A radar extended target detection method based on model-driven deep neural network

Abstract

The application provides a radar extended target detection method based on a model-driven deep neural network, and belongs to the technical field of radar target detection. Firstly, under a likelihood ratio detection criterion, the radar extended target detection problem is modeled as an optimization problem of binary variables; then a deep unfolding network driven by model information and data information is constructed, offline training is performed, and a decision threshold is set according to a false alarm probability requirement; finally, in the online detection stage, real-time decision making is performed, so that target detection is realized. The method simultaneously utilizes model information and data information, and has stronger generalization ability. Taking a detection probability as an evaluation performance index, simulation experiments show that the application has higher detection performance than existing typical detection methods, and can guarantee constant false alarm characteristics, and is suitable for practical application.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target detection, and in particular relates to a radar extended target detection method in a non-Gaussian clutter environment. Background Art

[0002] Radar detection performance is primarily limited by clutter, noise, and other interference. Especially for high-resolution radar systems, the statistical characteristics of clutter deviate from a Gaussian distribution, exhibiting distinct non-Gaussian characteristics. Furthermore, when the radar's range resolution unit is smaller than the target size, the target echo signal occupies multiple range units, a situation known as a distributed target or extended target. For extended-range targets, traditional point target detection methods suffer from shadowing effects from adjacent range units and cannot effectively accumulate energy across multiple range units, leading to degraded detection performance or even complete failure. Therefore, improving radar extended-target detection performance in the presence of non-Gaussian clutter is crucial.

[0003] Many research institutions at home and abroad have conducted research on radar extended target detection methods in non-Gaussian clutter. Common detector design criteria include the generalized likelihood ratio (GLRT) criterion and the Wald and Rao detection criteria. For example, for compound Gaussian clutter, the U.S. Naval Research Laboratory designed the SDD-GLRT algorithm using density information of target scattering points (K. Gerlach, “Spatially distributed target detection in non-Gaussian clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. 35, no. 3, pp. 926-934, July 1999). The Naval University of Aeronautics and Astronautics designed the OS-GLRT detection algorithm using ordered statistics under the GLRT detection criterion. (Y.He, T.Jian, F.Su, et al., "Novel range-spread target detectors in non-Gaussian clutter," IEEE Trans.Aerosp.Electron.Syst., vol.46, no.3, pp.1312-1328, 2010.)

[0004] Most of the aforementioned detection methods are designed based on the GLRT criterion. Although GLRT has a theoretical basis, it lacks a uniform maximum potential and therefore does not possess an optimal structure. Furthermore, most existing detection algorithms are based on statistical theory, and their performance is heavily dependent on the statistical properties of the target and clutter. In real-world scenarios, when the pre-defined target / clutter statistical properties do not match the actual situation, the detection performance of these methods degrades significantly. Therefore, designing detection algorithms with improved performance and robustness based on alternative strategies is of great value in the field of radar target detection.

[0005] Deep learning technology, with its powerful data processing capabilities and ability to learn the valuable information contained in data, is gradually being applied to radar target detection, achieving some success. However, existing deep neural networks are mostly "black-box" processing, lacking interpretability. Furthermore, conventional deep networks involve a large number of parameters to learn, requiring vast amounts of training data, and their performance is heavily data-dependent. In the field of radar signal processing, particularly for non-cooperative targets, labeled data containing the target is limited. Therefore, it is worthwhile to further investigate how to combine statistical signal processing theory with deep learning techniques to design new radar extended target detection algorithms using small sample sizes. Summary of the Invention

[0006] To address the above issues, the present invention provides a radar extended target detection method based on a model-driven deep neural network, suitable for non-Gaussian clutter environments. First, under the LRT detection criterion, the radar extended target detection problem is modeled as an optimization problem of binary variables. Then, a deep extended network driven by both model and data information is constructed for offline training, with a decision threshold set based on the false alarm probability requirement. Finally, during the online detection phase, a real-time decision is made to determine whether the target signal is present, thereby achieving target detection.

[0007] The technical solutions of the present invention are as follows:

[0008] A radar extended target detection method based on a model-driven deep neural network, the method comprising the following steps:

[0009] Step 1: Under the likelihood ratio detection criterion, the radar extended target detection problem is modeled as an optimization problem;

[0010] The application scenario is a distributed MIMO radar with M transmitting elements and K receiving elements. This means there are M-M paths from the transmitter to the receiver. Each antenna sends L pulses within a coherent processing interval, and each antenna on the transmitter sends mutually orthogonal waveforms.

[0011] Assume that the target echo signal occupies H range units, let y mk,hRepresents the received signal vector of the mk-th path and the h-th range unit. The radar extended target detection under non-Gaussian clutter is expressed as:

[0012]

[0013] Among them, α mk,h represents the sum of target scattering and channel propagation effects at the hth range unit of the mkth path;

[0014] p mk is the corresponding target-oriented vector, denoted as p mk =[1,exp(j2πf mk T r ),…,exp(j2π(L-1)f mk T r )] T ,f mk is the target Doppler shift, T r is the pulse repetition time, c mk,h is the clutter vector, which is generally modeled using a composite Gaussian model and expressed as a slowly varying component τ mk,h and a rapidly varying component g mk,h The product of where g mk,h It obeys a complex Gaussian distribution with a mean of zero and a variance of Σ. H0 means there is no target, and H1 means there is a target.

[0015] For the above detection problem, the likelihood ratio detection criterion is:

[0016]

[0017] Among them, η is the decision threshold based on the likelihood ratio detection criterion, α mk represents the sum of target scattering and channel propagation effects of the mkth path, τ mk represents the slowly varying component of the mkth path, y mk represents the received signal vector of the mk-th path;

[0018] Introducing a discrete binary variable ω∈{0,1}, the above detection criterion is expressed as:

[0019]

[0020] Among them, ω = 0 means the target does not exist, and ω = 1 means the target exists;

[0021] Step 2: Build a deep unfolding network driven jointly by model information and data

[0022] To solve the above optimization problem, the conventional projected gradient algorithm is expanded into a deep neural network. Each expanded layer in the network is a fully connected neural network structure, consisting of an input layer, a hidden layer, and an output layer.

[0023] The input of the t-th layer network is:

[0024]

[0025] in, is the estimated value of ω by the previous layer network, It is the dimension-raising vector of the t-th layer network used to increase the dimension of the input layer. is the optimization objective function f b The gradient with respect to ω;

[0026] The hidden layer of the t-th layer network is expressed as:

[0027] h t =f ReLU (W1(t)x(t)+b1(t))

[0028] where f ReLU (·) represents the nonlinear activation function of the hidden layer;

[0029] The output layer of the t-th layer network is expressed as:

[0030]

[0031] where f sigmoid (·) is a nonlinear activation function, {W1(t),b1(t),W2(t),b2(t),W3(t),b3(t)} are the network parameters of the tth layer;

[0032] Step 3: Set the decision threshold

[0033] First, the target-free data in the training set is preprocessed, then input into the trained network, and the network output results are sorted in descending order. Finally, the decision threshold is determined according to the false alarm probability.

[0034] Step 4: Online detection and judgment

[0035] The test set is preprocessed and input into the trained network, and the judgment is made according to the following formula;

[0036]

[0037] If the network output value is greater than the threshold, it is judged that the target exists, otherwise it is judged that it does not exist.

[0038] Furthermore, the decision threshold set in step 3 in After the non-target data in the training set is input into the network and sorted in descending order, Results, P fa is the preset false alarm probability.

[0039] The beneficial effects of the present invention are:

[0040] The present invention proposes a radar extended target detection method based on a model-driven deep neural network. The method models the radar extended target detection problem as a minimization problem that satisfies binary constraints, and constructs a deep expansion network to determine whether the target signal exists. Compared with the conventional deep neural network mentioned in the technical background, the network is interpretable and can converge quickly under small sample data. Compared with the traditional detection method based on statistical models, this method utilizes both model information and data information and has stronger generalization ability. Taking detection probability as the performance evaluation index, simulation experiments show that in a non-Gaussian clutter environment, the present invention has higher detection performance than existing detection methods, and can ensure constant false alarm characteristics, which is suitable for practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The network architecture diagram for the tth layer of the deep expansion network is shown below.

[0042] Figure 2 Flowchart of the detection method based on model-data joint driven deep expansion network.

[0043] Figure 3 Radar system configuration diagram used in simulation experiments.

[0044] Figure 4 Detection probability of the present invention and the classic detection algorithm varies with the signal-to-noise ratio curve.

[0045] Figure 5 This is the curve showing the change of detection probability versus false alarm probability for the present invention and the classic detection algorithm. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0047] A radar extended target detection method based on a model-driven deep neural network includes the following steps:

[0048] 1. Model the radar extended target detection problem as an optimization problem

[0049] For a distributed MIMO radar equipped with M transmitting elements and K receiving elements, both the transmitting and receiving ends are configured with antennas that are far apart. Assume that each antenna sends L pulses within a coherent processing interval, and that each transmitting antenna transmits mutually orthogonal waveforms.

[0050] Assume that the target echo signal occupies H range units, let y mk,h Denotes the received signal vector of the mkth path and the hth range unit. The radar extended target detection problem under non-Gaussian clutter is expressed as

[0051]

[0052] where α mk,h reflects the target scattering and channel propagation effects of the mkth path, p mk is the corresponding steering vector, expressed as p mk =[1,exp(j2πf mk T r ),…,exp(j2π(L-1)f mk T r )] T ,f mk is the target Doppler shift, T r is the pulse repetition time. c mk,h is the clutter vector, which is generally modeled using a composite Gaussian model and expressed as a slowly varying component τ mk,h and a rapidly varying component g mk,h The product of where g mk,h It follows a complex Gaussian distribution with mean zero and variance Σ.

[0053] For the above detection problem, the likelihood ratio detection criterion is expressed as

[0054]

[0055] Where η is the decision threshold based on the likelihood ratio detection criterion.

[0056] Introducing a discrete binary variable ω∈{0,1}, the above detection criterion is expressed as

[0057] The scattering coefficient of the target signal and the texture component of the clutter are estimated using the maximum likelihood estimation method. The above problem can be further equivalently expressed as

[0058]

[0059] Here, ω=0 indicates that the target does not exist, and ω=1 indicates that the target exists.

[0060] 2. Build a deep unfolding network

[0061] For the above optimization problem, the conventional projected gradient algorithm is expanded into a deep neural network. Figure 1As shown in the figure, each expanded layer in the network is a fully connected neural network structure, consisting of an input layer, a hidden layer, and an output layer.

[0062] The input of the t-th layer network is recorded as

[0063]

[0064] in is the estimated value of ω by the previous layer network, is the raw dimension vector of the t-th layer network used to increase the dimension of the input layer. is the optimization objective function f b The gradient with respect to ω.

[0065] The hidden layer of the t-th network is expressed as

[0066] h t =f ReLU (W1(t)x(t)+b1(t))

[0067] where f ReLU (·) represents the nonlinear activation function of the hidden layer.

[0068] The output layer of the t-th layer network is expressed as

[0069]

[0070] where f sigmoid (·) is a nonlinear activation function, and {W1(t),b1(t),W2(t),b2(t),W3(t),b3(t)} are the network parameters of the tth layer.

[0071] 3. Set the decision threshold

[0072] like Figure 2 As shown, let is the untargeted data in the training set, Preprocessing to obtain Input it into the trained network, and the output result is recorded as Ψ={ω z ,z=1,…,Z}. Sort the network output results in descending order to obtain According to the false alarm probability P fa , take the The detection statistic is used as the decision threshold, which is expressed as

[0073] 4. Online detection and judgment

[0074] like Figure 2 As shown in the figure, in the online detection stage, the test set is preprocessed, and then the test set is input into the trained network and judged according to the following formula.

[0075]

[0076] If the network output value is greater than the threshold in step 3, it is judged that the target exists, otherwise it is judged that it does not exist. Specific embodiment:

[0078] Parameter settings: such as Figure 3 As shown, a specific embodiment of the present invention is implemented in a distributed MIMO radar system. Assume that the number of transmitting antenna elements and the number of receiving antenna elements are both 2, wherein the angles between the two transmitting antennas and the target are 0° and 65°, and the angles between the two receiving antennas and the target are -30° and 40°. Assume that the number of pulses sent within a coherent processing interval is M=10, the repetition frequency of the transmitting pulse is 500Hz, the carrier frequency is set to 1GHz, the moving speed of the measured target is 108km / h, and the number of distance units occupied by the target echo signal is 6. The statistical characteristics of the clutter amplitude are modeled as K distribution, the texture component obeys the gamma distribution, the speckle component obeys the mean zero, and the covariance matrix is ​​Σ=Σ0+I M The complex Gaussian distribution of , and Σ0 is modeled as an exponential form,

[0079]

[0080] Where ρ is the first-order delay correlation coefficient of the clutter, which is set to ρ = 0.9; is the noise ratio, set to 10dB; f dc is the clutter normalized Doppler frequency, set to 0.05.

[0081] Example 1:

[0082] Assume the signal-to-clutter ratio (SCR) is -21dB to 0dB, the false alarm probability is 0.001, and the target echo signal energy is evenly scattered in each range unit, and the target scattering amplitude follows a Gaussian distribution. Figure 4 The detection probability of the existing technology and the present invention changes with the SCR. Figure 4 As shown in the figure, the technology of the present invention has a higher detection probability and significantly improves the radar extended target detection performance in non-Gaussian clutter environment.

[0083] Example 2:

[0084] Assume that the false alarm probability is 10 -3 The signal-to-noise ratio (SCR) is set to -12dB. Figure 5 The results show that the detection probability of the existing technology and the present invention varies with the false alarm probability. The results show that the detection probability of the present invention is significantly improved under different false alarm probabilities.

[0085] Example 3:

[0086] Table 1 below shows the actual false alarm probability of the present technology and the change of the preset false alarm probability under different clutter shape parameters, where the clutter shape parameters are set to 1 and 5, and the preset false alarm probability is 10 -5 ,10 -4 ,10 -3 ,10 -2 As shown in the results in the table, the actual false alarm probability of the present technology is almost the same as the setting. In addition, the false alarm probability changes slightly under different clutter shape parameters, which indicates that the present technology can guarantee constant false alarm probability with respect to the shape parameter of the clutter.

[0087] Table 1

[0088]

Claims

1. A radar extended target detection method based on a model-driven deep neural network, the method comprising the following steps: Step 1: Under the likelihood ratio detection criterion, the radar extended target detection problem is modeled as an optimization problem; The application scenario is a distributed MIMO radar with M transmitting elements and K receiving elements. This means there are M-M paths from the transmitter to the receiver. Each antenna sends L pulses within a coherent processing interval, and each antenna on the transmitter sends mutually orthogonal waveforms. Assume that the target echo signal occupies H range units, let y mk,h Represents the received signal vector of the mk-th path and the h-th range unit. The radar extended target detection under non-Gaussian clutter is expressed as: Among them, α mk,h represents the sum of target scattering and channel propagation effects at the hth range unit of the mkth path; p mk is the corresponding target-oriented vector, denoted as p mk =[1,exp(j2πf mk T r ),…,exp(j2π(L-1)f mk T r )] T ,f mk is the target Doppler shift, T r is the pulse repetition time, c mk,h is the clutter vector, which is generally modeled using a composite Gaussian model and expressed as a slowly varying component τ mk,h and a rapidly varying component g mk,h The product of where g mk,h It obeys a complex Gaussian distribution with a mean of zero and a variance of Σ. H0 means there is no target, and H1 means there is a target. For the above detection problem, the likelihood ratio detection criterion is: Among them, η is the decision threshold based on the likelihood ratio detection criterion, α mk represents the sum of target scattering and channel propagation effects of the mkth path, τ mk represents the slowly varying component of the mkth path, y mk represents the received signal vector of the mk-th path; Introducing a discrete binary variable ω∈{0,1}, the above detection criterion is expressed as: Among them, ω = 0 means the target does not exist, and ω = 1 means the target exists; Step 2: Build a deep unfolding network driven jointly by model information and data To solve the above optimization problem, the conventional projected gradient algorithm is expanded into a deep neural network. Each expanded layer in the network is a fully connected neural network structure, consisting of an input layer, a hidden layer, and an output layer. The input of the t-th layer network is: in, is the estimated value of ω by the previous layer network, It is the dimension-raising vector of the t-th layer network used to increase the dimension of the input layer. is the optimization objective function f b The gradient with respect to ω; The hidden layer of the t-th layer network is expressed as: h t =f ReLU (W1(t)x(t)+b1(t)) where f ReLU (·) represents the nonlinear activation function of the hidden layer; The output layer of the t-th layer network is expressed as: where f sigmoid (·) is a nonlinear activation function, {W1(t),b1(t),W2(t),b2(t),W3(t),b3(t)} are the network parameters of the tth layer; Step 3: Set the decision threshold First, the targetless data in the training set is preprocessed, then input into the trained network, and the network output results are sorted in descending order. Finally, the decision threshold is determined according to the false alarm probability. Step 4: Online detection and judgment The test set is preprocessed and input into the trained network, and the judgment is made according to the following formula; If the network output value is greater than the decision threshold, it is judged that the target exists, otherwise it is judged that it does not exist.

2. The radar extended target detection method based on a model-driven deep neural network according to claim 1, characterized in that: The decision threshold set in step 3 in After the non-target data in the training set is input into the network and sorted in descending order, Results, P fa is the preset false alarm probability.

Images