Airborne radar target display method and device based on 3D convolutional neural network

Abstract

The present invention discloses an airborne radar target display method and device based on a 3D convolutional neural network. The method comprises: acquiring spatiotemporal observation data; then performing Fourier transform and MVDR transform processing on the spatiotemporal observation data to obtain a training data set; then training the constructed deep learning network based on the training data set to obtain a deep learning network model; finally, predicting the newly acquired spatiotemporal observation data based on the deep learning network model to obtain a prediction result, and determining the end-to-end target display of the airborne radar. The present invention uses a small number of samples to increase the probability of target detection, which solves the problem of insufficient samples in non-uniform and non-stationary clutter environments; furthermore, the present invention does not need to suppress clutter through the traditional STAP algorithm to display and detect targets, but directly implements end-to-end target display of the airborne radar based on deep learning, further improving the detection accuracy, and can be widely used in the field of computer technology.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an airborne radar target display method and device based on a 3D convolutional neural network. Background Art

[0002] Radar is a device used to detect and locate targets. Its 24 / 7, all-weather operation makes it widely used in both military and civilian applications. In the military, radar, as one of the core equipment in modern warfare, plays an irreplaceable role in air surveillance, land / sea surveillance and reconnaissance, and target search, tracking, and identification, making it an indispensable component of joint combat systems. In civilian applications, radar is widely used in meteorological observation, counterterrorism, air traffic control, and other fields. Radar is primarily categorized as ground-based, airborne, and satellite-based, depending on the platform on which it is mounted. Ground-based radar, as an early warning radar, boasts relatively mature technology and is widely deployed worldwide. However, ground-based radar, located on the Earth's surface, is susceptible to the effects of Earth's curvature and surface obstructions. Its detection range is limited, and it has blind spots at low altitudes. This results in a short detection time for low- or ultra-low-altitude combat aircraft and cruise missiles, limiting the response time provided by ground combat systems. Airborne and satellite-based radar platforms operate at a higher altitude, effectively overcoming the effects of Earth's curvature and surface obstructions, resulting in a detection range far greater than that of ground-based radar. Airborne radar also offers the advantage of high mobility, enabling it to quickly reach designated areas. Satellite-borne radar, on the other hand, can transcend land area limitations and provide global surveillance coverage. In modern warfare, ground-based, airborne, and satellite-borne radars are often used in conjunction to support each other and jointly achieve battlefield early warning, surveillance, and reconnaissance.

[0003] For ground-based radars, clutter is typically distributed near zero Doppler due to the stationary platform. The Doppler of a moving target is proportional to its radial velocity. Therefore, the difference in the Doppler dimension can generally be used to effectively distinguish between clutter and moving targets. Common Doppler processing techniques include Moving Target Indication (MTI) and pulse Doppler processing. MTI improves the signal-to-clutter ratio of the target by filtering out clutter signals in the time domain through designing filters, while pulse Doppler processing improves the signal-to-clutter ratio by filtering out clutter outside the Doppler channel through Doppler transform. For airborne and spaceborne radars, due to the motion of the platform, clutter scatterers stationary relative to the ground have a radial velocity proportional to the platform's velocity relative to the moving platform. Therefore, clutter for airborne and spaceborne radars is no longer confined to zero Doppler but is significantly broadened in the Doppler dimension, often making it impossible to effectively distinguish between targets and clutter in the Doppler dimension. It is worth noting that the Doppler frequency of clutter scatterers for airborne and spaceborne radars is coupled with their angle of arrival. It is precisely based on this that related technologies propose Space Time Adaptive Processing (STAP) technology, which achieves clutter suppression by designing filters in a two-dimensional space-time plane. The optimal filters of MTI and STAP require the known noise covariance matrix of the unit to be detected. This matrix is ​​generally unknown in practical applications and needs to be estimated using training samples. The method of adaptively estimating this matrix and constructing a filter is called adaptive clutter suppression. In order to obtain good adaptive clutter suppression performance, the training samples and the samples to be detected need to have the same noise statistical characteristics, that is, the training samples and the samples to be detected are required to meet the independent and identically distributed (IID) condition.

[0004] When the airborne radar is operating in a non-positive direction, the relationship between the clutter Doppler frequency and spatial orientation changes with distance, that is, the clutter non-stationary phenomenon exists. At the same time, since airborne early warning radars often operate in medium / high pulse repetition frequency (PRF) mode, the radar echo often has serious range ambiguity. The above non-stationary clutter and range ambiguity are intertwined, which leads to the following problems in the subsequent traditional two-dimensional STAP (2D-STAP) processing: (1) The clutter distribution characteristics of each range unit are different, which makes the number of IID training samples used to estimate the clutter covariance matrix (CCM) seriously insufficient, causing the STAP performance to deteriorate; (2) There are multiple azimuth mainlobe clutters with different Doppler frequencies in the cell under test (CUT), which greatly increases the blind speed range. Affected by the actual non-stationary environment, the IID condition may be destroyed, resulting in a decrease in the performance of adaptive clutter suppression, and then causing a loss in target display and detection performance. Therefore, studying the adaptive clutter suppression and target display technology in non-stationary environments has important theoretical significance and practical application value. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides an airborne radar target display method and device based on a 3D convolutional neural network with high detection accuracy.

[0006] An aspect of an embodiment of the present invention provides an airborne radar target display method based on a 3D convolutional neural network, comprising:

[0007] Obtain spatiotemporal observation data;

[0008] Performing Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set;

[0009] Training the constructed deep learning network according to the training data set to obtain a deep learning network model;

[0010] The newly acquired spatiotemporal observation data is predicted based on the deep learning network model to obtain the prediction results and determine the end-to-end target display of the airborne radar.

[0011] Optionally, performing Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set includes:

[0012] Discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data;

[0013] performing space-time power spectrum estimation on the space-time observation data;

[0014] Determining the covariance matrix of the spatiotemporal observation data by maximum likelihood estimation;

[0015] Performing a Fourier spectrum transform on the spatiotemporal observation data according to the covariance matrix to obtain the spectrum intensity of the corresponding grid unit;

[0016] The target data with accurate angle and Doppler are converted into high-resolution angle-Doppler spectrum through MVDR spectrum transformation;

[0017] The angle-Doppler spectrum is obtained by superimposing the spectrum intensity of each grid;

[0018] Complete the construction of the training dataset.

[0019] Optionally, the method further comprises the step of constructing a deep learning network structure, which comprises:

[0020] Using a 3D convolutional neural network, the first layer extracts features from the low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in the training samples to obtain the first-layer features.

[0021] The second to fourth layers of the 3D convolutional neural network are all nonlinear mappings of features, and the extracted feature maps are nonlinearly mapped into the transformed high-dimensional space in the second to fourth layers;

[0022] The fifth layer is the image reconstruction layer, which generates a high-resolution output image.

[0023] Optionally, in the step of discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data, the discretization formula for the pitch data is:

[0024] N e =ρ e N

[0025] The discrete formula of the orientation data is:

[0026] N a =ρ a M

[0027] The discrete formula of the Doppler frequency data is:

[0028] N d =ρ d K

[0029] Among them, N e represents the pitch data; ρ e represents the pitch dispersion coefficient; N represents the number of elements in the uniform plane airborne phased array radar; Na represents the orientation data; ρ a represents the azimuth dispersion coefficient; M represents the number of uniform plane airborne phased array radar elements; N d represents Doppler frequency data; ρ d represents the Doppler frequency dispersion coefficient; K represents the number of pulses.

[0030] Optionally, in the step of performing space-time power spectrum estimation on the space-time observation data, the expression of the space-time power spectrum estimation is:

[0031] Y=P[X]

[0032] Where Y represents the clutter power spectrum obtained by space-time power spectrum estimation; P[.] represents the power spectrum estimation operator; and X represents the training sample of space-time observation data.

[0033] Optionally, in the step of using a 3D convolutional neural network to extract features of a low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in a training sample in the first layer to obtain first-layer features, the feature extraction expression is:

[0034] F1=max(0,W1*Y+b1)

[0035] Where F1 represents the first-layer features; W1*Y represents the 3D convolution operation on the first-layer input feature map. W1 represents the convolution kernel of dimensions c×f1×f1×d1×n1, where c represents the number of input image channels, f1 represents the length and width of the convolution kernel, d1 represents the number of input frames, n1 represents the number of convolution kernels, and b1 is the n1-dimensional bias vector.

[0036] Optionally, the expression for nonlinearly mapping the extracted feature maps to the transformed high-dimensional space in the second to fourth layers is:

[0037] F i =max(0,W i *F i-1 +b i )

[0038] Among them, F i represents the i-th layer feature; W i Indicates dimension n i-1 ×f i ×f i ×d i ×n i The convolution kernel, f i Represents the length and width of the convolution kernel, d i Indicates the number of input frames, b i is the n1-dimensional bias vector.

[0039] Optionally, the method further comprises the step of configuring an implementation scenario, which comprises:

[0040] Configure the carrier platform to fly at a constant speed at the target altitude;

[0041] The radar antenna is configured as a rectangular uniform array, where the distance between any adjacent array elements is half a wavelength;

[0042] In each coherent processing interval, the radar transmits multiple narrowband frequency-modulated coherent pulse trains with a constant PRF through the transmit / receive antenna and receives the echoes.

[0043] Determine the azimuth and elevation angles between the array and the ground clutter block, and determine the angle between the antenna array placement direction and the carrier aircraft flight direction.

[0044] Another aspect of the present invention further provides an airborne radar target display device based on a 3D convolutional neural network, comprising:

[0045] The first module is used to obtain spatiotemporal observation data;

[0046] The second module is used to perform Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set;

[0047] The third module is used to train the constructed deep learning network according to the training data set to obtain a deep learning network model;

[0048] The fourth module is used to predict the newly acquired spatiotemporal observation data based on the deep learning network model, obtain the prediction results, and determine the end-to-end target display of the airborne radar.

[0049] Another aspect of an embodiment of the present invention further provides an electronic device, including a processor and a memory;

[0050] The memory is used to store programs;

[0051] The processor executes the program to implement the method described above.

[0052] The present invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0053] The embodiment of the present invention first obtains spatiotemporal observation data; then performs Fourier transform processing and MVDR transform processing on the spatiotemporal observation data to obtain a training data set; then trains the constructed deep learning network based on the training data set to obtain a deep learning network model; finally, predicts the newly acquired spatiotemporal observation data based on the deep learning network model to obtain a prediction result and determine the end-to-end target display of the airborne radar. The present invention uses a small number of samples to improve the probability of target detection, which solves the problem of insufficient samples in non-uniform and non-stationary clutter environments; furthermore, the present invention does not need to suppress clutter through the traditional STAP algorithm to display and detect targets, but directly implements end-to-end target display of the airborne radar based on deep learning, further improving the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0055] Figure 1 A geometric illumination diagram of a plane array airborne radar provided by an embodiment of the present invention;

[0056] Figure 2 An overall implementation flow chart provided for an embodiment of the present invention;

[0057] Figure 3 The clutter plus target three-dimensional spectrum provided by the embodiment of the present invention;

[0058] Figure 4 The target three-dimensional spectrum provided by the embodiment of the present invention;

[0059] Figure 5 An overall block diagram provided for an embodiment of the present invention;

[0060] Figure 6 Schematic diagrams of prediction results and label comparison results for targets of different speeds provided in an embodiment of the present invention, wherein (a), (b), (c), and (d) respectively represent schematic diagrams of moving targets of different speeds. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0062] In order to solve the problems existing in the prior art, an embodiment of the present invention provides an airborne radar target display method based on a 3D convolutional neural network, comprising:

[0063] Obtain spatiotemporal observation data;

[0064] Performing Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set;

[0065] Training the constructed deep learning network according to the training data set to obtain a deep learning network model;

[0066] The newly acquired spatiotemporal observation data is predicted based on the deep learning network model to obtain the prediction results and determine the end-to-end target display of the airborne radar.

[0067] Optionally, performing Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set includes:

[0068] Discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data;

[0069] performing space-time power spectrum estimation on the space-time observation data;

[0070] Determining the covariance matrix of the spatiotemporal observation data by maximum likelihood estimation;

[0071] Performing a Fourier spectrum transform on the spatiotemporal observation data according to the covariance matrix to obtain the spectrum intensity of the corresponding grid unit;

[0072] The target data with accurate angle and Doppler are converted into high-resolution angle-Doppler spectrum through MVDR spectrum transformation;

[0073] The angle-Doppler spectrum is obtained by superimposing the spectrum intensity of each grid;

[0074] Complete the construction of the training dataset.

[0075] Optionally, the method further comprises the step of constructing a deep learning network structure, which comprises:

[0076] Using a 3D convolutional neural network, the first layer extracts features from the low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in the training samples to obtain the first-layer features.

[0077] The second to fourth layers of the 3D convolutional neural network are all nonlinear mappings of features, and the extracted feature maps are nonlinearly mapped into the transformed high-dimensional space in the second to fourth layers;

[0078] The fifth layer is the image reconstruction layer, which generates a high-resolution output image.

[0079] Optionally, in the step of discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data, the discretization formula for the pitch data is:

[0080] N e =ρ e N

[0081] The discrete formula of the orientation data is:

[0082] N a =ρ a M

[0083] The discrete formula of the Doppler frequency data is:

[0084] N d =ρ d K

[0085] Among them, N e represents the pitch data; ρ e represents the pitch dispersion coefficient; N represents the number of elements in the uniform plane airborne phased array radar; N a represents the orientation data; ρ a represents the azimuth dispersion coefficient; M represents the number of uniform plane airborne phased array radar elements; N d represents Doppler frequency data; ρ d represents the Doppler frequency dispersion coefficient; K represents the number of pulses.

[0086] Optionally, in the step of performing space-time power spectrum estimation on the space-time observation data, the expression of the space-time power spectrum estimation is:

[0087] Y=P[X]

[0088] Where Y represents the clutter power spectrum obtained by space-time power spectrum estimation; P[.] represents the power spectrum estimation operator; and X represents the training sample of space-time observation data.

[0089] Optionally, in the step of using a 3D convolutional neural network to extract features of a low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in a training sample in the first layer to obtain first-layer features, the feature extraction expression is:

[0090] F1=max(0,W1*Y+b1)

[0091] Where F1 represents the first-layer features; W1*Y represents the 3D convolution operation on the first-layer input feature map. W1 represents the convolution kernel of dimensions c×f1×f1×d1×n1, where c represents the number of input image channels, f1 represents the length and width of the convolution kernel, d1 represents the number of input frames, n1 represents the number of convolution kernels, and b1 is the n1-dimensional bias vector.

[0092] Optionally, the expression for nonlinearly mapping the extracted feature maps to the transformed high-dimensional space in the second to fourth layers is:

[0093] F i =max(0,W i *F i-1 +b i )

[0094] Among them, F i represents the i-th layer feature; W i Indicates dimension n i-1 ×f i ×f i ×d i ×n i The convolution kernel, f i Represents the length and width of the convolution kernel, d i Indicates the number of input frames, b i is the n1-dimensional bias vector.

[0095] Optionally, the method further comprises the step of configuring an implementation scenario, which comprises:

[0096] Configure the carrier platform to fly at a constant speed at the target altitude;

[0097] The radar antenna is configured as a rectangular uniform array, where the distance between any adjacent array elements is half a wavelength;

[0098] In each coherent processing interval, the radar transmits multiple narrowband frequency-modulated coherent pulse trains with a constant PRF through the transmit / receive antenna and receives the echoes.

[0099] Determine the azimuth and elevation angles between the array and the ground clutter block, and determine the angle between the antenna array placement direction and the carrier aircraft flight direction.

[0100] Another aspect of the present invention further provides an airborne radar target display device based on a 3D convolutional neural network, comprising:

[0101] The first module is used to obtain spatiotemporal observation data;

[0102] The second module is used to perform Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set;

[0103] The third module is used to train the constructed deep learning network according to the training data set to obtain a deep learning network model;

[0104] The fourth module is used to predict the newly acquired spatiotemporal observation data based on the deep learning network model, obtain the prediction results, and determine the end-to-end target display of the airborne radar.

[0105] Another aspect of an embodiment of the present invention further provides an electronic device, including a processor and a memory;

[0106] The memory is used to store programs;

[0107] The processor executes the program to implement the method described above.

[0108] The present invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0109] The specific implementation process of the present invention is described in detail below with reference to the accompanying drawings:

[0110] To address the problem of insufficient short-range clutter samples due to non-stationary clutter and range ambiguity, this paper proposes a deep learning-based target display method for three-dimensional short-range clutter environments. First, a three-dimensional clutter and target signal model is introduced. Next, an overall framework for a three-dimensional convolutional neural network is proposed. The network structure is then established, and the required training and validation sets are created. Finally, comprehensive experimental results verify that the proposed method can achieve end-to-end target display for airborne radars with few samples.

[0111] like Figure 1 As shown, the carrier platform flies at a constant speed v at an altitude H. The radar antenna is a rectangular uniform array with M rows and N columns, and the spacing d between any adjacent array elements is half a wavelength. Assuming that the area to be detected is in the far field of the antenna, the incoming echoes from different array elements can be considered to have the same angle of incidence. In each coherent processing interval, the radar transmits K narrowband frequency-modulated coherent pulse trains at a constant PRF through the transmit / receive antenna and receives the echoes. θ and θ represent the azimuth and elevation angles between the array and the ground clutter block, respectively. crab It is the angle between the antenna array placement direction and the aircraft flight direction.

[0112] The overall implementation steps of the present invention are as follows Figure 2As shown in the figure, it mainly includes: (1) obtaining space-time observation data; (2) obtaining input training data set; (3) establishing deep learning network structure; (4) training network model; (5) testing network model.

[0113] Step (1): Obtain space-time observation data X

[0114] X=[x1,x2,…,x L ]∈C NMK×L (1)

[0115] Where L represents the total number of distance units.

[0116] Step (2): Get the input training dataset

[0117] Optionally, the steps are:

[0118] Step (2.1): First, this embodiment can discretize the pitch, azimuth and Doppler frequency into N e =ρ e N, N a =ρ a M and N d =ρ d K grid cells. e , ρ s and ρ d are the dispersion coefficients of pitch, azimuth and Doppler frequency respectively. Then, the set of all steering vectors in three dimensions of space and time is given by the following formula.

[0119]

[0120] Among them, f e,i ,1≤i≤N e 、f a,i ,1≤j≤N a and f d,k ,1≤k≤N d denote the normalized pitch frequency of the i-th grid cell, the normalized azimuth frequency of the j-th grid cell, and the Doppler frequency of the k-th grid cell, respectively.

[0121] Step (2.2): For the space-time observation data, i.e. the training sample X=[x1,x2,…,x L ]∈C NMK×L Perform space-time power spectrum estimation:

[0122] Y=P[X] (3)

[0123] in, is the clutter power spectrum, and P[·] is the power spectrum estimation operator.

[0124] Step (2.3): Perform Fourier spectrum transformation on the space-time observation data (1), P(f e,i ,f a,j ,f d,k ) is the spectral intensity of the corresponding grid cell. Therefore, the Fourier spectrum transform can be defined as:

[0125]

[0126] in, is the covariance matrix, obtained by maximum likelihood estimation (MLE):

[0127]

[0128] Where P is the number of training units. p Represents the STO data of the p-th training unit.

[0129] Here, Fourier spectral transform plays an important role in the input of the network, which converts the data of STO clutter plus target into the form of low-resolution angle-Doppler spectrum to prepare the input data for training the network.

[0130] Step (2.4): MVDR spectrum transform converts the target data with accurate angle and Doppler into a high-resolution angle-Doppler spectrum to prepare the label for training the network. The minimum variance distortion-free response (MVDR) spectrum transform is expressed as:

[0131]

[0132] Step (2.5): The final angle-Doppler spectrum is obtained by superimposing the spectrum intensity of each grid. Therefore, the angle-Doppler spectrum can be expressed as:

[0133]

[0134] Step (2.6): In the proposed method, this embodiment first applies beamforming technology, i.e., Fourier spectrum transform, to the clutter plus target echo data (1) X using equations (2), (4), (5), and (7), constructs the input data of the network, and then sets the low-resolution clutter plus target pitch-azimuth-Doppler three-dimensional spectrum Y=P[X] after the Fourier spectrum transform. Then, the MVDR spectrum transform is performed on the echo data (1) containing only the target X using equations (2), (5), (6), and (7), constructs the label data of the network, and sets the high-resolution target pitch-azimuth-Doppler three-dimensional spectrum after the MVDR spectrum transform to The input of the network uses the clutter plus target covariance matrix of equation (5), and the label of the network uses the target covariance matrix in equation (5).

[0135] Therefore, input and tags should be included in the training dataset, and T is the total number of training samples. The training dataset is defined as:

[0136]

[0137] Step (3): Establish a deep learning network structure:

[0138] 3D convolution uses the 3D receptive field to perform 3D convolution on each unit in the local block set of the 3D image to obtain the average value of local and non-local pixels. As the network layer deepens, the convolution layer continuously extracts information from the block set. As the 3D receptive field expands, the information of the entire block set is obtained, and finally the local and non-local characteristics of the entire block set are obtained. In this technology, the non-stationary environment of clutter and distance ambiguity are taken into account. Therefore, the clutter and target are no longer a simple two-dimensional space-time plane relationship, but rather a Figure 3 The three-dimensional spatial relationship. Because the repetition frequency is too high, distance ambiguity is generated, so the clutter is distributed at the pitch frequency corresponding to the blurred distance, that is, different clutter distributions will appear in the pitch dimension. Convolution is usually represented as 2D convolution in the field of super-resolution, which is used to extract two-dimensional information of image blocks. If there is no distance ambiguity, and it is only in the two-dimensional plane of space and time, then 2D convolution is sufficient to process two-dimensional information. However, for three-dimensional information, 3D convolution can be processed more efficiently. In this technology, a 3D convolutional neural network (CNN) is used to process the set of space-time three-dimensional matrix blocks to be processed, where the first two dimensions, the azimuth-Doppler dimension, preserve the local information of the matrix, and the third dimension, the pitch dimension, preserves the non-local information. This ensures that while learning local information through 3D convolution, non-local information is also retained.

[0139] In essence, the airborne radar target display method based on 3D convolutional neural network can be regarded as a classification problem, in which moving targets with the same normalized pitch frequency, normalized azimuth frequency and normalized Doppler frequency are considered to be the same category, such as Figure 4 In addition, the target three-dimensional spectrum can be Figure 3As can be seen, clutter and targets are separable in the space-time domain. Therefore, through the mapping properties of convolutional neural networks, targets and clutter can be distinguished in three-dimensional space. Therefore, in this technology, clutter is actually filtered out first, and then the target can be better displayed, thereby improving detection results. Therefore, to improve the display performance of targets in non-stationary clutter and thus increase the target detection rate, this embodiment applies 3DCNN to intermediate reconstruction and filtering.

[0140] The network has five convolutional layers. The input is the low-resolution spatial-temporal Fourier spectrum of clutter containing the target, and the output is the spatial-temporal spectrum of the target after filtering out clutter and interference.

[0141] The network design uses a fully convolutional 3D network. The entire network consists of a nonlinear structure and 3D convolutional layers, using convolution kernels to extract features as input to the next layer. After the convolution operation, the data size is reduced. When training the network, zero padding is performed between each input and data, which effectively ensures that the input and output sizes of the convolution layer are the same. This embodiment applies 3DCNN to intermediate reconstruction and filtering, inputting low-resolution clutter plus the target pitch-azimuth-Doppler three-dimensional spectrum Y, and outputting a high-resolution target angle-Doppler spectrum Z according to the following formula.

[0142] Z=F[Y] (9)

[0143] in, The high-resolution pitch-azimuth-Doppler three-dimensional spectrum of the target obtained at the output of the neural network is Characterizing 3D neural network operators.

[0144] Step (3.1): The low-resolution pitch-azimuth-Doppler three-dimensional spectrum of the clutter containing the target contains rough information about the true position and energy distribution of the clutter and target. Its characteristics are relatively intuitive and effective. Therefore, the features of the training samples can be extracted in the first layer:

[0145] F1=max(0,W1*Y+b1) (1)

[0146] Where W1*Y represents a 3D convolution operation on the first-layer input feature map. W1 represents a convolution kernel of dimensions c×f1×f1×d1×n1, where c represents the number of input image channels, f1 represents the kernel's length and width, d1 represents the number of input frames, n1 represents the number of kernels (the number of output feature channels), and b1 represents the n1-dimensional bias vector. This represents extracting multiple image patches from a low-resolution image, with each patch undergoing convolution to produce a high-dimensional feature matrix. All five convolutional layers use the ReLU activation function and zero-padding on the edges.

[0147] Step (3.2): The second to fourth layers are all nonlinear mappings of features, which nonlinearly map the extracted feature maps to the transformed high-dimensional space:

[0148] F i =max(0,W i *F i-1 +b i ), i=2,3,4 (2)

[0149] Among them, i represents the number of convolutional layers, W i Indicates dimension n i-1 ×f i ×f i ×d i ×n i The convolution kernel, f i Represents the length and width of the convolution kernel, d i Indicates the number of input frames, b i n i dimensional bias vector, each n of the fourth output i Each dimensional feature vector represents a high-resolution image block and is used for reconstruction.

[0150] The fifth layer of step (3.3) is the image reconstruction layer, which generates a high-resolution output image. The specific expression is:

[0151] Z=W5*F4+b5 (3)

[0152] Where W5 represents a convolution kernel of dimensions n4×f5×f5×d5×c, f5 represents the length and width of the convolution kernel, d5 represents the number of input frames, and b5 is the c-dimensional bias vector. This layer is essentially a deconvolution process.

[0153] Step (4): Train the network model:

[0154] Assume a set of low-resolution clutter plus target pitch-azimuth-Doppler three-dimensional images and its corresponding known high-resolution target pitch-azimuth-Doppler three-dimensional spectrum image By minimizing the model MSE loss function, a nonlinear mapping relationship between input and output is obtained:

[0155]

[0156] Among them, ||·|| F represents the Frobenius norm, M is the number of training data, Θ={W i ,b i i = 1, 2, ..., 5 are network parameters. During the model training phase, stochastic gradient descent is used to update the parameters.

[0157] Step (5): Test the network model (attached Figure 5 )

[0158] During the testing phase, a sample first undergoes Fourier transform (FT) and then is fed into the network testing model. The trained 3DCNN outputs the target’s angular Doppler spectrum, ultimately obtaining the target’s prediction result.

[0159] It should be noted that if Figure 5 As shown in the overall block diagram, the present invention can obtain the target display result.

[0160] It should be noted that Figure 6 The figure shows the comparison between the prediction results and labels of moving targets at different speeds in a non-stationary and non-uniform environment. It can be seen that in this technology, the use of 3D convolutional neural networks can effectively achieve end-to-end target display.

[0161] In summary, the present invention has the following characteristics:

[0162] (1) First, considering the target display and detection issues caused by non-stationary clutter and range ambiguity, a clutter-plus-target training dataset was simulated, which took into account various non-ideal factors such as aircraft yaw, array errors, and clutter internal motion. This dataset takes into account various realistic situations, making the proposed method more robust.

[0163] (2) Then, a three-dimensional five-layer convolutional neural network is designed to learn the characteristic distribution of clutter and targets. The proposed network structure can predict targets with high resolution and achieve end-to-end moving target display with high detection accuracy. In this network, the input is constructed by the angle-Doppler three-dimensional spectrum of the clutter plus the target, which has a very low resolution and is estimated from only a few samples. The label is given by the target angle-Doppler three-dimensional spectrum, whose high-resolution spectrum is obtained by the accurate angle and Doppler estimation of the target.

[0164] (3) Only a few samples are needed to effectively suppress non-stationary clutter and display targets.

[0165] Compared to existing technologies, this method uses fewer samples to improve target detection probability, which solves the problem of insufficient samples in non-uniform and non-stationary clutter environments. Furthermore, the proposed method does not require the traditional STAP algorithm to suppress clutter to display and detect targets. Instead, it directly achieves end-to-end target display for airborne radar based on deep learning.

[0166] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0167] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0168] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0169] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0170] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0171] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0172] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0173] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0174] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. Airborne radar target display method based on 3D convolutional neural network, characterized in that: include: Obtain spatiotemporal observation data; Performing Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set; Training the constructed deep learning network according to the training data set to obtain a deep learning network model; The newly acquired spatiotemporal observation data is predicted based on the deep learning network model to obtain prediction results and determine the end-to-end target display of the airborne radar; The Fourier transform and MVDR transform processing are performed on the spatiotemporal observation data to obtain a training data set, including: Discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data; performing space-time power spectrum estimation on the space-time observation data; Determining the covariance matrix of the spatiotemporal observation data by maximum likelihood estimation; Performing a Fourier spectrum transform on the spatiotemporal observation data according to the covariance matrix to obtain the spectrum intensity of the corresponding grid unit; The target data with accurate angle and Doppler are converted into high-resolution angle-Doppler spectrum through MVDR spectrum transformation; The angle-Doppler spectrum is obtained by superimposing the spectrum intensity of each grid; Complete the construction of the training dataset; The method further comprises the step of constructing a deep learning network structure, which comprises: Using a 3D convolutional neural network, the first layer extracts features from the low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in the training samples to obtain the first-layer features. The second to fourth layers of the 3D convolutional neural network are all nonlinear mappings of features, and the extracted feature maps are nonlinearly mapped into the transformed high-dimensional space in the second to fourth layers; The fifth layer is the image reconstruction layer, which generates a high-resolution output image.

2. The airborne radar target display method based on 3D convolutional neural network according to claim 1, characterized in that: In the step of discretizing the spatiotemporal observation data to obtain pitch data, azimuth data, and Doppler frequency data, the discretization formula of the pitch data is: The discrete formula of the orientation data is: The discrete formula of the Doppler frequency data is: in, Represents pitch data; represents the pitch dispersion coefficient; Represents the number of array elements of uniform planar airborne phased array radar; Represents position data; represents the coefficient of orientation dispersion; Represents the number of uniform plane airborne phased array radar row elements; represents Doppler frequency data; represents the Doppler frequency dispersion coefficient; Represents the number of pulses.

3. The airborne radar target display method based on 3D convolutional neural network according to claim 1, characterized in that: In the step of performing space-time power spectrum estimation on the space-time observation data, the expression of the space-time power spectrum estimation is: in, represents the clutter power spectrum obtained by space-time power spectrum estimation; represents the power spectrum estimation operator; Training samples representing spatiotemporal observations.

4. The airborne radar target display method based on 3D convolutional neural network according to claim 1, characterized in that: The 3D convolutional neural network is used to extract features from the low-resolution pitch-azimuth-Doppler three-dimensional spectrum of clutter containing targets in the training samples in the first layer to obtain the first layer features. The expression for the feature extraction is: in, Represents the first layer features; Represents the 3D convolution operation on the first layer input feature map, Indicates the dimension The convolution kernel, Indicates the number of input image channels, represents the length and width of the convolution kernel, Indicates the number of input frames, represents the number of convolution kernels, for dimensional bias vector.

5. The airborne radar target display method based on 3D convolutional neural network according to claim 1, characterized in that: The expression for nonlinearly mapping the extracted feature maps from the second to fourth layers to the transformed high-dimensional space is: in, Indicates the Layer characteristics; Indicates the dimension The convolution kernel, Represents the length and width of the convolution kernel, Indicates the number of input frames, for dimensional bias vector.

6. The airborne radar target display method based on 3D convolutional neural network according to claim 1, characterized in that: The method further comprises the step of configuring an implementation scenario, which comprises: Configure the carrier platform to fly at a constant speed at the target altitude; The radar antenna is configured as a rectangular uniform array, where the distance between any adjacent array elements is half a wavelength; In each coherent processing interval, the radar transmits multiple narrowband frequency-modulated coherent pulse trains with a constant PRF through the transmit / receive antenna and receives the echoes. Determine the azimuth and elevation angles between the array and the ground clutter block, and determine the angle between the antenna array placement direction and the carrier aircraft flight direction.

7. A device for implementing the airborne radar target display method based on a 3D convolutional neural network as described in any one of claims 1 to 6, characterized in that: include: The first module is used to obtain spatiotemporal observation data; The second module is used to perform Fourier transform and MVDR transform on the spatiotemporal observation data to obtain a training data set; The third module is used to train the constructed deep learning network according to the training data set to obtain a deep learning network model; The fourth module is used to predict the newly acquired spatiotemporal observation data based on the deep learning network model, obtain the prediction results, and determine the end-to-end target display of the airborne radar.

8. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 6.

Images