Radar target classification method based on uncertainty and data rejection

Abstract

The application discloses a radar target classification method based on uncertainty and data rejection, and the implementation steps are as follows: a training set is constructed, and then a loss function with uncertainty representation ability is generated to train a convolutional neural network; a radar echo to be tested is processed; an uncertainty index of the processed radar echo to be tested is calculated; whether the radar echo to be tested is rejected is judged; if the radar echo to be tested is rejected, the radar echo sample is regarded as a low-quality radar echo signal polluted; if the radar echo to be tested is not rejected, a target classification result of the radar echo to be tested is output. The method rejects low-quality radar echo signals such as clutter and deceptive jamming, and improves the classification recognition rate of a radar automatic recognition system. Meanwhile, compared with other rejection methods, the uncertainty index is used as a rejection criterion, and the low-quality radar echo can be more effectively distinguished.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and more specifically relates to a radar target classification method based on uncertainty and data rejection in the field of radar signal processing technology. This invention can reliably classify radar echoes in environments with clutter, low signal-to-noise ratios, and deceptive tactics. Background Technology

[0002] In practical applications, radar tracks targets at long range and continuously obtains their pulse echo data. Existing target identification methods make decisions on each pulse echo without rejecting any. However, the target echo data acquired in actual engineering applications is affected by various factors, such as clutter from other targets, weather conditions, and enemy aircraft interference. This interfered, low-quality data has a severely mismatched signal support region compared to the data in the training database, lacking extractable classifiable features, making it impossible for the RATR system to accurately classify and identify it. Without filtering, these contaminated radar echoes, mixed with separable, high-quality radar echoes, will significantly reduce the final classification and recognition rate. Therefore, how to effectively remove low-quality radar echo data is one of the challenges in radar target classification.

[0003] In his paper "Target-attentional CNN for Radar Automatic Target Recognition with HRRP" (Journal: Singal Processing, DOI: 10.1016 / j.sigpro.2022.108497), Jian Chen proposed a radar target recognition method based on attention-weighted CNN convolutional recognition networks. This method extracts deep semantic features for HRRP using convolutional layers, then sequentially feeds these features into an LSTM model to obtain weights for different distance units. The original features are then weighted using these weights to obtain the final features, which are then fed into a classification layer for HRRP target discrimination. While this attention-weighted approach achieves good recognition performance in high-quality radar signal identification, it still has shortcomings. Because it relies on the premise that the training and test sets are equally distributed to test network performance, it lacks the ability to reject incorrect targets. Therefore, in real environments with clutter and deceptive interference, radar echoes that are interfered with can still be misclassified as other targets. This leads to misclassification by the classification method, resulting in a decrease in the recognition rate and seriously affecting the reliability of the RATR system.

[0004] Xi'an University of Electronic Science and Technology proposed a data rejection target recognition method based on random forest in its patent application "Ground Target Classification Method Based on Random Forest and Data Rejection" (Patent Application No.: 201810874485.5, Authorization Announcement No.: CN 109190673 B). The specific steps of this method are as follows: First, preprocess the training sample set; second, extract the feature matrix from the training sample set; third, train the random forest classifier; the above is the training stage. In the testing stage, first, preprocess the test samples using the same method; second, extract the feature vector of the test samples; third, calculate the output probability vector; finally, compare the output probability vector with a pre-set threshold. If the output probability vector is greater than the preset threshold, the test sample category is output according to the probability vector; otherwise, the radar echo is rejected. While this method can reject low-quality radar echo samples, it suffers from several drawbacks. The class probability vector only represents the model's judgment result, not its confidence level. Furthermore, the model exhibits overconfidence, making it difficult to effectively distinguish between low-quality and high-quality radar echo samples using only the class probability vector as the rejection criterion. In addition, the pre-set threshold lacks robustness and heavily relies on the researcher's prior knowledge, making it difficult to fine-tune. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a radar target classification method based on uncertainty and data rejection. This aims to solve two key problems: low target recognition rate in real-world testing environments and the difficulty of existing rejection methods effectively identifying low-quality radar echoes using rejection criteria.

[0006] The idea behind this invention is to correctly classify high-quality radar echo signals with distributions not significantly different from the training set, while rejecting low-quality radar echo signals such as clutter and deceptive interference, thus addressing the problem of low target recognition rates in real-world testing environments. Simultaneously, this invention utilizes subjective logic to quantify the uncertainty index of radar echoes, consciously widening the gap between high-quality and low-quality radar echo samples during network training. During testing, the uncertainty index of the network on the test samples is used as the rejection criterion, more accurately distinguishing between low-quality and high-quality radar echo samples. By solving these two problems, it can effectively and reliably classify different radar targets in real-world testing environments with interference.

[0007] The specific steps of this invention are as follows:

[0008] Step 1, Generate the training set:

[0009] Step 1.1: Compile a sample set of radar echoes of at least two different categories and their corresponding category labels received by the radar;

[0010] Step 1.2: Use the centroid alignment method and the L2 normalization method to preprocess the data of each HRRP radar echo in the sample set in sequence, and use the preprocessed sample set as the training set.

[0011] Step 2, construct the convolutional neural network:

[0012] Step 2.1: Construct a 17-layer convolutional neural network, with the following structure: first convolutional layer, first batch normalization (BN) layer, first activation layer, first pooling layer, second convolutional layer, second BN layer, second activation layer, second pooling layer, third convolutional layer, third BN layer, third activation layer, third pooling layer, fourth convolutional layer, fourth BN layer, fourth activation layer, fourth pooling layer, and fully connected layer; the fully connected layer is the classification layer, and the remaining layers are feature extraction layers.

[0013] Step 2.2: Set the kernel size of the first to fourth convolutional layers to 9, and set the number of kernels to 32, 64, 128, and 128 respectively. The first to fourth activation layers are all implemented using the ReLU activation function. The first to fourth pooling layers are all set to max pooling. The number of output nodes of the fully connected layer is set to K, where K represents the total number of aircraft target categories.

[0014] Step 3, generate the loss function with uncertainty representation capability as follows:

[0015]

[0016] Where N represents the total number of samples in the training set, i represents the sample number in the training set, and α i λ represents the Dirichlet distribution parameters corresponding to the i-th sample in the training set, and μ represents the weight coefficients, where λ and μ are arbitrarily chosen numbers in the range [0,1]. i L represents the radar echo uncertainty index corresponding to the i-th sample in the training set. ace (·) represents the loss function used to model the uncertainty index of radar echo results by generating different radar echo results using the Dirichlet distribution. EUC1 (·) represents the loss function that uses KL divergence to better model the uncertainty index of radar echo results, L EUC2 (·) represents u i The loss function is incorporated into the squared term loss function, which uses different operations on training radar echoes that are correctly and incorrectly classified to widen the gap between low-quality and high-quality radar echoes.

[0017] Step 4: Train the convolutional neural network using the generated loss function that has the ability to represent uncertainty;

[0018] Step 5: Process the radar echo to be tested;

[0019] Step 5.1: Use the target HRRP echo signal received by the high-resolution radar in real time as the test radar echo;

[0020] Step 5.2: Using the same method as in step 1.2, process the real-time received HRRP signal to obtain the processed test radar echo;

[0021] Step 6: Calculate the uncertainty index of the processed test radar echo;

[0022] Step 6.1: Input the processed test radar echo into the trained convolutional neural network, and the fully connected layer of the network outputs the evidence vector.

[0023] Step 6.2: Calculate the uncertainty index of the test radar echo;

[0024] Step 7: Determine whether the uncertainty index of the test radar echo is greater than the threshold of 0.5. If yes, proceed to step 8; otherwise, proceed to step 9.

[0025] Step 8: The test radar echo is treated as a contaminated, low-quality radar echo signal and rejected.

[0026] Step 9: The category corresponding to the maximum value in the processed test radar echo evidence vector is taken as the target classification result of the test radar echo.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] First, because this invention uses a loss function with uncertainty representation capabilities to quantify the uncertainty index of radar echoes, it overcomes the shortcomings of existing technologies that directly use the probability vector output by the network to effectively distinguish between low-quality and high-quality radar echo samples. This allows the invention to more accurately describe the model's confidence in the discrimination results, and at the same time, to more accurately mine low-quality radar echo samples that are mismatched with the training library.

[0029] Second, this invention determines whether the uncertainty index of the radar target output is less than a threshold. If so, the test sample is considered a contaminated, low-quality radar echo signal, and the test sample is rejected for classification in the next frame of radar echo test samples. Otherwise, the category corresponding to the maximum value in the test radar echo evidence vector is taken as the target classification result of the test sample. This overcomes the shortcomings of existing technologies in real battlefield environments with clutter and deceptive interference, where the low-quality radar echoes are inconsistent with the distribution of training library samples, resulting in low recognition rates. This invention has the advantages of reducing false positives and improving classification recognition rates. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0031] The following is combined Figure 1 The implementation steps of the embodiments of the present invention will be further described below.

[0032] Step 1: Generate the training set.

[0033] Step 1.1: From the one-dimensional high-resolution range profile (HRRP) echo signals of three different aircraft models (Cessna, An-26, and Yark-42), through target attitude calculation, 35,840 HRRP echoes from segments 2 and 5 of the Yark-42, 42,512 HRRP echoes from segments 6 and 7 of the Cessna, and 51,200 HRRP echoes from segments 5 and 6 of the An-26 are selected, resulting in a total of 129,552 HRRP radar echoes and their corresponding category labels forming a sample set.

[0034] Step 1.2: The centroid alignment method and the L2 normalization method are used sequentially to preprocess each HRRP radar echo in the sample set, thereby eliminating the attitude and amplitude sensitivity of the radar echoes in the sample set. The preprocessed sample set is then used as the training set.

[0035] Step 2: Construct a convolutional neural network.

[0036] Step 2.1: Construct a 17-layer convolutional neural network, whose structure is as follows: first convolutional layer, first batch normalization (BN) layer, first activation layer, first pooling layer, second convolutional layer, second BN layer, second activation layer, second pooling layer, third convolutional layer, third BN layer, third activation layer, third pooling layer, fourth convolutional layer, fourth BN layer, fourth activation layer, fourth pooling layer, and fully connected layer; where the fully connected layer is the classification layer, and the remaining layers are the feature extraction layers.

[0037] The parameters of the convolutional neural network are set as follows. The kernel size of the first to fourth convolutional layers is set to 9, and the number of kernels is set to 32, 64, 128, and 128 respectively. The first to fourth activation layers are all implemented using the ReLU activation function, and the first to fourth pooling layers are all set to max pooling. The number of output nodes of the fully connected layers is set to K, where K represents the total number of aircraft target categories. In this embodiment, K = 3.

[0038] Step 3, generate the loss function with uncertainty representation capability as follows:

[0039]

[0040] Where N represents the total number of samples in the training set, the subscript i represents the sample number in the training set, and α i λ represents the Dirichlet distribution parameters corresponding to the i-th sample in the training set; λ and μ represent the weight coefficients, which are generally constants less than or equal to 1 and greater than or equal to 0, and are both 1 in this embodiment; L ace (·) represents the cross-entropy loss function based on the Dirichlet distribution. By incorporating the Dirichlet distribution, it completes the modeling of the uncertainty distribution, enabling the network output to represent uncertainty. Its role is to help the network extract separable features and achieve correct classification. L EUC1 (·) Loss function and L EUC2 (·) All loss functions are regularization loss functions, used to assist L ace (·) Learn better features and avoid overfitting. EUC1 The loss function (·) better models radar echo uncertainty by treating paired and misclassified training samples differently in the current iteration, prompting the network to learn strongly separable features while ignoring some weakly separable features that are prone to misjudgment, thus solving the problem of overconfidence in traditional neural networks. EUC2 The loss function treats correctly classified HRRP training samples and misclassified HRRP training samples separately. For correctly classified HRRP training samples, the uncertainty index is set below 0.5, while for misclassified HRRP training samples, the uncertainty index is set above 0.5. This approach avoids situations where some samples, despite being correctly classified, have a high uncertainty index due to insufficient network optimization, and similarly avoids situations where some misclassified, difficult-to-classify samples have a low uncertainty index. This conscious introduction of uncertainty during training widens the difference in uncertainty index between easily classified and difficult-to-classify samples, making it reasonable to use u=0.5 as the threshold during testing, thus solving the problem of difficulty in determining the threshold for different datasets.

[0041] The α i with u iThe specific calculation formula is as follows. Let the value of the i-th training HRRP sample in the training set, output by the fully connected layer of the neural network, be E. i =(e i1 ,e i2 ,...,e iK ), define α i =(α i1 ,α i2 ,...,α iK )=(e1+1,e2+1,...,e K +1), let The uncertainty index of the training radar echo is then: K represents the number of radar echo categories; in this embodiment, K = 3.

[0042] The L ace The specific form of the loss function is shown below:

[0043]

[0044] in, Represents the double gamma function, y i y represents the one-hot vector of the class label corresponding to the i-th training HRRP sample. ij Represents y i The value corresponding to the j-th element.

[0045] The L EUC1 The specific form of the (·) function is shown below:

[0046]

[0047] in, KL[·] represents the one-hot vector of the class label predicted by the radar echo network for the i-th training HRRP sample, and KL[·] represents the Kullback-Leibler divergence function. The mask Dirichlet distribution parameters represent the parameters corresponding to the i-th training HRRP sample in the training set. ⊙ represents the vector dot product operation. D(p i |α i ) and D(p i |1) respectively represent the α i The Dirichlet distribution with parameters of a K-dimensional all-1 vector.

[0048] The L EUC2 The specific form of the (·) function is shown below:

[0049]

[0050] Among them, u i The uncertainty index represents the training HRRP sample with the i-th number.

[0051] Step 4: Train the convolutional neural network using the generated loss function that has the ability to represent uncertainty.

[0052] The training set is input into the convolutional neural network. The loss function with uncertainty representation capability generated in step 3 is used to calculate the error between the output of the fully connected layer in the convolutional neural network and the class label. This error is propagated down from the top layer of the convolutional neural network to update the parameters of the convolutional neural network. The above process is iterated until the loss function converges, and the trained convolutional neural network is obtained.

[0053] Step 5: Process the radar echo to be tested.

[0054] Step 5.1: Use the target HRRP echo signal received in real time by the high-resolution radar as the test radar echo.

[0055] Step 5.2: Using the same process as in step 1.2, process the HRRP signal received in real time to obtain the processed test radar echo.

[0056] Step 6: Calculate the uncertainty index of the processed test radar echo.

[0057] Step 6.1: Input the processed test radar echo into the trained convolutional neural network, and the fully connected layer of the convolutional neural network outputs the evidence vector E = (e1, e2, ..., e...). K The radar echo category index corresponding to the subscript index of the maximum value of the elements in the evidence vector is used as the processed test radar echo category. In this embodiment of the invention, K=3, assuming the output evidence vector E=(1,1,10), the third element has the largest value, then the processed test radar echo belongs to the third type of radar echo target corresponding to the third element.

[0058] Step 6.2: Calculate the uncertainty index of the test radar echo according to the following formula.

[0059]

[0060] Where u represents the uncertainty index of the test radar echo, K represents the total number of radar echo categories, and in this embodiment K = 3, e j This represents the value corresponding to the j-th element in the evidence vector.

[0061] Step 7: Determine whether the uncertainty index of the test radar echo is greater than the threshold of 0.5. If yes, proceed to step 8; otherwise, proceed to step 9.

[0062] Step 8: The test radar echo is treated as a contaminated, low-quality radar echo signal and rejected.

[0063] Step 9, process the test radar echo evidence vector E = (e1, e2, ..., e K The category corresponding to the maximum value in the test radar echo is taken as the target classification result.

[0064] The effects of the present invention will be further explained below with reference to simulation experiments.

[0065] 1. Simulation experimental conditions:

[0066] The software platform of this invention is the Windows 10 operating system and the PyTorch framework.

[0067] 2. Simulation Experiment Content:

[0068] The simulation experiment of this invention uses a ground-based inverse synthetic aperture (ISAR) C-band radar, transmitting linear frequency modulated pulse signals. The radar center frequency is 5520MHz, the pulse repetition frequency is 400Hz, the signal bandwidth is 400MHz, and the range resolution is 0.375 meters. HRRP echo data of three different aircraft types (Yark-42, Cessna, and An-26) were acquired using this radar. The training set consisted of radar echo data from segments 2 and 5 of the Yak-42, segments 6 and 7 of the Cessna, and segments 5 and 6 of the An-26. The remaining radar echoes were used as test data. The test samples contained a small amount of clutter, untracked targets, and low-quality target radar echoes with low signal-to-noise ratios. During the experiment, the method of this invention and conventional classification methods were used to classify the test samples. Conventional classification methods refer to traditional machine learning algorithms and deep learning algorithms that do not have the ability to output uncertainties and do not reject classification results.

[0069] The existing conventional identification method, TARAN classification, refers to the HRRP classification method proposed by Bo Chen et al. in "Target-aware recurrent attentional network for radar HRRP target recognition[J].SignalProcessing,2019,155:268-280.", abbreviated as TARAN method.

[0070] The existing conventional recognition method, TACNN classification method, refers to the HRRP classification method proposed by Jian Chen et al. in "Target-attentionalCNN for Radar Automatic Target Recognition with HRRP[J].Signal Processing,2022,196:108497.", abbreviated as TACNN method.

[0071] 3. Simulation Result Analysis:

[0072] To evaluate the method of this invention compared to conventional classification methods, the test recognition rate of each class of ground targets in the simulation experiment of this invention is calculated according to the following formula. The calculation method of the class average recognition rate index is given below:

[0073]

[0074] Where Tr i Q represents the number of correctly identified test samples in the i-th class. i N represents the total number of test samples in the i-th class. C This indicates the number of types of test radar targets. For conventional methods, its Q... i Let Q be the total number of test samples in class i. For this method, Q... i This represents the total number of test samples where the uncertainty index is below a set threshold of 0.5. A higher index indicates better performance of the target recognition method.

[0075] Table 1 shows the recognition rate of different methods on test radar echo datasets with clutter, target not tracked, and low signal-to-noise ratio, as follows:

[0076] Table 1. Comparison of results from conventional classification methods and the test recognition rate of this invention.

[0077] Model Method Average recognition rate TARAN 90.04％ TACNN 93.14％ Ours (not refused to be judged) 90.12％ Ours (refusal to sentence) 99.17％

[0078] As can be seen from the average recognition rate in Table 1, the method of this invention has significant advantages over conventional classification methods. Specifically, compared with the TARAN method based on the RNN model, the test recognition rate of radar targets using this invention is improved by about 10%, resulting in a significant improvement in overall recognition performance. Even compared with the latest TACNN method based on the CNN model, its recognition rate is still improved by about 6%, demonstrating the effectiveness of the proposed algorithm. Since the method of this invention and the conventional method use the same data preprocessing methods, and the recognition performance is 90.12% without rejection, the performance improvement is mainly due to data rejection. That is, data rejection can effectively remove low-quality radar echo test samples such as clutter and deceptive interference, while conventional methods accept low-quality radar echo test samples such as clutter and deceptive interference and perform random classification, which leads to a decrease in the recognition rate and average recognition rate of each type of radar target.

Claims

1. A radar target classification method based on uncertainty and data rejection, characterized in that, The method comprises the following steps: Step 1, generating a training set: Step 1.1, at least two different categories of radar echoes received by the radar and their corresponding category labels are combined to form a sample set; Step 1.2, the center of gravity alignment method and the two-norm normalization method are used in turn to preprocess each HRRP radar echo in the sample set, and the preprocessed sample set is used as the training set; Step 2, constructing a convolutional neural network: Step 2.1, a 17-layer convolutional neural network is built, and the structure is as follows: a first convolutional layer, a first BN layer, a first activation layer, a first pooling layer, a second convolutional layer, a second BN layer, a second activation layer, a second pooling layer, a third convolutional layer, a third BN layer, a third activation layer, a third pooling layer, a fourth convolutional layer, a fourth BN layer, a fourth activation layer, a fourth pooling layer, and a full connection layer; wherein the full connection layer is a classification layer, and the remaining layers are feature extraction layer parts; Step 2.2, the convolution kernel size of the first to fourth convolutional layers is set to 9, and the number of convolution kernels is set to 32, 64, 128, and 128 in turn; the first to fourth activation layers are realized by using the ReLU activation function; the first to fourth pooling layers are set to maximum pooling; and the output node number of the full connection layer is set to K, wherein K represents the total number of aircraft target categories; Step 3, the loss function with uncertainty representation capability is generated as follows: ； wherein, represents the total number of samples in the training set, denotes the number of samples in the training set, represents the Dirichlet distribution parameter corresponding to the th sample in the training set, and represents the weight coefficient, and is a number randomly selected in the range of , represents the radar echo uncertainty indicator corresponding to the th sample in the training set, represents the loss function for modeling the uncertainty indicator of the radar echo result by means of the Dirichlet distribution to generate different radar echo results, represents the loss function for modeling the uncertainty indicator of the radar echo result by means of the divergence, represents the loss function for incorporating the into the square term loss function, and taking different operations on the training radar echoes according to the correct and incorrect classification to thereby widen the gap between the low-quality radar echoes and the high-quality radar echoes. The loss function for modeling the uncertainty index of the radar echo result is as follows: ； wherein, is the total number of radar echo types, is the class label one-hot vector corresponding to the th sample in the training set, represents the value corresponding to the th element in the vector, represents the double gamma function, represents the sum of Dirichlet parameters corresponding to the th sample in the training set, , represents the value corresponding to the th element in the vector; The method comprises the steps of: The loss function modeling the uncertainty indicator of the radar echo result is as follows: ； in, Representative training set The one-hot vector of the category label predicted by the radar echo network for each sample. Represents the Kullback-Leibler divergence function. Representative training set The mask Dirichlet distribution parameters corresponding to each sample , Represents the vector dot product operation. , as well as Representing respectively , and A Dirichlet distribution with parameters as an all-1 vector; The loss function for taking different operations on the training radar echoes according to the correct and incorrect division to widen the gap between low-quality radar echoes and high-quality radar echoes is as follows: ； wherein, represents an uncertainty indicator of the radar echo corresponding to the i-th sample in the training set, ;​ Step 4, training the convolutional neural network using the generated loss function with uncertainty representation capability; Step 5, processing the radar echo to be tested; Step 5.1, taking a target HRRP echo signal received by a high-resolution radar in real time as the test radar echo; Step 5.2, processing the real-time received HRRP signal by using the same method as step 1.2 to obtain the processed test radar echo; Step 6, calculating the uncertainty index of the processed test radar echo; Step 6.1, inputting the processed test radar echo into the trained convolutional neural network to output an evidence vector from the full connection layer of the network; Step 6.2, calculating the uncertainty index of the test radar echo; Step 7, determining whether the uncertainty index of the test radar echo is greater than the threshold value 0.5, if yes, executing step 8, otherwise, executing step 9; Step 8, regarding the test radar echo as a low-quality radar echo signal contaminated, and rejecting the test radar echo; Step 9, taking the category corresponding to the maximum value in the evidence vector of the processed test radar echo as the target classification result of the test radar echo.

Images