Intelligent recognition method of radar target one-dimensional range profile

Abstract

The present invention discloses a method for intelligent recognition of one-dimensional range profiles of radar targets, which belongs to the field of radar signal processing. In order to solve the problem of poor accuracy of HRRP classification and recognition of radar targets under small sample conditions, the theoretical idea of ​​meta-learning is rationally used to decouple the characteristics of tasks from the commonalities between tasks, and the task characteristics and the commonalities between tasks are modeled and learned by the basic learner and the meta-learner. Through decoupling, the basic learner and the meta-learner simultaneously achieve the best results in these two aspects. By transferring the learning experience of task characteristics and commonalities between tasks, the classification and recognition performance of multiple types of radar targets under small sample conditions is effectively improved, and the classification and recognition ability of radar in complex environments is enhanced, which has the value of promotion and application.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and in particular relates to a method for intelligently identifying one-dimensional range profiles of radar targets. Background Art

[0002] A high-resolution range profile (HRRP) is the projection of a target's scattering center onto the radar's line of sight. It reflects the target's location and, to a certain extent, some of the target's structural information. It offers advantages such as ease of storage, processing, and access. For cooperative targets, acquiring a high-resolution range profile (HRRP) is relatively simple, allowing for a large number of target samples. However, in actual combat environments, radar targets are often non-cooperative. During radar detection, the situation changes rapidly, suitable detection angles become fleeting, and non-cooperative targets may perform evasive maneuvers while being detected by our radar. This makes it impossible for our radar to track and detect non-cooperative targets for extended periods of time. This significantly increases the difficulty of acquiring target data, significantly reducing the quantity and completeness of the acquired HRRP data.

[0003] With the advent of the big data era, deep learning methods have been widely used in radar target recognition in recent years. Deep learning models can autonomously learn the inherent patterns and representational hierarchies of sample data, but they often rely on large amounts of sample data and a deep network structure to achieve effective recognition results. In actual combat environments, data on non-cooperative targets is scarce and from a single perspective. For such small sample data, overly deep deep learning networks can lead to model overfitting. Therefore, HRRP classification and recognition of radar targets under small sample conditions has become a research hotspot in radar target recognition.

[0004] In view of the poor HRRP classification and recognition accuracy of radar targets under small sample conditions, how to improve the classification and recognition rate of targets under small sample conditions is the key to improving radar target recognition capabilities and is also one of the research focuses of HRRP recognition technology. Summary of the Invention

[0005] To address the problem of poor HRRP classification and recognition accuracy of radar targets under small sample conditions, how to reasonably utilize the theoretical idea of ​​meta-learning to decouple the characteristics of tasks from the commonalities between tasks, and transfer the experience of task characteristics and commonalities between tasks through basic learners and meta-learners to enhance the recognition accuracy of new tasks under small sample conditions.

[0006] The method for intelligently identifying a radar target one-dimensional range profile according to the present invention comprises the following steps:

[0007] Step 1: Construct a dataset for deep meta-learning. Obtain a high-resolution range image of the radar target, divide the simulated and measured range image data into a meta-training set and a meta-test set, respectively, and then divide the data in the meta-training set and meta-test set into a training set and a test set.

[0008] In terms of data structure, the data structure of traditional machine learning is generally divided into training set and test set, while deep meta-learning is similar to traditional machine learning and is also divided into meta-training set and meta-test set. The difference is that the meta-training set and meta-test set are not sample data, but a collection of various tasks, and each task is further divided into training set and test set.

[0009] Step 2: Build a basic learner. The input of the basic learner is the high-resolution range image of each task, and the output is the model parameters. The model structure is built using a convolutional neural network.

[0010] The base learner models the characteristics of the task and is required to discover the inherent patterns in each task dataset. Deep learning models rely on powerful feature representation capabilities to extract deep-level features of the target and provide a good fit to the data. Therefore, based on the dimensionality of HRRP, a one-dimensional convolutional neural network was designed as the training model for the base learner.

[0011] The basic learner consists of an input layer, three convolutional layers (Conv), three pooling layers (POOL), a batch normalization layer (BN), two fully connected layers (FC), and an output layer. The specific connection method is:

[0012] Conv1→POOL1→Conv2→POOL2→Conv3+BN→POOL3→FC1→FC2

[0013] Among them, the input layer imports a small batch of HRRP target data in the training set of each task; the first two convolutional layers are standard convolutional layers, and the third convolutional layer is a standardized convolutional layer. A batch normalization layer is added on the basis of the convolutional layer to enhance the gradient change of the loss function E. The three convolutional layers all use one-dimensional convolution kernels with sizes of A1, A2 and A3 respectively, and numbers of B1, B2 and B3 respectively, with a step size of D, and the filling method is the same; the pooling layer is an average pooling layer with a step size of H; the number of neurons in the first fully connected layer is F, and the number of neurons in the second fully connected layer is consistent with the number of target categories C; the output layer uses the softmax function to calculate the category probability of each sample, and calculates the loss function value based on the calculation results.

[0014] Because there is often an imbalance in the number of samples between categories under small sample conditions, during the backpropagation process, the category with more samples accounts for a higher proportion of the loss function, which easily guides the model to optimize in the direction that is conducive to outputting the classification results of that category. To solve this problem, we draw on the cosine loss function and the Focal loss function, and on this basis, design a multi-class balanced cosine loss function. This loss function balances the proportion of each category in the loss function by reducing the weight of the easy-to-classify category in the loss function. While improving the sample imbalance problem, it increases the inter-class distance of the feature, reduces the intra-class distance of the feature, and improves the separability of the feature. The specific expression of the loss function of the i-th sample is as follows:

[0015]

[0016] Among them, x i Represents the fully connected layer output feature corresponding to the i-th sample; y i is the true label of the i-th sample data; W j It is the j-th column of the weight matrix W of the fully connected layer, which represents the weight vector corresponding to the j-th type of target; W j and x i The angle between is the cosine value of the angle; parameter a is greater than 0, and its function is to enhance the feature x i With weight W j The real parameter s solves the problem of non-convergence of the loss function; the parameter γ is used to adjust the output weight.

[0017] Then use the Adam optimizer to initialize the parameters θ of the basic learner B Perform optimization and update, subscript B represents the basic learner, and obtain the updated parameter set Where K is the total number of tasks, θ Bk The characteristic parameters of the kth task are finally passed to the meta-learner.

[0018] Step 3: Build a meta-learner: The input of meta-learning is the output parameters of the base learner, the output is the updated parameters of the meta-learner, and the model is the parameter update algorithm.

[0019] The meta-learner is a model that models the commonalities between tasks. Its function is to summarize the training experience of all tasks after each base learner is trained, synthesize new experience, and feed it back to the base learner. The meta-learner can take various forms. It can be any model based on stochastic gradient descent or a parameter update algorithm. Based on the model parameter form passed by the base learner, a parameter update method based on gradient descent is used to make the initial model parameter θ of the meta-learner MUpdate and move closer to the common direction of all tasks.

[0020] The focus of meta-learning is to model the commonality of tasks. The ultimate goal is to find the common direction of all tasks so that the initialization model parameters θ of the meta-learner M The characteristic parameters θ of each task can be learned by the base learner Bk is close, that is, the parameter θ M and characteristic parameter θ Bk The distance between them is getting shorter and shorter. However, sometimes the task types of the meta-training set and the meta-test set may differ significantly. This difference can cause the model to overlearn during meta-learning, hindering the model from finding a better update direction. Therefore, to avoid model bias on certain tasks, we borrow the idea of ​​inequality minimization to minimize task bias. The learning objectives of the meta-learner are designed as follows:

[0021]

[0022] Among them, min is the minimization function, T k is the inequality measure, which is as follows

[0023]

[0024] Among them, min(·) is the minimization function, The initial parameters of the model are θ M The loss of the kth task on the test set of the meta-training set is minimized by calculating the gradient with respect to , i.e. And D k is the characteristic parameter θ Bk and the meta-learner parameters θ M For example, the Euclidean distance is used to represent this distance as follows:

[0025]

[0026] In order to facilitate the subsequent derivation, the distance D k The coefficient is 1 / 2. According to Lagrange's theorem, the distance can be calculated at the parameter θ M The gradient of the distance D is minimized k Therefore, for the distance D k Taking the derivative we get:

[0027]

[0028] And the parameter θ M The update direction is the direction of the objective function gradient descent, that is, Assume that the influence of each task on the meta-learner parameter θ MThe influence of each task on the meta-learner parameters θ is the same. M By superimposing the influence of for:

[0029]

[0030] The parameter update equation of the meta-learner is:

[0031]

[0032] Among them, ε is the update step size of the meta-learner, and λ is the balance coefficient. Finally, the parameters updated by the meta-learner are Feedback is given to the basic learner, and the new initialization model parameters of the basic learner are used to carry out a new round of optimization training for each task, that is,

[0033] Step 4 uses the built model for classification and recognition; the classification and recognition method based on the deep meta-learning model is mainly divided into pre-training stage and testing stage.

[0034] The pre-training stage is structurally divided into two loop structures: the inner loop and the outer loop. The inner loop refers to the cyclic update of parameters between the base learner and the meta-learner, while the outer loop refers to the training rounds of the inner loop. First, the model parameters of the base learner are initialized, and the recognition tasks are extracted from the meta-training set and imported into the base learner for learning. The base learner parameters are updated according to the loss function value; secondly, the model parameters learned by the base learner are imported into the meta-learner for parameter update, and the updated parameters are fed back to the base learner for re-parameter assignment. The cycle repeats until the outer loop number is reached and the final model parameters are saved.

[0035] The testing phase uses the training parameters from the training phase to initialize the model parameters, transferring previous learning experience to the meta-test set. Finally, the initialized model is used for classification and recognition to verify the effectiveness of the method. First, the model parameters saved in the pre-training phase are used as the model's initial parameters and imported into the training set in the meta-test set for model training until the loss function converges. Then, the trained model and parameters are used to classify and recognize the test set of the meta-test set. Finally, the classification recognition rate for each type of target in the test set is calculated.

[0036] 3. Beneficial effects

[0037] Compared with the background technology, the beneficial effects of the present invention are: 1) through the multi-classification cosine loss function, the proportion of each category in the loss function can be balanced by reducing the weight of the easy-to-classify category in the loss function, improving the sample imbalance problem, while increasing the inter-class distance of the feature, reducing the intra-class distance of the feature, and improving the separability of the feature; 2) for the target recognition problem under small sample conditions, on the basis of transfer learning, the commonality between tasks is included in the scope of experience transfer; 3) the characteristics of the task are decoupled from the commonality between tasks by combining the basic learner and the meta-learner. Through decoupling, the two achieve the best results on both the basic learner and the meta-learner; 4) in the pre-training process, based on the convolutional neural network, a pre-training model combining the basic learner and the meta-learner is designed, and the learning experience of the pre-training model is used to perform classification and recognition of radar target HRRP on the small sample test task set, which improves the recognition accuracy of radar targets under small sample conditions and enhances the recognition ability of radar targets in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Implementation flow chart of the method of the present invention.

[0039] Figure 2 It is a simulated HRRP of a certain type of ship target.

[0040] Figure 3 It is the measured HRRP of a certain type of ship target.

[0041] Figure 4 Classification confusion matrix of previous transfer learning methods under small sample conditions with a total number of samples of 25, 50, 100, and 250.

[0042] Figure 5 Classification confusion matrix of the method of the present invention under small sample conditions with a total number of samples of 25, 50, 100 and 250.

[0043] Figure 6 The classification recognition rate of each model changes with the total number of samples. DETAILED DESCRIPTION

[0044] The present invention will be further described below with reference to the accompanying drawings. The embodiments of the present invention are intended to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the scope of protection of the claims shall fall within the scope of protection of the present invention.

[0045] In order to verify the effectiveness of the method of the present invention, this specific implementation method provides an example for identifying five types of ship targets.

[0046] Example 1:

[0047] The specific implementation process of Example 1 is as follows Figure 1 As shown, it can be divided into the following steps:

[0048] Step A-1: ​​Acquire HRRP simulation data and measured data, and divide the data set into a meta-training set and a meta-test set.

[0049] Training phase

[0050] Step A-2: Initialize the model parameters θ of the base learner B , specifically including the weights of the convolutional layer {w conv1 ,b conv1 ,w conv2 ,b conv2 , w conv3 ,b conv3}, the weights of the fully connected layer {w fc ,b fc ,w fc2 ,b fc2 The structural parameter values ​​of the basic learner are shown in Table 1.

[0051] Table 1 Structural parameters of the basic learner in Example 1

[0052] parameter Value parameter Value parameter Value <![CDATA[A1]]> 9 A2 9 <![CDATA[A3]]> 9 <![CDATA[B1]]> 8 <![CDATA[B2]]> 16 <![CDATA[B3]]> 32 D 1 H 2 F 100 C 5

[0053] Step A-3: Randomly select a batch of tasks from the meta-training set and import them into the base learner for base learner training and update. For each task in the batch, import the training set of each task in batches for training, and use the Adam optimizer to optimize and update the parameters to obtain the updated parameter set.

[0054] Step A-4: Import the updated parameter set of the base learner into the meta-learner, and use the parameter update equation of formula (7) to update the parameters to obtain the updated parameters of the meta-learner.

[0055] Step A-5: Feed the updated parameters of the meta-learner back to the base learner and assign them to the base learner as initialization parameters. Repeat steps A-3 and A-4 until the epoch task is randomly selected. End the loop and save the parameters of the final model.

[0056] Testing phase

[0057] Step A-6: Assign the model parameters finally saved in step A-5 during the training phase to the convolutional neural network model in the base learner as the model initial parameters.

[0058] Step A-7: For the tasks in the meta-test set, import the training set of the tasks in batches to perform model training and update until the loss function E converges. End the training and save the model and parameters.

[0059] Step A-7: Use the model and parameters obtained by training the training set to classify and identify the test set of the meta-test set, and calculate the classification and recognition rate of each type of ship target in the test set.

[0060] The effectiveness of the radar target one-dimensional range profile intelligent recognition method provided by the present invention can be further illustrated by the following simulation results.

[0061] Simulation data description: The simulation data used in this example is HRRP data of 10 types of ships obtained by computer simulation software. The simulation software radar parameters are set as follows: radar center frequency 10GHz, bandwidth 80MHz, azimuth range 0-360 degrees, interval 1 degree, 200 range units, resolution 1.875 meters. Finally, 360 azimuth sample data of each type of target are obtained, and the data length of each azimuth is 200. Figure 2 The simulated HRRP of a certain type of ship target is shown in Figure 2. The test data is the measured HRRP data of 5 types of ship targets, and the length of the data is also 500. Figure 3 is the measured HRRP of a certain type of ship target.

[0062] Figure 4 (a)(b)(c)(d) show the classification confusion matrix of previous transfer learning (DTL model) under small sample conditions with a total number of samples of 25, 50, 100 and 250. Figure 5 Figures (a), (b), (c), and (d) show the classification confusion matrices of the method of the present invention (DML model) under the small sample conditions of 25, 50, 100, and 250 samples. It can be clearly seen from the figures that with the increase in the total number of samples, the recognition performance of the DTL model and the DML model are improved. Figure 4 and Figure 5 In (a), under the small sample size of 25, the recognition performance of both the DTL model and the DML model is poor, and the recognition effect on the third category of targets is not good. However, the DML model is still better than the DTL model in recognizing other categories of targets. Figure 4 and Figure 5 In (b) and (c), under the small sample conditions of 50 and 100 samples, the DTL model still has poor recognition effect on the third type of target, while the DML model can perform more accurate classification and recognition due to the transfer of common task experience, and its recognition performance is significantly better than the DTL model. Figure 4 and Figure 5 (d) Under the small sample condition of 250 total samples, both the DTL model and the DML model can maintain good recognition performance due to the increase in the number of samples.

[0063] Figure 6The graph shows how the classification recognition rate of each model changes with the total number of samples. As can be seen, the recognition performance of all three models increases with the increasing number of samples. The DML model achieves the best overall recognition performance, maintaining an average classification recognition rate of over 90%. The CNN model has the worst overall recognition performance. The DTL model is slightly inferior to the DML model, but still outperforms the CNN model. Furthermore, under conditions with fewer samples, the DTL and DML models significantly improve recognition performance under small sample sizes by leveraging previous experience, significantly outperforming the CNN model. However, as the number of samples increases, this improvement diminishes, and the gap in recognition performance between the three models narrows and eventually disappears.

[0064] In summary, the method of the present invention is superior to other learning models as a whole, showing that the transfer method that combines task characteristics and commonalities between tasks is better than the transfer of task characteristics alone. This shows that the method of the present invention is more suitable for the classification and recognition of small sample targets than previous transfer learning methods, verifying the superiority of the method of the present invention.

Claims

1. A radar target one-dimensional range profile intelligent recognition method, characterized in that: The following steps are involved: Step 1: Obtain a high-resolution range image of the radar target, divide the simulated and measured range image data into a meta-training set and a meta-test set, and then divide the data in the meta-training set and meta-test set into a training set and a test set for each recognition task; Step 2: Pre-training phase: Initialize the model parameters of the base learner, extract the recognition task from the meta-training set and import it into the base learner for learning, and update the base learner parameters according to the loss function value; The model parameters learned by the base learner are imported into the meta-learner for parameter update, and the updated parameters are fed back to the base learner for re-assignment of parameters. The cycle is repeated until the epoch recognition task is randomly selected to complete the training, the cycle ends, and the final model parameters are saved. In sub-step 2.1, we design the loss function of the base learner in the pre-training phase. Specifically, we design a multi-class balanced cosine loss function: Among them, x i Represents the fully connected layer output feature corresponding to the i-th sample; y i is the true label of the i-th sample data; W j It is the j-th column of the weight matrix W of the fully connected layer, which represents the weight vector corresponding to the j-th type of target; W j and x i The angle between is the cosine value of the angle; parameter a>0 enhances feature x i With weight W j The angle constraint of ; the real parameter s solves the problem of non-convergence of the loss function; the parameter γ is used to adjust the output weight; Sub-step 2.2: Design of the meta-learner parameter update method in the pre-training phase. Specifically, the learning objectives of the meta-learner are designed as follows: Among them, min(·) is the minimization function, ||·|| represents the Euclidean space distance, θ M is the initialization parameter of the meta-learner, θ Bk is the model parameter of the base learner after learning the kth task, The initial parameters of the model are θ M The loss of the kth task on the test set of the meta-training set is , and K is the total number of training tasks. The parameters of meta-learning are updated by minimizing this learning objective. The specific parameter update equation is as follows: Among them, ε is the update step size of the meta-learner, λ is the balance coefficient, are the updated parameters of the meta-learner, T K For parameter θ M gradient; Step 3: Testing phase: The model parameters saved in the pre-training phase are used as the initial parameters of the model. The training set of the recognition task in the meta-test set composed of measured data is imported for model training until the loss function converges and the training is terminated. The model and parameters obtained from the training are used to classify and recognize the test set of the recognition task in the meta-test set, and the classification recognition rate of each type of target in the test set is calculated.

2. The radar target one-dimensional range profile intelligent recognition method according to claim 1, characterized in that: The model structure of the basic learner is as follows: The basic learner is a basic convolutional neural network. Its layer structure includes an input layer, three convolutional layers Conv, three pooling layers POOL, a batch normalization layer BN, two fully connected layers FC and an output layer. The connection method is: Conv1→POOL1→Conv2→POOL2→Conv3+BN→POOL3→FC1→FC2 Among them, the first two convolutional layers are standard convolutional layers, and the third convolutional layer is a standardized convolutional layer. A batch normalization layer is added on the basis of the convolutional layer to enhance the gradient change of the loss function. The three convolutional layers all use one-dimensional convolution kernels and the filling method is the same; the pooling layer is an average pooling layer; the number of neurons in the second fully connected layer is consistent with the number of target categories; finally, the softmax function is used to calculate the category probability of each sample.

Images