Airborne radar moving target display method based on data preprocessing and CNN
By adopting dynamic target display methods based on data preprocessing and CNN in airborne radar, the problem of poor clutter suppression ability in low signal-to-mismatch environments is solved, and a higher detection probability and accuracy of dynamic target display are achieved.
Patent Information
- Application Number
- CN202510253504.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-05
- Publication Date
- 2025-06-06
AI Technical Summary
The prior art has poor ability to suppress clutter in low signal-to-miscellaneous environments, resulting in difficulty in displaying targets onboard radar.
The airborne radar dynamic target display method based on data preprocessing and CNN is adopted. By constructing an echo signal model, estimating the clutter covariance matrix, building an improved CNN model (including SE attention mechanism and residual network) and performing data preprocessing, the dynamic target display is finally achieved.
In a low-information and miscellaneous environment, the detection probability is improved, the main characteristics of the moving target are effectively extracted, and the accuracy and efficiency of the dynamic target display are improved.
Smart Images
Figure SMS_1 
Figure SMS_3 
Figure SMS_4
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of airborne radar signal processing, and in particular relates to an airborne radar moving target display method based on data preprocessing and CNN. Background Art
[0002] On the modern battlefield, airborne radar plays an irreplaceable role. One of its important functions is moving target indication (MTI). Moving target indication focuses on whether there is a moving target with a certain relative speed in a certain scene, also known as a detection unit (CUT). However, the working environment of modern radar is very complex. The targets to be detected are often in various complex environments, resulting in a large number of background reflection echoes in the echo signal. The target is often annihilated and detection is difficult. Therefore, how to suppress clutter in these strong clutter backgrounds is a problem that radar systems must face.
[0003] Space-time adaptive processing (STAP) technology uses the space-time coupling characteristics of ground clutter to perform two-dimensional filtering on radar clutter in the spatial and temporal dimensions, and realizes real-time adaptive implementation due to the fluctuations of the clutter environment. However, in the actual clutter environment, the rapid movement of the airborne platform causes the ground scattering characteristics in the echo signal to only guarantee local stability, so that the independent and identically distributed training samples required to estimate the clutter covariance matrix are seriously insufficient. Therefore, the estimated clutter covariance matrix cannot effectively reflect the true distribution of clutter in the detection unit, the clutter suppression capability is reduced, and a high false alarm rate appears in the output detection results, making this technology difficult to implement in engineering. To solve the above problems, researchers have successively proposed a variety of methods such as the auxiliary channel method (ACR) based on dimensionality reduction, the local joint processing method (JDL) based on matrix dimensionality reduction, and the sum and difference beam STAP processing structure, as well as a number of measured STAP projects. The relevant research on the STAP-MTI method is essentially to achieve the method that is closest to the optimal performance. But in fact, since different methods have different prerequisites and characteristics, each method has its own limitations. Therefore, there is no perfect algorithm, but it is necessary to select the corresponding algorithm according to the specific application environment.
[0004] Recently, some researchers have proposed moving target display methods related to deep learning. One is the CNN-STAP method, which uses a multi-layer two-dimensional convolutional neural network to predict the high-resolution clutter angle-Doppler spectrum from the low-resolution clutter angle-Doppler spectrum, thereby obtaining a spatiotemporal weight vector and realizing moving target display. Another is the end-to-end moving target display method (ETE-MTI), which uses deep CNN to learn the distribution of clutter and target features. Its input is low-resolution clutter plus target angle-Doppler spectrum, and the label is high-resolution target angle-Doppler spectrum. This method directly implements ETE-MTI to suppress clutter without supporting the traditional STAP algorithm. These methods have great advantages over other STAP-MTI methods, but the currently implemented CNN network is small in scale and has relatively poor reconstruction capabilities. Summary of the invention
[0005] Aiming at the problems in the prior art that the airborne radar has poor clutter suppression capability in a low signal-to-clutter ratio environment and it is difficult to realize moving target display, the present invention provides an airborne radar moving target display method based on data preprocessing and CNN. The method has a higher detection probability than other methods under the same conditions, and can also well extract the main features of the moving target in a low signal-to-clutter ratio environment, so as to better realize the moving target display of the airborne radar.
[0006] The present invention is achieved through the following technical solutions:
[0007] The airborne radar moving target display method based on data preprocessing and CNN specifically includes the following steps:
[0008] Step 1: Construct an echo signal model of the airborne radar and generate echo signals under different normalized Doppler frequencies NDF and signal-to-noise ratios SCR;
[0009] Step 2: Select different distance units from the echo signal as detection units, protection units and training samples, use the training samples to estimate the clutter covariance matrix of clutter data containing moving targets and clutter data without moving targets, and obtain the minimum variance distortion-free response MVDR spectrum set of the two;
[0010] Step 3: Construct an improved CNN model: The improved CNN model includes a SE attention mechanism and a residual network;
[0011] Step 4: Perform data preprocessing on the MVDR spectrum to obtain the final MVDR spectrum set, and use it as the input of the improved CNN model, taking the pairing of the normalized Doppler frequency NDF and the normalized spatial frequency NSF of the moving target as the label;
[0012] Step 5: Train the improved CNN model using the training set to obtain an optimized network model;
[0013] Step 6: The optimized network model is trained through a test set to predict the information of the moving target, namely the Doppler frequency and the spatial frequency.
[0014] Furthermore, in step one, the echo signal model is constructed when the radar array is a positive side-view uniform linear array, and the space-time snapshot vector of the echo signal includes a clutter space-time snapshot vector, a moving target space-time snapshot vector and a noise signal that obeys a complex Gaussian distribution.
[0015] Furthermore, in step 2, in the process of estimating the clutter covariance matrix of the detection unit, data of 2-4 distance units far away from the detection unit are used as training samples; the unit data between the training samples and the detection unit contains the moving target component to be detected, which is used as a protection unit.
[0016] Furthermore, in step 2, the clutter covariance matrix is estimated by using the training samples, specifically by maximum likelihood estimation MLE, and the calculation formula is:
[0017]
[0018] in, is the clutter covariance matrix; L is the number of training samples; x l , 1≤l≤L represents the data of the lth training sample.
[0019] Furthermore, in step 2, the minimum variance distortion-free response (MVDR) spectrum transformation formula is:
[0020]
[0021] Where p is the angle Doppler spectrum, s is the space-time steering vector, and f d，i and f s，i are the normalized Doppler frequency and normalized spatial frequency of the i-th clutter block, respectively.
[0022] Furthermore, in step three, the improved CNN model includes two identical residual modules with attention mechanism, a fully connected layer and a softmax layer, which are used to convert the output into a result corresponding to the label one by one.
[0023] Furthermore, in step 4, the data preprocessing of the MVDR spectrum is to normalize and subtract the MVDR spectra of the radar echo data containing the moving target and the radar echo data not containing the target, perform a singular value decomposition SVD operation, select an appropriate number to determine the retained singular values, and then set the remaining singular values to zero, thereby retaining part of the principal components, and using the retained part to reconstruct a new matrix to obtain the final MVDR spectrum.
[0024] Furthermore, the training set and the test set are both composed of different normalized Doppler frequencies and signal-to-noise ratios, which are combined and generated repeatedly for different times.
[0025] Compared with the prior art, the advantages of the present invention are as follows:
[0026] The airborne radar moving target display method based on data preprocessing and CNN of the present invention adopts the data preprocessing method of SVD decomposition, which can retain the main components of the MVDR spectrum and improve the quality of the data set; more information features are extracted by improving the residual structure in the network model; adding the attention mechanism can more effectively focus on important features; this method can improve the classification accuracy and more effectively achieve the moving target display task. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.
[0028] Figure 1 This is a flow chart of an airborne radar moving target display method based on data preprocessing and CNN according to Embodiment 1 of the present invention;
[0029] Figure 2 This is a structural diagram of an improved network model according to Example 1 of the present invention.
[0030] Among them, (a) is the residual network module, (b) is the SE attention mechanism module, and (c) is the structural diagram of the proposed improved network model.
[0031] Figure 3 This is a flow chart of data preprocessing in Example 1 of the present invention;
[0032] Figure 4 This is a confusion matrix diagram of the airborne radar moving target display method based on data preprocessing and CNN in Example 1 of the present invention. DETAILED DESCRIPTION
[0033] In order to clearly and completely describe the technical solution and its specific working process of the present invention, the specific implementation methods of the present invention are as follows in conjunction with the accompanying drawings of the specification:
[0034] Example 1
[0035] like Figure 1 As shown, this embodiment provides an airborne radar moving target display method based on data preprocessing and CNN, which specifically includes the following steps:
[0036] Step 101: construct an echo signal model of the airborne radar, and generate echo signals under different normalized Doppler frequencies and signal-to-noise ratios;
[0037] Step 102: Select different distance units from the echo signal as detection units, protection units and training samples, use the training samples to estimate the clutter covariance matrix of clutter data containing moving targets and clutter data not containing targets, and obtain the MVDR spectrum set of the two;
[0038] Step 103: construct an improved CNN model: the network model includes a SE attention mechanism and a residual network;
[0039] Step 104: preprocess the MVDR spectrum to obtain a final MVDR spectrum set, and use it as the input of the improved CNN model, and use the pairing of the normalized Doppler frequency and the normalized spatial frequency of the moving target as a label;
[0040] Step 105: training the improved CNN model using a training set to obtain an optimized network model;
[0041] Step 106: training the optimized network through a test set to predict information of the moving target, namely, Doppler frequency and spatial frequency;
[0042] Assume that the radar array is a positive side-view uniform linear array, and the receiving array N = 10, the radar uniformly transmits and receives pulses K = 10 in the coherent processing interval, and samples L = 2000 times in each pulse repetition interval.
[0043] In step 101, the spatiotemporal snapshot vector of the echo signal includes the clutter spatiotemporal snapshot vector, the moving target spatiotemporal snapshot vector, and the noise signal that obeys the complex Gaussian distribution. Assuming that the normalized spatial frequency of the moving target is known, the normalized spatial frequency is assumed to be 0, and the normalized Doppler frequency is taken as [-0.5, 0.5], which is divided into eleven categories and used as labels. To achieve a low signal-to-noise ratio environment, the value of the signal-to-noise ratio is set to [-30, 0]. Generate data sets under different normalized Doppler frequencies and signal-to-noise ratios;
[0044] In step 102, to estimate the clutter covariance matrix of the detection unit, selectn-3 and x n+3 Initially, the data of ten units far from the detection unit are used as training samples. The four units adjacent to the detection unit are selected as protection units because they may contain moving target components to be detected. The clutter covariance matrix of clutter data containing moving targets and excluding targets is estimated using the training samples, using the maximum likelihood estimation (MLE) method, and its formula is:
[0045]
[0046] in, is the clutter covariance matrix; L is the number of training samples; x l , 1≤l≤L represents the data of the lth training sample;
[0047] Furthermore, the minimum variance distortionless response (MVDR) spectrum transformation formula is:
[0048]
[0049] Where p is the angle Doppler spectrum, s is the space-time steering vector, and f d，i and f s，i are the normalized Doppler frequency and normalized spatial frequency of the i-th clutter block respectively;
[0050] The MVDR spectrum set of moving targets and clutter data without targets can be obtained;
[0051] In step 103, if Figure 2 , the improved CNN model includes SE attention mechanism and residual network.
[0052] The SE attention mechanism module compresses the feature map into a feature vector through global average pooling, reduces the number of channels through an FC fully connected layer, and then restores the dimension using the ReLU activation function and another fully connected layer. Finally, the Sigmoid activation function is used for normalization to characterize the weights in the graph features.
[0053] The residual module has two 5×5 convolutional layers with the same number of output channels. Each convolutional layer is followed by a batch normalization (BN) layer and a rectified linear unit (ReLU) function, and the input is added before the final ReLU activation function through a short connection. In order to add the output of the convolutional layer to the input, an additional 1×1 convolutional layer is introduced.
[0054] The improved network structure is composed of two identical residual modules with attention mechanism, and a fully connected layer and a softmax layer are added to convert the output into a result corresponding to 11 labels one by one.
[0055] In step 104, the MVDR spectrum is preprocessed, such as Figure 3 , normalize them separately, then subtract them and perform SVD operation. Select an appropriate number to determine the retained singular values, and set the remaining singular values to zero. Use the retained part to reconstruct a new matrix and output the final MVDR spectrum, while retaining a small part of the principal components and reducing the complexity of the data.
[0056] The final MVDR spectrum set is used as the input of the network model, and the pair of normalized Doppler frequency and normalized spatial frequency of the moving target is used as the label;
[0057] The training set and test set in step 105 and step 106 are composed of MVDR spectrum sets generated by combining 11 NDF values and 6 SCR values. Under the same NDF and SCR conditions for each group, the training set is generated 20 times. In the same way, each group is generated 5 times to form a test set. The network model is trained with the training set, and the specific training setting batch size is 40 and the learning rate is 0.001. The optimized network is trained with the test set to predict the information of the moving target, that is, the Doppler frequency and the spatial frequency.
[0058] After training the improved CNN model, a confusion matrix can be obtained through the testing phase to intuitively display the classification of the model in each category. Figure 4 As shown in the figure, one of the classification experiments was conducted using a data set based on data preprocessing and the improved CNN model proposed in this paper, and an accuracy of about 90.6% was obtained on the test set. In order to avoid the influence of accidental differences and improve the reliability and accuracy of the results, ten experiments were conducted under the same conditions and the average value was taken, resulting in a result of 90.576%. The high accuracy shows that the model has good performance and can effectively classify and predict in the Doppler data task of detecting targets, thereby further realizing the display of moving targets.
[0059] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.
[0060] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.
[0061] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.
Claims
1. The method for displaying moving targets of airborne radar based on data preprocessing and CNN is characterized in that: The specific steps include: Step 1: Construct an echo signal model of the airborne radar and generate echo signals under different normalized Doppler frequencies NDF and signal-to-noise ratios SCR; Step 2: Select different distance units from the echo signal as detection units, protection units and training samples, use the training samples to estimate the clutter covariance matrix of clutter data containing moving targets and clutter data without moving targets, and obtain the minimum variance distortion-free response MVDR spectrum set of the two; Step 3: Construct an improved CNN model: The improved CNN model includes a SE attention mechanism and a residual network; Step 4: Perform data preprocessing on the MVDR spectrum to obtain the final MVDR spectrum set, and use it as the input of the improved CNN model, taking the pairing of the normalized Doppler frequency NDF and the normalized spatial frequency NSF of the moving target as the label; Step 5: Train the improved CNN model using the training set to obtain an optimized network model; Step 6: The optimized network model is trained through a test set to predict the information of the moving target, namely the Doppler frequency and the spatial frequency.
2. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step one, the echo signal model is constructed when the radar array is a positive side-view uniform linear array, and the space-time snapshot vector of the echo signal includes a clutter space-time snapshot vector, a moving target space-time snapshot vector, and a noise signal that obeys a complex Gaussian distribution.
3. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step 2, in the process of estimating the clutter covariance matrix of the detection unit, data of 2-4 distance units far away from the detection unit are used as training samples; The unit data between the training sample and the detection unit contains the moving target component to be detected, which serves as a protection unit.
4. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step 2, the clutter covariance matrix is estimated by using the training samples, specifically by maximum likelihood estimation MLE, and the calculation formula is: in, is the clutter covariance matrix; L is the number of training samples; x l , 1≤l≤L represents the data of the lth training sample.
5. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step 2, the minimum variance distortion-free response (MVDR) spectrum transformation formula is: Where p is the angle Doppler spectrum, s is the space-time steering vector, and f d，i and f s，i are the normalized Doppler frequency and normalized spatial frequency of the i-th clutter block, respectively.
6. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step three, the improved CNN model includes two identical residual modules with attention mechanism, a fully connected layer and a softmax layer, which are used to convert the output into a result corresponding to the label one by one.
7. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: In step 4, the data preprocessing of the MVDR spectrum is to normalize and subtract the MVDR spectra of the radar echo data containing the moving target and the radar echo data not containing the target, perform singular value decomposition SVD operation, select an appropriate number to determine the retained singular values, and then set the remaining singular values to zero, thereby retaining part of the principal components, and use the retained part to reconstruct a new matrix to obtain the final MVDR spectrum.
8. The airborne radar moving target display method based on data preprocessing and CNN as claimed in claim 1, characterized in that: The training set and the test set are both composed of different normalized Doppler frequencies and signal-to-noise ratios, which are combined and generated repeatedly for different times.