Airborne radar moving target detection method and system based on intelligent sample selection

Abstract

The invention discloses an airborne radar moving target detection method and system based on intelligent sample selection, and the method comprises the steps: obtaining space-time echo signals of radar distance units, including space-time echo signals of different distance units and space-time echo signals of uniform distance units; constructing a training data set through the space-time echo signals of the uniform distance units; constructing a complex memory enhanced auto-encoder, and inputting the training data set into the complex memory enhanced auto-encoder to realize off-line training of the auto-encoder; inputting space-time echo signals of different distance units into the trained complex memory enhanced auto-encoder to obtain a reconstruction error; and selecting uniform distance units according to the reconstruction error, and detecting and removing polluted distance units. According to the method, the better sample selection performance can be obtained under the condition of reasonable operation complexity, and the target detection performance is better.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing and deep learning technology, and in particular to an airborne radar moving target detection method and system based on intelligent sample selection. Background Art

[0002] In order to design space-time filters for adaptive clutter suppression and detection of low-altitude or ground moving targets, the traditional airborne radar space-time adaptive processing (STAP) method usually requires a large number of independent and identically distributed (IID) training distance units (i.e., samples) to estimate the clutter plus noise covariance matrix (CNCM) of the range unit to be tested. In practice, some training range units may be contaminated by interference signals, resulting in the degradation of STAP performance and the omission or cancellation of moving targets. To address this problem, a non-homogeneous detector (NHD) was proposed for STAP, which uses specific indicators to detect and eliminate training range units contaminated by interference signals, and only selects uniform range unit data for processing.

[0003] At present, the most typical NHD is GIP-NHD based on the generalized inner product (GIP) criterion. When the GIP statistic of a specific distance unit deviates significantly from the GIP average of all available distance units, it can be regarded as a contaminated distance unit. However, when the number of available training distance units for covariance estimation is insufficient or contains significant outliers, the performance of GIP-NHD will drop sharply. Based on the diagonal loading technology, the LGIP-NHD method can perform covariance estimation under the condition of a small number of distance units for uniform distance unit selection. However, how to set the appropriate diagonal loading in practical applications is still an unresolved problem. The GIP-NHD method based on cross-spectral metric smoothing (CSMS) can reduce the impact of significant outliers on covariance estimation, but this method requires accurate preset of the steering vector of the interference signal, otherwise the mismatched covariance matrix will be estimated.

[0004] Therefore, it is an urgent problem for technical personnel in this field to propose an airborne radar moving target detection method and system based on intelligent sample selection to solve the difficulties existing in the prior art. Summary of the invention

[0005] The purpose of the present invention is to provide an airborne radar moving target detection method and system based on intelligent sample selection, which can obtain better sample selection performance and better target detection performance under the condition of reasonable computational complexity.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for detecting moving targets by airborne radar based on intelligent sample selection comprises the following steps:

[0008] S1. Acquire space-time echo signals of radar range units, including space-time echo signals of different range units and space-time echo signals of uniform range units;

[0009] S2, constructing a training data set through space-time echo signals of uniform distance units;

[0010] S3, constructing a complex memory enhanced autoencoder, and inputting the training data set into the complex memory enhanced autoencoder to implement offline training of the autoencoder;

[0011] S4, inputting the space-time echo signals of different distance units into the trained complex memory enhanced autoencoder to obtain the reconstruction error;

[0012] S5. Select uniform distance units according to the reconstruction error, and detect and remove contaminated distance units.

[0013] Preferably, the complex memory enhanced autoencoder constructed in S3 includes three processing steps, namely encoding, reconstruction and decoding.

[0014] Preferably, in the reconstruction step, the memory module reconstructs the encoded output through the sparse memory items It is expressed as:

[0015]

[0016] Where M R ∈R R×Q and M I ∈R R×Q Represent the real and imaginary parts of the memory cell, R and Q represent the number of rows and columns of the memory cell, respectively. R +jw I ∈C 1×R is a sparse weight vector, M = M R +jM I The row vector of represents the memory item, where the rth row is represented as

[0017] Preferably, the autoencoder training in S3 includes:

[0018] For the space-time echo signal xg , the reconstruction error in the complex memory enhanced autoencoder is expressed as:

[0019]

[0020] In the formula, For x g The normalized signal, for The normalized signal reconstructed by the complex memory enhanced autoencoder, Θ CM ={Θ En ,Θ Mem ,Θ De} is the network parameter, Θ En is the encoding parameter, Θ De is the decoding parameter, Θ Mem ={M R ,M I} is the reconstruction parameter;

[0021] The loss function is defined as:

[0022]

[0023] In the formula, G represents the size of the training dataset.

[0024] Optimal parameters for complex memory-augmented autoencoders It is obtained by back propagation method and is expressed as:

[0025]

[0026] Preferably, the uniform distance unit selection index in S5 is expressed as:

[0027]

[0028] In the formula, E Rc (x l ) represents the space-time echo signal x of the lth training range unit l The reconstruction error of is the normalized signal of the lth training distance unit, The optimal parameters for the complex memory enhanced autoencoder after training The reconstructed normalized signal is Rc (x l )<η, the lth training distance unit is a uniform sample; when E Rc (x l )≥η, the lth training distance unit is a non-uniform sample, and η is the detection threshold of the non-uniform distance unit.

[0029] Preferably, the method also includes the basic principle of using complex memory enhanced autoencoder for non-uniform sample detection:

[0030] In the offline training stage, the memory module of the complex memory enhanced autoencoder reconstructs the encoding output of the training data set based on the sparse addressing operation, and updates the parameters of the memory unit to ensure that the training data set is reconstructed in the complex memory enhanced autoencoder; in the online application stage, the parameters of the memory unit are fixed, and reconstruction errors occur because the encoding output of the non-uniform sample cannot obtain matching memory items from the memory unit.

[0031] An airborne radar moving target detection system based on intelligent sample selection, applied to any of the above-mentioned airborne radar moving target detection methods based on intelligent sample selection, comprising a data acquisition module, a training set construction module, an autoencoder training module, a reconstruction error acquisition module, and a distance unit selection module connected in sequence; wherein:

[0032] A data acquisition module, used to acquire space-time echo signals of radar range units, including space-time echo signals of different range units and space-time echo signals of uniform range units;

[0033] A training set construction module, used for constructing a training data set through space-time echo signals of uniform distance units;

[0034] An autoencoder training module is used to construct a complex memory enhanced autoencoder and input the training data set into the complex memory enhanced autoencoder to realize offline training of the autoencoder;

[0035] A reconstruction error acquisition module is used to input the space-time echo signals of different distance units into the trained complex memory enhanced autoencoder to obtain the reconstruction error;

[0036] The range unit selection module is used to select uniform range units according to the reconstruction error and detect and remove contaminated range units.

[0037] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned methods for detecting moving targets by airborne radar based on intelligent sample selection.

[0038] According to the specific embodiment provided by the present invention, the present invention discloses the following technical effects: the complex memory-augmented autoencoder constructed by the present invention forms a non-uniform detector based on deep learning (DL), called the DL-NHD intelligent sample selection method, which is used for airborne radar clutter suppression and moving target detection; first, a complex memory-augmented autoencoder (CV-MemAE) is constructed; then, the CV-MemAE network is trained based on a data set constructed only from space-time echo signals of uniform distance units; finally, according to the reconstruction error of space-time echo signals of different distance units in the trained CV-MemAE, the contaminated training distance units are detected and eliminated. Compared with the existing typical GIP-NHD method, the DL-NHD method can obtain better sample selection performance and better target detection performance under the condition of reasonable computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0040] Figure 1 A schematic flow chart of the method for detecting moving targets by airborne radar based on intelligent sample selection provided by the present invention;

[0041] Figure 2 It is the network structure diagram of CV-MemAE of the present invention;

[0042] Figure 3 The non-uniform distance unit detection result diagram of different methods;

[0043] Among them, 3a is the non-uniform distance unit detection result of the GIP-NHD method, 3b is the non-uniform distance unit detection result of the LGIP-NHD method, 3c is the non-uniform distance unit detection result of the CSMS-GIP-NHD method; 3d is the non-uniform distance unit detection result of the DL-NHD method;

[0044] Figure 4 The non-uniform distance unit detection results of different methods when the training distance unit is insufficient;

[0045] Among them, 4a is the non-uniform distance unit detection result of the GIP-NHD method when the training distance unit is insufficient, 4b is the non-uniform distance unit detection result of the LGIP-NHD method when the training distance unit is insufficient, 4c is the non-uniform distance unit detection result of the CSMS-GIP-NHD method when the training distance unit is insufficient, and 4d is the non-uniform distance unit detection result of the DL-NHD method when the training distance unit is insufficient;

[0046] Figure 5 This is the non-uniform distance unit detection result diagram of LGIP-NHD under unreasonable diagonal loading;

[0047] Among them, 5a is a non-uniform distance unit detection result diagram when the diagonal loading μ changes from -3dB to -10dB, and 5b is a non-uniform distance unit detection result diagram when the diagonal loading μ changes from -3dB to -15dB;

[0048] Figure 6 The non-uniform distance unit detection results of different methods under strong interference conditions;

[0049] Among them, 6a is the non-uniform distance unit detection result of the GIP-NHD method under strong interference conditions, 6b is the non-uniform distance unit detection result of the LGIP-NHD method under strong interference conditions, 6c is the non-uniform distance unit detection result of the CSMS-GIP-NHD method under strong interference conditions, and 6d is the non-uniform distance unit detection result of the DL-NHD method under strong interference conditions;

[0050] Figure 7 The signal to clutter-plus-noise ratio (SCNR) loss diagram of each NHD method under different conditions;

[0051] Among them, 7a is Figure 3 The SCNR loss of each method is 7b. Figure 4 The SCNR loss of each method is 7c. Figure 6 SCNR loss of each method;

[0052] Figure 8 The relationship between the computational complexity of different NHD methods and the number of distance units. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] The problem of airborne radar moving target detection is simplified to a binary hypothesis testing problem as shown below:

[0056]

[0057] Where N represents the number of airborne radar array elements, K represents the number of pulses in the coherent processing interval, x represents the space-time echo signal of the distance unit to be measured, and x c represents the clutter signal, x n represents the noise signal, ξ and s T They represent the complex amplitude and space-time steering vector of the moving target respectively.

[0058] The typical STAP method usually requires the radar space-time signals of multiple IID training range cells to estimate the CNCM of the range cell to be measured, so as to suppress clutter and noise and perform target detection. To avoid missed target detection, the radar space-time signals of the training range cells should obey the assumption H0, that is, only contain clutter and noise. However, in practice, some training range cells may obey the assumption H1, that is, they also contain interference signals similar to the target. Therefore, a key point in the application of STAP in practice is to eliminate the range cells polluted by interference (i.e., non-uniform range cells or non-uniform samples) from the available training range cells and only select uniform range cells. The classic NHD method (such as the GIP-NHD method) uses the statistical characteristics of the training range cells to achieve this goal. Although it may achieve better results in specific cases, its performance usually depends on the number and quality of available training range cells.

[0059] like Figure 1 As shown, the present invention provides an airborne radar moving target detection method based on intelligent sample selection, comprising the following steps:

[0060] S1. Acquire space-time echo signals of radar range units, including space-time echo signals of different range units and space-time echo signals of uniform range units;

[0061] S2, constructing a training data set through space-time echo signals of uniform distance units;

[0062] S3. Construct a complex memory enhanced autoencoder (CV-MemAE), and input the training data set into the complex memory enhanced autoencoder to realize offline training of the autoencoder;

[0063] S4, inputting the space-time echo signals of different distance units into the trained complex memory enhanced autoencoder to obtain the reconstruction error;

[0064] S5. Select uniform distance units according to the reconstruction error, and detect and remove contaminated distance units.

[0065] Specifically, the processing framework of the non-uniform sample detection method (i.e., intelligent sample selection method) based on deep learning of the present invention is as follows: Figure 1 As shown, the normalization module N(·) performs normalization processing on the radar space-time signal x and outputs x N =x / ||x||2 (the purpose of normalization is to avoid the influence of signal power of units at different distances and improve network convergence), which is used as the input of the data reconstruction network CV-MemAE; CV-MemAE outputs the reconstructed normalized signal Mean Square Error Loss (MSE Loss) module calculation and x N The error between them is used as a detection indicator for non-uniform samples.

[0066] The basic idea of ​​the method of the present invention is to train a DL network so that it can only accurately reconstruct the space-time echo signals from uniform distance units (i.e., data within the distribution), so as to use the reconstruction error of the space-time echo signals from the contaminated distance units (i.e., data outside the distribution) in the trained DL network as a detection indicator. Therefore, the processing flow of the method of the present invention can be summarized as follows: 1) Offline training of CV-MemAE based on a data set constructed only from space-time echo signals of uniform distance units; 2) Input space-time echo signals of different distance units into the trained CV-MemAE to obtain the reconstruction error; 3) Select uniform distance units according to the reconstruction error, and detect and eliminate contaminated distance units.

[0067] 1. CV-MemAE Network

[0068] As a typical DL network, autoencoder (AE) can learn effective features from unlabeled data sets with specific distribution. Ideally, AE has a low reconstruction error only for data within the distribution. However, due to its strong generalization ability, AE can sometimes reconstruct certain out-of-distribution data more accurately. By adding a unique memory module (including a memory unit and a sparse addressing unit based on an attention mechanism), the memory-augmented autoencoder (MemAE) can effectively solve the above problems. However, the existing MemAE can only be used for processing real-valued data. In order to realize STAP non-uniform sample detection, the present invention constructs a complex memory-augmented autoencoder MemAE (Complex-ValuedMemAE, CV-MemAE) suitable for airborne radar space-time signal processing, and its network structure is as follows Figure 2 shown.

[0069] The basic principle of CV-MemAE for out-of-distribution data detection is as follows: 1) In the offline training stage, the memory module of CV-MemAE reconstructs the encoded output of the in-distribution data based on sparse addressing operations, and continuously updates the parameters of the memory unit to ensure that the in-distribution data is accurately reconstructed in CV-MemAE; 2) In the online application stage, the parameters of the memory unit are fixed. Since the encoded output of the out-of-distribution data cannot obtain matching memory items from the memory unit, a high reconstruction error will be generated. Figure 2 It can be seen that CV-MemAE mainly includes three processing steps, namely encoding, reconstruction and decoding, as shown below:

[0070] (1) Coding

[0071] like Figure 2 As shown in Figure 1, the encoder of CV-MemAE is composed of E cascaded complex-valued linear (CV-Linear) layers, and there is a complex-valued batch normalization (CV-BN) layer between adjacent CV-Linear layers.

[0072] The parameters of the encoder are and denote the real and imaginary parts of the e-th weight matrix, respectively. and denote the real and imaginary parts of the e-th bias vector, respectively, Q e-1 and Q e Respectively represent the input dimension and output dimension of the e-th layer, satisfying Q0=P and Q E =Q, P is the dimension of the encoder input signal, Q is the dimension of the encoder output signal. Given input u = u R +ju I ∈C 1×P And the encoder transformation function is expressed as F En (·), the output of the encoder can be expressed as:

[0073] v=F En (Θ En ,u)∈C 1×Q

[0074] Specifically, the operations of each layer of the encoder are:

[0075] CV-Linear-(1) is the input layer, and its output can be expressed as:

[0076]

[0077] CV-Linear-(e) is a hidden layer, where e = 2, 3, L, E-1, and its input is the output of the previous layer, that is, Its output can be expressed as:

[0078]

[0079] CV-Linear-(E) is the output layer, and its output can be expressed as:

[0080]

[0081] With E(·) and Var(·) representing the expectation and variance, the output of the e-th (e=1,2,...,E-1) CV-BN layer can be expressed as:

[0082]

[0083] (2) Reconstruction

[0084] The parameter of the CV-MemAE memory module is Θ Mem ={M R ,M I},M R ∈R R×Q and M I ∈R R×Q Represent the real and imaginary parts of the memory unit, M=M R +jM I The row vector of represents the memory item, which is expressed as Where r = 1, 2, ..., R. The memory module reconstructs the encoded output through sparse memory items It is expressed as:

[0085]

[0086] Where w = w R +jw I ∈C 1×R is a sparse weight vector, whose rth element is

[0087] For sparse addressing based on attention mechanism, m r The weight w r By w r =F Sp (F At (m r ,v),λ) is calculated, where F At (·) represents the addressing operation based on the attention mechanism, expressed as:

[0088]

[0089] FSp (·) is a sparse contraction operation, expressed as:

[0090]

[0091] In the formula, Indicates m r and the cosine similarity of v, λ is the shrinkage threshold, which is set in the range of [1 / R,3 / R].

[0092] (3) Decoding

[0093] Similar to the encoder, the decoder of CV-MemAE consists of D CV-Linear layers and D-1 CV-BN layers with parameters in and denote the real and imaginary parts of the d-th weight matrix, respectively. and Represent the real and imaginary parts of the d-th bias vector, d = 1, 2, ..., D, P d-1 and P d denote the input dimension and output dimension of the dth layer, respectively, satisfying P0 = Q and P d =P. Given input The transform function of the decoder is expressed as F De (·), its output can be expressed as:

[0094]

[0095] (4) Network training

[0096] As a typical self-supervised learning network, AE can be trained based on unlabeled data. Similarly, CV-MemAE can also adopt a self-supervised learning strategy. For STAP non-uniform detection, to detect the contaminated distance unit (i.e., out-of-distribution data), its reconstruction error in CV-MemAE should be greater than the reconstruction error of the uniform distance unit (i.e., in-distribution data). Therefore, given the parameters of the airborne radar system, a large number of space-time echo signals can be generated based on the uniform distance unit that meets the assumption H0 of equation (1) as an unlabeled training data set.

[0097] For the space-time echo signal x g , where g = 1, 2, ..., G, G is the size of the training data set, and its reconstruction error in CV-MemAE can be expressed as:

[0098]

[0099] In the formula, For x g The normalized signal, for The normalized signal reconstructed by the complex memory enhanced autoencoder, Θ CM ={Θ En ,Θ Mem ,Θ De} is the network parameter;

[0100] The loss function is defined as:

[0101]

[0102] Optimal parameters for complex memory-augmented autoencoders It is obtained by back propagation method and is expressed as:

[0103]

[0104] 2. Sample selection

[0105] The trained CV-MemAE can be used for intelligent sample selection method, which is expressed as:

[0106]

[0107] In the formula, E Rc (x l ) represents the space-time echo signal x of the lth training range unit l The reconstruction error of (l=1,2,...,L), is the normalized signal of the lth training distance unit, The optimal parameters for the complex memory enhanced autoencoder after training The reconstructed normalized signal is Rc (x l )<η, the lth training distance unit is a uniform sample; when E Rc (x l )≥η, the lth training distance unit is a non-uniform sample, η is the non-uniform distance unit detection threshold, which is set by the Median Absolute Deviation (MAD) method and is expressed as:

[0108] η=median(E Rc )+κmedian(|E Rc -median(E Rc )|) / α

[0109] In the formula, median(·) represents the median, α=0.6745 is the right boundary of the 50% bilateral confidence interval of the standard normal distribution, κ represents the prior parameter with a value range of [2,3], and E RcRepresents the reconstruction error vector of the space-time echo signals of all training range units;

[0110] In the specific embodiment, the performance of the DL-NHD method is verified by simulation experiments, and compared with three typical NHD methods, GIP-NHD, LGIP-NHD and CSMS-GIP-NHD. In the simulation, the signal wavelength, uniform linear array (ULA) element spacing, number of ULA elements, 3dB main lobe width, pulse repetition period, and number of pulses of the airborne radar system are set to 0.2m, 0.1m, N=10, 30°, 0.5ms, and K=10, respectively. In order to simulate the complex clutter environment, it is assumed that the parameters related to the radar space-time echo signal are all subject to random distribution: carrier speed v~U[70,120]m / s, carrier height H~U[8,15]km, detection distance R~U[15,75]km, angle between ULA and carrier motion direction α~U[-30,30]°, and noise-to-noise ratio CNR~U[40,60]dB. In addition, assuming that the noise power is 1, each range cell contains 181 clutter scattering cells uniformly distributed in the azimuth range [0,π], whose scattering coefficient obeys the complex Gaussian distribution and whose amplitude depends on the CNR. Then, the simulation parameters are randomly set, and the radar space-time echo signal is generated based on the range cell obeying the assumption H0 to obtain the training data of CV-MemAE. Where G=32768.

[0111] In the constructed CV-MemAE, the dimensions of the complex linear layer in the encoder are set to (100×256), (256×128), (128×64) and (64×64), respectively, and the dimensions of the complex linear layer in the decoder are set to (64×64), (64×128), (128×256) and (256×100), respectively. The dimensions of the memory unit in the memory module are set to R×Q=128×64, and the threshold of sparse addressing is set to λ=1 / 64. To train CV-MemAE, the parameters of the encoder and decoder are initialized using the Glorot method, and the memory unit is initialized to a random matrix whose elements follow the uniform distribution U[-8,8]. Other training parameters are: batch size 1024, training number 1000, initial learning rate 10-3, and learning rate decay rate 0.999.

[0112] 1. Test results

[0113] After training CV-MemAE, the performance and advantages of the DL-NHD method are verified based on the non-uniform sample detection results of different methods. The specific simulation parameters are: carrier speed v = 100m / s, carrier height H = 11.5km, angle α between ULA and carrier motion direction = 10°, CNR = 50dB. Assume that there are 200 training range unit data, the target is located at the 100th range unit, its normalized Doppler frequency is -0.25, signal-to-noise ratio SNR = 15dB, 12 interference signals are located at the 52nd to 92nd and 108th to 148th range units at intervals of 8 range units, the interference-to-noise ratio (INR) is 15dB, and the normalized Doppler frequency is evenly distributed in the range of [-0.45, -0.05].

[0114] Figure 3 a- Figure 3 Figure d shows the detection results of non-uniform distance units using different methods. It can be found that: although the GIP-NHD method can correctly detect all non-uniform distance units containing targets and interference (as shown in the blue box), a large number of false alarms occur; under the condition of appropriate diagonal loading (diagonal loading μ = -3dB), the number of distance units misdetected by the LGIP-NHD method is relatively small, and its performance is improved compared with the GIP-NHD method; since the target's steering vector is used as the interference's steering vector to calculate the CSM, the CSMS-GIP-NHD method cannot effectively detect interference signals whose Doppler frequencies are far away from the target (as shown in the red circle); the DL-NHD method can detect all distance units containing targets and interference without missed detection or false detection.

[0115] To further compare the non-uniform sample detection performance of different methods, the number of available distance units is reduced from 200 to 100 while keeping other parameters unchanged. The results are shown in the figure below. Figure 4 a-4d. It can be found that: due to the insufficient number of training distance units for estimating the covariance matrix, both the GIP-NHD and CSMS-GIP-NHD methods cannot correctly detect the non-uniform distance units containing targets and interference; the LGIP-NHD and DL-NHD methods can still work normally, and the DL-NHD method does not have the problem of false detection. In addition, LGIP-NHD needs to set a suitable diagonal loading μ, otherwise its performance will be greatly reduced. Figure 5 a- Figure 5 As shown in (b), changing the diagonal loading μ of LGIP-NHD from -3dB to -10dB and -15dB will result in missed detection of non-uniform samples and false detection of uniform samples.

[0116] Finally, the INR of the interference is increased from 15dB to 25dB to simulate the situation of strong interference. The other parameters are set the same as Figure 3 The same, different methods can be used to obtain the test results such as Figure 6 a- Figure 6 It can be found that: the three typical NHD methods all have the problem of missed detection and / or false detection; the DL-NHD of the present invention can still correctly detect all the expected distance units.

[0117] 2. Signal-to-noise ratio loss

[0118] Based on the above detection results, this section uses the signal-to-clutter-to-noise ratio (SCNR) loss as an indicator to evaluate the clutter suppression and target detection performance of different methods. Figure 3 , 4 The detection results in 6 are shown in Figure 1. Different NHD methods remove the detected non-uniform distance units respectively, and use the remaining distance unit data to calculate the CNCM of the distance unit to be measured. Then, the spatial frequency of the target is kept at 0, and its normalized Doppler frequency is made to change uniformly in the range of [-0.5, 0.5] to calculate the SCNR loss. The results are shown in Figure 1. Figure 7 a- Figure 7 c. It can be found that: under different conditions, the DL-NHD method of the present invention can obtain the SCNR loss curve closest to the condition without interference pollution, and obtain better STAP performance than other typical NHD methods.

[0119] 3. Computational complexity

[0120] The computational complexity of the DL-NHD method of the present invention is analyzed and compared with other typical NHD methods. Taking the number of complex multiplications as an indicator, the computational complexity of different NHD methods is shown in Table 1. Where Q1, Q2 and Q3 are equal to 256, 128 and 64 respectively. Let the number of distance units L change from 100 to 1000 in steps of 100, and the computational complexity of different NHD methods can be obtained as follows: Figure 8 As shown in the figure, it can be seen that the computational complexity of the DL-NHD method is higher than that of CSMS-GIP-NHD, but much lower than that of GIP-NHD and LGIP-NHD. Considering that DL-NHD has higher performance and does not require a preset interference steering vector, it can be considered that the computational complexity of DL-NHD exceeding that of CSMS-GIP-NHD is acceptable.

[0121] Table 1

[0122]

[0123] Aiming at the problem of uniform sample (i.e., IID training distance unit) selection faced by the airborne radar moving target detection method based on STAP method, the present invention converts the non-uniform sample detection problem into the out-of-distribution data detection problem, constructs a reconstruction network based on complex memory enhanced autoencoder, and uses the reconstruction error of radar space-time echo signal of different distance units as the detection index to detect non-uniform samples, thereby proposing an intelligent sample selection method based on deep learning. Simulation results show that compared with the existing non-uniform detector based on statistical analysis method, the DL-NHD method of the present invention achieves better non-uniform detection performance under different conditions, thereby improving clutter suppression and target detection performance, and its computational complexity is relatively moderate.

[0124] An airborne radar moving target detection system based on intelligent sample selection, applied to any of the above-mentioned airborne radar moving target detection methods based on intelligent sample selection, comprising a data acquisition module, a training set construction module, an autoencoder training module, a reconstruction error acquisition module, and a distance unit selection module connected in sequence; wherein:

[0125] A data acquisition module, used to acquire space-time echo signals of radar range units, including space-time echo signals of different range units and space-time echo signals of uniform range units;

[0126] A training set construction module, used for constructing a training data set through space-time echo signals of uniform distance units;

[0127] An autoencoder training module is used to construct a complex memory enhanced autoencoder and input the training data set into the complex memory enhanced autoencoder to realize offline training of the autoencoder;

[0128] A reconstruction error acquisition module is used to input the space-time echo signals of different distance units into the trained complex memory enhanced autoencoder to obtain the reconstruction error;

[0129] The range unit selection module is used to select uniform range units according to the reconstruction error and detect and remove contaminated range units.

[0130] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned methods for detecting moving targets by airborne radar based on intelligent sample selection.

[0131] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0132] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for detecting moving targets by airborne radar based on intelligent sample selection, characterized in that: The following steps are involved: S1. Acquire space-time echo signals of radar range units, including space-time echo signals of different range units and space-time echo signals of uniform range units; S2. constructing a training data set through the space-time echo signals of the uniform distance units; S3, constructing a complex memory enhanced autoencoder, and inputting the training data set into the complex memory enhanced autoencoder to implement offline training of the autoencoder; S4, inputting the space-time echo signals of the different distance units into the trained complex memory enhanced autoencoder to obtain a reconstruction error; S5. Select uniform distance cells according to the reconstruction error, and detect and remove contaminated distance cells.

2. The method for detecting moving targets by airborne radar based on intelligent sample selection according to claim 1, characterized in that: The complex memory-augmented autoencoder constructed in S3 consists of three processing steps, namely encoding, reconstruction, and decoding.

3. The method for detecting moving targets by airborne radar based on intelligent sample selection according to claim 2, characterized in that: In the reconstruction step, the memory module reconstructs the encoding output through the sparse memory items. It is expressed as: Where M R ∈R R×Q and M I ∈R R×Q Represent the real and imaginary parts of the memory cell, R and Q represent the number of rows and columns of the memory cell, respectively. R +jw I ∈C 1×R is a sparse weight vector, M = M R +jM I The row vector of represents the memory item, where the rth row is represented as 4. The method for detecting moving targets by airborne radar based on intelligent sample selection according to claim 1, characterized in that: The autoencoder training in S3 includes: For the space-time echo signal x g , the reconstruction error in the complex memory enhanced autoencoder is expressed as: In the formula, For x g The normalized signal, for The normalized signal reconstructed by the complex memory enhanced autoencoder, Θ CM ={Θ En ,Θ Mem ,Θ De } is the network parameter, Θ En is the encoding parameter, Θ De is the decoding parameter, Θ Mem ={M R ,M I } is the reconstruction parameter; The loss function is defined as: In the formula, G represents the size of the training dataset. Optimal parameters for complex memory-augmented autoencoders It is obtained by back propagation method and is expressed as:

5. The method for detecting moving targets by airborne radar based on intelligent sample selection according to claim 1, characterized in that: The selection index of uniform distance units in S5 is expressed as: In the formula, E Rc (x l ) represents the space-time echo signal x of the lth training range unit l The reconstruction error of is the normalized signal of the lth training distance unit, The optimal parameters for the complex memory enhanced autoencoder after training The reconstructed normalized signal is Rc (x l )<η, the lth training distance unit is a uniform sample; when E Rc (x l )≥η, the lth training distance unit is a non-uniform sample, and η is the detection threshold of the non-uniform distance unit.

6. The method for detecting moving targets by airborne radar based on intelligent sample selection according to claim 1, characterized in that: The method also includes the basic principle of complex memory enhanced autoencoder for non-uniform sample detection: In the offline training stage, the memory module of the complex memory enhanced autoencoder reconstructs the encoding output of the training data set based on the sparse addressing operation, and updates the parameters of the memory unit to ensure that the training data set is reconstructed in the complex memory enhanced autoencoder; in the online application stage, the parameters of the memory unit are fixed, and reconstruction errors occur because the encoding output of the non-uniform sample cannot obtain matching memory items from the memory unit.

7. An airborne radar moving target detection system based on intelligent sample selection, applied to an airborne radar moving target detection method based on intelligent sample selection as described in any one of claims 1 to 6, characterized in that: It includes a data acquisition module, a training set construction module, an autoencoder training module, a reconstruction error acquisition module, and a distance unit selection module connected in sequence; wherein: The data acquisition module is used to acquire space-time echo signals of radar distance units, including space-time echo signals of different distance units and space-time echo signals of uniform distance units; The training set construction module is used to construct a training data set through space-time echo signals of uniform distance units; The autoencoder training module is used to construct a complex memory enhanced autoencoder, and input the training data set into the complex memory enhanced autoencoder to realize offline training of the autoencoder; The reconstruction error acquisition module is used to input the space-time echo signals of different distance units into the trained complex memory enhanced autoencoder to obtain the reconstruction error; The distance unit selection module is used to select uniform distance units according to the reconstruction error, and detect and remove contaminated distance units.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting moving targets by airborne radar based on intelligent sample selection as described in any one of claims 1 to 6 is implemented.

Images