Zero-auxiliary data target detection method based on FISTA deep unfolding network

Abstract

The application discloses a zero auxiliary data target detection method based on a FISTA deep unfolding network, and comprises the following steps: obtaining radar echo data; according to the angle-Doppler domain sparse characteristics of airborne radar clutter, the clutter is expressed as the product of a space-time steering dictionary and a clutter angle-Doppler spectrum; a binary hypothesis testing target detection model under the condition of zero auxiliary data is constructed based on this; a test statistic is constructed based on the model, and the clutter angle-Doppler spectrum estimation values under the target existing hypothesis and the target non-existing hypothesis are respectively solved by using the FISTA deep unfolding network, and the detection statistic of a to-be-detected unit is calculated; the detection statistic is compared with a preset detection threshold to determine whether a target exists, wherein the detection threshold is determined by the detection statistic of a pure clutter unit and a preset false alarm rate. The application realizes adaptive target detection under the condition of no auxiliary data.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a zero-assistance data target detection method based on FISTA deep unfolding network. Background Technology

[0002] To achieve wide-area coverage from high to very low altitudes, airborne radar antennas typically employ a wide-beam elevation pattern, with the beam center pointing horizontally or slightly downwards, illuminating both the ground and the airspace. Therefore, the radar echoes received not only contain target signals but also a large amount of high-intensity ground object scattering echoes. Because ground object scatterers in different directions have varying radial velocities, clutter exhibits significant broadening in the Doppler frequency domain, causing moving targets and clutter to overlap in the spectrum, making them difficult to distinguish effectively.

[0003] To address the problem of insufficient auxiliary data, researchers have proposed improvements by utilizing prior clutter information and reducing system dimensionality. However, when the prior information mismatches with the actual clutter environment, the performance of prior-based detection methods deteriorates significantly. When auxiliary data is extremely scarce, the accuracy of clutter covariance matrix estimation is insufficient, and the dimensionality reduction projection direction shifts to some extent. Especially in the absence of auxiliary data, dimensionality reduction algorithms struggle to reliably separate the signal subspace from the noise subspace, easily misclassifying weak target features as noise and discarding them, ultimately leading to severe degradation in detection performance.

[0004] In recent years, sparse recovery methods have provided a new technical approach to alleviate dependence on auxiliary data. Airborne radar clutter exhibits sparsity in the angle-Doppler domain, meaning that the true clutter energy is concentrated only near the clutter ridge determined by the spatiotemporal coupling relationship. Sparse recovery methods model the clutter signal as the product of an overcomplete spatiotemporal dictionary and a sparse coefficient vector, modeling the spatiotemporal spectrum super-resolution problem as a signal recovery problem. By solving an underdetermined sparse optimization problem, the spatiotemporal spectrum of the clutter is directly estimated from the data of the target cell itself, ultimately achieving high-performance spatiotemporal filtering under conditions of minimal training samples. However, existing sparse recovery methods employ fixed-parameter iterative algorithms, which are cumbersome to tune, have weak generalization ability, and lack interpretability, making it difficult to adaptively match different clutter scenarios. Deep unfolding algorithms, by establishing a specific and systematic connection between iterative algorithms and deep neural networks, overcome the limitations of fixed parameters while preserving the physical interpretability of the underlying signal model. With the help of a moderately sized training dataset, an efficient, high-performance, and interpretable network architecture is constructed. This fusion approach has been widely explored and practiced.

[0005] However, existing deep unfolding spatiotemporal adaptive processing methods are mainly used for clutter suppression tasks, optimizing spatiotemporal filters to maximize the output signal-to-clutter-to-noise ratio under the assumption of only clutter and noise. Object detection, on the other hand, is essentially a binary hypothesis testing problem that requires constructing a likelihood ratio statistic under the assumptions of both target presence and target absence. How to introduce deep unfolding networks into adaptive detection, addressing both hypothesis tests and fusing target priors to solve the target detection problem under zero-auxiliary data, urgently requires in-depth exploration and innovation. Summary of the Invention

[0006] The main objective of this invention is to provide a zero-assistance data target detection method based on the FISTA deep unfolding network.

[0007] Another objective of this invention is to propose a zero-assistance data target detection system based on the FISTA deep unfolding network.

[0008] The third objective of this invention is to provide a computer device.

[0009] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above objectives, a first aspect of the present invention proposes a zero-assistance data target detection method based on a FISTA deep unfolded network, comprising: The radar echo data is acquired, and based on the sparse characteristics of the angle-Doppler domain of airborne radar clutter, the clutter in the radar echo is represented as the product of the space-time steering dictionary and the clutter angle-Doppler spectrum. Based on the product representation of the space-time guidance dictionary and the clutter angle-Doppler spectrum, a target detection model under zero-auxiliary data conditions is constructed. The target detection model is a binary hypothesis testing model, corresponding to the target existence hypothesis and the target non-existence hypothesis, respectively. Based on the binary hypothesis testing model, a test statistic is constructed. The clutter angle-Doppler spectrum estimates under the target presence hypothesis and the target non-existence hypothesis are solved using the FISTA deep unfolding network. The detection statistic of the unit to be detected is calculated based on the clutter angle-Doppler spectrum estimates. The detection statistics are compared with a preset detection threshold, and the presence of a target is determined based on the comparison result. The detection threshold is determined by the detection statistics of the pure clutter detection unit and the preset false alarm rate.

[0011] In one embodiment of the present invention, the FISTA deep unfolding network includes: The input layer is used to receive radar echo data, a fully complete space-time guidance dictionary, and a target space-time guidance vector from the unit to be detected, and to construct a target orthogonal projection matrix based on the target space-time guidance vector. The initialization layer is used to initialize the clutter angle-Doppler spectrum estimates, momentum variables, and acceleration parameters under the assumptions of target presence and target non-existence, respectively. The FISTA iterative expansion layers are connected sequentially with a preset number of layers. Each FISTA iterative expansion layer is used to calculate the clutter angle-Doppler spectrum reconstruction residual under two assumptions, update the momentum variable by gradient descent, and perform soft thresholding on the gradient update result to obtain the clutter angle-Doppler spectrum estimate of the current layer. The momentum is extrapolated based on the difference between the current layer and the previous layer estimate to obtain the momentum variable required for the next iteration. After completing the preset number of iterations, the clutter angle-Doppler spectrum estimation results under the two assumptions are output.

[0012] In one embodiment of the present invention, the iteration step size and sparse regularization parameter in the FISTA iterative unrolling layer are set as learnable parameters and trained and optimized by the backpropagation algorithm.

[0013] In one embodiment of the present invention, the step of using the FISTA depth unfolding network to solve for clutter angle-Doppler spectrum estimates under two assumptions includes: By introducing a residual term, the problem of minimizing the squared residuals in the detector is transformed into a sparse optimization problem. The FISTA algorithm is used to solve the clutter angle-Doppler spectrum estimates under the assumptions of target presence and target absence, respectively.

[0014] In one embodiment of the present invention, comparing the detection statistic with a preset detection threshold and determining whether a target exists based on the comparison result includes: If the detection statistic is greater than or equal to the detection threshold, it is determined that a target exists in the unit to be detected. If the detection statistic is less than the detection threshold, it is determined that there is no target in the unit to be detected.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a zero-assistance data target detection system based on a FISTA deep unfolding network, comprising: The data acquisition module is used to acquire radar echo data, a fully complete space-time guidance dictionary, and target space-time guidance vectors. The data extraction module is used to extract the data of the unit to be detected from the radar echo data to form a training dataset; The network initialization module is used to initialize the clutter angle-Doppler spectrum estimates, momentum variables, and acceleration parameters under the assumptions of target presence and target non-existence, and input the initialization results into the FISTA depth unfolding network. The training module is used to preprocess the dataset and then train the FISTA deep unfolded network using the backpropagation algorithm to obtain the trained target detection model. The detection module is used to input the data to be detected into the trained target detection model, output the calculated test statistic, compare the test statistic with a preset threshold, and determine whether the target is contained based on the comparison result.

[0016] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, for implementing a zero-assistance data target detection method based on a FISTA deep unfolding network as described in the first aspect embodiment.

[0017] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a zero-assistance data target detection method based on a FISTA deep unfolded network as described in the first aspect embodiment.

[0018] The embodiments of the present invention have the following beneficial effects: This application presents a zero-auxiliary data adaptive target detection method based on a FISTA deep unfolding network. By leveraging the sparsity of airborne radar clutter in the angle-Doppler domain, a zero-auxiliary data detector, DU-GLRT, based on sparse recovery is obtained. A FISTA deep unfolding network is constructed, transforming the detector solution into a clutter angle-Doppler spectrum estimation problem under H0 and H1 hypotheses. Through deep unfolding, the gradient update, soft threshold shrinkage, and momentum acceleration processes of FISTA are mapped to a multi-layer network structure, and the iteration step size and sparse regularization parameters are set as learnable parameters, fully integrating the interpretability of model-driven methods with the adaptive learning capability of data-driven methods. The method constructs test statistics using only the radar echo data of the target cell itself, without relying on auxiliary data from adjacent range cells to estimate the clutter covariance matrix. Addressing the problem of decreased covariance matrix estimation accuracy, performance degradation, or even failure in non-uniform clutter environments due to insufficient auxiliary data and mismatch between sample and target cell statistical characteristics, this method improves target detection performance in zero-auxiliary data and strongly non-uniform clutter environments. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1A flowchart illustrating a zero-assistance data target detection method based on a FISTA deep unfolding network, provided in an embodiment of the present invention; Figure 2 This is a diagram of the FISTA deep unfolded network structure according to an embodiment of this application; Figure 3 This is a comparison chart of the detection performance of embodiments of this application; Figure 4 This is a structural diagram of a zero-assistance data target detection system based on the FISTA deep unfolding network, provided in an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes a zero-assistance data target detection method based on a FISTA deep unfolded network according to an embodiment of the present invention.

[0023] Example 1 This embodiment provides a zero-assistance data target detection method based on the FISTA deep unfolding network, such as... Figure 1 As shown, the method includes the following steps: S1. Acquire radar echo data. Based on the sparse characteristics of the angle-Doppler domain of airborne radar clutter, represent the clutter in the radar echo as the product of the space-time steering dictionary and the clutter angle-Doppler spectrum.

[0024] S2, Based on the product representation of the space-time guidance dictionary and the clutter angle-Doppler spectrum, a target detection model under zero-auxiliary data conditions is constructed. The target detection model is a binary hypothesis testing model, corresponding to the target existence hypothesis and the target non-existence hypothesis, respectively.

[0025] S3. Based on the binary hypothesis testing model, construct the test statistic, use the FISTA deep unfolding network to solve the clutter angle-Doppler spectrum estimates under the target existence hypothesis and the target non-existence hypothesis respectively, and calculate the detection statistic of the unit to be detected based on the clutter angle-Doppler spectrum estimates.

[0026] S4, compare the detection statistics with a preset detection threshold, and determine whether a target exists based on the comparison result, wherein the detection threshold is determined by the detection statistics of the pure clutter detection unit and the preset false alarm rate.

[0027] In this embodiment, the above method is based on the real need for airborne radar to have scarce or even missing auxiliary data in non-uniform clutter environments. It deeply integrates the sparse recovery problem of clutter angle-Doppler spectrum with the binary hypothesis testing framework and uses the FISTA deep unfolding network to achieve adaptive target detection under zero auxiliary data.

[0028] Specifically, step S1 includes: The echo data of the unit under test received by the airborne radar within one coherent processing time is acquired. The echo data corresponding to each channel and each pulse of the array are arranged in a preset order to form the space-time echo vector of the unit under test. .

[0029] Specifically, let the number of array elements of the airborne radar be... M The number of pulses within one coherent processing time is N Then the space-time echo vector of the unit to be detected is expressed as:

[0030] in, Represents the data to be measured, and represents the target signal. Indicates clutter. Represents noise. The target is represented as... ,in The target time-domain guiding vector, Guide vector for the target space. Indicates the target spacetime steering vector. and These are the target Doppler frequency and the target spatial frequency, respectively. The maximum range of the normalized Doppler frequency is... .

[0031] Based on the sparse characteristics of the angle-Doppler domain of airborne radar clutter, the clutter in the radar echo is represented as the product of the space-time steering dictionary and the clutter angle-Doppler spectrum.

[0032] Specifically, by utilizing the sparsity characteristics of airborne radar clutter in the angle-Doppler domain, the spatial frequency and Doppler frequency are discretized into... and grid points (of which) and (Representing the scale factor). Then the clutter signal can be expressed as:

[0033] The The spacetime-guided dictionary can be represented as:

[0034] in Indicates the first Temporal steering vector of each grid, Indicates the first Spatial guidance vector of each grid, .

[0035] clutter signal In the angle-Doppler domain (by spacetime-guided dictionary) The complex amplitude on the (representation), i.e., the angle-Doppler image.

[0036] Specifically, step S2 includes: In this embodiment, the binary hypothesis for target detection under zero auxiliary data conditions can be expressed as:

[0037] in This indicates that the mean is 0 and the variance is . Complex white Gaussian noise.

[0038] Specifically, step S3 includes: In this embodiment, the data of the unit under test exist The conditional probability density function (PDF) under the assumption is:

[0039] The assumed PDF is:

[0040] according to and Under the given PDF conditions, the expression for the test statistic in this application embodiment can be written as:

[0041] in Indicates the detection threshold.

[0042] Then the target complex amplitude Maximum likelihood estimation for:

[0043] Taking the derivative and setting it to zero, we get The maximum likelihood estimate is:

[0044] Substitute The following is obtained from the PDF under the given conditions:

[0045] right Find the maximum likelihood estimate:

[0046] in, Represents the orthogonal projection matrix. Represents the identity matrix. Indicates the unit under test In vector Orthogonal projection on.

[0047] In this embodiment, and Under the assumption that maximizing the log-likelihood function to obtain the test statistic can be transformed into minimizing the sum of squared residuals, that is:

[0048]

[0049] Preferably, the linear problem can be simplified to a sparse estimation problem, that is, estimating under sparse constraints. and Therefore, minimizing the problem is equivalent to:

[0050] Solving the clutter angle-Doppler spectrum sparsity recovery minimization problem using the FISTA algorithm: In this embodiment, using The norm transforms the problem of minimizing the squared residuals under different hypothesis tests in the detector into a continuously differentiable convex function, introducing a residual term. and The unconstrained optimization problem is transformed into a constrained problem, and an augmented Lagrangian function is constructed. The clutter angle-Doppler spectrum estimation under both assumptions H and H is transformed into a sparse optimization problem, and the FISTA method is used to solve for the clutter angle-Doppler spectrum estimates under each hypothesis test.

[0051] Specifically, in Under the assumption that the sparse recovery optimization problem is solved based on the data of the cells to be detected and the normalized space-time guided dictionary:

[0052] Specifically, under the H1 assumption, the sparse recovery optimization problem is solved based on the data of the cells to be detected, the orthogonal projection matrix, and the spatiotemporal guidance dictionary:

[0053] Constructing a FISTA deep unfolded network; In this embodiment, as Figure 2 The model constructs a deep FISTA unfolded network, which consists of an input layer, an initialization layer, and 10 sequentially connected FISTA iterative unfolding layers. Each FISTA iterative unfolding layer includes a residual calculation layer, a gradient update layer, a momentum update layer, and a clutter angle-Doppler spectrum reconstruction residual calculation layer. The iteration step size and sparse regularization parameter in each FISTA iterative unfolding layer are set as learnable parameters.

[0054] In this embodiment, the initial clutter angle-Doppler spectrum under both the H0 and H1 assumptions is set to a zero vector. The initial momentum extrapolation variables for the two paths are set to the corresponding initial clutter angle-Doppler spectra. And set the initial acceleration parameters of FISTA to... .

[0055] Specifically, the clutter angle-Doppler spectrum reconstruction layer update formula can be expressed as:

[0056] in, and These represent the gradient update step sizes under the H0 and H1 hypotheses in the k-th FISTA expansion layer, respectively, which are used to control the gradient update magnitude during the gradient descent process of the current layer. and These represent the sparse regularization parameters for the corresponding paths, used to control the soft threshold shrinkage intensity.

[0057] Specifically, based on the Nesterov momentum formula, the momentum extrapolation variables are updated using the sparse coefficients of the current layer and the previous layer:

[0058]

[0059] The parameter update method is as follows:

[0060] The FISTA depth unfolding model is trained using the backpropagation algorithm, resulting in the trained FISTA depth unfolding model: In this embodiment, the FISTA deep unfolded network optimizes the network's learnable parameters through data-driven training.

[0061] Specifically, setting up the training dataset Given a network layer number K, the Normalized Root Mean Square Error (NMSE) is defined as the network loss function, and the optimal parameters are... The following equation can be solved using the backpropagation (BP) method:

[0062] Specifically, the expression is as follows:

[0063] in and Indicates , , , , As input, with The output of the clutter angle-Doppler spectrum update layer in the Kth layer of the FISTA-Net (FISTA Deep Unfolding Network) with parameters is given. The NMSE is calculated by multiplying the clutter angle-Doppler spectrum obtained after sparse recovery with the space-time steering dictionary, which effectively avoids the problems caused by the loss function calculation. and Excessive sparsity leads to gradient explosion.

[0064] Specifically, the loss function of the network can be written as:

[0065] in, and These represent the two hypothesis tests under the first... Sparse recovery loss on each sample.

[0066] Preferably, during backpropagation, the training processes of the two network paths are independent and can be optimized in parallel. When the NMSE loss is small, the difference between the sparse recovery result of the network output and the real signal is small. By reducing the NMSE value of the network and updating the network parameters, the estimation effect of the clutter angle-Doppler spectrum of the network is improved.

[0067] In this embodiment, after obtaining the optimal parameters through training, they can be used for target detection of actual unit signals. Regarding the detection data... Its clutter angle-Doppler spectrum is and The estimate under the assumptions can be expressed as:

[0068] in, , Indicates , and , and , and As input, with The output of the clutter angle-Doppler spectrum update layer is obtained from the Kth layer of the FISTA depth expansion network with parameters .

[0069] Data of the unit to be detected is acquired, and the clutter angle-Doppler spectrum of the unit to be detected is estimated by using a FISTA-based deep unfolding network. The sparsely recovered clutter angle-Doppler spectrum is output. The detection statistic of the unit to be detected is calculated based on the proposed detector, and the detection statistic is compared with the detection threshold. The presence of a target is determined based on the comparison result.

[0070] Specifically, step S4 includes: The pre-trained FISTA deep unfolding network model weights are loaded. To determine the detection threshold under a specified false alarm probability (Pfa), a Monte Carlo simulation method based on pure clutter (no target) data is adopted. A batch of radar echo data without targets is input into the trained FISTA deep unfolding network model, and the test statistic obtained from the clutter data according to the embodiments of this application is calculated. Based on the preset false alarm probability Pfa, the corresponding quantile in the statistic distribution is selected as the final detection threshold.

[0071] In this embodiment, the data of the unit to be detected is input into the trained FISTA depth unfolding model, and the test statistic for the current sample obtained according to the embodiments of this application is calculated according to the steps. The calculated test statistic is compared with a preset detection threshold to make a final decision. When the test statistic is greater than or equal to the detection threshold, it is determined that there is a target in the unit to be detected; when the test statistic is less than the detection threshold, it is determined that there is no target in the unit to be detected.

[0072] The proposed deep learning model-based assisted data filtering method is compared with the performance of traditional Generalized Likelihood Ratio Test (GLRT), Adaptive Matched Filter (AMF), and Persymmetric Generalized Likelihood Ratio Test (PS-GLRT) for unassisted target detection based on airborne radar simulation data. The signal-to-clutter-plus-noise ratio (SCNR) in the simulation experiments is defined as follows:

[0073] Specifically, for the simulation data, five protection units were set on each side of the unit to be detected, and several adjacent distance gate data outside the protection units were selected as auxiliary data. Based on the airborne radar simulation signal, the radar echo data was constructed into a two-dimensional space-time data matrix containing N array elements and M pulses. Simultaneously, targets were added according to the target guidance vector to complete the detection task. In all experiments, the simulation parameters used were as follows: clutter-to-noise ratio (CNR) was set to 30 dB, and the number of received pulses was... N The number of array elements is 10. M The grid number is 10, representing the number of grids in the spatial frequency and Doppler frequency directions. and All are set to 100. The target is located in the 101st range cell, and the target Doppler frequency is... Set to -0.18, target spatial frequency Set it to 0.3.

[0074] The detection probability was defined as the percentage of correctly detected targets out of the total number of targets. Performance was evaluated through a Monte Carlo experiment with 1000 independent trials. Since the GLRT, AMF, and PS-GLRT detectors all require auxiliary data to calculate the test statistic, detection and comparison experiments were conducted using K=101, K=101, and K=51 auxiliary data points, respectively.

[0075] The performance comparison results of the embodiments in this application with traditional generalized likelihood ratio test, adaptive matched filter, and skew-symmetric generalized likelihood ratio test detector are as follows: Figure 3 As shown, the detection probability of the embodiments of this application gradually increases with the increase of the signal-to-noise ratio (SNR), and outperforms the GLRT, AMF, and PS-GLRT methods. When the added target SNR is 29 dB, the detection probability of the embodiments of this application can reach 0.9. Compared with GLRT, DU-GLRT achieves a performance gain of approximately 2.7 dB; compared with AMF, it achieves a performance gain of approximately 2.8 dB; and compared with PS-GLRT, it achieves a gain of over 10 dB. This indicates that under the condition of achieving the same detection probability, the proposed method can detect weaker targets, demonstrating that the embodiments of this application can achieve effective target detection under zero auxiliary data conditions.

[0076] In summary, this embodiment expands the iterative solution process of FISTA into a deep neural network and unifies the clutter angle-Doppler spectrum estimation under the two hypothesis testing methods into the same network architecture. It fully integrates the interpretability of the model-driven method and the adaptive learning capability of the data-driven method. It constructs the test statistic using only the radar echo data of the unit to be detected, without relying on auxiliary data to estimate the clutter covariance matrix, and achieves effective target detection in environments with zero auxiliary data and strong non-uniform clutter.

[0077] Example 2 This invention also provides a zero-assistance data target detection system based on the FISTA deep unfolding network, such as... Figure 4 As shown, the system 10 includes: The data acquisition module 100 is used to acquire radar echo data, a highly complete space-time guidance dictionary, and target space-time guidance vector. The data extraction module 200 is used to extract the data of the unit to be detected from the radar echo data to form a training dataset; The network initialization module 300 is used to initialize the clutter angle-Doppler spectrum estimates, momentum variables, and acceleration parameters under the assumptions of target presence and target non-existence, and input the initialization results into the FISTA depth unfolding network. The training module 400 is used to preprocess the dataset and then train the FISTA deep unfolding network through the backpropagation algorithm to obtain the trained target detection model. The detection module 500 is used to input the data to be detected into the trained target detection model, output the calculated test statistic, compare the test statistic with a preset threshold, and determine whether the target is contained based on the comparison result.

[0078] Example 3 To implement the methods of the above embodiments, the present invention also provides a computer device, which includes a memory and a processor; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, so as to implement the various steps of the methods described above.

[0079] Example 4 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0081] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions 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 one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0082] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A zero-assistance data target detection method based on FISTA deep unfolded network, characterized in that, Includes the following steps: The radar echo data is acquired, and based on the sparse characteristics of the angle-Doppler domain of airborne radar clutter, the clutter in the radar echo is represented as the product of the space-time steering dictionary and the clutter angle-Doppler spectrum. Based on the product representation of the space-time guidance dictionary and the clutter angle-Doppler spectrum, a target detection model under zero-auxiliary data conditions is constructed. The target detection model is a binary hypothesis testing model, corresponding to the target existence hypothesis and the target non-existence hypothesis, respectively. Based on the binary hypothesis testing model, a test statistic is constructed. The clutter angle-Doppler spectrum estimates under the target presence hypothesis and the target non-existence hypothesis are solved using the FISTA deep unfolding network. Based on the clutter angle-Doppler spectrum estimates, the detection statistic of the unit to be detected is calculated. The detection statistics are compared with a preset detection threshold, and the presence of a target is determined based on the comparison result. The detection threshold is determined by the detection statistics of the pure clutter detection unit and the preset false alarm rate.

2. The method according to claim 1, characterized in that, The FISTA deep unfolded network includes: The input layer is used to receive radar echo data, an overcomplete space-time guidance dictionary, and a target space-time guidance vector from the unit to be detected, and to construct a target orthogonal projection matrix based on the target space-time guidance vector. The initialization layer is used to initialize the clutter angle-Doppler spectrum estimates, momentum variables, and acceleration parameters under the assumptions of target presence and target non-existence, respectively. The FISTA iterative expansion layers are connected sequentially with a preset number of layers. Each FISTA iterative expansion layer is used to calculate the clutter angle-Doppler spectrum reconstruction residual under two assumptions, update the momentum variable by gradient descent, and perform soft thresholding on the gradient update result to obtain the clutter angle-Doppler spectrum estimate of the current layer. The momentum is extrapolated based on the difference between the current layer and the previous layer estimate to obtain the momentum variable required for the next iteration. After completing the preset number of iterations, the clutter angle-Doppler spectrum estimation results under the two assumptions are output.

3. The method according to claim 2, characterized in that, The iteration step size and sparse regularization parameters in the FISTA iterative unrolling layer are set as learnable parameters and are trained and optimized using the backpropagation algorithm.

4. The method according to claim 1, characterized in that, The method of using the FISTA depth unfolding network to solve for clutter angle-Doppler spectrum estimates under two assumptions includes: By introducing a residual term, the problem of minimizing the squared residuals in the detector is transformed into a sparse optimization problem. The FISTA algorithm is used to solve the clutter angle-Doppler spectrum estimates under the assumptions of target presence and target absence, respectively.

5. The method according to claim 1, characterized in that, The step of comparing the detection statistics with a preset detection threshold and determining whether a target exists based on the comparison result includes: If the detection statistic is greater than or equal to the detection threshold, it is determined that a target exists in the unit to be detected. If the detection statistic is less than the detection threshold, it is determined that there is no target in the unit to be detected.

6. A zero-assistance data target detection system based on FISTA deep unfolded network, used to implement the method according to any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire radar echo data, a fully complete space-time guidance dictionary, and target space-time guidance vector; The data extraction module is used to extract the data of the unit to be detected from the radar echo data to form a training dataset; The network initialization module is used to initialize the clutter angle-Doppler spectrum estimates, momentum variables, and acceleration parameters under the assumptions of target presence and target non-existence, and input the initialization results into the FISTA depth unfolding network. The training module is used to preprocess the dataset and then train the FISTA deep unfolded network using the backpropagation algorithm to obtain the trained target detection model. The detection module is used to input the data to be detected into the trained target detection model, output the calculated test statistic, compare the test statistic with a preset threshold, and determine whether the target is contained based on the comparison result.

7. A computer device, characterized in that, Including processor and memory; The processor reads the executable program code stored in the memory to run the program corresponding to the executable program code, so as to implement the zero-assistance data target detection method based on FISTA deep unfolding network as described in any one of claims 1-5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a zero-assistance data target detection method based on the FISTA deep unfolding network as described in any one of claims 1-5.

Images