Underwater target DOA estimation method based on CNN under ray sound field model

Abstract

The invention discloses a CNN-based underwater target DOA estimation method under a ray sound field model, and the method comprises the steps: firstly calculating the time delay of an array element for receiving a sound signal through a ray sound field model, constructing a multipath signal data set received by an array in a shallow sea environment and a deep sea environment through Bellhop simulation, and calculating a covariance matrix of the multipath signal data set; building a CNN network based on shallow sea and deep sea ray sound field models, and training the CNN network by using the covariance matrix data set; and finally, realizing DOA estimation of the underwater target based on the trained CNN network. According to the method, a ray sound field model is applied, the method for calculating the delay inequality of the multipath sound source signals received by the uniform linear array is provided, the method is closer to an array signal processing scene in an actual marine environment, the algorithm robustness problem that the estimation performance of a traditional underwater DOA estimation method is sharply reduced along with reduction of the signal-to-noise ratio of the received signals is solved, and meanwhile the calculation efficiency is high; the method can be suitable for a scene in which real-time DOA estimation is carried out on an underwater target when an array receives multipath low-signal-to-noise-ratio signals in engineering.

Description

Technical Field

[0001] The invention belongs to the field of signal processing, and in particular relates to a CNN-based underwater target DOA estimation method under a ray acoustic field model. Background Art

[0002] Underwater target DOA estimation is a major function of sonar and plays a vital role in marine information research. Achieving high-precision and efficient underwater target DOA estimation in actual complex marine environments has always been an urgent problem to be solved in the field of underwater acoustics.

[0003] The underwater acoustic channel is a very complex environment composed of the seabed, sea surface and seawater. In the actual marine environment, the underwater acoustic channel has uncertainties in time, space, frequency and other aspects. In addition, the inherent measurement errors brought by the sensor array structure and the limitations of the DOA estimation algorithm lead to errors in most DOA estimation methods with good theoretical performance in the actual marine environment. One of the important factors is the influence of multipath effect. The multipath effect is that the signal emitted by the signal source reaches the receiving points of each element of the sensor array through different propagation paths. The energy and time of the signals from different paths to the receiving array are different, which will cause signal attenuation and waveform distortion; at the same time, the signals from different paths are related or even coherent, which will have a huge impact on the azimuth estimation of underwater targets; in addition, the current research on underwater target DOA estimation mainly focuses on the uniform shallow sea sound field environment, and the array signal modeling is based on the plane wave sound field model. However, in the actual marine environment, the sound speed distribution in shallow seas is also uneven, such as the sound speed distribution in shallow seas in autumn and summer; at the same time, unlike shallow seas, typical deep sea sound fields often have non-uniform sound speed distribution. Due to the influence of factors such as multipath effect and uneven sound speed distribution, there are errors in DOA estimation based on the plane wave acoustic field model. Therefore, the research on DOA estimation of underwater targets under the ray acoustic field model has important theoretical and practical significance.

[0004] Traditional underwater target DOA estimation algorithms and their improved algorithms, such as the MUSIC algorithm and the spatial smoothing MUSIC algorithm, have achieved certain results in engineering applications. However, as experts and scholars continue to study the algorithms in depth, they have found that there are two difficult problems to overcome in underwater target DOA estimation: one is that the algorithm has poor robustness. When the signal-to-noise ratio of the signal is low, the performance of most traditional DOA estimation algorithms will drop sharply; the other problem is that the algorithm has a large amount of computation. The traditional high-resolution DOA estimation algorithm represented by the MUSIC algorithm involves matrix operations such as eigenvalue decomposition of the array signal covariance matrix, which greatly increases the amount of computation. In recent years, machine learning algorithms represented by deep learning have developed extremely rapidly and have achieved many results in many fields such as image recognition, speech recognition, and semantic analysis. Compared with traditional DOA estimation methods, deep neural networks, as a powerful nonlinear function fitter, have obvious advantages in solving the problem of underwater target DOA estimation. Deep neural networks regard DOA estimation as a classification or regression problem, and use a large amount of data to implement function fitting, thereby realizing underwater target DOA estimation. Xiong Xiao et al. used the lower triangular part of the covariance matrix as input to train a single hidden layer neural network to achieve DOA estimation. They first applied the idea of ​​deep neural network to DOA estimation. This method opened the prelude to DOA estimation based on deep learning. Takeda et al. improved this method and used a multi-hidden layer deep neural network to achieve DOA estimation, which promoted the application of deep learning in DOA estimation research. Soumitro Chakrabarty and Emanuel A. P. Habets first introduced convolutional neural network to the DOA estimation problem. They regarded the microphone array DOA estimation problem as a classification problem, trained the convolutional network with the phase of the signal spectrum, and output the classification of the target direction. Rong Fan et al. proposed a deep hybrid neural network for coherent source DOA estimation. By multiplying the array received signal with the corresponding ideal array manifold, a high-dimensional preprocessed feature vector is obtained and directly input into the autoencoder. The output of the autoencoder is then used as the input of the cascaded deep residual neural network. The direction of arrival of the signal is determined by analyzing the position of the normalized spectrum peak. Zhong Dongqiang et al. proposed a robust DOA estimation method based on neural network: a multi-channel CNN+DNN network model, which uses the covariance matrix of the array received signal as input to improve the accuracy and robustness of DOA estimation.

[0005] From the above introduction, it can be seen that few people have studied the DOA estimation of targets in actual underwater multi-path coherent channel scenarios. For the strategic goal of "deepening the near-shallow sea and opening up the deep sea" clearly proposed in my country's medium- and long-term science and technology development plan, it is of practical significance to study the DOA estimation method of underwater targets that is closer to the actual marine environment. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a CNN-based underwater target DOA estimation method under a ray acoustic field model. The method has better robustness in a low signal-to-noise ratio marine environment and high computational efficiency. It can adapt to the scenario of real-time DOA estimation of underwater targets when an array receives multi-path low signal-to-noise ratio signals in engineering.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0008] S10. Calculate the time delay of the acoustic signal received by the array element using the ray acoustic field model, and construct the multipath signal data set DATA-s received by the array in the shallow sea and the multipath signal data set DATA-d received by the array in the deep sea environment through Bellhop simulation, and calculate their covariance matrices, and obtain the shallow sea covariance matrix data set DATA-Rs and the deep sea covariance matrix data set DATA-Rd correspondingly, and proceed to S20.

[0009] S20. Build a CNN network based on the shallow sea and deep sea ray acoustic field models, and proceed to S30.

[0010] S30, using the covariance matrix data sets DATA-Rs and DATA-Rd to train the CNN network, obtain a trained CNN network, and proceed to S40.

[0011] S40. Realize DOA estimation of underwater targets based on the trained CNN network.

[0012] Compared with the prior art, the present invention has the following beneficial effects: the present invention proposes a time delay difference algorithm for receiving multipath signals by a uniform linear array under a ray acoustic field model, which fully utilizes the characteristics of array receiving signals in shallow sea and deep sea environments under the ray acoustic field model, and solves the problem that the performance of traditional underwater target DOA estimation methods drops sharply as the signal-to-noise ratio decreases in a low signal-to-noise ratio environment. At the same time, the present invention has high efficiency and is suitable for scenarios in engineering where real-time DOA estimation of underwater targets is performed when an array receives multipath low signal-to-noise ratio signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The figure is a flow chart of the method of the present invention.

[0014] Figure 2 This is a diagram of the CNN network structure of the present invention under the shallow sea sound field model.

[0015] Figure 3 This is a diagram of the CNN network structure of the present invention under the deep sea acoustic field model.

[0016] Figure 4 1 is a performance diagram of the underwater target DOA estimation method in a shallow sea environment according to an embodiment of the present invention.

[0017] Figure 5 1 is a performance diagram of the underwater target DOA estimation method in a deep sea environment according to an embodiment of the present invention.

[0018] Figure 6 Classic CNN network diagram. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0020] At present, the traditional underwater target DOA estimation method based on the ray acoustic field model is mainly based on the conventional beamforming algorithm (CBF) and the MUSIC algorithm. The present invention proposes a CNN-based underwater target DOA estimation method under the ray acoustic field model, which utilizes the CNN network to extract the frequency domain characteristics of the array received signal. The DOA estimation of this method is highly robust under low signal-to-noise ratio conditions and has high computational efficiency. It is suitable for the scenario of real-time DOA estimation of underwater targets when the array receives multi-path low signal-to-noise ratio signals in engineering.

[0021] Combination Figure 1 The method for estimating DOA of underwater targets based on CNN under a ray acoustic field model of the present invention comprises the following steps:

[0022] S10, using the ray acoustic field model to calculate the delay of the acoustic signal received by the array element, and constructing the multipath signal data set DATA-s received by the array in the shallow sea and the multipath signal data set DATA-d received by the array in the deep sea environment through Bellhop simulation, and calculating their covariance matrices, and correspondingly obtaining the shallow sea covariance matrix data set DATA-Rs and the deep sea covariance matrix data set DATA-Rd, as follows:

[0023] S11. Consider the array signal model of K narrowband sound sources incident on a uniform linear array composed of M array elements. The first array element is taken as the reference array element. The transmitted signal of each sound source has L multipath propagation. k If the number of sound rays reaches the receiving array element, k = 1, 2, ..., K, then the received signal of the mth array element at time t is expressed as x m (t):

[0024]

[0025] Among them, b k,l represents the amplitude of the lth sound line of the kth sound source; s k (t) represents the transmission signal of the kth sound source, represents the time delay difference of the lth sound ray of the kth sound source reaching the mth array element relative to the reference array element, n m (t) represents the noise signal received by the mth array element.

[0026] make:

[0027] X=[x1(t),…,x M (t)] T ,N=[n1(t),…,n M (t)] T

[0028] S=[s1(t),…,s K (t)] T ,A=[A1(θ),…,A K (θ)]

[0029]

[0030] Where X represents the received signal x m (t), the array element receiving vector is composed of m=1,…,M, N represents the noise signal n n (t) composed of a noise vector, the serial number n = 1, ..., M, S represents the emission signal s of the sound source s (t), the sequence number s = 1, ..., K, A represents the signal vector composed of the steering vector group A a (θ) composed of a steering matrix, serial number a = 1, ..., K, θ q Indicates the angle between the incident direction of the sound source and the normal direction of the array, serial number q=1,…,L k , represents the time delay difference of the lth sound line of the kth signal source reaching the mth array element relative to the reference array element, the serial number m1=2,…,M, j represents the imaginary part, T represents the transpose, f0 represents the signal frequency, A k (θ) represents the steering vector corresponding to the kth sound source, A k (θ l ) represents the steering vector corresponding to the lth sound ray.

[0031] In the ray acoustic field model, the time delay difference between multipath sound source signals reaching adjacent array elements is Calculated as:

[0032]

[0033] Where ω is the angular frequency of the sound wave, r m is the horizontal distance from the sound source to the mth array element, z s Indicates the depth of the sound source, z r represents the depth of the receiving array, k l r,m represents the horizontal wave number of the lth sound ray reaching the mth array element, k l z,mRepresents the vertical wave number of the lth sound ray reaching the mth array element.

[0034] k l r,m for:

[0035] k l r,m = k(z s )cosα0 l,m

[0036] Among them, k(z s ) is the wave number at the depth of the sound source, α0 l,m is the exit angle of the lth sound ray reaching the mth array element.

[0037] k l z,m for:

[0038]

[0039] make:

[0040]

[0041] Among them, B is the diagonal array B divided by the sound line b The signal diagonal matrix composed of b=1,…,K,B k The sound line amplitude b k,l1 The sound line amplitude diagonal matrix composed of l1=l,…,L k .

[0042] Then the signal X received by the array can be expressed as:

[0043] X=A·B·S+N

[0044] According to the uniform sound velocity distribution and non-uniform sound velocity distribution in shallow water, the Bellhop acoustic field model is used to generate two noisy array receiving signal data sets DATA-s1 and DATA-s2, which constitute the array receiving signal data set DATA-ss in the shallow sea acoustic field environment; the generation parameters include the sea depth H, the sampling frequency f s , the number of snapshots is N s , the number of uniform linear array elements is M, the element spacing is d, and the depth of the receiving linear array is H m , the depth of the signal source is H s , the distance between the signal source and the reference element of the uniform linear array is r, and the number of sound lines is N s , the angle range is [θ l ,θ h ],θ l Represents the lower bound of the angle, θ h represents the upper limit of the angle, and the signal-to-noise ratio range is [SNR l,SNR h ], SNR l Represents the lower bound of the signal-to-noise ratio, SNR h represents the upper bound of the signal-to-noise ratio; the data set contains the multipath signal X received by the array x , the signal dimension is M d1 ×N d1 , calculate its autocorrelation covariance matrix R xx ,Right now:

[0045] R xx =E{X x ·X x H}

[0046] Here, H represents the conjugate transpose.

[0047] The covariance matrix is ​​a complex matrix R ij , each element of which can be expressed as ij and the imaginary part C ij It is expressed as:

[0048] R ij =R ij +jC ij

[0049] Multiple sets of data are uniformly taken within the range of angle and signal-to-noise ratio, and the R xx The shallow water multipath signal data set DATA-s is composed of;

[0050] S12, the array receiving signal model is the same as that described in S11. According to the deep-sea MUNK sound velocity distribution, the Bellhop sound field model is used to generate the noisy array receiving signal data set DATA-dd, and the signal dimension is M d ×N d The meaning of the generated parameters is the same as that of S11. The autocorrelation covariance matrix of the array received signal is calculated in the same way as S11. Multiple sets of data are uniformly taken within the angle and signal-to-noise ratio range set by S11 to obtain the autocorrelation covariance matrix R of the received multipath signal. xx The deep-sea multipath signal dataset DATA-d is composed of

[0051] Transfer to S20.

[0052] S20. Build a CNN network based on the shallow sea and deep sea ray acoustic field models, as follows:

[0053] S21, based on the classic CNN structure (such as Figure 6 ) Build a CNN network in shallow sea environment; the input of the network is the multipath signal data set DATA-s, and the real part R of the complex covariance matrix in the input data set is converted into s and the imaginary part C sExtract, where:

[0054]

[0055] Real part of the n1th complex covariance matrix and the imaginary part The component dimension is N d ×M d ×M d ×2 dual-channel two-dimensional matrix is ​​used as the network input; each two-dimensional matrix input is extracted through multiple layers of convolution-Dropout layers to extract the covariance matrix features, and the convolution layer dimension is C L ×C W ×N C , the feature map dimension obtained after convolution of each convolution layer is (M d -C L +1)×(M d -C L +1)×N C , the Dropout layer randomly discards the feature map with a data rate of p. After the Dropout layer, the amount of data becomes p times the original, and the dimension of the feature map remains unchanged; after the convolution-Dropout layer, the feature map is flattened by the Flatten layer. The dimension of the feature map before flattening is C Li ×C Wi ×N Ci , the dimension after flattening is 1×(C L ×C W ×N C ); Finally, the predicted value of the angle is obtained through two layers of fully connected layers. The fully connected layer uses linear connection, and the input-output relationship is expressed as:

[0056] N o =WN i +b

[0057] Among them, N i represents the fully connected layer input, W is the weight matrix, b is the bias matrix, N o is the output of the fully connected layer, and the final output is a scalar.

[0058] S22. Based on the classic CNN structure, a CNN network is built in a deep-sea environment. The input of the network is the multi-path signal dataset DATA-d. The real and imaginary parts of the complex covariance matrix in the input dataset are extracted. The extraction process is the same as S21. Each two-dimensional matrix input is subjected to multiple layers of convolution-Dropout layers to extract the covariance matrix features. After the convolution-Dropout layer, the feature map is flattened by the Flatten layer, and finally the scalar prediction value of the angle is obtained through two layers of fully connected layers. The dimensional change is the same as S21.

[0059] Transfer to S30.

[0060] S30, using the covariance matrix data sets DATA-Rs and DATA-Rd to train the CNN network, to obtain a trained CNN network, as follows:

[0061] S31, dimension N d ×M d ×M d Each row of the ×2 dual-channel two-dimensional matrix input is normalized using the maximum and minimum value normalization method, namely:

[0062]

[0063] Among them, min(x) and max(x) represent the minimum and maximum values ​​of each row of the feature sequence x respectively; Represents the normalized network input.

[0064] S32. Use the normalized dual-channel two-dimensional matrix as the input of network training, the angle predicted by the CNN network as the output of network training, and the actual orientation corresponding to the sample generation as the label of network training. Train the shallow sea and deep sea CNN networks respectively, and save the corresponding model parameters.

[0065] Transfer to S40.

[0066] S40, based on the trained CNN network, the DOA estimation of underwater targets is realized as follows:

[0067] S41, performing covariance calculation on the array received signal, and performing normalization preprocessing consistent with S31.

[0068] S42, using the preprocessed dual-channel two-dimensional matrix as the input of the trained CNN network, and using the model parameters saved in S32 to realize the DOA estimation of the underwater target.

[0069] Example 1

[0070] The shallow sea CNN network structure is as follows Figure 2 As shown in Figure 1, the input layer is a two-channel two-dimensional matrix composed of the real and imaginary parts of the covariance matrix, a composite three-layer convolution-Dropout layer, followed by a Flatten layer and two fully connected layers, and finally outputs the predicted value of the angle. The number of kernel function sizes, the number of size channels of each layer, and the Dropout rate are shown in Figure 1. Figure 2 As shown in the figure, the activation function of the three convolutional layers and the first fully connected layer is Relu, and its function expression is:

[0071]

[0072] Among them, x is the input variable of the Relu function. The activation function of the second fully connected layer is the Logistic function, and its function expression is:

[0073]

[0074] Among them, x is the input variable of the Logistic function.

[0075] The loss function of the network uses the mean squared error function (MSE), which is expressed as:

[0076]

[0077] Among them, N c is the number of samples, y i and They represent the label of the i-th sample and the predicted value of the model respectively.

[0078] The simulation signal parameters are set as follows: shallow sea acoustic field, sea depth of 103m, sampling frequency of 625Hz, number of snapshots of 1024, number of uniform linear array elements of 16, element spacing of half the wavelength corresponding to the signal center frequency, depth of the receiving linear array of 103m, depth of the signal source of 10m, distance between the signal source and the reference element of the uniform linear array of 20km, signal-to-noise ratio (SNR) of -20 to 20dB, direct wave angle of 60°, signal-to-noise ratio range of -20dB to 20dB, and 50 Monte Carlo experiments were performed on each group of signal-to-noise ratios.

[0079] Figure 4 The figure shows the variation of DOA estimation deviation and root mean square error with signal-to-noise ratio for the method of the present invention under non-uniform sound velocity distribution in shallow water. From the results of the embodiment, it can be seen that the recognition method can still have good DOA estimation performance when the signal-to-noise ratio is as low as -20dB.

[0080] Example 2

[0081] The deep sea CNN network structure is as follows Figure 3 As shown in , the input layer is a two-channel two-dimensional matrix composed of the real and imaginary parts of the covariance matrix, and a composite five-layer convolution-Dropout layer, where the Dropout rate of the first three layers is 0.1, and the Dropout rate of the last two layers is 0.5. It then passes through a Flatten layer and two fully connected layers, and finally outputs the predicted value of the angle. The number of kernel function sizes, the number of size channels of each layer, and the Dropout rate are shown in Figure 3As shown, the activation functions of the five convolutional layers and the first fully connected layer are Relu functions, the activation function of the second fully connected layer is Logistic function, and the loss function of the network uses the mean square error function, and its expressions are the same as those in Example 1.

[0082] The simulation signal parameters are set as follows: deep sea acoustic field, sea depth of 5000m, sampling frequency of 625Hz, number of snapshots of 1024, number of array elements of uniform linear array of 16, array element spacing of half the wavelength corresponding to the signal center frequency, depth of receiving linear array of 300m, depth of signal source of 300m, distance between signal source and reference element of uniform linear array of 20km, signal-to-noise ratio of -20 to 20dB random number, direct wave angle of 60°, signal-to-noise ratio range of -20dB to 20dB, 50 Monte Carlo experiments were performed on each group of signal-to-noise ratio.

[0083] Figure 5 The figure shows the variation of DOA estimation deviation and root mean square error with signal-to-noise ratio under deep-sea MUNK sound velocity distribution by the method of the present invention. It can be seen from the results of the embodiment that the recognition method has better DOA estimation performance in deep-sea environment than in shallow sea, and can still achieve effective DOA estimation when the signal-to-noise ratio is as low as -20dB.

Claims

1. A method for estimating underwater target DOA based on CNN under ray acoustic field model, characterized in that: The following steps are involved: S10, using the ray acoustic field model to calculate the time delay of the acoustic signal received by the array element, and constructing the multipath signal data set DATA-s received by the array in the shallow sea and the multipath signal data set DATA-d received by the array in the deep sea environment through Bellhop simulation, and calculating their covariance matrices, and correspondingly obtaining the shallow sea covariance matrix data set DATA-Rs and the deep sea covariance matrix data set DATA-Rd, and then proceeding to S20; S20, build a CNN network based on the shallow sea and deep sea ray acoustic field models, and then proceed to S30; S30, using the covariance matrix data sets DATA-Rs and DATA-Rd to train the CNN network, obtaining a trained CNN network, and proceeding to S40; S40. Realize DOA estimation of underwater targets based on the trained CNN network.

2. The underwater target DOA estimation method based on CNN under a ray acoustic field model according to claim 1 is characterized in that: S10 includes the following steps: S11. Consider the array signal model of K narrowband sound sources incident on a uniform linear array composed of M array elements. The first array element is taken as the reference array element. The transmitted signal of each sound source has L multipath propagation. k If the number of sound rays reaches the receiving array element, k = 1, 2, ..., K, then the received signal of the mth array element at time t is expressed as x m (t): Among them, b k,l represents the amplitude of the lth sound line of the kth sound source; s k (t) represents the transmission signal of the kth sound source, represents the time delay difference of the lth sound ray of the kth sound source reaching the mth array element relative to the reference array element, n m (t) represents the noise signal received by the mth array element; let: X=[x1(t),…,x M (t)] T ,N=[n1(t),…,n M (t)] T S=[s1(t),…,s K (t)] T ,A=[A1(θ),…,A K (θ)] Where X represents the received signal x m (t), the array element receiving vector is composed of m=1,…,M, N represents the noise signal n n (t) composed of a noise vector, the serial number n = 1, ..., M, S represents the emission signal s of the sound source s (t), the sequence number s = 1, ..., K, A represents the signal vector composed of the steering vector group A a (θ) composed of a steering matrix, serial number a = 1, ..., K, θ q Indicates the angle between the incident direction of the sound source and the normal direction of the array, serial number q=1,…,L k , represents the time delay difference of the lth sound line of the kth signal source reaching the mth array element relative to the reference array element, the serial number m1=2,…,M, j represents the imaginary part, T represents the transpose, f0 represents the signal frequency, A k (θ) represents the steering vector corresponding to the kth sound source, A k (θ l ) represents the steering vector corresponding to the lth sound ray; In the ray acoustic field model, the time delay difference between multipath sound source signals reaching adjacent array elements is Calculated as: Where ω is the angular frequency of the sound wave, r m is the horizontal distance from the sound source to the mth array element, z s Indicates the depth of the sound source, z r represents the depth of the receiving array, k l r,m represents the horizontal wave number of the lth sound ray reaching the mth array element, k l z,m represents the vertical wave number of the lth sound ray reaching the mth array element; k l r,m for: k l r,m =k(z s )cosα0 l,m Among them, k(z s ) is the wave number at the depth of the sound source, α0 l,m is the exit angle of the lth sound ray reaching the mth array element; k l z,m for: make: Among them, B is the diagonal array B divided by the sound line b The signal diagonal matrix composed of b=1,…,K,B k The sound line amplitude b k,l1 The sound line amplitude diagonal matrix composed of l1=l,…,L k ; Then the signal X received by the array can be expressed as: X=A·B·S+N According to the uniform sound velocity distribution and non-uniform sound velocity distribution in shallow water, the Bellhop acoustic field model is used to generate two noisy array receiving signal data sets DATA-s1 and DATA-s2, which constitute the array receiving signal data set DATA-ss in the shallow sea acoustic field environment; the generation parameters include the sea depth H, the sampling frequency f s , the number of snapshots is N s , the number of uniform linear array elements is M, the element spacing is d, and the depth of the receiving linear array is H m , the depth of the signal source is H s , the distance between the signal source and the reference element of the uniform linear array is r, and the number of sound lines is N s , the angle range is [θ l ,θ h ],θ l Represents the lower bound of the angle, θ h represents the upper limit of the angle, and the signal-to-noise ratio range is [SNR l ,SNR h ], SNR l Represents the lower bound of the signal-to-noise ratio, SNR h represents the upper bound of the signal-to-noise ratio; the data set contains the multipath signal X received by the array x , the signal dimension is M d1 ×N d1 , calculate its autocorrelation covariance matrix R xx ,Right now: R xx =E{X x ·X x H } Where H represents conjugate transpose; The covariance matrix is ​​a complex matrix R ij , each element of which can be expressed as ij and the imaginary part C ij It is expressed as: R ij =R ij +jC ij Multiple sets of data are uniformly taken within the range of angle and signal-to-noise ratio, and the R xx The shallow sea multipath signal data set DATA-s is composed of; S12, the array receiving signal model is the same as that described in S11. According to the deep-sea MUNK sound velocity distribution, the Bellhop sound field model is used to generate the noisy array receiving signal data set DATA-dd, and the signal dimension is M d ×N d The meaning of the generated parameters is the same as that of S11. The autocorrelation covariance matrix of the array received signal is calculated in the same way as S11. Multiple sets of data are uniformly taken within the angle and signal-to-noise ratio range set by S11 to obtain the autocorrelation covariance matrix R of the received multipath signal. xx The deep-sea multipath signal dataset DATA-d is composed of 3. The underwater target DOA estimation method based on CNN under a ray acoustic field model according to claim 2 is characterized in that: S20 includes the following steps: S21. Build a CNN network in shallow sea environment based on the classic CNN structure. The input of the network is the multipath signal data set DATA-s. The real part R of the complex covariance matrix in the input data set is converted into s and the imaginary part C s Extract, where: Real part of the n1th complex covariance matrix and the imaginary part The component dimension is N d ×M d ×M d ×2 dual-channel two-dimensional matrix is ​​used as the network input; each two-dimensional matrix input is extracted through multiple layers of convolution-Dropout layers to extract the covariance matrix features, and the convolution layer dimension is C L ×C W ×N C , the feature map dimension obtained after convolution of each convolution layer is (M d -C L +1)×(M d -C L +1)×N C , the Dropout layer randomly discards the feature map with a data rate of p. After the Dropout layer, the amount of data becomes p times the original, and the dimension of the feature map remains unchanged; after the convolution-Dropout layer, the feature map is flattened by the Flatten layer. The dimension of the feature map before flattening is C Li ×C W i×N Ci , the dimension after flattening is 1×(C L ×C W ×N C ); Finally, the predicted value of the angle is obtained through two layers of fully connected layers. The fully connected layer uses linear connection, and the input-output relationship is expressed as: N o =WN i +b Among them, N i represents the fully connected layer input, W is the weight matrix, b is the bias matrix, N o is the output of the fully connected layer, and the final output is a scalar; S22. Based on the classic CNN structure, a CNN network is built in a deep-sea environment. The input of the network is the multi-path signal dataset DATA-d. The real and imaginary parts of the complex covariance matrix in the input dataset are extracted. The extraction process is the same as S21. Each two-dimensional matrix input is subjected to multiple layers of convolution-Dropout layers to extract the covariance matrix features. After the convolution-Dropout layer, the feature map is flattened by the Flatten layer, and finally the scalar prediction value of the angle is obtained through two layers of fully connected layers. The dimensional change is the same as S21.

4. The underwater target DOA estimation method based on CNN under a ray acoustic field model according to claim 3 is characterized in that: S30 includes the following steps: S31, dimension N d ×M d ×M d Each row of the ×2 dual-channel two-dimensional matrix input is normalized using the maximum and minimum value normalization method, namely: Among them, min(x) and max(x) represent the minimum and maximum values ​​of each row of the feature sequence x respectively; Represents the normalized network input; S32. Use the normalized dual-channel two-dimensional matrix as the input of network training, the angle predicted by the CNN network as the output of network training, and the actual orientation corresponding to the sample generation as the label of network training. Train the shallow sea and deep sea CNN networks respectively, and save the corresponding model parameters.

5. The underwater target DOA estimation method based on CNN under the ray acoustic field model according to claim 4 is characterized in that: S40 includes the following steps: S41, performing covariance calculation on the array received signal, and performing normalization preprocessing consistent with S31; S42, using the preprocessed dual-channel two-dimensional matrix as the input of the trained CNN network, and using the model parameters saved in S32 to realize the DOA estimation of the underwater target.

Images