Acoustic signal estimation method of high subsonic flying target based on acoustic vector uniform linear array

Through the processing of acoustic vector uniform linear array and neural network, the problem of difficult tracking of the acoustic signals of high subsonic flight targets is solved, and efficient acoustic detection and real-time estimation are achieved.

CN114200392BActive Publication Date: 2024-11-19CHINESE PEOPLES LIBERATION ARMY UNIT 91550
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111409850.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-24
Publication Date
2024-11-19
Estimated Expiration
2041-11-24

AI Technical Summary

Technical Problem

The prior art is difficult to effectively track the acoustic signals of high subsonic flight targets, and because they exhibit broadband, transient and non-stationary characteristics, there are few researches on acoustic tracking methods.

Method used

Using a method based on acoustic vector uniform linear array, acoustic vector sensor array is formed by spatially uniformly distributed multiple acoustic vector sensors, discrete Fourier transform and principal component analysis neural network processing is performed, and high subsonic flight is estimated in combination with BP neural network. The target sound signal direction.

Benefits of technology

It effectively reduces the system calculation complexity, improves the real-time nature of high-sonic flight targets, expands the application range of DOA estimation method based on artificial neural networks, and realizes acoustic detection of high-sonic flight targets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114200392B_ABST
    Figure CN114200392B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of acoustic measurement, and in particular to a method for estimating acoustic signals of high-subsonic flying targets based on a uniform linear array of acoustic vectors. The present invention constructs a time-domain broadband signal expression of a uniform linear spatial acoustic cross acoustic vector array, and adopts a time-domain broadband DOA fast estimation method based on PCA‑BP, thereby effectively reducing the computational complexity of the system and improving the real-time performance of high-subsonic flying target estimation. The present invention effectively expands the scope of application of the DOA estimation method based on artificial neural networks by constructing a broadband receiving model in the time domain, and provides a technical basis for effectively realizing acoustic detection of high-subsonic flying targets.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of acoustic measurement, and in particular to a high subsonic flying target acoustic signal estimation method based on an acoustic vector uniform linear array. Background Art

[0002] At present, due to the height restriction of the sea surface platform, the over-the-horizon detection of low-altitude flying targets on the sea surface is a difficult point of conventional radar detection technology. At the same time, the complex electromagnetic environment near the sea surface platform will lead to increased electromagnetic background noise. The sea surface platform radar is in a low elevation mode when detecting the sea, which will be subject to unfavorable factors such as severe sea clutter and multipath effect. The above unfavorable factors make it difficult to detect short-range low-altitude flying targets on the sea surface.

[0003] In recent years, the detection and positioning technology based on target acoustic signals has received extensive attention as a supplementary solution to radar detection technology. Acoustic detection technology determines the target position by passively receiving the sound waves emitted by the target. It has the advantages of strong concealment, not susceptible to electromagnetic interference, and the ability to detect low-altitude targets. Especially when applied to sea platforms, the background and environmental noise are relatively simple and there is less clutter interference. Therefore, this method has unique advantages in detecting low-altitude flying targets on the sea surface.

[0004] Based on the speed of sound, low-altitude flying targets can be divided into two categories: supersonic targets and high subsonic targets. When flying, supersonic targets will disturb the surrounding air and form Mach waves. The wavefront of Mach waves is also called shock waves. Shock waves propagate in a cone and have obvious time domain characteristics. Therefore, there are relatively mature theories and technologies for acoustic tracking of supersonic targets. For the acoustic signals of high subsonic flying targets, they show broadband, transient and non-stationary characteristics, and there are relatively few studies on acoustic tracking methods for such targets. Summary of the invention

[0005] In order to solve the problem that the acoustic signals of high-subsonic flying targets present broadband, transient and non-stationary characteristics, making them difficult to be acoustically tracked, the present invention discloses a method for estimating the acoustic signals of high-subsonic flying targets based on an acoustic vector uniform linear array, which is implemented by using an acoustic vector sensor array composed of multiple acoustic vector sensors uniformly distributed in space.

[0006] The technical solution of the present invention:

[0007] The acoustic signal estimation method of a high subsonic flying target based on an acoustic vector uniform linear array has the following specific steps:

[0008] S1: Performing discrete Fourier transform (FFT) on the received signal of the acoustic vector sensor array, segmenting the frequency domain signal to obtain multiple frequency domain signals, and calculating the covariance matrix of each frequency domain signal to obtain multiple covariance matrices;

[0009] S2: Each covariance matrix obtained in step S1 is input into the principal component analysis neural network, i.e., the PCA neural network. The PCA neural network performs eigenvalue decomposition on its input matrix to obtain the eigenvector A (f j ), j = 0, 1, 2, ..., J;

[0010] S3: Using the eigenvectors of each covariance matrix, construct the focusing matrix of each frequency domain signal respectively, and then use each focusing matrix to focus the corresponding frequency domain signal respectively, calculate the total covariance matrix of all frequency domain signals, and use the PCA neural network to perform matrix eigenvalue decomposition on the total covariance matrix;

[0011] S4: Using the matrix eigenvalues ​​obtained by PCA neural network processing, a training data set is constructed to train the BP neural network;

[0012] S5: If the training error of the BP neural network is less than the set error threshold, the training is considered to be completed and the trained BP neural network is obtained. Otherwise, the next round of training is performed until the training error of the BP neural network is less than the set error threshold.

[0013] S6: For the signal collected by the acoustic vector sensor, calculate the total covariance matrix of all frequency domain signals, input the total covariance matrix into the PCA neural network to obtain its eigenvalues ​​and eigenvectors, construct a matrix with the eigenvectors corresponding to the eigenvalues ​​of the total covariance matrix that are greater than the threshold, and input the matrix into the trained BP neural network. At this time, the output of the BP neural network is the estimated value of the direction of arrival DOA of the high subsonic flying target.

[0014] The step S1 specifically includes: using an acoustic vector sensor array that is uniformly linearly distributed in space and located on a spatial cross array to achieve the step, establishing a three-dimensional rectangular coordinate system with the array baseline of the spatial cross array, the reference sensor in the acoustic vector sensor array is located at the coordinate origin of the three-dimensional rectangular coordinate system, the acoustic vector sensors are uniformly distributed on the positive and negative semi-axes of the x-axis, y-axis and z-axis, and the distance between adjacent sensors, that is, the baseline length is d.

[0015] For K far-field sound sources, The incident direction enters the acoustic vector sensor array, k = 1, 2, ..., K, and the position coordinates of the mth acoustic vector sensor are (x m ,y m ,z m ), then at time t, after K sound sources are incident on the mth sound vector sensor, the output x of the sound vector sensor is (m) (t) is expressed as:

[0016]

[0017] In the formula, a k is the direction vector of the incident signal from the kth sound source, a k =[u k ,v k ,w k ], q m (θ k )=exp(-j2πfτ km ), τ km represents the time delay between the kth incident signal to the mth acoustic vector sensor and to the reference sensor, τ km =(x m u k +y m v k +z m w k ) / c, c is the speed of sound, s k (t) is the sound pressure signal of the incident signal of the kth sound source, n (m) (t) is the measurement noise vector of the mth acoustic vector sensor.

[0018] For K far-field sound sources, the received signal vector x(t) of the acoustic vector sensor array is expressed as:

[0019]

[0020] In the formula, is the acoustic vector array steering vector, q(θ k )=[q1(θ k ),…,q M (θ k )] T ,symbol represents the Kronecker product operation, M is the number of acoustic vector sensors. After discrete sampling of the signal collected by the acoustic vector sensor, L snapshot discrete signals are obtained, then the corresponding acoustic vector sensor array receives discrete signal matrix X of M×K dimensions, and n(t) is the measurement noise vector of the acoustic vector sensor array.

[0021] Perform FFT on the signal collected by the acoustic vector sensor and divide the obtained frequency domain signal into J segments. j ), whose covariance matrix R X (f j ) is calculated as:

[0022] R X (f j )=E[X(f j )X H (f j )],j=0,1,2,…,J

[0023] In the step S2, the PCA neural network is used to realize matrix eigenvalue decomposition, which specifically includes: the PCA neural network is realized by using a single-layer forward propagation neural network with unsupervised learning, and its network weight vector W satisfies WW H =I, the discrete signal matrix X is input into the PCA neural network, and the output of the PCA neural network is Y=WX. After the kth iteration, the cost function of the PCA neural network is:

[0024] L(W)=(XW H (k)Y) H (XW H (k)Y)

[0025] Among them, the fastest gradient descent algorithm is used to solve the optimal weight vector W of the neural network through multiple iterations. opt , when the difference between the network weight vectors of adjacent iterations is less than the threshold ε, that is, ||W(k)-W(k-1)|| 2 ≤ε, judging that the PCA neural network converges, thus obtaining the optimal weight vector W opt , which is the eigenvector of the input matrix X.

[0026] The step S3 specifically includes: the covariance matrix of the j-th frequency domain signal is R X (f j ), through R X (f j ) to perform eigenvalue decomposition and obtain the focusing matrix, where j = 0, 1, 2, ..., J, and the covariance matrix R of each frequency domain signal X (f j ) is input into the PCA neural network for eigenvalue decomposition, and the characteristic matrix U composed of its typical eigenvalues ​​is obtained. S (f j ), for the covariance matrix R at the reference frequency point f0 X (f0), whose characteristic matrix is ​​denoted by V S (f0), so the focusing matrix of the j-th frequency domain signal is obtained as:

[0027] T(f j )=U s (f j )V s H (f0),

[0028] The total covariance matrix of all frequency domain signals after focusing is:

[0029]

[0030] Step S4, using the eigenvectors corresponding to the eigenvalues ​​greater than the threshold of the total covariance matrix of all frequency domain signals to form a matrix U, constructing a feature training matrix and the target DOA parameter vector in The dataset As a training data set, the number of neurons in each layer of the BP neural network is determined, and the connection weights between layers and the thresholds of each layer are initialized.

[0031] In step S4, the BP neural network is specifically as follows: the number of neurons in the input layer, hidden layer and output layer of the BP neural network are I, H and J respectively. The activation function of the hidden layer is the sigmoid function, and the activation function of the output layer is the purelin function. The training data set is input into the BP neural network, and the output of the hth hidden layer neuron b=[b1,b2,…,b H ], where b h =[b1,b2,…,b H ], represents the nth output of the hth hidden layer neuron, Among them, α h represents the input of the hth hidden layer neuron, w ih represents the weight value between the hth hidden layer neuron and the ith input layer neuron, represents the threshold of the hth neuron in the hidden layer, and f1 represents the sigmoid function. The output result of the output layer neuron is represents the output of the jth neuron in the output layer, The formula for calculating the output of the jth neuron in the output layer is: Among them, β j is the input of the jth output layer neuron, v hj is the weight value between the hth hidden layer neuron and the jth output layer neuron, χ j represents the threshold of the jth neuron in the output layer, and f2 represents the purelin function.

[0032] The step S5 is as follows: Calculate the output of the output neuron of the BP neural network The mean square error between the expected output y and the expected output y is calculated, and the sum of the mean square errors of all samples is calculated, that is, the training error. The weights and thresholds of each layer are updated using the fastest gradient descent algorithm. When the training error is less than the set error threshold, the training is considered complete, otherwise the next round of training will be continued.

[0033] The beneficial effects of the present invention are:

[0034] The present invention constructs a time-domain broadband signal expression of a uniform linear spatial acoustic cross vector array and adopts a time-domain broadband DOA fast estimation method based on PCA-BP, which effectively reduces the computational complexity of the system and improves the real-time performance of high subsonic flight target estimation. The present invention constructs a broadband receiving model in the time domain, effectively expands the application scope of the DOA estimation method based on artificial neural networks, and provides a technical basis for effectively realizing acoustic detection of high subsonic flight targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a schematic diagram of an acoustic vector sensor array of the present invention;

[0036] Figure 2 A single-layer feedforward neural network model diagram for unsupervised learning of the present invention;

[0037] Figure 3 It is a schematic diagram of the topological structure of the BP neural network of the present invention;

[0038] Figure 4 The present invention is a flowchart for implementing the method of the present invention. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0040] The present invention discloses a high-subsonic flying target acoustic signal estimation method based on an acoustic vector uniform linear array. The method is implemented by adopting an acoustic vector sensor array which is uniformly linearly distributed in space and is located on a spatial cross array. A three-dimensional rectangular coordinate system is established with the array baseline of the spatial cross array. The reference sensor in the acoustic vector sensor array is located at the coordinate origin of the three-dimensional rectangular coordinate system. The acoustic vector sensors are uniformly distributed on the positive and negative semi-axes of the x-axis, y-axis and z-axis. The distance between adjacent sensors, that is, the baseline length is d.

[0041] For K far-field sound sources, The incident direction enters Figure 1 The acoustic vector sensor array shown in the figure has k = 1, 2, ..., K. The propagation medium is isotropic. The sensor at the origin of the coordinate is taken as the reference. The position coordinate of the mth sensor is (x m ,y m ,z m ), then at time t, K sound sources are incident on the output x of the mth sound vector sensor (m) (t) can be expressed as:

[0042]

[0043] In the formula, a k is the direction vector of the kth incident signal, ak =[u k ,v k ,w k ], q m (θ k )=exp(-j2πfτ km ), τ km represents the time delay between the kth incident signal to the mth array element and to the reference array element, τ km =(x m u k +y m v k +z m w k ) / c,s k (t) is the sound pressure signal of the kth incident signal, n (m) (t) is the measurement noise vector of the mth acoustic vector sensor.

[0044] According to the output expression of the acoustic vector sensor mentioned above, the received signal vector of the uniform linear space cross array based on the acoustic vector sensor can be obtained as follows:

[0045]

[0046] In the formula, is the acoustic vector array steering vector, q(θ k )=[q1(θ k ),…,q M (θ k )] T ,symbol represents the Kronecker product operation, and M is the number of acoustic vector sensors. After discrete sampling of the signal collected by the acoustic vector sensor, L snapshot discrete signals are obtained, and the corresponding discrete signal matrix X received by the acoustic vector sensor array is M×K dimensional.

[0047] The principal component analysis method is to represent high-dimensional data with several unrelated low-dimensional vectors, so as to achieve the purpose of reducing the dimension of the high-dimensional data vector. The unrelated low-dimensional vectors are called the principal components of the original high-dimensional data. After PCA dimensionality reduction, the original data vector reduces the number of characteristic variables and removes the redundant information and noise in the original characteristic variables. For an N-dimensional vector X, assuming that its mean is 0, the projection of vector X on the unit vector v is:

[0048] a=X T v=v T X,

[0049] The mean of variable a is 0 and the variance is:

[0050] σ 2 =E[a 2]=E[(v T X)(X T v)]=v T E[XX T ]v=v T R x v,

[0051] Among them, R x Represents the covariance matrix of vector X. For a unit vector v, its variance function The basic process of PCA method is to find Get the maximum value of v, at this time v is the matrix R x The characteristic vector of x v = λv, λ is the eigenvalue of the matrix. Get the unit vector of the first N maximum values, which is the matrix R x The first N eigenvectors of , and the corresponding eigenvalues ​​are the first N largest eigenvalues, thus completing the process of eigenvalue decomposition of the matrix.

[0052] Assuming that the number of acoustic vector sensors is M, the broadband incident signal is divided into J segments for FFT. j ), whose covariance matrix R X (f j ) is calculated as:

[0053] R X (f j )=E[X(f j )X H (f j )]=A(f j )R s (f j )A H (f j )+R N (f j ),

[0054] Among them, A(f j ) represents the focusing matrix of the jth frequency domain signal, R N (f j ) represents the covariance matrix of the noise in the jth frequency domain signal, R s (f j ) represents the cross-correlation matrix of the j-th frequency domain signal.

[0055] The corresponding data covariance matrix R at the reference frequency point f0 X (f0) is:

[0056] R X (f0) = E[X(f0)X H(f0)]=A(f0)R s (f0)A H (f0)+R N (f0),

[0057] For DOA estimation of broadband signals, a focusing matrix is ​​needed to focus all frequency signals on f0. X (f j ) to perform eigenvalue decomposition to obtain a focusing matrix, where j = 0, 1, 2, ..., J. The eigenvalue decomposition process is implemented using the PCA neural network. Since the data after FFT transformation is complex data, and the general neural network model only processes real data, the cross-correlation complex matrix is ​​transformed here. For the M×M dimensional cross-correlation complex matrix R X , construct a 2M×2M dimensional real matrix The real number matrix is ​​input into the PCA neural network.

[0058] Substitute the cross-correlation matrix X into the PCA neural network, and the network weight W satisfies WW H =I, the output of the PCA neural network is Y = WX, and the output is used to reconstruct the input and obtain Therefore, the cost function of constructing the PCA network is:

[0059] L(W)=||e|| 2 =(XW H Y) H (XW H Y)

[0060] =X H X-2X H W H WX+X H W H WW H WX,

[0061] By taking the derivative of the cost function, we can get the negative gradient of L(W):

[0062]

[0063] Thanks WW H →I, the above formula can be written as:

[0064]

[0065] Use the fastest gradient descent algorithm to solve the optimal weight vector W opt , during the solution process, the weight vector is updated along the negative gradient direction, and its update formula is:

[0066]

[0067] Using the GHA algorithm, only the matrix W is used H The network weights can be obtained by the lower triangular part of (k)Y(k), and thus the complex weight iteration formula of the PCA neural network can be obtained as follows:

[0068]

[0069] When the difference between the network weight vectors of adjacent iterations is less than the threshold, that is, ||w(k)-w(k-1)|| 2 ≤ε, the PCA neural network is judged to converge, so as to obtain the optimal weight vector, which is the eigenvector of the cross-correlation matrix X. For the learning efficiency η, its value is generally set to satisfy 0<η<1 / λ1, where λ1 is the maximum eigenvalue of the cross-correlation matrix X. For the PCA neural network, an unsupervised learning single-layer forward propagation neural network is used to implement it. Its basic structure is as follows Figure 2 shown.

[0070] like Figure 2 As shown, for M input components, the network can be used to extract P principal components. X (f j ) is input into the PCA neural network for eigenvalue decomposition, and the characteristic matrix U composed of its typical eigenvalues ​​is obtained. S (f j ), for the data covariance matrix R at the reference frequency point f0 X (f0), whose characteristic matrix is ​​denoted by V S (f0), so the focusing matrix is:

[0071] T(f j )=U s (f j )V s H (f0)

[0072] The covariance matrix of the received data after focusing is:

[0073]

[0074] After getting the received data covariance matrix R y After that, continue to input it into the PCA neural network to obtain R y The complex weights of the PCA neural network are the eigenvectors of the cross-correlation matrix of the signal collected by the acoustic vector sensor.

[0075] In the current PCA-based DOA estimation method, after the signal subspace is obtained using the PCA method, the MUSIC method is still needed to construct the spatial spectrum, and then find the maximum point of the spatial spectrum as the result of the spectrum estimation. For the two-dimensional DOA estimation problem, it is necessary to conduct a combined search of the 360° azimuth angle range and the 90° elevation angle range. At the same time, in order to obtain high-precision search accuracy, the search step size needs to be small enough. If the above azimuth and elevation angle ranges are searched with a step size of 0.1°, approximately 106 spatial spectrum calculations are required, and the computational complexity of each spatial spectrum calculation is approximately O(M 2 +M), so the computational complexity of the spatial spectrum search process is also huge. In order to avoid this process, the present invention proposes a PCA-BP algorithm, after obtaining the signal subspace using the PCA method, further using the BP neural network to estimate the two-dimensional DOA.

[0076] BP neural network is a multi-layer feedforward neural network trained according to the error back propagation algorithm. The topological structure of BP neural network is as follows: Figure 3 As shown. The learning process of BP neural network is mainly divided into two processes: signal forward propagation and error back propagation. In the process of signal forward propagation, the input signal goes from the input layer through the hidden layer to the output layer, and the result of the output layer is judged. If the result is not the expected output value, the error back propagation process is entered. In the process of error back propagation, from the output layer through the hidden layer to the input layer, the weights and biases from the hidden layer to the output layer, and the weights and biases from the input layer to the hidden layer are adjusted in turn.

[0077] exist Figure 3 middle, represents the i-th feature of the n-th training sample, that is, the i-th component in the n-th snapshot data vector, w ih is the weight value between the i-th input layer neuron and the h-th hidden layer neuron, is the output of the hth hidden layer neuron corresponding to the nth training sample, v hj is the weight value between the hth hidden layer neuron and the jth output layer neuron, is the output value of the jth output layer neuron corresponding to the nth training sample.

[0078] Using the received data covariance matrix R y The eigenvectors corresponding to the larger eigenvalues ​​of form the matrix U, and construct the feature training matrix and the target DOA parameter vector in The dataset As the training data set, determine the number of neurons in each layer of the BP neural network, and initialize the connection weights between each layer and the threshold of each layer. The number of neurons contained in the input layer, hidden layer and output layer are I, H and J respectively. The activation function of the hidden layer is the sigmoid function, and the activation function of the output layer is the purelin function. Input the training data set into the BP neural network, and the output of the hth hidden layer neuron is b = [b1, b2, ..., b H ], where b h =[b1,b2,…,b H ], represents the nth output of the hth hidden layer neuron, Among them, α h represents the input of the hth hidden layer neuron, w ih represents the weight value between the hth hidden layer neuron and the ith input layer neuron, represents the threshold of the hth neuron in the hidden layer, and f1 represents the sigmoid function. The output result of the output layer neuron is represents the output of the jth neuron in the output layer, The formula for calculating the output of the jth neuron in the output layer is: Among them, β j is the input of the jth output layer neuron, v hj is the weight value between the hth hidden layer neuron and the jth output layer neuron, χ j represents the threshold of the jth neuron in the output layer, and f2 represents the purelin function. Calculate the output of the output neuron of the neural network The mean square error between the expected output y and the mean square error of all samples is calculated, which is the training error. The weights and thresholds of each layer are updated using the fastest gradient descent algorithm. When the training error is less than the set error, the training is considered complete, otherwise the next round of training will continue. The signal collected by the acoustic vector sensor is processed by PCA and input into the trained BP neural network. At this time, the output of the network is the estimated value of DOA.

[0079] According to the above description, Figure 4 The implementation flow chart of the PCA-BP algorithm for two-dimensional DOA estimation is given.

[0080] like Figure 4 As shown in Figure 1, the basic steps of the PCA-BP algorithm to estimate the two-dimensional DOA of a high subsonic flying target are:

[0081] 1) Perform FFT on the received signal of the acoustic vector sensor array, segment the frequency domain signal, and calculate the covariance matrix of each frequency domain signal;

[0082] 2) Input each covariance matrix into the PCA neural network to obtain its eigenvector respectively;

[0083] 3) Using the eigenvectors of each covariance matrix, the focusing matrix of each frequency band is constructed respectively, the frequency domain signal is focused, and then the total covariance matrix of the frequency domain data is calculated and PCA is performed on it;

[0084] 4) Using the feature matrix obtained by PCA processing, a training data set is constructed to train the BP neural network;

[0085] 5) If the training error of the BP neural network is less than the set error threshold, the training is considered completed, otherwise the next round of training is carried out;

[0086] 6) Using the trained BP neural network, the acoustic vector sensor collected signal after PCA processing is processed to obtain the two-dimensional DOA estimation value of the target.

[0087] The above description is only an embodiment of the present invention and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A high subsonic flying target acoustic signal estimation method based on an acoustic vector uniform linear array, characterized in that: The specific steps are as follows: S1: Perform discrete Fourier transform FFT on the received signal of the acoustic vector sensor array, segment the frequency domain signal to obtain multiple frequency domain signals, and calculate the covariance matrix of each frequency domain signal to obtain multiple covariance matrices; S2: Each covariance matrix obtained in step S1 is input into the principal component analysis neural network, i.e., the PCA neural network. The PCA neural network performs eigenvalue decomposition on its input matrix to obtain the eigenvector A (f j ), j = 0, 1, 2, ..., J; S3: Using the eigenvectors of each covariance matrix, construct the focusing matrix of each frequency domain signal respectively, and then use each focusing matrix to focus the corresponding frequency domain signal respectively, calculate the total covariance matrix of all frequency domain signals, and use the PCA neural network to perform matrix eigenvalue decomposition on the total covariance matrix; S4: Using the matrix eigenvalues ​​obtained by PCA neural network processing, a training data set is constructed to train the BP neural network; S5: If the training error of the BP neural network is less than the set error threshold, the training is considered to be completed and the trained BP neural network is obtained. Otherwise, the next round of training is performed until the training error of the BP neural network is less than the set error threshold. S6: For the signal collected by the acoustic vector sensor, calculate the total covariance matrix of all frequency domain signals, input the total covariance matrix into the PCA neural network to obtain its eigenvalues ​​and eigenvectors, construct a matrix with the eigenvectors corresponding to the eigenvalues ​​of the total covariance matrix that are greater than the threshold, and input the matrix into the trained BP neural network. At this time, the output of the BP neural network is the estimated value of the direction of arrival DOA of the high subsonic flying target.

2. The method for estimating acoustic signals of high subsonic flying targets based on acoustic vector uniform linear array according to claim 1, characterized in that: The step S1 specifically includes: using an acoustic vector sensor array that is uniformly linearly distributed in space and located on a spatial cross array to achieve the goal, establishing a three-dimensional rectangular coordinate system with the array baseline of the spatial cross array, wherein the reference sensor in the acoustic vector sensor array is located at the origin of the coordinate system of the three-dimensional rectangular coordinate system, and the acoustic vector sensors are uniformly distributed on the positive and negative semi-axes of the x-axis, y-axis, and z-axis, and the distance between adjacent sensors, i.e., the baseline length, is d; For K far-field sound sources, The incident direction enters the acoustic vector sensor array, k = 1, 2, ..., K, and the position coordinates of the mth acoustic vector sensor are (x m ,y m ,z m ), then at time t, after K sound sources are incident on the mth sound vector sensor, the output x of the sound vector sensor is (m) (t) is expressed as: In the formula, a k is the direction vector of the incident signal from the kth sound source, a k =[u k ,v k ,w k ], q m (θ k )=exp(-j2πfτ km ), τ km represents the time delay between the kth incident signal to the mth acoustic vector sensor and to the reference sensor, τ km =(x m u k +y m v k +z m w k ) / c, c is the speed of sound, s k (t) is the sound pressure signal of the incident signal of the kth sound source, n (m) (t) is the measurement noise vector of the mth acoustic vector sensor; For K far-field sound sources, the received signal vector x(t) of the acoustic vector sensor array is expressed as: In the formula, is the acoustic vector array steering vector, q(θ k )=[q1(θ k ),…,q M (θ k )] T ,symbol represents the Kronecker product operation, M is the number of acoustic vector sensors; after discrete sampling of the signal collected by the acoustic vector sensor, L snapshot discrete signals are obtained, then the corresponding acoustic vector sensor array receives discrete signal matrix X of M×K dimensions, and n(t) is the measurement noise vector of the acoustic vector sensor array; Perform FFT on the signal collected by the acoustic vector sensor and divide the obtained frequency domain signal into J segments. j ), whose covariance matrix R X (f j ) is calculated as: R X (f j )=E[X(f j )X H (f j )],j=0,1,2,…,J。 3. The high subsonic flying target acoustic signal estimation method based on acoustic vector uniform linear array according to claim 1 or 2, characterized in that: In the step S2, the PCA neural network is used to realize matrix eigenvalue decomposition, which specifically includes: the PCA neural network is realized by using a single-layer forward propagation neural network with unsupervised learning, and its network weight vector W satisfies WW H =I, the discrete signal matrix X is input into the PCA neural network, and the output of the PCA neural network is Y=WX. After the kth iteration, the cost function of the PCA neural network is: L(W)=(X-W H (k)Y) H (X-W H (k)Y) Among them, the fastest gradient descent algorithm is used to solve the optimal weight vector W of the neural network through multiple iterations. opt , when the difference between the network weight vectors of adjacent iterations is less than the threshold ε, that is, ||W(k)-W(k-1)|| 2 ≤ε, judging that the PCA neural network converges, thus obtaining the optimal weight vector W opt , which is the eigenvector of the input matrix X.

4. The high subsonic flying target acoustic signal estimation method based on acoustic vector uniform linear array according to claim 1 or 2, characterized in that: The step S3 specifically includes: the covariance matrix of the j-th frequency domain signal is R X (f j ), through R X (f j ) to perform eigenvalue decomposition and obtain the focusing matrix, where j = 0, 1, 2, ..., J, and the covariance matrix R of each frequency domain signal X (f j ) is input into the PCA neural network for eigenvalue decomposition, and the characteristic matrix U composed of its typical eigenvalues ​​is obtained. S (f j ), for the covariance matrix R at the reference frequency point f0 X (f0), whose characteristic matrix is ​​denoted by V S (f0), so the focusing matrix of the j-th frequency domain signal is obtained as: T(f j )=U s (f j )V s H (f0), The total covariance matrix of all frequency domain signals after focusing is:

5. The method for estimating acoustic signals of high subsonic flying targets based on acoustic vector uniform linear array according to claim 1 or 2, characterized in that: Step S4, using the eigenvectors corresponding to the eigenvalues ​​greater than the threshold of the total covariance matrix of all frequency domain signals to form a matrix U, constructing a feature training matrix and the target DOA parameter vector in The dataset As a training data set, determine the number of neurons in each layer of the BP neural network, and initialize the connection weights between layers and the thresholds of each layer; The specific structure of BP neural network is as follows: the number of neurons in the input layer, hidden layer and output layer of BP neural network are I, H and J respectively; the activation function of hidden layer is sigmoid function, and the activation function of output layer is purelin function; input the training data set into BP neural network, the output of the hth hidden layer neuron is b=[b1,b2,…,b H ], where b h =[b1,b2,…,b H ], represents the nth output of the hth hidden layer neuron, Among them, α h represents the input of the hth hidden layer neuron, w ih represents the weight value between the hth hidden layer neuron and the ith input layer neuron, represents the threshold of the hth neuron in the hidden layer, f1 represents the sigmoid function; the output result of the output layer neuron is represents the output of the jth neuron in the output layer, The formula for calculating the output of the jth neuron in the output layer is: Among them, β j is the input of the jth output layer neuron, v hj is the weight value between the hth hidden layer neuron and the jth output layer neuron, χ j represents the threshold of the jth neuron in the output layer, and f2 represents the purelin function.

6. The method for estimating acoustic signals of high subsonic flying targets based on acoustic vector uniform linear array according to claim 1 or 2, characterized in that: The step S5 is as follows: Calculate the output of the output neuron of the BP neural network The mean square error between the output and the expected output y is calculated, and the sum of the mean square errors of all samples, that is, the training error, is calculated; the weights and thresholds of each layer are updated using the fastest gradient descent algorithm; when the training error is less than the set error threshold, the training is considered complete, otherwise the next round of training is continued.