Construction method of novel radar target intelligent identification model
The dynamic RCS sequence of radar target recognition is trained through the one-dimensional convolutional neural network model, which solves the problem of insufficient recognition performance of existing radar targets and achieves higher recognition accuracy and performance.
Patent Information
- Application Number
- CN202411591202.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-01-10
AI Technical Summary
The recognition performance of existing radar target recognition methods is not strong and cannot meet the practical application needs, especially the accurate identification of targets in complex environments.
Using the one-dimensional convolutional neural network model, through static and dynamic RCS data simulation calculation and sliding window interception, a dynamic RCS sequence for radar target recognition is constructed, a one-dimensional convolutional neural network structure is designed for training and testing, and a crossover is used. Entropy loss and stochastic gradient descent algorithm optimize network model.
It improves the accuracy and recognition performance of radar target recognition, can effectively identify flight targets in complex environments, and enhances the recognition capabilities of the radar system.
Smart Images
Figure CN119291642A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of radar target recognition, and in particular to a method for constructing a novel radar target intelligent recognition model. Background Art
[0002] Among various sensors, radar has become an important part of the combat defense system with its advantages of all-weather, all-day, long detection distance, high measurement accuracy, and strong real-time performance. It is an indispensable eye in the defense system. The work of radar plays an important role in the detection, early warning, tracking, identification, guidance, guided interception, and damage effect evaluation of the anti-missile system. Traditional radar target recognition methods usually extract target features to obtain target feature information such as radar cross section (RCS) features, micro-motion features, and motion features, and then use decision trees, Bayesian, and K-nearest neighbor methods to identify the target. However, the current radar target recognition methods have problems such as weak recognition performance or inability to be applied in practice.
[0003] In recent years, with the improvement of computer computing performance, a classification and regression technology that requires high-performance computing capabilities, deep learning technology, has achieved great success in various classification recognition and regression fitting tasks. Deep learning uses artificial neural networks with very deep levels to learn the distribution of problems that need to be solved under the training of a large amount of data, so that it can predict the category to which the test sample belongs or the regression value to which it should be mapped. Deep learning technology can usually achieve the best recognition performance or regression accuracy at present. Research on key technologies of radar target recognition based on deep learning can deepen our understanding of deep learning technology and radar target recognition technology, and at the same time can effectively improve the accuracy of target recognition, provide the necessary prerequisites for further observation and evaluation of targets by broadband radar, change the battlefield situation, improve the victory rate of war, maintain national security, and provide strong security assistance for the country's peaceful development. Summary of the invention
[0004] In view of the deficiencies of the prior art, the present invention aims to provide a method for constructing a novel radar target intelligent recognition model.
[0005] In order to achieve the above object, the present invention adopts the following technical solution:
[0006] A method for constructing a novel radar target intelligent recognition model, wherein constructing the recognition model comprises the following steps:
[0007] Step 1: Static RCS data simulation calculation; after building a three-dimensional model of the target, set the radar parameters, then mesh the target, and finally perform electromagnetic scattering calculation;
[0008] Step 2, dynamic RCS sequence calculation: according to the azimuth and pitch angle information of the object to be identified during the actual flight process, the RCS values at the corresponding azimuth and pitch angles are extracted from the static RCS database by table lookup to form a dynamic RCS sequence under a specific track;
[0009] Step 3, design of recognition network model; construct a one-dimensional convolutional neural network structure to directly classify and recognize targets from the RCS sequence obtained in step 2;
[0010] Step 4: Use the dynamic RCS sequence to train the network model; input the RCS sequences in the training set into the network model in batches, and use the corresponding training algorithm to complete the network training.
[0011] It should be noted that in the step 1, the CADFEKO module in the FEKO software is used to perform three-dimensional modeling of five typical aircraft targets; wherein, the radar operating frequency is set to 3 GHz, 5.6 GHz and 9.4 GHz, the physical optics method is selected as the target scattering characteristic calculation method, the grid side length range is λ / 6-λ / 5 during grid generation, and the rest are set to the default values for grid generation; the radar adopts horizontal polarization, the pitch angle range is 0°-180°, the interval is 0.1°, the azimuth angle range is 0°-360°, the interval is 0.1°, and electromagnetic scattering calculation is performed.
[0012] It should be noted that in step 2, the generated RCS sequence fragments are cut into 512 points by sliding window clipping, where the sliding window step size is 100; after sliding window clipping, a total of 153,363 RCS sequence fragments are generated, of which 138,030 RCS sequence fragments are used as training sets, and the remaining 15,333 RCS sequence fragments are used as test sets.
[0013] It should be noted that in step 3, the network model takes a 512-point long RCS sequence as input, and has 6 convolutional layers, 3 pooling layers, 7 activation layers, 2 random dropout layers, 2 fully connected layers and one classification layer.
[0014] It should be noted that the convolution layer and the pooling layer do not expand the input during the convolution and pooling process to avoid extracting incorrect information; in order to extract as much useful information as possible from the 512-point RCS sequence, the translation step size of the convolution kernel in the convolution layer is 1.
[0015] It should be noted that in step 4, the network model uses cross entropy loss as the loss function, uses stochastic gradient descent for network training, and the initial learning rate is 0.0005. Training is performed until the loss function converges and no longer decreases, and finally a weight file is obtained.
[0016] The beneficial effect of the present invention is that the present invention automatically extracts features from a dynamic RCS sequence based on a self-created one-dimensional convolutional neural network, which is used for the task of intelligent recognition of flying targets, thereby improving the accuracy of radar target recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is the RCSnet network structure diagram in the present invention. DETAILED DESCRIPTION
[0018] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.
[0019] Example
[0020] The present invention is a method for constructing a novel radar target intelligent recognition model. The novel radar target intelligent recognition model is constructed by the following method.
[0021] Step 1: Static RCS data simulation calculation.
[0022] The present invention uses FEKO software to simulate the RCS of a target, firstly establishes a three-dimensional model of the target, then sets radar parameters, then meshes the target, and finally performs electromagnetic scattering calculation.
[0023] The present invention adopts the software module CADFEKO of FEKO software to perform three-dimensional modeling on five typical aircraft targets. The radar operating frequency is set to 3GHz, 5.6GHz and 9.4GHz, the target scattering characteristic calculation method selects the physical optics method (Physical Optics, PO), the grid side length range is λ / 6-λ / 5 during grid generation, and the rest are set to the default value for grid generation. The radar adopts horizontal polarization mode, the pitch angle range is 0°-180° (interval is 0.1°), and the azimuth angle range is 0°-360° (interval is 0.1°) to perform electromagnetic scattering calculation.
[0024] Step 2: Dynamic RCS sequence calculation.
[0025] According to the azimuth and pitch angle information of the aircraft (object to be identified) during actual flight, the RCS values at the corresponding azimuth and pitch angles are extracted from the static RCS database by table lookup to form a dynamic RCS sequence under a specific track. The present invention cuts the generated RCS sequence fragments into 512 points by sliding window interception, where the sliding window step size is 100. After sliding window cutting, a total of 153,363 RCS sequence fragments are generated, of which 138,030 RCS sequence fragments are used as training sets, and the remaining 15,333 RCS sequence fragments are used as test sets.
[0026] Step 3: Identify the model design.
[0027] The present invention designs a one-dimensional convolutional neural network (CNN) structure for directly classifying and identifying targets from RCS sequences. Figure 1 As shown in the figure, the network model takes a 512-point long RCS sequence as input, and has 6 convolutional layers (Conv), 3 pooling layers (MaxPooling), 7 activation layers (Activation ReLU), 2 random dropout layers (Dropout), 2 fully connected layers (FC) and a classification layer (Softmax).
[0028] The first layer (Conv_1) is a convolutional layer, the number and size of the convolution kernels (that is, the length in one-dimensional CNN) are 16 and 3 respectively; the second layer is an activation layer, and its activation function is the ReLU function; the third layer (Conv_2) is a convolutional layer, the number and size of the convolution kernels are 16 and 3 respectively; the fourth layer is a ReLU function activation layer; the fifth layer (Max Pooling_1) is a pooling layer, which performs a "2-choose-1" with a step size of 2 on the data of the previous layer; the sixth layer (Conv_3) is a convolutional layer, the number and size of the convolution kernels are 32 and 3 respectively; the seventh layer is a ReLU function activation layer; the eighth layer (Conv_4) is a convolutional layer, the number and size of the convolution kernels are 32 and 3 respectively; the ninth layer is a ReLU function activation layer; the tenth layer (Max The first layer (Conv_5) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_6) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_7) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_8) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_9) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_10) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_11) is a convolution layer, and the number and size of convolution kernels are 64 and 3 respectively; the first layer (Conv_12) is a convolution layer The first layer (Pooling_3) is a pooling layer, which performs a "2-choose-1" operation with a step size of 2 on the data in the previous layer; the 16th layer (FC_1) is a fully connected layer with an output size of 50; the 17th layer is a ReLU function activation layer; the 18th layer (Dropout_1) is a random dropout layer with a random dropout probability set to 0.5; the 19th layer (FC_2) is a fully connected layer with an output size of 5; the 20th layer (Dropout_2) is a random dropout layer with a random dropout probability set to 0.5; the 21st layer (Softmax) is a classification layer, which outputs the predicted probability of each type of target.
[0029] The convolution layer and pooling layer do not expand the input during the convolution and pooling process (i.e., do not fill zeros at the edges) to avoid extracting incorrect information. In order to extract as much useful information as possible from the 512-point RCS sequence, the translation step size of the convolution kernel in the convolution layer is 1.
[0030] Step 4: Train and test the network model using the dynamic RCS sequence.
[0031] In the training stage, the present invention inputs the RCS sequences in the training set into the network model in batches, and completes the network training using the corresponding training algorithm. Specifically, the network model uses cross entropy loss (Cross Entropy Loss) as the loss function, uses stochastic gradient descent (SGD) for network training, and the initial learning rate is 0.0005. The training is carried out until the loss function converges and no longer decreases, and finally the weight file is obtained. In the testing stage, the present invention uses the weight file obtained by training to initialize the network model, and then directly inputs the RCS sequence in the test set to complete the RCS data test.
[0032] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.
Claims
1. A method for constructing a new radar target intelligent recognition model, characterized in that: Constructing the recognition model includes the following steps: Step 1: Static RCS data simulation calculation: after building a three-dimensional model of the target, set the radar parameters, then mesh the target, and finally perform electromagnetic scattering calculation; Step 2, dynamic RCS sequence calculation: according to the azimuth and pitch angle information of the object to be identified during the actual flight process, the RCS values at the corresponding azimuth and pitch angles are extracted from the static RCS database by table lookup to form a dynamic RCS sequence under a specific track; Step 3, design of recognition network model; construct a one-dimensional convolutional neural network structure to directly classify and recognize targets from the RCS sequence obtained in step 2; Step 4: Use the dynamic RCS sequence to train the network model; input the RCS sequences in the training set into the network model in batches, and use the corresponding training algorithm to complete the network training.
2. The method for constructing a new radar target intelligent recognition model according to claim 1, characterized in that: In the step 1, the module CADFEKO in the FEKO software is used to perform three-dimensional modeling on five typical aircraft targets; wherein, the radar operating frequency is set to 3 GHz, 5.6 GHz and 9.4 GHz, the target scattering characteristic calculation method selects the physical optics method, the grid side length value range is λ / 6-λ / 5 during grid generation, and the rest are set to default values for grid generation; the radar adopts horizontal polarization mode, the pitch angle value range is 0°-180°, the interval is 0.1°, the azimuth angle value range is 0°-360°, the interval is 0.1°, and the electromagnetic scattering calculation is performed.
3. The method for constructing a new radar target intelligent recognition model according to claim 1, characterized in that: In the step 2, the generated RCS sequence segments are cut into 512 points by sliding window clipping, where the sliding window step size is 100; after sliding window clipping, a total of 153,363 RCS sequence segments are generated, of which 138,030 RCS sequence segments are used as training sets, and the remaining 15,333 RCS sequence segments are used as test sets.
4. The method for constructing a new radar target intelligent recognition model according to claim 1, characterized in that: In step 3, the network model takes a 512-point long RCS sequence as input and has 6 convolutional layers, 3 pooling layers, 7 activation layers, 2 random dropout layers, 2 fully connected layers and one classification layer.
5. The method for constructing a new radar target intelligent recognition model according to claim 4 is characterized in that: The convolution layer and pooling layer do not expand the input during the convolution and pooling process to avoid extracting incorrect information; in order to extract as much useful information as possible from the 512-point RCS sequence, the translation step size of the convolution kernel in the convolution layer is 1.
6. The method for constructing a new radar target intelligent recognition model according to claim 1, characterized in that: In step 4, the network model uses cross entropy loss as the loss function, uses stochastic gradient descent for network training, and the initial learning rate is 0.0005. The training is carried out until the loss function converges and no longer decreases, and finally the weight file is obtained.