic target classification and orientation based on deep convolutional neural network with array element correlation features, comprising:

[0007] Step 1: Obtain the underwater acoustic target radiation noise signal x(t), generate a data sample set X from the underwater acoustic target radiation noise signal x(t), and obtain the spectrum Y of the underwater acoustic target radiation noise signal x(t) by Fourier transforming the data sample set X. i ;

[0008] Step 2: The spectrum Y i Calculate the covariance along the frequency dimension of the spectrum to obtain the spectrum Y i The covariance matrix Rx i ;

[0009] Step 3: The covariance matrix Rx i After the complex component dimension expansion, the covariance matrix Rx' after dimension expansion is obtained i ;

[0010] Step 4: The covariance matrix Rx' i Input is a deep convolutional neural network model based on array element correlation features. Each branch in the deep convolutional neural network model is associated with different target types, and the predicted direction of the underwater acoustic target is output.

[0011] As a further limitation of the present invention, the step 1 comprises:

[0012] Step (11) obtains the underwater acoustic target radiation noise signal x(t) received by the sonar array system, which is expressed as:

[0013] x(t)=x1(t)+x2(t)+…+x n (t)+noise(t) formula (1)

[0014] In formula (1), x1(t) represents the radiation noise signal of underwater acoustic target 1, x2(t) represents the radiation noise signal of underwater acoustic target 2, and x n (t) represents the radiation noise signal of the underwater acoustic target n, noise(t) represents the ocean noise signal; x1(t), x2(t), x n The expression of (t) is:

[0015]

[0016] In formula (2), f1 represents the center frequency of the underwater acoustic target 1 radiated noise signal, s1(t) represents the noise source signal of the underwater acoustic target 1, a1(f1, θ1) represents the direction vector of the underwater acoustic target 1 radiated noise signal, f2 represents the center frequency of the underwater acoustic target 2 radiated noise signal, s2(t) represents the noise source signal of the underwater acoustic target 2, a2(f2, θ2) represents the direction vector of the underwater acoustic target 2 radiated noise signal, f n represents the center frequency of the underwater acoustic target n radiated noise signal, s n (t) represents the noise source signal of the underwater acoustic target n, a n (f n ,θ n ) represents the direction vector of the underwater acoustic target n radiated noise signal, d represents the array element spacing, c represents the underwater sound speed, M represents the number of array elements, θ1 represents the direction of the underwater acoustic target 1 radiated noise signal relative to the sonar array system, θ2 represents the direction of the underwater acoustic target 2 radiated noise signal relative to the sonar array system, and θ n represents the direction of the underwater acoustic target n radiated noise signal relative to the sonar array system.

[0017] Step (12) sets the number of array elements of the sonar array system as M, the sampling frequency of the underwater acoustic target radiated noise signal x(t) as fs, the signal length as T, and the dimension size of the underwater acoustic target radiated noise signal received by each array element of the sonar array system as 1×N (N=T·fs). Therefore, the dimension of the underwater acoustic target radiated noise signal x(t) is M×N.

[0018] Step (13) sets the value range of the incident direction θ of the underwater acoustic target radiated noise signal x(t) as The value range of the signal-to-noise ratio SNR of the underwater acoustic target radiated noise signal x(t) is [S sta ,S end ]. The data sample set X is generated within the range of the incident direction θ and the signal-to-noise ratio SNR, and the expression is as follows:

[0019] X={X1,X2,…X K} Formula (3)

[0020] In formula (3), X1 represents the underwater acoustic target radiated noise signal x(t) received by the sonar array system under condition 1, X2 represents the underwater acoustic target radiated noise signal x(t) received by the sonar array system under condition 2, and X K represents the underwater acoustic target radiated noise signal x(t) received by the sonar array system under condition K. The incident direction θ and the signal-to-noise ratio SNR of condition 1, condition 2, and condition K are different.

[0021] Step (14) sets the number of data samples Xi is a data matrix of M rows and N columns, the data sample set X is respectively Fourier transformed along the dimension of the channel of the array element in the sonar array system to obtain M frequency spectrum vectors of 1x F (F=N / 2+1) in dimension, wherein F represents the number of frequency points δ of the frequency dimension, and the data sample X i The expression is:

[0022] X i , i=1, 2, …, K, formula (4)

[0023] In formula (4), the data sample X i has a dimension size of MxN, and K represents the number of data samples X i .

[0024] As a further limitation of the present application, the step two comprises:

[0025] Step (21) stacks the M frequency spectrum vectors of 1x F to obtain the frequency spectrum Y i of each data sample X i in the data sample set X i , and the frequency spectrum Y i has a dimension size of Mx F;

[0026] Step (22) calculates the covariance matrix Rx i (δ) of the frequency spectrum data Y i (δ) corresponding to each frequency point δ along the dimension of the frequency spectrum Y i , wherein the calculation formula of the covariance matrix Rx i (δ) is:

[0027] Rx i (δ) = E{Y i (δ)·Y H (δ)} formula (5)

[0028] In formula (5), E{·} represents mathematical expectation, Y i H (δ) represents the conjugate transpose of the frequency spectrum data Y i (δ), the dimension size of the covariance matrix Rx i (δ) is MxM, a total of F MxM covariance matrices Rx i (δ) are obtained, and the covariance matrix Rx i (δ) is a Hermitian matrix, and the Hermitian matrix is a complex matrix.

[0029] Step (23) stacks the F MxM covariance matrices Rx i (δ) to obtain the frequency spectrum Yi covariance matrix Rx i , the covariance matrix Rx i has a dimension size of MxMxF.

[0030] As a further limitation of the present application, step three includes:

[0031] Step (31) performs complex dimension expansion on the covariance matrix Rx i extracts the real part and the imaginary part of the covariance matrix Rx i respectively. Stacks the real part and the imaginary part to form a new dimension, obtaining the expanded covariance matrix Rx' i , wherein the covariance matrix Rx' i has a dimension size of MxMx2xF.

[0032] Step (32) performs normalization processing on the covariance matrix Rx' i , the expression is:

[0033] Rx' i = Rx' i / max(|Rx' i |).

[0034] As a further limitation of the present application, step four includes:

[0035] Step (41) constructs and trains a deep convolutional neural network model based on the array correlation feature, obtaining the trained deep convolutional neural network model.

[0036] Step (42) inputs the covariance matrix Rx' i obtained in step three into the deep convolutional neural network model, and the deep convolutional neural network model deeply excavates the underwater acoustic target class features and azimuth features contained in the covariance matrix Rx' i , each branch in the multi-branch output structure of the deep convolutional neural network model is associated with a different target type, and simultaneously outputs the predicted azimuth of each underwater acoustic target

[0037] As a further limitation of the present application, step (41) includes:

[0038] Step (411) sets the azimuth label L i,p of the underwater acoustic target radiated noise signal.

[0039] Step (412) obtains a data set X for underwater acoustic target classification and orientation task data The data set X data is divided into a training set X train and a validation set X val according to a:b ratio;

[0040] Step (413) constructs a Correlation Feature Deep Convolutional Neural Network (CFDCNN) model based on array element correlation features, constructs a PFFM module, sets an activation function, constructs a multi-branch network output structure, and sets a loss function corresponding to each branch; Specifically, the CFDCNN model performs feature extraction and mining on the covariance matrix Rx i through the PFFM module, then uses the activation function to perform nonlinear transformation on the extracted features to highlight key features and suppress invalid features, thereby enhancing the nonlinear expression ability of the CFDCNN model; Then, the CFDCNN model calculates the difference between the predicted result and the true label through the loss function, optimizes the network performance by minimizing the loss, and finally realizes the bearing prediction of different targets through the multi-branch network output structure;

[0041] Step (414) divides the training set X train into I train batch data sets B train containing N train samples, selects a small batch of training data sets as the input of the CFDCNN network model, calculates the output of each branch of the CFDCNN network model under the action of network training parameters, and updates the parameters Ω1 of the CFDCNN network model according to the loss calculated by each branch; Wherein, J represents the Jth batch of data, J=1, 2,…, I train ; y p represents the label corresponding to each branch, represents the predicted bearing output by each branch of the CFDCNN network model.

[0042] Step (415) inputs the validation set X val into the CFDCNN network model through the parameters Ω1, and outputs the accuracy of the validation set X val under the action of the parameters Ω1, wherein the accuracy is the ratio of the number of correctly predicted samples to the total number of predicted samples;

[0043] Step (416) repeats step (414) and step (415). When the accuracy of the output calculated by the CFDCNN network model no longer improves after S consecutive iterations, or when the number of iterations e satisfies e=E, the iteration is stopped to obtain the trained CFDCNN network model.

[0044] As a further limitation of the present invention, the position label L in step (411) i,p , where i represents the i-th input sample (i=1,2,…,K), p represents the p-th underwater acoustic target (p=1,2,…,n), and each underwater acoustic target radiation signal has a corresponding azimuth label;

[0045] The position label L in step (411) i,p , whose dimension depends on the angular range of spatial scanning [Φ sta °,Φ end °] and spatial scanning angle resolution ΔΦ; orientation label L i,p The dimension size is 1×Z(Z=(Φ end °-Φ sta °) / ΔΦ+1);

[0046] The position label L in step (411) i,p , defined in the form of one-hot encoding, specifically, L i,p =[01,02,13,…,0 Z ,0 Z+1 ], (h=1,2,…,Z,Z+1), orientation label L i,p It is a sparse vector with only one position being 1 and the rest being 0, and the orientation label L i,p The first Z index positions of represent the direction information of the underwater acoustic target radiation noise signal, and the last index position (h=Z+1) represents the situation of the underwater acoustic target radiation noise signal in the sonar array system receiving signal, which is 1 if there is a underwater acoustic target radiation noise signal and 0 if there is no such signal.

[0047] The position label L in step (411) i,p , whose dimension size is 1×Z' (Z'=Z+1).

[0048] As a further limitation of the present invention, in step (412), the data set X data The expression is:

[0049]

[0050] In formula (6), K represents the dataset X data Sample data, n represents the number of categories of underwater acoustic targets, Rx' idenotes the extended covariance matrix, i denotes the i-th data set sample pair, L i,1 denotes the bearing label corresponding to the i-th sample of the underwater target 1 radiated noise signal, L i,2 denotes the bearing label corresponding to the i-th sample of the underwater target 2 radiated noise signal, L i,n denotes the bearing label corresponding to the i-th sample of the underwater target n radiated noise signal.

[0051] As a further limitation of the application, the CFDCNN model is constructed, comprising:

[0052] The constructed CFDCNN model is composed of B phase frequency feature mining modules, P three-dimensional convolution layers, P global average pooling layers, P fully connected layers, and P output layers, wherein the dimension size of the three-dimensional convolution layer is g x g x 1, and P = n.

[0053] The PFFM module is constructed, comprising:

[0054] The constructed PFFM module is composed of one three-dimensional convolution layer and one normalization layer, wherein the dimension size of the three-dimensional convolution layer is m x m x 1.

[0055] The activation function is set, comprising:

[0056] The tanh function is used as the activation function in the three-dimensional convolution layer, and the expression of the activation function is:

[0057]

[0058] In formula (7), χ represents the output of the previous layer.

[0059] The ReLU function is used as the activation function in the fully connected layer, and the expression of the activation function is:

[0060] ReLU (γ) = max (0, γ) formula (8)

[0061] In formula (8), γ represents the output of the global average pooling layer.

[0062] The The activation function of the output layer is adopted, z represents the output of the last fully connected layer, h represents the h-th neuron in the output layer, wherein h = 1, 2, …, Z', and the output of the first branch corresponds to the bearing label L

[0063] The multi-branch network output structure is constructed, comprising:

[0064] The first branch outputs the bearing label L i,1 corresponding to the underwater target 1 radiated noise signal, and the second branch outputs the bearing label L i,2The direction label L corresponding to the nth branch output corresponds to the radiation noise signal of the nth underwater target i,n , a total of n output branches, i represents the ith input sample; each branch corresponds to a different type of underwater target, and the predicted direction of the underwater target is output simultaneously;

[0065] The loss function corresponding to each branch is set, including:

[0066] The loss function loss-f corresponding to each branch is set p ,(p=1, 2,..., n), and the weight weight of each branch loss function is set p ;

[0067] The loss function Loss of the CFDCNN model is obtained as weight p ⊙loss-f p , each branch is calculated according to the corresponding label L i,p and the respective loss function loss-f p , the corresponding loss is calculated, and the parameters Omega of the CFDCNN model are updated and optimized.

[0068] As a further limitation of the application, in the step (414), under the action of the network training parameters, the network training parameters include:

[0069] The network training parameters of the CFDCNN model are set, wherein the network training parameters include: initial learning rate lr0, batch size N train of the training set, batch size N val of the validation set, iteration number Epoch, early stopping step number S, optimizer type, and initial network model parameter Omega0.

[0070] The advantages of the application are:

[0071] 1. The application uses the covariance matrix of the underwater target radiation noise signal spectrum as the network input, extracts the phase frequency difference features between the underwater target radiation noise signals through the phase frequency feature mining module, constructs a multi-branch network output structure, and realizes the classification and orientation of the underwater target;

[0072] 2. The underwater target classification and orientation method based on the array element correlation feature deep convolutional neural network model uses the covariance matrix of the underwater target radiation noise signal spectrum as the input of the deep convolutional neural network model, associates each branch with different types of underwater targets through the phase frequency feature mining module and the constructed multi-branch network output structure, and simultaneously outputs the direction of the underwater target, realizing the corresponding association between the target direction and the target type.

[0073] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and the attendant drawings or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0074] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the accompanying drawings.

[0075] Figure 1 The application provides a water acoustic target classification and orientation method based on an array element correlation feature deep convolutional neural network;

[0076] Figure 2 The application provides a water acoustic target classification and orientation method based on an array element correlation feature deep convolutional neural network;

[0077] Figure 3 The application provides a complex component dimension expansion intention;

[0078] Figure 4 The application provides a structure diagram of the array element correlation feature CFDCNN model;

[0079] Figure 5 When the water acoustic target radiation noise signal only contains the water acoustic target 1 radiation noise signal, the application provides a mean absolute error result distribution diagram of the output prediction result of the first branch;

[0080] Figure 6 When the water acoustic target radiation noise signal only contains the water acoustic target 2 radiation noise signal, the application provides a MAE result distribution diagram of the output prediction result of the second branch;

[0081] Figure 7 When the water acoustic target radiation noise signal contains the water acoustic target 1 and the water acoustic target 2 radiation noise signal, the application provides a MAE result distribution diagram of the output prediction result of the two branches. DETAILED DESCRIPTION

[0082] Embodiments of the application are described in detail below, which are exemplary and intended to explain the application, and cannot be understood as a limitation of the application.

[0083] Reference should be made to Figure 1The embodiment of the present application provides a method for underwater acoustic target classification and orientation based on an array element correlation feature deep convolutional neural network, first, a phase frequency feature mining module based on a spectrum covariance matrix of underwater acoustic target radiation noise signals is constructed to extract different underwater acoustic target spectrum difference features, then, a multi-branch network output structure is designed, different branches correspond to different underwater acoustic target types, and the bearing of the underwater acoustic target is output simultaneously, finally, the classification and orientation of the underwater acoustic target are realized according to the type and bearing of the underwater acoustic target. Figure 1 and Figure 2 The underwater acoustic target classification and orientation method of the embodiment of the present application specifically comprises the following steps:

[0084] Step one, underwater acoustic target radiation noise signals are acquired, underwater acoustic target radiation noise signal data sample set X is generated, and the spectrum Y of the underwater acoustic target radiation noise signal is obtained through Fourier transform of the data sample set X. i .

[0085] The step one of the embodiment of the present application comprises the following steps (11)-(14):

[0086] Step (11) acquires the underwater acoustic target radiation noise signals x(t) received by the sonar array system, and the expression is:

[0087] x(t) = x1(t) + x2(t) + … + xn(t) + noise(t) Formula (1) n

[0088] In formula (1), x1(t) represents the radiation noise signal of the underwater acoustic target 1, x2(t) represents the radiation noise signal of the underwater acoustic target 2, x n (t) represents the radiation noise signal of the underwater acoustic target n, and noise(t) represents the ocean noise signal; wherein:

[0089]

[0090] In formula (2), f1 represents the center frequency of the radiation noise signal of the underwater acoustic target 1, s1(t) represents the noise source signal of the underwater acoustic target 1, a1(f1, θ1) represents the direction vector of the radiation noise signal of the underwater acoustic target 1, f2 represents the center frequency of the radiation noise signal of the underwater acoustic target 2, s2(t) represents the noise source signal of the underwater acoustic target 2, a2(f2, θ2) represents the direction vector of the radiation noise signal of the underwater acoustic target 2, f n n represents the center frequency of the radiation noise signal of the underwater acoustic target n, s n (t) represents the noise source signal of the underwater acoustic target n, a n (f n ,θ n ​) represents the direction vector of the underwater target n radiated noise signal, d represents the array element spacing, c represents the underwater sound speed, M represents the number of array elements, θ1 represents the azimuth of the underwater target 1 radiated noise signal relative to the sonar array system, θ2 represents the azimuth of the underwater target 2 radiated noise signal relative to the sonar array system, θ n represents the azimuth of the underwater target n radiated noise signal relative to the sonar array system; preferably f1=88Hz, f2=121Hz, d=3.75m, c=1500m / s.

[0091] Step (12) sets the number of array elements of the sonar array system as M, the sampling frequency of the underwater target radiated noise signal x(t) as fs, the signal length as T, and the dimension size of the underwater target radiated noise signal received by each array element as 1×N (N=T·fs=1×512=512), so the dimension of the underwater target radiated noise signal x(t) is M×N=12×512.

[0092] Considering the complexity of the real marine environment, the signal x(t) received by the sonar array system does not always contain the target radiated noise signal, and the ocean noise signal always exists. Therefore, the underwater target radiated noise signal x(t) has four cases: (1) only contains ocean noise signal x(t)=n(t); (2) contains underwater target 1 radiated noise signal x(t)=x1(t)+n(t); (3) contains underwater target 2 radiated noise signal x(t)=x2(t)+n(t); (4) contains underwater target 1 and underwater target 2 radiated noise signals x(t)=x1(t)+x2(t)+n(t).

[0093] Step (13) sets the value range of the incident direction θ of the underwater target radiated noise signal x(t) as [30°, 150°], and the value range of the signal-to-noise ratio SNR of the underwater target radiated noise signal x(t) as [-10dB, 5dB], and generates a data sample set X within the incident direction θ and the signal-to-noise ratio SNR range, expressed as:

[0094] X={X1,X2,…X K} Formula (3)

[0095] In formula (3), X1 represents the underwater target radiated noise signal x(t) received by the sonar array system under condition 1, X2 represents the underwater target radiated noise signal x(t) received by the sonar array system under condition 2, and X K represents the underwater target radiated noise signal x(t) received by the sonar array system under condition K; the incident direction θ and the signal-to-noise ratio SNR of condition 1, condition 2 and condition K have different values; preferably K=73,600, K represents the number of data samples.

[0096] Step (14) each data sample X in the data sample set X i , (i=1, 2, …, K) is a data matrix of M rows and N columns, the data sample set X is respectively Fourier transformed along the dimension where the element channel is located, 12 spectral vectors of 1x257 are obtained, wherein 257 represents the frequency point number of the frequency dimension (δ=1, 2, …, 257), and the data sample X i The expression is:

[0097] X i , (i=1, 2, …, K) formula (4)

[0098] In formula (4), the 12 spectral vectors of 1x257 are stacked along the “row” dimension, and K represents the number of data samples X i . i The spectrum Y i of the data sample X i is obtained, and the dimension size is 12x257.

[0099] In actual application, in step one of the embodiment of the application, the underwater acoustic target radiation noise signal x(t) is a simulated underwater acoustic target radiation noise signal using python, the sonar array system is a uniform linear array of M=12 elements, the sampling frequency fs=512Hz, the signal duration T=1s, and the target category number n=2.

[0100] Step two, the spectrum Y i of the underwater acoustic target radiation noise signal is calculated along the frequency dimension, and the covariance matrix Rx i of the signal spectrum Y i is obtained.

[0101] The above step two of the embodiment of the application includes steps (21)-(23):

[0102] Step (21) stacks the M spectral vectors of 1x F to obtain the spectrum Y i of the data sample X i , and the dimension size of the spectrum Y i is MxF;

[0103] Step (22) calculates the covariance matrix Rx i (δ) of the spectral data Y i (δ) corresponding to each frequency point δ along the dimension of the spectrum Y i , wherein the calculation formula of the covariance matrix Rx i (δ) is:

[0104] Rx i (δ) = E{Y i (δ)·Y iH (δ)} Equation (5)

[0105] In Equation (5), E{·} represents mathematical expectation, also known as mean value, Y i H (δ) represents the conjugate transpose of the spectrum data Y i (δ), the covariance matrix Rx i (δ) has a dimension of 12x12, and a total of 257 covariance matrices Rx i (δ) of 12x12 are obtained, and the covariance matrix Rx i (δ) is a Hermitian matrix, and the Hermitian matrix is a complex matrix.

[0106] Step (23) stacks the 257 covariance matrices Rx i (δ) of 12x12, to obtain the covariance matrix Rx i of the spectrum Y i , the covariance matrix Rx i has a dimension of 12x12x257.

[0107] Step three, the covariance matrix Rx i of the spectrum of the underwater target radiated noise signal is subjected to complex component dimension expansion, to obtain the dimension-expanded covariance matrix Rx' i . The specific complex component dimension expansion is shown in Figure 3 .

[0108] The step three of the embodiment of the application comprises steps (31) to (32):

[0109] Step (31) performs complex component dimension expansion on the covariance matrix Rx i of the spectrum Y i of the underwater target radiated noise signal x(t), to respectively extract the real component i and the imaginary component of the covariance matrix Rx Stack the real component and the imaginary component to form a new dimension, to obtain the expanded covariance matrix Rx' i , the dimension of the covariance matrix Rx' i is 12x12x2x257.

[0110] Step (32) performs normalization processing on the expanded covariance matrix Rx' i , and the expression is:

[0111] Rx' i = Rx' i / max(|Rx’ i |), wherein Rx'i represents an input sample of the deep convolutional neural network model.

[0112] Step four, inputting the covariance matrix Rx'i into the deep convolutional neural network model of the element correlation feature, associating different target types with each branch of the deep convolutional neural network model to realize classification of the underwater acoustic target, and outputting a predicted direction of the corresponding underwater acoustic target by each branch

[0113] Step four of the embodiment of the present application includes steps (41) and (42):

[0114] Step (41) constructs and trains the deep convolutional neural network model based on the element correlation feature to obtain the trained deep convolutional neural network model.

[0115] The above step (41) of the embodiment of the present application includes:

[0116] Step (411) sets the direction label L of the underwater acoustic target radiation noise signal i,p ; the direction label L in the above step (411) of the embodiment of the present application i,p , i represents the i-th input sample (i=1, 2, …, 73600), and p represents the p-th underwater acoustic target (p=1, 2). Each underwater acoustic target radiation signal corresponds to one direction label. The direction label L in the above step (411) of the embodiment of the present application i,p , the dimension size depends on the angle range [0°, 180°] of spatial scanning and the spatial scanning angle resolution ΔΦ=1°; the dimension size of the direction label L i,p is 1×Z (Z=180 / 1+1=181); the direction label L in the above step (411) of the embodiment of the present application i,p is defined in the form of one-hot encoding, specifically, L i,p =[01, 02, 13, …, 0 Z , 0 Z+1 ], (h=1, 2, …, Z, Z+1), the direction label L i,p is a sparse vector with only one position being 1 and the rest being 0, the first Z index positions of the direction label L i,p represent the direction information of the underwater acoustic target radiation noise signal, and the last index position (h=Z+1) represents the presence of the underwater acoustic target radiation noise signal in the sonar array system receiving signal, 1 if present, and 0 if not; the dimension size of the direction label L in the above step (411) of the embodiment of the present application i,p is 1×Z'=1×182 (Z'=Z+1=182).

[0117] Step (412) obtains a dataset X for underwater acoustic target classification and orientation tasks data , the dataset X data Divide the training set into X according to the ratio of 4:1 train and validation set X val ; Dataset X data The number of samples K = 73,600, the training set X train The number of samples is 58,880, and the validation set X val The number of samples is 14,720.

[0118] Specifically, in step (412) of the embodiment of the present invention, the data set X data The expression is:

[0119]

[0120] In formula (6), K represents the dataset X data Sample data, K = 73600, n represents the number of categories of underwater acoustic targets, Rx' i represents the expanded covariance matrix, i represents the i-th dataset sample pair, L i,1 Indicates the azimuth label of the noise signal radiated by underwater acoustic target 1 in the i-th sample, L i,2 represents the azimuth label of the noise signal radiated by underwater acoustic target 2 in the i-th sample, L i,n Represents the azimuth label of the noise signal radiated by underwater acoustic target n in the i-th sample.

[0121] Step (413) constructs a Correlation Feature Deep Convolutional Neural Network (CFDCNN) model based on the element correlation feature, constructs a PFFM module, sets the activation function, constructs a multi-branch network output structure, and sets the loss function corresponding to each branch. Specifically, the CFDCNN model uses the PFFM module to calculate the covariance matrix Rx' i Perform feature extraction and mining, and then use the activation function to perform nonlinear transformation on the extracted features to highlight key features and suppress invalid features, thereby enhancing the nonlinear expression ability of the CFDCNN model; then, the CFDCNN model calculates the difference between the predicted results and the true labels through the loss function, and optimizes the network performance by minimizing the loss, and finally realizes the orientation prediction of different targets through the multi-branch network output structure. The specific structure of the CFDCNN model of the embodiment of the present invention is as follows Figure 4 shown.

[0122] Furthermore, step (413) of the embodiment of the present invention includes:

[0123] (413-1) The above constructing the CFDCNN model of the embodiment of the application comprises: the constructed CFDCNN model is composed of B phase frequency feature mining modules, P three-dimensional convolution layers, P global average pooling layers, P full connection layers and P output layers, wherein the dimension size of the three-dimensional convolution layer is 3*3*1, and P = n.

[0124] (413-2) The above constructing the PFFM module of the embodiment of the application comprises: the constructed PFFM module is composed of one three-dimensional convolution layer and one normalization layer, wherein the dimension size of the three-dimensional convolution layer is m*m*1.

[0125] (413-3) The above setting the activation function of the embodiment of the application comprises: adopting a tanh function as the activation function in the three-dimensional convolution layer, and the expression of the activation function is:

[0126]

[0127] In formula (7), χ represents the output of the previous layer.

[0128] (413-4) The above adopting a ReLU function as the activation function in the full connection layer of the embodiment of the application, and the expression of the activation function is:

[0129] ReLU (γ) = max (0, γ) formula (8)

[0130] In formula (8), a represents the output of the global average pooling layer.

[0131] (413-5) The above adopting As the activation function of the output layer, z represents the output of the last full connection layer, and h represents the hth neuron in the output layer, wherein h = 1, 2, … 182.

[0132] (413-6) The above constructing the multi-branch network output structure of the embodiment of the application comprises: the first branch outputs the bearing label L i,1 of the underwater acoustic target 1 radiation noise signal (i represents the ith input sample), the second branch outputs the bearing label L i,2 of the underwater acoustic target 2 radiation noise signal, the output result of the first branch is the bearing prediction result of the underwater acoustic target 1 radiation noise signal , and the output result of the second branch is the bearing prediction result of the underwater acoustic target 2 radiation noise signal

[0133] (413-7) The above setting the loss function (Cross Entropy-function, CE-f) corresponding to each branch of the embodiment of the application comprises: setting the loss function loss-fp =[loss-f1,loss-f2]=[CE-f,CE-f](p=1,2), set the weight of each branch loss function p ; Obtain the loss function of the network model Loss = weight p ⊙loss-f p , each branch follows the corresponding label L i,p And their respective loss functions loss-f p Calculate the corresponding loss and update and optimize the parameters Ω of the CFDCNN model.

[0134] Step (414) sets the training set X train Divided into I train =460 including N train =Batch dataset B of 128 samples train , select a mini-batch training dataset As the input of the CFDCNN network model, under the action of the network training parameters, the output of each branch of the CFDCNN network model is calculated, and the loss obtained is calculated based on each branch Update the parameters Ω1 of the CFDCNN network model; where J represents the Jth batch of data, J = 1, 2, ..., 460; y p Indicates the labels corresponding to each branch, Indicates the predicted orientation of each branch output of the CFDCNN network model.

[0135] In the above step (414) of the embodiment of the present invention, under the action of the network training parameters, the network training parameters include: setting the network training parameters of the CFDCNN model, wherein the network training parameters include: initial learning rate lr0 = 0.001, batch size N of the training set train =128, validation set batch size N val =50, number of iterations Epoch, number of early stopping steps Step=40, optimizer type Adam, initialize network model parameters Ω0.

[0136] Step (415) sets the validation set X val Through the parameter Ω1 of the CFDCNN network model, the accuracy of the validation set is output under the action of the parameter Ω1, where the accuracy is the ratio of the number of correctly predicted samples to the total number of predicted samples.

[0137] Step (416) repeats step (414) and step (415). When the accuracy of the output calculated by the CFDCNN network model no longer improves after S consecutive iterations, or when the number of iterations e satisfies e=50, the iteration is stopped to obtain a trained CFDCNN network model.

[0138] Step (42) converts the spectrum Y of the underwater acoustic target radiation noise signal x(t) into i The covariance matrix Rx' i After inputting the CFDCNN model, each branch of the CFDCNN model is associated with different underwater acoustic target types, and each branch outputs the predicted direction of each underwater acoustic target. The classification and orientation tasks of underwater acoustic targets are achieved. In the range of SNR = -10dB ~ 5dB, the incident azimuth θ∈[30°,150°], a data set for CFDCNN model testing is generated. The test data set contains 121 angle conditions, each angle corresponds to 50 samples, which are used to evaluate the performance of the underwater acoustic target classification and orientation method of the embodiment of the present invention. Specifically, when the underwater acoustic target radiation noise signal only contains the radiation noise of underwater acoustic target 1, the predicted azimuth output of the first branch is focused on. In the range of SNR=-10dB~5dB, The average absolute error results are concentrated in the range of 0° to 1°. For detailed experimental results, see Figure 5 When the underwater acoustic target radiation noise signal only contains the radiation noise of underwater acoustic target 2, the predicted direction output by the second branch is focused on. In the range of SNR=-10dB~5dB, The mean absolute error results are concentrated in the range of 0° to 0.75°. For detailed experimental results, see Figure 6 When the underwater acoustic target radiation noise signal contains the radiation noise of underwater acoustic target 1 and underwater acoustic target 2, the first branch outputs the predicted direction of the radiation noise signal of underwater acoustic target 1. The second branch outputs the predicted direction of the noise signal radiated by underwater acoustic target 2 The average absolute error of the predicted azimuth is concentrated between 0° and 2.5°. Detailed experimental results are shown in Figure 7 shown.

[0139] The embodiment of the present invention uses the covariance matrix of the spectrum of the noise signal radiated by underwater acoustic targets as the network input, extracts the phase-frequency difference characteristics between the noise signals radiated by underwater acoustic targets through a phase-frequency feature mining module, and constructs a multi-branch network output structure to achieve the classification and orientation tasks of underwater acoustic targets. The embodiment of the present invention is based on a deep convolutional neural network model with array element correlation features. The method for underwater acoustic target classification and orientation uses the covariance matrix of the spectrum of the noise signal radiated by underwater acoustic targets as the input of the deep convolutional neural network model. Through the phase-frequency feature mining module and the constructed multi-branch network output structure, each branch is associated with a different underwater acoustic target type, and the direction of the underwater acoustic target is output simultaneously, realizing the association and correspondence between the target direction and the target type.

[0140] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application.

Claims

1. A method for underwater acoustic target classification and orientation based on deep convolutional neural network with array element correlation features, characterized in that: include: Step 1: Obtain the underwater acoustic target radiation noise signal x(t), generate a data sample set X from the underwater acoustic target radiation noise signal x(t), and obtain the spectrum Y of the underwater acoustic target radiation noise signal x(t) by Fourier transforming the data sample set X. i ; Step 2: The spectrum Y i Calculate the covariance along the frequency dimension of the spectrum to obtain the spectrum Y i The covariance matrix Rx i ; Step 3: The covariance matrix Rx i After the complex component dimension expansion, the covariance matrix Rx is obtained i '; Step 4: The covariance matrix Rx i Input a deep convolutional neural network model based on array element correlation features. Each branch in the deep convolutional neural network model is associated with different target types, and the predicted position of the underwater acoustic target is output.

2. The underwater acoustic target classification and orientation method based on deep convolutional neural network with array element correlation characteristics according to claim 1 is characterized in that: The step one comprises: Step (11) obtains the underwater acoustic target radiation noise signal x(t) received by the sonar array system, which is expressed as: x(t)=x1(t)+x2(t)+…+x n (t)+noise(t) Formula (1) In formula (1), x1(t) represents the radiation noise signal of underwater acoustic target 1, x2(t) represents the radiation noise signal of underwater acoustic target 2, and x n (t) represents the radiation noise signal of the underwater acoustic target n, noise(t) represents the ocean noise signal; x1(t), x2(t), x n The expression of (t) is: In formula (2), f1 represents the center frequency of the noise signal radiated by underwater acoustic target 1, s1(t) represents the noise source signal of underwater acoustic target 1, a1(f1,θ1) represents the direction vector of the noise signal radiated by underwater acoustic target 1, f2 represents the center frequency of the noise signal radiated by underwater acoustic target 2, s2(t) represents the noise source signal of underwater acoustic target 2, a2(f2,θ2) represents the direction vector of the noise signal radiated by underwater acoustic target 2, and f n represents the center frequency of the noise signal radiated by underwater acoustic target n, s n (t) represents the noise source signal of underwater acoustic target n, a n (f n ,θ n ) represents the direction vector of the noise signal radiated by underwater acoustic target n, d represents the array element spacing, c represents the underwater sound speed, M represents the number of array elements, θ1 represents the azimuth of the noise signal radiated by underwater acoustic target 1 relative to the sonar array system, θ2 represents the azimuth of the noise signal radiated by underwater acoustic target 2 relative to the sonar array system, and θ n represents the azimuth of the noise signal radiated by underwater acoustic target n relative to the sonar array system, and k represents the kth array element; Step (12) sets the number of array elements of the sonar array system to M, the sampling frequency of the underwater acoustic target radiation noise signal x(t) to fs, the signal duration to T, and the dimension of the underwater acoustic target radiation noise signal received by each array element of the sonar array system to be 1×N, N=T·fs, then the dimension of the underwater acoustic target radiation noise signal x(t) is M×N; Step (13) sets the value range of the incident azimuth θ of the underwater acoustic target radiation noise signal x(t) to The signal-to-noise ratio SNR of the underwater acoustic target radiation noise signal x(t) ranges from [S sta ,S end ], a data sample set X is generated within the range of the incident azimuth θ and the signal-to-noise ratio SNR, and the expression is: X={X1,X2,…X K } Formula (3) In formula (3), X1 represents the noise signal x(t) radiated by the underwater acoustic target received by the sonar array system under condition 1, X2 represents the noise signal x(t) radiated by the underwater acoustic target received by the sonar array system under condition 2, and X K represents the underwater acoustic target radiation noise signal x(t) received by the sonar array system under condition K; wherein, the incident azimuth θ and the signal-to-noise ratio SNR values ​​are different for conditions 1, 2, and K; Each data sample X in the data sample set X in step (14) i is a data matrix with M rows and N columns. The data sample set X is Fourier transformed along the dimension of the array element channel in the sonar array system to obtain M frequency spectrum vectors with a dimension of 1×F, F=N / 2+1, where F represents the number of frequency points δ in the frequency dimension. The data sample X i The expression is: X i ,i=1,2,…,K Formula (4) In formula (4), the data sample X i The dimensions are all M×N, K represents the data sample X i The number of 3. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 2 is characterized in that: The second step includes: Step (21) stacks M frequency spectrum vectors of dimension 1×F to obtain each data sample X in the data sample set X. i Spectrum Y i , the spectrum Y i The dimension size is M×F; Step (22) along the spectrum Y i The dimension of each frequency point δ is used to calculate the corresponding spectrum data Y i The covariance matrix Rx of (δ) i (δ), where the covariance matrix Rx i The calculation formula of (δ) is: Rx i (δ) = E{Y i (δ)·Y i H (δ)} Equation (5) In formula (5), E{·} represents the mathematical expectation, Y i H (δ) represents the spectrum data Y i The conjugate transpose of (δ), the covariance matrix Rx i The dimension of (δ) is M×M, and a total of F M×M covariance matrices Rx are obtained i (δ), and the covariance matrix Rx i (δ) is the Hermitian matrix, which is a complex matrix; Step (23) converts F M×M covariance matrices Rx i (δ) are stacked to obtain the spectrum Y i The covariance matrix Rx i , the covariance matrix Rx i The dimensions are M×M×F.

4. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 3 is characterized in that: The step three includes: Step (31) is to calculate the covariance matrix Rx in step 2. i Perform complex component dimension expansion and extract the covariance matrix Rx i The real part of and the imaginary component The real part and the imaginary component Stacking forms a new dimension to obtain the expanded covariance matrix Rx i ', where the covariance matrix Rx i 'The dimension size is M×M×2×F; Step (32) is to calculate the covariance matrix Rx i 'Perform normalization, the expression is: Rx’ i =Rx’ i / max|Rx’ i |。 5. The underwater acoustic target classification and orientation method based on deep convolutional neural network with array element correlation characteristics according to claim 4 is characterized in that: The fourth step includes: Step (41) constructs and trains a deep convolutional neural network model based on array element correlation features to obtain the trained deep convolutional neural network model; Step (42) converts the covariance matrix Rx obtained in step 3 into i After inputting the deep convolutional neural network model, the deep convolutional neural network model deeply mines the covariance matrix Rx i The underwater acoustic target category features and azimuth features contained in the ', each branch in the multi-branch output structure of the deep convolutional neural network model is associated with a different target type, and simultaneously outputs the predicted azimuth of each underwater acoustic target 6. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 5 is characterized in that: Step (41) comprises: Step (411) sets the azimuth label L of the underwater acoustic target radiation noise signal i,p ; Step (412) obtains a dataset X for underwater acoustic target classification and orientation tasks data , the dataset X data Divide the training set X into the ratio of a:b train and validation set X val ; Step (413) constructs a CFDCNN model, constructs a PFFM module, sets an activation function, constructs a multi-branch network output structure, and sets a loss function corresponding to each branch; specifically, the CFDCNN model uses the PFFM module to calculate the covariance matrix Rx i 'Perform feature extraction and mining, and then use the activation function to perform nonlinear transformation on the extracted features to highlight key features and suppress invalid features, thereby enhancing the nonlinear expression ability of the CFDCNN model; then, the CFDCNN model calculates the difference between the predicted results and the true labels through the loss function, and optimizes the network performance by minimizing the loss, and finally realizes the orientation prediction of different targets through the multi-branch network output structure; Step (414) converts the training set X train Divided into I train Contains N train A batch dataset B of samples train , select a small batch of training data As the input of the CFDCNN model, under the action of the network training parameters, the output of each branch of the CFDCNN model is calculated, and the loss obtained by calculating each branch is calculated. Update the parameters Ω1 of the CFDCNN model; where J represents the Jth batch of data, J = 1, 2, ..., I train ;y p Indicates the labels corresponding to each branch, Indicates the predicted orientation of each branch output of the CFDCNN model; Step (415) sets the validation set X val Through the parameter Ω1 of the CFDCNN model, the verification set X is output under the action of the parameter Ω1 val The accuracy rate is the ratio of the number of correctly predicted samples to the total number of predicted samples; Step (416) repeats step (414) and step (415). When the accuracy of the output calculated by the CFDCNN model no longer improves after S consecutive iterations, or when the number of iterations e satisfies e=E, the iteration is stopped to obtain the trained CFDCNN model.

7. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 6 is characterized in that: The position label L in step (411) i,p , where i represents the i-th input sample i=1,2,…,K, p represents the p-th underwater acoustic target p=1,2,…,n, and each underwater acoustic target radiation signal has a corresponding azimuth label; The position label L in step (411) i,p , whose dimension depends on the angular range of spatial scanning [Φ sta °,Φ end °] and spatial scanning angle resolution ΔΦ; orientation label L i,p The dimension size is 1×Z, Z=(Φ end °-Φ sta °) / ΔΦ+1; The position label L in step (411) i,p , defined in the form of one-hot encoding, specifically, L i,p =[01,02,13,…,0 Z ,0 Z+1 ],h=1,2,…,Z,Z+1, orientation label L i,p It is a sparse vector with only one position being 1 and the rest being 0, and the orientation label L i,p The first Z index positions represent the direction information of the underwater acoustic target radiation noise signal, and the last index position represents the situation of the underwater acoustic target radiation noise signal in the sonar array system receiving signal, which is 1 if there is a noise signal, and 0 if there is no noise signal; The position label L in step (411) i,p , whose dimension size is 1×Z', Z'=Z+1.

8. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 6 is characterized in that: In the step (412), the data set X data The expression is: In formula (6), K represents the dataset X data Sample data, n represents the number of categories of underwater acoustic targets, Rx i ' represents the expanded covariance matrix, i represents the i-th data set sample pair, L i,1 represents the azimuth label of the noise signal radiated by underwater acoustic target 1 in the i-th sample, L i,2 represents the azimuth label of the noise signal radiated by underwater acoustic target 2 in the i-th sample, L i,n Represents the azimuth label of the noise signal radiated by underwater acoustic target n in the i-th sample.

9. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 6 is characterized in that: The construction of the CFDCNN model includes: The constructed CFDCNN model consists of B phase-frequency feature mining modules, P three-dimensional convolutional layers, P global average pooling layers, P fully connected layers, and P output layers. The dimension of the three-dimensional convolutional layer is g×g×1, and P=n. The construction of the PFFM module includes: The constructed PFFM module consists of one 3D convolutional layer and one normalization layer, where the dimension of the 3D convolutional layer is m×m×1; The setting of the activation function includes: The tanh function is used as the activation function in the three-dimensional convolutional layer. The expression of the activation function is: In formula (7), χ represents the output of the previous layer; The ReLU function is used as the activation function in the fully connected layer. The expression of the activation function is: ReLU(γ)=max(0,γ) Formula (8) In formula (8), γ represents the output of the global average pooling layer; use As the activation function of the output layer, z represents the output of the last fully connected layer, and h represents the h-th neuron in the output layer, where h = 1, 2, ..., Z'; The constructing of a multi-branch network output structure includes: The first branch outputs the azimuth label L corresponding to the noise signal radiated by underwater acoustic target 1 i,1 , the second branch outputs the azimuth label L corresponding to the noise signal radiated by underwater acoustic target 2 i,2 , the nth branch outputs the azimuth label L corresponding to the noise signal radiated by underwater acoustic target n i,n , there are n output branches in total, i represents the i-th input sample; each branch corresponds to a different underwater acoustic target type and outputs the predicted direction of the underwater acoustic target at the same time; The setting of the loss function corresponding to each branch includes: Set the loss function loss-f corresponding to each branch p ,p=1,2,…,n, set the weight of each branch loss function p ; Get the loss function Loss=weight of the CFDCNN model p ⊙loss-f p , each branch follows the corresponding label L i,p And their respective loss functions loss-f p Calculate the corresponding loss and update and optimize the parameters Ω of the CFDCNN model.

10. The underwater acoustic target classification and orientation method based on deep convolutional neural network with element correlation characteristics according to claim 6 is characterized in that: In the step (414), under the effect of the network training parameters, the network training parameters include: Set the network training parameters of the CFDCNN model, including the initial learning rate lr0, the batch size N of the training set train , batch size N of the validation set val , number of iterations Epoch, number of early stopping steps S, optimizer type, and initialized network model parameters Ω0.

Images