Radar moving target detection method based on small sample transfer learning and attention mechanism

Abstract

According to the radar moving target detection method based on small sample transfer learning and the attention mechanism, a small amount of actual radar echoes and high-fidelity scene simulation data are used as a data training set, and direct detection of a moving target under a strong clutter background can be realized. An attention mechanism is introduced by utilizing the point / line distribution characteristics of a moving target and a scene clutter in a space-time two-dimensional plane, the clutter sorting and recognition capability of the network can be ensured while the network training speed is improved, meanwhile, a small sample migration twin network is further combined, a small amount of actually measured data can be utilized to correct a target detection network, and the target detection accuracy is improved. The adaptability of the network to actual working conditions is improved, and the generalization ability is improved. According to airborne flight calibration test data analysis, under the condition of a low signal-clutter-noise ratio, compared with a traditional STAP + CFAR technology, the method has higher detection performance and lower false alarm rate. Meanwhile, based on network pre-training and a small sample transfer learning mechanism, the method has practicability in the aspect of early warning radar real-time processing and has high engineering application value.

Description

Technical Field

[0001] The invention discloses a radar moving target detection method based on small sample transfer learning and attention mechanism, belonging to the technical field of radar. Background Art

[0002] Land and sea clutter suppression and moving target detection are key issues in the application of space-based early warning radar. A variety of methods have been proposed to solve this problem, such as space-time adaptive clutter suppression technology (STAP), direct data domain (DDD) method, knowledge-assisted method (KA-STAP), sparse reconstruction method (SR-STAP), sample reconstruction method based on deep learning, etc. Among them, the DDD STAP method only uses the echo data of the unit to be detected in the range, which completely eliminates the problem of insufficient IID samples in non-stationary / non-uniform clutter environment. However, it is easily affected by noise and has serious aperture loss, which makes the performance unable to reach suboptimal. At the same time, the spatial smoothing processing also determines that it is only applicable to uniform linear / area array airborne radars. The KA STAP method uses the prior information of the carrier and radar system to pre-estimate the clutter covariance matrix (CCM) and pre-whiten the clutter. However, due to the existence of internal / external unknown non-ideal factors such as antenna spatial error and clutter fluctuation, the estimated CCM has a large deviation from the true value. In addition, the reasonable allocation of weight coefficients between the CCM constructed by prior knowledge and the CCM estimated by maximum likelihood has not been well solved. SR STAP uses the sparse distribution characteristics of clutter in the space-time spectrum plane. By imposing sparse constraints on the space-time spectrum coefficients, it can reconstruct the clutter super-resolution space-time spectrum and calculate the corresponding CCM under the condition of a small number of observation samples. However, this type of method is usually solved by iterative optimization technology, which often has problems such as long running time, low execution efficiency and difficulty in selecting hyperparameters, and is therefore not conducive to practical applications. Nevertheless, sparse recovery methods provide a new idea for improving the performance of STAP under small sample conditions.

[0003] In the field of deep learning, the detection and classification of moving targets mainly rely on the extraction of deep features of data by convolutional neural networks (CNNs). Their powerful generalization ability can weaken the impact of non-uniform clutter on target detection to a certain extent. Some experts have proposed a CNN network model for moving target detection, which can realize moving target detection in the case of non-uniform radar echo data; in addition, the application of autoencoder networks in radar signal processing can reconstruct and simulate radar echo signals. Some experts use autoencoder networks to generate radar echo signals and find the area where the target signal is located by comparing them with the actual radar echo signals, so as to accurately complete the detection of moving target signals; in addition, some experts use adversarial generative networks to suppress sea clutter and improve the detection performance of moving targets at sea. Some scholars use CNN to reconstruct high-resolution angle Doppler profiles, and then reconstruct the covariance matrix based on the obtained space-time two-dimensional clutter spectrum. Finally, the reconstructed covariance matrix is ​​used to calculate the STAP weights and achieve clutter suppression. Although convolutional neural networks can overcome the impact of non-uniform environments, the training of convolutional networks requires a lot of sample data, which makes it difficult to apply them to actual airborne radar systems. At the same time, existing CNN-based methods require online learning of radar clutter features, which requires a very high amount of computation and is not conducive to real-time processing of airborne radars.

[0004] At present, the target detection technology in the background of land and sea clutter at home and abroad mainly focuses on STAP clutter suppression technology, STAP processing method based on prior, clutter reconstruction technology based on statistical information, and moving target detection technology based on convolutional neural network model. However, under the condition of small sample, the research on the real-time detection network model of moving targets in non-uniform and non-stationary clutter environment has not been made public. Therefore, there is an urgent need for a moving target detection method in the clutter background with small sample conditions and strong generalization ability. Summary of the invention

[0005] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, to provide a radar moving target detection method based on small sample transfer learning and attention mechanism, to solve the problems that the non-uniform and non-stationary clutter covariance matrix is ​​difficult to accurately construct and the applicability and effectiveness of the moving target detection network are low in actual engineering.

[0006] The technical solution of the present invention is: a radar moving target detection method based on small sample transfer learning and attention mechanism, comprising:

[0007] Generate ground clutter background radar echo data for different application scenarios and early warning radar system conditions;

[0008] Add moving target signals of different powers and speeds to the ground clutter background radar echo data to generate the training data set required for model training;

[0009] A deep learning feature extraction network based on the attention mechanism, namely the FENet network model, is used to calculate the feature distribution of the FENet network model and its corresponding attention mechanism parameters based on the generated training data set;

[0010] Based on the twin network model, the measured data samples obtained by radar are used to modify the parameters of the FENet network model, thereby obtaining the SSM-Net model under small samples;

[0011] The SSM-Net model is applied to the real-time radar echo. The real-time radar echo data is used as the input of the SSM-Net model. The output of the SSM-Net model directly corresponds to the moving target detection result of the radar, that is, the distance and speed parameters of the moving target are detected.

[0012] The generation of simulated echo data of ground clutter background under different application scenarios and early warning radar system conditions includes:

[0013] Assume that the orbital height of the space-based radar platform is H, the platform movement speed is V, the coordinate axis is established with the northeast sky, and the platform movement direction is the Y axis; the array is a positive side-view uniform linear array composed of N array elements, θ and are the azimuth and elevation angles of the antenna beam pointing to the current clutter block; the spacing between array elements is half a wavelength, and the pulse repetition interval is T r , the pulse repetition frequency is f r =1 / T r , a total of M pulses are emitted within the coherent processing interval.

[0014] Assume that there are N distance units to be detected. c clutter units are evenly distributed in azimuth, then the normalized Doppler frequency corresponding to the clutter scattering unit is for:

[0015]

[0016] Where λ is the radar operating wavelength, f r is the pulse repetition frequency;

[0017] Normalized spatial frequency corresponding to the clutter block for:

[0018]

[0019] For a uniform linear array, the time domain steering vector corresponding to any clutter block is and the airspace steering vector They are:

[0020]

[0021] Considering the range ambiguity, array element error and clutter fluctuation problems in radar ground observation scenarios, assuming that each clutter block is statistically independent, the clutter space-time snapshot data of any range unit is represented as the superposition of the echo signals of each clutter block at different ambiguity distances:

[0022]

[0023] In the formula, x c Represents the clutter echo data of the unit to be detected, α i Indicates the echo amplitude of the i-th clutter block;

[0024] γ=[γ1,γ2,…,γ M ] T Indicates the clutter fluctuation component between coherent pulses;

[0025] represents the array element error component; ⊙ and Respectively represent the Hadamard product and Kronecker product between matrices;

[0026] Similarly, the echo signal x of the moving target t It is expressed as:

[0027]

[0028] In the formula, α t represents the complex amplitude of the signal; f t ,θ t and They represent the normalized Doppler frequency, azimuth and elevation angle of the moving target respectively;

[0029] The ground clutter background radar echo data y received by the radar consists of scene clutter, moving targets and noise:

[0030] y=x t +x c +n;

[0031] Where n represents the Gaussian white noise component.

[0032] The method of adding moving target signals of different powers and different moving speeds to the ground clutter background radar echo data to generate a training data set required for model training includes:

[0033] Assume that the L1th range gate containing the moving target is regarded as the unit to be detected, and there are T Pprotection units, the range gate where the training data is selected is located outside the protection unit, and each training selects T E Range gates;

[0034] The number of range gates selected when expanding data is T E , the signal-to-noise ratio range is -15dB~20dB, the number of moving target categories is K, and the expansion multiple is K C ; The expanded data set in each case reaches T E ×K×K C ; During the training process, the batch Epochs is set to 13;

[0035] The expanded moving target signal is divided into K categories, of which category 0 is pure background clutter signal, that is, no moving target signal is added; for the category where the target exists, each category corresponds to the normalized Doppler frequency f of the moving target t,k for:

[0036]

[0037] Where K represents the total number of categories of different moving targets;

[0038] When detecting moving targets, it is assumed that the target is located at the center of the radar beam and its spatial frequency is fixed. Extended target data corresponding to different speeds is generated. The extended training data Y k,h as follows:

[0039] Y k,h =x t,k,h +x c +n;

[0040] where x t,k,h The corresponding normalized Doppler frequency is w k , with an amplitude of α t,h The extended motion target data is as follows:

[0041]

[0042] The FENet network model consists of 4 convolutional layers, 2 pooling layers, 4 attention modules, 2 fully connected layers and 1 recognition layer; the input of the FENet network model is the radar training data Y k,h , the output is the moving target speed corresponding to the detection range gate.

[0043] The method uses a deep learning feature extraction network based on the attention mechanism, namely a FENet network model, to extract the feature distribution of the training data set and its corresponding attention mechanism parameters; including:

[0044] 1) Design each layer of the FENet network model as follows:

[0045] The convolution layer is used to extract features and extract the main features of the moving target from the input data:

[0046] F c =Relu(W c *Y k,h +b c );

[0047] Among them, Relu is the activation function used by the convolutional layer, F c Represents the feature matrix after the convolutional network, W c represents a convolution kernel of dimension f×f×c; f represents the length and width of the convolution kernel; c represents the number of convolution kernels, that is, the number of output feature channels; * represents the convolution operation; b c is the bias vector;

[0048] The attention module adopts a space-channel joint attention mechanism, including a space attention module and a channel attention module; the space attention module targets the beam-Doppler two-dimensional space of the radar data, and the channel attention module targets the real and imaginary parts of the radar data; the channel attention module first performs global average pooling and global maximum pooling on the convolutional feature map of the channel to obtain two 1×1×C feature maps;

[0049]

[0050] In the formula, represents the feature vector obtained after the average pooling operation, represents the feature vector obtained by the maximum pooling operation, mean(·) and max(·) represent the maximum pooling operation and the average pooling operation respectively;

[0051] The obtained feature vector and They are sent to a two-layer neural network MLP respectively. The size of the feature matrix output by the neural network MLP is the same as the size of the input feature matrix. The activation function used in the network is Relu, and the updated feature vector is obtained. and

[0052]

[0053] In the formula, represents the average pooling feature parameters extracted by the neural network, Represents the maximum pooling feature parameters extracted by the neural network;

[0054] For the feature vector and Perform the sum operation and then use the sigmoid function for normalization to generate the final channel attention feature vector M C (F):

[0055]

[0056] Where the channel attention feature vector M C (F) represents the weight of each channel of the input data;

[0057] The spatial attention module performs global average pooling and global maximum pooling on the feature vector and Perform concatenation and convolution operations, and finally perform global normalization of weights through the sigmoid function to obtain the updated weight vector matrix M s (F):

[0058]

[0059] Where [;] represents the splicing operation, f s×s (·) represents the convolution operation, s is the size of the convolution kernel, and softmax is the classification identifier; M S (F) is the weighted convolutional network of a single channel.

[0060] The pooling layer is used to reduce the dimension of the data and compress the features after convolution:

[0061] F P =γ P (F c ; S p ,F p );

[0062] Among them, F P is the feature matrix after pooling, γ P represents the downsampling operator, S p is the window length of the downsampling operator, F P is the downsampling operation step size;

[0063] The fully connected layer is used to flatten the feature matrix extracted by the convolution operation and compress the flattened eigenvalues:

[0064] y f =γ f ( <W f ,vev(F P )>+b f );

[0065] Among them, y fis the output feature of the fully connected layer, vec(·) represents the vectorization of the matrix, converting the three-dimensional feature vector into a two-dimensional vector, <·> represents the inner product of two vectors, and W f represents the weight vector, b f represents the bias vector, γ f Represents the activation function, where the first fully connected layer of the FENet network uses the tanh activation function, and the second layer uses the sigmoid activation function;

[0066] The recognition layer uses a softmax classifier. The probability of belonging to the kth class target is obtained based on the output data of the fully connected layer as follows:

[0067]

[0068] The prediction value of FENet is the probability value of multiple categories. In order to achieve better convergence effect, the loss function used in the experiment is the mean square error. The specific mathematical expression is:

[0069]

[0070] Where M total Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample.

[0071] Solving the FENet network model, the process is as follows:

[0072] After completing the construction of the FENet network model, it is necessary to solve the parameters of the network model through training data; set the network feature distribution in the FENet model and its corresponding attention mechanism parameter set as Θ FEN , the training process of the model is described as follows:

[0073] Θ FEN =argmin Loss(Θ FEN ).

[0074] The feature matrix F after the convolutional network c The output feature dimension is h×w×c, where h and w are the length and width of the output data, respectively, and their values ​​are related to the convolution step length s. c The details are as follows:

[0075]

[0076] in This is a round-up operation.

[0077] The training process of the model is:

[0078] Θ FEN=argmin Loss(Θ FEN );

[0079] The above training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.001. The data set is divided into a training data set and a test data set in a ratio of 8:2. The network model is trained for 10 batches, and the test data set is used for target detection prediction and accuracy evaluation. The termination condition of the training iteration is set to the prediction deviation of the test data set before and after being less than 0.01% or the training batch reaches 500 times.

[0080] Based on the twin network model, the measured data samples obtained by radar are used to modify the parameters of the FENet network model, so as to obtain the SSM-Net model under small samples, including:

[0081] set up To extract radar simulation data using FENet network, that is, the features of the supporting dataset; The FENet network is used to extract the real radar echo data, that is, the characteristics of the query data set; the nearest centroid cosine similarity function is used to characterize the similarity of the above two types of data after passing through the feature extraction network;

[0082] The cosine similarity y between the feature vector of sample Y and the centroid of class k k-ssm The expression is:

[0083]

[0084] Where A k represents the samples belonging to category k in the support data set; <·,·> represents the cosine similarity, τ is the scaling value of the cosine similarity, which is used to scale the interval of the cosine similarity; w k Represents the centroid of each class, and its mathematical expression is:

[0085]

[0086] The loss function of the small sample network still uses the mean square error function, which is:

[0087]

[0088] Where M total-SSM Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample.

[0089] After completing the construction of the model based on the twin network idea, the parameters of the SSM-Net network model are trained and solved through the radar simulation data level and real-time data set, and all parameter sets in the SSM-Net model are set to Θ SSM , the training process of the model is described as follows:

[0090] Θ SSM =argmin Loss(Θ SSM ).

[0091] The τ is set to 0.1.

[0092] The parameters of the SSM-Net network model are trained and solved through radar simulation data level and real-time data set, including: the training process of the model is:

[0093] Θ SSM =argmin Loss(Θ SSM );

[0094] The above training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.001. The data set is divided into a training data set and a test data set in a ratio of 8:2. The network model is trained for 10 batches, and the test data set is used for target detection prediction and accuracy evaluation. The termination condition of the training iteration is set to the prediction deviation of the test data set before and after being less than 0.01% or the training batch reaches 500 times.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] The method of the present invention belongs to the problem of moving target detection under non-uniform and non-stationary clutter background under small sample conditions. At present, a variety of methods have been proposed to solve this problem, among which clutter suppression technology mainly includes traditional space-time STAP and KA-STAP, DDD-STAP, SR-STAP and other methods under small sample conditions. However, the above methods have problems such as scene mismatch, aperture loss and hyperparameter selection, and there are performance losses or scene constraints in practical applications. The moving target detection algorithm uses the statistical feature differences of the target, residual clutter and background noise after clutter suppression to perform target detection processing in a statistical sense, such as CA-CFAR, OS-CFAR, etc. The above detection algorithm is very sensitive to the statistical distribution of energy after clutter suppression, and has high requirements for energy distribution characteristics. And the step-by-step processing method based on clutter suppression + target detection has the problem of information loss introduced by nonlinear processing, and has certain performance losses.

[0097] Under the condition that the clutter samples are non-uniform and non-stationary, the present invention uses a small amount of actual radar echoes and high-fidelity scene simulation data as the data training set, and can realize direct detection of moving targets in a strong clutter background. It mainly utilizes the point / line distribution characteristics of moving targets and scene clutter in the space-time two-dimensional plane, introduces the attention mechanism, and can ensure the network's ability to sort and identify clutter while improving the network training speed. At the same time, it is further combined with the small sample migration twin network, and a small amount of measured data can be used to correct the target detection network, thereby improving the network's adaptability to actual working conditions and improving generalization ability. At the same time, the network integrates the clutter suppression and target detection processing processes, which can avoid the information loss introduced by step-by-step processing. Compared with the traditional step-by-step processing strategy, it still has higher detection performance under low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 This is a schematic diagram of the space-based early warning radar signal model.

[0099] Figure 2 Populate the diagram with training data.

[0100] Figure 3 Extract network model diagram for FENet network model features.

[0101] Figure 4 This is a performance comparison chart of the SSM-Net small sample network and the traditional STAP+CFAR method.

[0102] Figure 5 This is the moving target detection result diagram output by the SSM-Net model. DETAILED DESCRIPTION

[0103] The implementation and effects of the present invention are described in further detail below.

[0104] like Figure 1 As shown, the use scenario of the present invention is: applied to the problem of moving target detection in space-based / airborne early warning radars facing non-uniform / non-stationary clutter interference in wide and complex land and sea scenes. Combined with high-precision early warning radar simulated echo data, by introducing an attention mechanism, and coordinating with space and channel networks, the model's ability to recognize and classify clutter features can be enhanced. By performing migration corrections on the model through small samples received in real time, moving target detection in complex clutter environments can be achieved under small sample conditions, thereby improving the feasibility of artificial intelligence networks in real-time processing of moving target detection in early warning radars. The main steps are as follows:

[0105] Step 1 Generate echo data of ground clutter and moving targets of space-based radar

[0106] like Figure 2As shown in the figure, it is assumed that the orbital height of the space-based radar platform is H, the platform movement speed is V, the coordinate axis is established with the northeast sky, and the platform movement direction is the Y axis; the array is a positive side-view uniform linear array composed of N array elements, θ and are the azimuth and elevation angles of the antenna beam pointing to the current clutter block; the spacing between array elements is half a wavelength, and the pulse repetition interval is T r , the pulse repetition frequency is f r =1 / T r , a total of M pulses are emitted within the coherent processing interval.

[0107] Assume that there are N distance units to be detected. c clutter units are evenly distributed in azimuth, then the normalized Doppler frequency corresponding to the clutter scattering unit is for:

[0108]

[0109] Where λ is the radar operating wavelength, f r is the pulse repetition frequency;

[0110] Normalized spatial frequency corresponding to the clutter block for:

[0111]

[0112] For a uniform linear array, the time domain steering vector corresponding to any clutter block is and the airspace steering vector They are:

[0113]

[0114] Considering the range ambiguity, array element error and clutter fluctuation problems in radar ground observation scenarios, assuming that each clutter block is statistically independent, the clutter space-time snapshot data of any range unit is represented as the superposition of the echo signals of each clutter block at different ambiguity distances:

[0115]

[0116] In the formula, x c Represents the clutter echo data of the unit to be detected, α i Indicates the echo amplitude of the i-th clutter block;

[0117] γ=[γ1,γ2,…,γ M ] T Indicates the clutter fluctuation component between coherent pulses;

[0118] represents the array element error component; ⊙ and Respectively represent the Hadamard product and Kronecker product between matrices;

[0119] Similarly, the echo signal x of the moving target t It is expressed as:

[0120]

[0121] In the formula, α t represents the complex amplitude of the signal; f t ,θ t and They represent the normalized Doppler frequency, azimuth and elevation angle of the moving target respectively;

[0122] The echo data y received by the radar consists of scene clutter, moving targets and noise:

[0123] y=x t +x c +n (6)

[0124] Where n represents the Gaussian white noise component.

[0125] The generated simulated echo data under different radar system scales, application scenarios and targets can be used as training samples for the characteristic parameters of the FENet attention mechanism network model.

[0126] Step 2: Dataset expansion: Based on the radar echo data generated by simulation, it is used as background clutter data (the echo data of moving targets is not considered here), and moving target signals of different powers and speeds are added to the background clutter data to expand the training data and generate the data set required for model training.

[0127] 1) Assume that the 33rd range gate containing the moving target is regarded as the unit to be detected. There are 5 protection units on both sides of the unit to be detected. The range gate where the training data is selected is located outside the protection unit. We select 8 range gates for each training.

[0128] 2) When expanding the data, the number of range gates selected is 8, the signal-to-noise ratio range is -15dB~20dB, the number of moving targets is 8, and the expansion multiple is 101. In each case, the expanded data set reaches 8×8×101; during the training process, Epochs is set to 13 batches

[0129] 3) The expanded moving target signal is divided into 8 categories, among which Category 0 is pure background clutter signal, that is, no moving target signal is added.

[0130] For the classes where the target exists, the normalized Doppler frequency of the moving target corresponding to each class is:

[0131]

[0132] Where K represents the total number of categories of different moving targets.

[0133] When detecting moving targets, it is assumed that the target is located at the center of the radar beam and its spatial frequency is fixed. It is only necessary to generate extended target data corresponding to different speeds. The extended training data Y k,h as follows:

[0134] Y k,h =x t,k,h +x c +n (8)

[0135] where x t,k,h The corresponding normalized Doppler frequency is w k , with an amplitude of α t,h The extended motion target data is expressed as follows:

[0136]

[0137] Step 3 constructs a deep learning feature extraction network (FENet) model based on the attention mechanism, and trains the network parameter set of the FENet model based on the expanded radar simulation data.

[0138] like Figure 3 As shown in FIG, the FENet network model consists of 4 convolutional layers, 2 pooling layers, 4 attention modules, 2 fully connected layers and 1 recognition layer; the input of the FENet network model is the radar training data Y k,h , the output is the moving target speed corresponding to the detection range gate.

[0139] 1) The design of each layer of the FENet network model is as follows:

[0140] The convolution layer is used to extract features and extract the main features of the moving target from the input data:

[0141] F c =Relu(W c *Y k,h +b c ) (10)

[0142] Among them, Relu is the activation function used by the convolutional layer, F c Represents the feature matrix after the convolutional network, W c represents a convolution kernel of dimension f×f×c; f represents the length and width of the convolution kernel; c represents the number of convolution kernels, that is, the number of output feature channels; * represents the convolution operation; b c is the bias vector; F c The output feature dimension is h×w×c, where h and w are the length and width of the output data, respectively, and their values ​​are related to the convolution step length s.c The details are as follows:

[0143]

[0144] in This is a round-up operation.

[0145] The attention module adopts a spatial-channel joint attention mechanism: the spatial attention module mainly targets the beam-Doppler two-dimensional space of radar data, and the channel attention module corresponds to the real and imaginary parts of radar data.

[0146] The channel attention module first performs global average pooling and global maximum pooling on the convolutional feature map of the channel to obtain two 1×1×C feature maps;

[0147]

[0148] In the formula, represents the feature vector obtained after the average pooling operation, represents the feature vector obtained by the maximum pooling operation, mean(·) and max(·) represent the maximum pooling operation and the average pooling operation respectively;

[0149] The obtained feature vector and They are sent to a two-layer neural network MLP respectively. The size of the feature matrix output by the neural network MLP is the same as the size of the input feature matrix. The activation function used in the network is Relu, and the updated feature vector is obtained. and

[0150]

[0151] In the formula, represents the average pooling feature parameters extracted by the neural network, Represents the maximum pooling feature parameters extracted by the neural network;

[0152] The feature vector output by the neural network MLP and Perform the sum operation and then use the sigmoid function for normalization to generate the final channel attention feature vector M C (F):

[0153]

[0154] Where the channel attention feature vector M C (F) represents the weight of each channel of the input data.

[0155] The spatial attention module performs global average pooling and global maximum pooling on the feature vector and Perform concatenation and convolution operations, and finally perform global normalization of weights through the sigmoid function to obtain the updated weight vector matrix M s (F):

[0156]

[0157] Where [;] represents the splicing operation, f s×s (·) represents the convolution operation, s is the size of the convolution kernel, and softmax is the classification identifier. S (F) is the weighted convolutional network of a single channel.

[0158] The pooling layer is used to reduce the dimension of the data and compress the features after convolution:

[0159] F P =γ P (F c ; S p ,F p ) (15)

[0160] Among them, F P is the feature matrix after pooling, γ P represents the downsampling operator, S p is the window length of the downsampling operator, F P is the downsampling operation step size. The fully connected layer flattens the feature matrix extracted by the convolution operation and compresses the flattened eigenvalues:

[0161] y f =γ f ( <W f ,vev(F P )>+b f ) (16)

[0162] Among them, y f is the output feature of the fully connected layer, vec(·) represents the vectorization of the matrix, converting the three-dimensional feature vector into a two-dimensional vector, <·> represents the inner product of two vectors, and W f represents the weight vector, b f represents the bias vector, γ f Represents the activation function, where the first fully connected layer of the FENet network uses the tanh activation function, and the second layer uses the sigmoid activation function;

[0163] The recognition layer uses a softmax classifier. The probability of belonging to the kth class target is obtained based on the output data of the fully connected layer as follows:

[0164]

[0165] The prediction value of FENet is the probability value of multiple categories. In order to achieve better convergence effect, the loss function used in the experiment is the mean square error. The specific mathematical expression is:

[0166]

[0167] Where M total Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample.

[0168] 2) FENet network solution:

[0169] After completing the construction of the FENet network model, it is necessary to solve the parameters of the network model through training data. Set all parameter sets in the FENet model to Θ FEN , the training process of the model is described as follows:

[0170] Θ FEN =argmin Loss(Θ FEN ) (19)

[0171] The training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.001. The dataset is divided into a training dataset and a test dataset in a ratio of 8:2. The network model is trained for 10 batches, and the test dataset is used for target detection prediction and accuracy evaluation. The termination condition of the training iteration is set to the prediction deviation of the test dataset before and after being less than 0.01% or the training batch reaches 500 times.

[0172] Step 4: Based on step 3, with the assistance of a small amount of measured data samples, the FENet model is fine-tuned using the idea of ​​transfer learning based on the twin network model to construct the SSM-Net model.

[0173] Considering factors such as system array error, equivalent yaw of rotation, attitude control, and scene scattering model deviation, there is often a deviation between the offline generated training data and the actual radar acquisition data. Therefore, the detection results of the network model generated by the simulated training data are less tolerant to errors. A small sample optimization migration SSM-Net model based on a twin network is designed. Figure 4 , 5 shown.

[0174] 1) Constructing the SSM-Net network model based on the twin network concept

[0175] set up To extract radar simulation data using FENet network, that is, the features of the supporting dataset; In order to extract the real radar echo data using FENet network, that is, to query the characteristics of the data set; the nearest centroid cosine similarity function is used to characterize the similarity between the above two types of data after passing through the feature extraction network.

[0176] The cosine similarity y between the feature vector of sample Y and the centroid of class k k-ssm The expression is as follows:

[0177]

[0178] In the above formula, A k represents the samples belonging to category k in the support data set; <·,·> represents the cosine similarity, and τ is the scaling value of the cosine similarity, which is generally set to 0.1 and is used to scale the interval of the cosine similarity. k Represents the centroid of each class, and its mathematical expression is:

[0179]

[0180] The loss function of the small sample network still uses the mean square error function, as shown below.

[0181]

[0182] Where M total-SSM Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample.

[0183] 2) Training process of SSM-Net model

[0184] After completing the construction of the model based on the twin network idea, the SSM-Net network model parameters can be trained and solved through radar simulation data level and real-time data sets.

[0185] Set all parameter sets in the SSM-Net model to Θ SSM , the training process of the model is described as follows:

[0186] Θ SSM =argmin Loss(Θ SSM );

[0187] The training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.001. The data set is divided into a training data set and a test data set in a ratio of 8:2. The network model is trained for 10 batches, and the test data set is used for target detection prediction and accuracy evaluation. The termination condition of the training iteration is set to the prediction deviation of the test data set before and after being less than 0.01% or the training batch reaches 500 times.

[0188] After completing the network parameter update through the real small sample data obtained by the radar, the SSM-Net network has the adaptability to the real radar scene echo and the system's spatiotemporal errors. The network can be directly applied to the real echo of the radar to detect moving targets and obtain moving target detection results at different distances and speeds.

[0189] In step 5, the echo data acquired by the radar is used as the model input of the SSM-Net obtained in step 4. The model output of the SSM-Net directly corresponds to the detection results of moving targets at different distances and speeds.

[0190] The effect of the present invention is further illustrated by simulation experiments below.

[0191] According to the system parameters of the airborne equivalent flight test, the full-link data simulation software is used to generate the training sample data required by the FENet network model. The simulation parameters are shown in the following table:

[0192] Table 1 Basic parameters of small sample simulation data

[0193] parameter Numeric Number of array elements in elevation 1 Azimuth array elements 14 Number of pulses 64 Pulse repetition frequency 2500HZ Element spacing 0.25m Signal-to-Noise Ratio 40dB Aircraft height 4000m Carrier speed 120m / s Equivalent yaw angle 2.5°

[0194] This paper conducts comparative experiments with traditional STAP+CFAR and small sample target detection network (MB-FSNet). The data used are 500 range gates of the validation data set. Moving target signals of different powers are added to each range gate, and the clutter-to-noise ratio in the background clutter is 40dB.

[0195] The experimental results are as follows Figure 4 As shown in the figure, when the SCR is below -10dB, the detection accuracy of the traditional STAP+CFAR method drops to about 0, but the prediction accuracy of SSM-Net and MB-FSNet remains above 90%; in the comparison between the SSM-Net model and the MB-FSNet model, when the SCR is between -30dB and -25dB, the prediction accuracy of SSM-Net is about 15% higher than that of MB-FSNet. This experiment fully proves the feasibility of the network.

[0196] In order to further verify the accuracy and superiority of SSM-Net in actual radar moving target detection, the 33 range gate data containing moving target signals in the simulated echo signal are used as the detection unit. The SCR of the echo signal is -25dB. The target is detected using SSM-Net and STAP+CFAR method. The results are as follows: Figure 5 The SSM-Net model can accurately detect moving targets without false alarms, while the traditional STAP+CFAR has different degrees of false alarms, which further shows that the SSM-Net model proposed in the present invention has better detection results.

Claims

1. A radar moving target detection method based on small sample transfer learning and attention mechanism, characterized in that: include: Generate ground clutter background radar echo data for different application scenarios and early warning radar system conditions; Add moving target signals of different powers and speeds to the ground clutter background radar echo data to generate the training data set required for model training; A deep learning feature extraction network based on the attention mechanism, namely the FENet network model, is used to calculate the feature distribution of the FENet network model and its corresponding attention mechanism parameters based on the generated training data set; Based on the twin network model, the measured data samples obtained by radar are used to modify the parameters of the FENet network model, thereby obtaining the SSM-Net model under small samples; The SSM-Net model is applied to the real-time radar echo. The real-time radar echo data is used as the input of the SSM-Net model. The output of the SSM-Net model directly corresponds to the moving target detection result of the radar, that is, the distance and speed parameters of the moving target are detected.

2. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 1 is characterized in that: The generation of simulated echo data of ground clutter background under different application scenarios and early warning radar system conditions includes: Assume that the orbital height of the space-based radar platform is H, the platform movement speed is V, the coordinate axis is established with the northeast sky, and the platform movement direction is the Y axis; the array is a positive side-view uniform linear array composed of N array elements, θ and are the azimuth and elevation angles of the antenna beam pointing to the current clutter block; the spacing between array elements is half a wavelength, and the pulse repetition interval is T r , the pulse repetition frequency is f r =1 / T r , a total of M pulses are emitted within the coherent processing interval; Assume that there are N distance units to be detected. c clutter units are evenly distributed in azimuth, then the normalized Doppler frequency corresponding to the clutter scattering unit is for: Where λ is the radar operating wavelength, f r is the pulse repetition frequency; Normalized spatial frequency corresponding to the clutter block for: For a uniform linear array, the time domain steering vector corresponding to any clutter block is and the airspace steering vector They are: Considering the range ambiguity, array element error and clutter fluctuation problems in radar ground observation scenarios, assuming that each clutter block is statistically independent, the clutter space-time snapshot data of any range unit is represented as the superposition of the echo signals of each clutter block at different ambiguity distances: In the formula, x c Represents the clutter echo data of the unit to be detected, α i Indicates the echo amplitude of the i-th clutter block; γ=[γ1,γ2,…,γ M ] T Indicates the clutter fluctuation component between coherent pulses; represents the array element error component; ⊙ and Respectively represent the Hadamard product and Kronecker product between matrices; Similarly, the echo signal x of the moving target t It is expressed as: In the formula, α t represents the complex amplitude of the signal; f t ,θ t and They represent the normalized Doppler frequency, azimuth and elevation angle of the moving target respectively; The ground clutter background radar echo data y received by the radar consists of scene clutter, moving targets and noise: y=x t +x c +n; Where n represents the Gaussian white noise component.

3. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 1 is characterized in that: The method of adding moving target signals of different powers and different moving speeds to the ground clutter background radar echo data to generate a training data set required for model training includes: Assume that the L1th range gate containing the moving target is regarded as the unit to be detected, and there are T P protection units, the range gate where the training data is selected is located outside the protection unit, and each training selects T E Range gates; The number of range gates selected when expanding data is T E , the signal-to-noise ratio range is -15dB~20dB, the number of moving target categories is K, and the expansion multiple is K C ; The expanded data set in each case reaches T E ×K×K C ; During the training process, the batch Epochs is set to 13; The expanded moving target signal is divided into K categories, of which category 0 is pure background clutter signal, that is, no moving target signal is added; for the category where the target exists, each category corresponds to the normalized Doppler frequency f of the moving target t,k for: Where K represents the total number of categories of different moving targets; When detecting moving targets, it is assumed that the target is located at the center of the radar beam and its spatial frequency is fixed. Extended target data corresponding to different speeds is generated. The extended training data Y k,h as follows: Y k,h =x t,k,h +x c +n; where x t,k,h The corresponding normalized Doppler frequency is w k , with an amplitude of α t,h The extended motion target data is as follows:

4. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 1 is characterized in that: The FENet network model consists of 4 convolutional layers, 2 pooling layers, 4 attention modules, 2 fully connected layers and 1 recognition layer; the input of the FENet network model is the radar training data Y k,h , the output is the moving target speed corresponding to the detection range gate.

5. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 4 is characterized in that: The deep learning feature extraction network based on the attention mechanism, namely the FENet network model, is used to extract the feature distribution of the training data set and its corresponding attention mechanism parameters, including: The various layers of the FENet network model are designed as follows: The convolution layer is used to extract features and extract the main features of the moving target from the input data: F c =Relu(W c *Y k,h +b c ); Among them, Relu is the activation function used by the convolutional layer, F c Represents the feature matrix after the convolutional network, W c represents a convolution kernel of dimension f×f×c; f represents the length and width of the convolution kernel; c represents the number of convolution kernels, that is, the number of output feature channels; * represents the convolution operation; b c is the bias vector; The attention module adopts a space-channel joint attention mechanism, including a space attention module and a channel attention module; the space attention module targets the beam-Doppler two-dimensional space of the radar data, and the channel attention module targets the real and imaginary parts of the radar data; the channel attention module first performs global average pooling and global maximum pooling on the convolutional feature map of the channel to obtain two 1×1×C feature maps; In the formula, represents the feature vector obtained after the average pooling operation, represents the feature vector obtained by the maximum pooling operation, mean(·) and max(·) represent the maximum pooling operation and the average pooling operation respectively; The obtained feature vector and They are sent to a two-layer neural network MLP respectively. The size of the feature matrix output by the neural network MLP is the same as the size of the input feature matrix. The activation function used in the network is Relu, and the updated feature vector is obtained. and In the formula, represents the average pooling feature parameters extracted by the neural network, Represents the maximum pooling feature parameters extracted by the neural network; For the feature vector and Perform the sum operation and then use the sigmoid function for normalization to generate the final channel attention feature vector M C (F): Where the channel attention feature vector M C (F) represents the weight of each channel of the input data; The spatial attention module performs global average pooling and global maximum pooling on the feature vector and Perform concatenation and convolution operations, and finally perform global normalization of weights through the sigmoid function to obtain the updated weight vector matrix M s (F): Where [;] represents the splicing operation, f s×s (·) represents the convolution operation, s is the size of the convolution kernel, and softmax is the classification identifier; M S (F) is to weight the convolutional network of a single channel; The pooling layer is used to reduce the dimension of the data and compress the features after convolution: F P =γ P (F c ;S p ,F p ); Among them, F P is the feature matrix after pooling, γ P represents the downsampling operator, S p is the window length of the downsampling operator, F P is the downsampling operation step size; The fully connected layer is used to flatten the feature matrix extracted by the convolution operation and compress the flattened eigenvalues: y f μγ f ( <W f ,widow(F P )>+b f ); Among them, y f is the output feature of the fully connected layer, vec(·) represents the vectorization of the matrix, converting the three-dimensional feature vector into a two-dimensional vector, <·> represents the inner product of two vectors, and W f represents the weight vector, b f represents the bias vector, γ f Represents the activation function, where the first fully connected layer of the FENet network uses the tanh activation function, and the second layer uses the sigmoid activation function; The recognition layer uses a softmax classifier, and the probability of belonging to the kth class target is obtained based on the output data of the fully connected layer as follows: The predicted value of FENet is the probability value of multiple categories. The loss function used in the experiment is the mean square error. The specific mathematical expression is: Where M total Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample; Solving the FENet network model, the process is as follows: After completing the construction of the FENet network model, it is necessary to solve the parameters of the network model through training data; set the network feature distribution in the FENet model and its corresponding attention mechanism parameter set as Θ FEN , the training process of the model is described as follows: I FEN =argmin Loss(Θ FEN )。 6. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 5 is characterized in that: The feature matrix F after the convolutional network c The output feature dimension is h×w×c, where h and w are the length and width of the output data, respectively, and their values ​​are related to the convolution step length s. c The details are as follows: in This is a round-up operation.

7. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 5 is characterized in that: The training process of the model is: I FEN =argmin Loss(Θ FEN ); The above training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.001; the data set is divided into a training data set and a test data set in a ratio of 8:2; the network model is trained for every 10 batches, and the test data set is used for target detection prediction and accuracy evaluation; The termination condition of the training iteration is set as the prediction deviation of the test data set before and after is less than 0.01% or the training batch reaches 500 times.

8. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 5 is characterized in that: Based on the twin network model, the measured data samples obtained by radar are used to modify the parameters of the FENet network model, so as to obtain the SSM-Net model under small samples, including: set up To extract radar simulation data using FENet network, that is, the features of the supporting dataset; The FENet network is used to extract the real radar echo data, that is, the characteristics of the query data set; the nearest centroid cosine similarity function is used to characterize the similarity of the above two types of data after passing through the feature extraction network; The cosine similarity y between the feature vector of sample Y and the centroid of class k k-ssm The expression is: Where A k represents the samples belonging to category k in the support data set; <·,·> represents the cosine similarity, τ is the scaling value of the cosine similarity, which is used to scale the interval of the cosine similarity; w k Represents the centroid of each class, and its mathematical expression is: The loss function of the small sample network still uses the mean square error function, which is: Where M total-SSM Indicates the total number of samples in batch training, I k Represents the classification label vector corresponding to the query sample. After completing the construction of the model based on the twin network idea, the parameters of the SSM-Net network model are trained and solved through the radar simulation data level and real-time data set, and all parameter sets in the SSM-Net model are set to Θ SSM , the training process of the model is described as follows: I SSM =argmin Loss(Θ SSM )。 9. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 8 is characterized in that: The τ is set to 0.

1.

10. The radar moving target detection method based on small sample transfer learning and attention mechanism according to claim 8, characterized in that: The parameters of the SSM-Net network model are trained and solved through radar simulation data level and real-time data set, including: the training process of the model is: I SSM =argmin Loss(Θ SSM ); The above training process calls the model.fit function in tensorflow, uses Adam as the optimization strategy, and sets the learning rate to 0.

001. The data set is divided into a training data set and a test data set in a ratio of 8:

2. The network model is trained for 10 batches, and the test data set is used for target detection prediction and accuracy evaluation. The termination condition of the training iteration is set to the prediction deviation of the test data set before and after being less than 0.01% or the training batch reaches 500 times.

Images