A robust radar target recognition method based on carrier frequency assisted attention network

Abstract

This invention discloses a robust radar target recognition method based on a carrier-assisted attention network, comprising: generating a training dataset composed of high-resolution range image data from multiple carrier frequencies and an auxiliary information dataset composed of corresponding carrier frequencies; constructing a carrier-assisted attention network, which includes an embedding layer generative subnetwork and a ViT subnetwork, wherein the embedding layer generative subnetwork is used to encode the carrier frequency information of the input high-resolution range image data, and the ViT subnetwork is used to recognize the input high-resolution range image data; training the carrier-assisted attention network using the auxiliary information dataset and the training dataset; and recognizing the high-resolution range image data using the pre-trained carrier-assisted attention network. This invention incorporates carrier frequency information into the network training process, enabling the network to learn the most separable features for the current carrier frequency, thereby improving the network's ability to recognize data from different carrier frequencies.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a robust radar target recognition method based on a carrier frequency-assisted attention network. Background Technology

[0002] High-resolution range images (HRRAs) are vector sums of the projections of target scattering point echoes acquired from broadband radar signals onto the radar's line-of-sight. They reflect the distribution of target scattering points along the range dimension and are important structural features of the target, useful for target identification. According to the scattering point model, carrier frequency variations significantly affect the amplitude of the HRA, impacting the echoes. Typically, radar databases are built at a specific carrier frequency; however, in practice, to combat interference, the carrier frequency needs adjustment. This leads to a mismatch between the training and test sets, causing a decline in the model's recognition performance or even loss of recognition capability. Attention mechanism network models, due to their powerful feature extraction capabilities, are widely used in image recognition. Similarly, they can be applied to the identification of high-resolution radar range images.

[0003] The patent document "A Radar High-Resolution Range Target Recognition Method Based on a One-Dimensional Convolutional Neural Network" (application number: 201710838721.3) proposes a radar high-resolution range target recognition method based on a one-dimensional convolutional neural network. However, the recognition capability of this method decreases when the radar carrier frequency changes. For scenarios with changing carrier frequencies, this method requires establishing a database of all carrier frequencies and retraining the model; however, the cost of establishing such a large database and retraining the model is extremely high.

[0004] The patent document "Robust Target Recognition Method for High-Resolution Range Image under Radar Carrier Frequency Change" (application number: 202110490722.X) proposes a robust target recognition method based on transfer learning for radar carrier frequency changes. This method requires fine-tuning the original carrier frequency model using data from the new carrier frequency for each new carrier frequency recognition task. It cannot use the same model to recognize samples from different carrier frequencies, and its performance degrades when the sample size of the original carrier frequency is small. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a robust radar target recognition method based on a carrier-assisted attention network. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a robust radar target recognition method based on a carrier-assisted attention network, comprising:

[0007] Step 1: Generate a training dataset consisting of radar high-resolution range image data from multiple carrier frequencies and an auxiliary information dataset consisting of corresponding carrier frequencies;

[0008] Step 2: Construct a carrier frequency-assisted attention network, which includes an embedding layer generation subnetwork and a ViT subnetwork. The embedding layer generation subnetwork is used to encode the carrier frequency information of the input radar high-resolution range image data to obtain auxiliary weight information, and the ViT subnetwork is used to identify the input radar high-resolution range image data.

[0009] Step 3: Train the carrier-assisted attention network using the auxiliary information dataset and the training dataset;

[0010] Step 4: Use the pre-trained carrier-assisted attention network to perform target recognition on the original high-resolution range image data.

[0011] In one embodiment of the present invention, step 1 includes:

[0012] Step 1.1: Obtain radar high-resolution range image echo data containing M types of targets under K carrier frequencies to form a training echo dataset, where K≥6 and M≥5;

[0013] Step 1.2: Perform sliding window processing on the training echo dataset to obtain a training dataset with a sequence length of N, wherein the sequence length N of the training dataset is determined by the dimension of the high-resolution range image echo data;

[0014] Step 1.3: Set class labels for the training dataset of M classes of targets under all K carrier frequencies. The label values ​​are: 0, 1, ..., M-1;

[0015] Step 1.4: Record the carrier frequency of each data in the training dataset separately as auxiliary information for the current data, and the carrier frequency information of all samples constitutes the auxiliary information dataset.

[0016] In one embodiment of the present invention, the embedded layer generating subnetwork includes a first fully connected layer, a second fully connected layer and a third fully connected layer connected in sequence.

[0017] In one embodiment of the present invention, the weight dimensions of the first fully connected layer, the second fully connected layer and the third fully connected layer are set to 1×8, 8×32 and 32×2560 respectively, and the activation function of each fully connected layer is the ReLU activation function.

[0018] In one embodiment of the present invention, the ViT sub-network includes an embedding layer, a position encoding module, two Transformer encoding layers, and a classification head connected in sequence, wherein,

[0019] The embedded layer is used for processing data generated by the radar high-resolution range image sliding window. Encode to obtain embedded features Where E is the weight parameter of the embedding layer;

[0020] The location encoding module is used to associate the embedded features with the classification label x. class The features are concatenated to obtain the concatenated embedded features. The concatenated embedded features and position encoding E pos Add them together to obtain the position-encoded features.

[0021] The two Transformer encoding layers are used to extract features from the position-encoded features;

[0022] The classification head is used to classify targets based on the extracted features.

[0023] In one embodiment of the present invention, S3 includes:

[0024] Step 3.1: Initialize the weights of the carrier-assisted attention network using Gaussian methods;

[0025] Step 3.2: Input the auxiliary information dataset corresponding to the training dataset into the embedding layer generation sub-network to generate auxiliary information weights;

[0026] Step 3.3: Use the auxiliary information weights obtained in Step 3.2 as the weights of the embedding layer in the ViT sub-network, and input the training data in the training dataset into the ViT sub-network to obtain the predicted classification label.

[0027] In one embodiment of the present invention, S3 further includes:

[0028] Pack the same carrier frequency data in the auxiliary information dataset corresponding to the training dataset into a carrier frequency mini-batch, and then input the mini-batch training data corresponding to each carrier frequency into the carrier frequency auxiliary attention network in sequence.

[0029] After the training data of all carrier frequencies in mini-batch are input into the carrier frequency-assisted attention network, the backpropagation algorithm is used to optimize the total loss function of all carrier frequencies in one mixed training, and the parameters of each layer of the carrier frequency-assisted attention network are iteratively updated until the network converges or reaches the maximum set number of training rounds, and finally the trained carrier frequency-assisted attention network is obtained.

[0030] In one embodiment of the present invention, the total loss function for a single hybrid training of all carrier frequencies is:

[0031]

[0032] Where K represents the total number of carrier frequencies in the training dataset, B represents the number of samples in each mini-batch after splitting, H represents the cross-entropy loss function, and loss k This represents the loss function for the k-th carrier mini-batch. This represents the label value of the i-th training sample in the k-th carrier mini-batch. Let Σ represent the predicted classification label of the i-th training sample in the k-th carrier mini-batch, and let Σ· represent the summation operation.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. The robust radar target recognition method of this invention incorporates carrier frequency auxiliary information into the network training process, improving the network's ability to recognize data from different carrier frequencies and accelerating network convergence. Furthermore, the use of an attention mechanism ViT network for high-resolution range images further enhances the network's recognition performance.

[0035] 2. The radar target robustness identification method of the present invention does not require the collection of complete datasets under multiple carrier frequencies for the identification problem of different carrier frequency data, which greatly reduces the cost of database construction, and only requires one training of the network, which reduces the time cost required for training the network.

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0037] Figure 1 This is a flowchart of a robust radar target recognition method based on a carrier frequency-assisted attention network provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the structure of a carrier-assisted attention network provided in an embodiment of the present invention;

[0039] Figure 3 This is a comparison chart of the accuracy results of different carrier frequency data identification between the method of this invention and the traditional method;

[0040] Figure 4 These are the loss function learning curves of the method of this invention and the traditional method on the training dataset. Detailed Implementation

[0041] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a robust radar target identification method based on a carrier frequency-assisted attention network proposed according to the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0042] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element.

[0044] See Figure 1 , Figure 1 This is a flowchart of a robust radar target recognition method based on a carrier-assisted attention network provided in an embodiment of the present invention. The method includes:

[0045] Step 1: Generate a training dataset consisting of radar high-resolution range image data from multiple carrier frequencies and an auxiliary information dataset consisting of the corresponding carrier frequencies.

[0046] In this embodiment, step 1 specifically includes:

[0047] Step 1.1: Obtain radar high-resolution range image echo data containing M types of targets under K carrier frequencies to form a training echo dataset. The number of training samples for each type of target under each carrier frequency is at least 800, K≥6, M≥5.

[0048] Step 1.2: Perform sliding window processing on the training echo dataset, with a window length and step size of 20 and 5 respectively, to obtain a training dataset with a sequence length of N.

[0049] The sequence length N of the training dataset is determined by the dimension of the high-resolution range image echo data. Assuming the dimension of the echo data is L, then:

[0050]

[0051] in, This indicates rounding down to the nearest integer.

[0052] Step 1.3: Set class labels for the training dataset of M class targets under all K carrier frequencies. The label values ​​are: 0, 1, ..., M-1.

[0053] Step 1.4: Record the carrier frequency of each data in the training dataset separately as auxiliary information for the current data, and the carrier frequency information of all samples constitutes the auxiliary information dataset.

[0054] Step 2: Construct a carrier-assisted attention network, which includes an embedding layer generation subnetwork and a ViT subnetwork.

[0055] See Figure 2 , Figure 2 This is a schematic diagram of a carrier-assisted attention network provided in an embodiment of the present invention. The carrier-assisted attention network includes an embedding layer generation subnetwork and a ViT (Visual Transformer) subnetwork. The embedding layer generation subnetwork is used to encode the carrier information of the input data to obtain auxiliary weight information; the ViT subnetwork is used to recognize the input data.

[0056] In this embodiment, the embedded layer generating subnetwork includes a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence. The weight dimensions of the first fully connected layer, the second fully connected layer, and the third fully connected layer are set to 1×8, 8×32, and 32×2560, respectively, and the activation function of each fully connected layer is the ReLU activation function.

[0057] Furthermore, the ViT sub-network includes an embedding layer, a location encoding module, two Transformer encoding layers, and a classification head. The embedding layer is used to process the data generated by the sliding window of the radar high-resolution range image. Encode to obtain embedded features Where E is the weight parameter of the embedding layer; the position encoding module is used to combine the embedding feature with the classification label x. class The features are concatenated to obtain the concatenated embedded features. The concatenated embedded features and position encoding E pos Add them together to obtain the position-encoded features. The two Transformer encoding layers are used to extract features from the position-encoded features; the classification head is used to classify the target based on the extracted features.

[0058] Specifically, the embedding layer is a fully connected layer with an embedding dimension of 128 and a weight dimension of 128×20; the encoding dimension of the position encoding module is (N+1)×128, where N is the sequence length of the training dataset. The two Transformer encoding layers are sequentially connected and have the same structure. The Transformer encoding layer consists of a layer normalization layer, a multi-head self-attention mechanism, another layer normalization layer, and a multilayer perceptron. The multi-head self-attention mechanism has 8 heads, and the multilayer perceptron contains two fully connected layers with weight dimensions of 128×2048 and 2048×128, respectively.

[0059] Furthermore, the classification head includes a layer normalization layer, a fully connected layer, a ReLU activation function, a fully connected layer, and a softmax layer. The weight dimensions of the two fully connected layers are 128×512 and 512×10, respectively. The softmax layer uses the softmax function to calculate the probability that the input data is identified as each class.

[0060] Step 3: Train the carrier-assisted attention network using the auxiliary information dataset and the training dataset;

[0061] In this embodiment, step 3 includes:

[0062] Step 3.1 Perform Gaussian initialization on the weights of the carrier-assisted attention network.

[0063] Step 3.2: Input the auxiliary information dataset corresponding to the training dataset into the embedding layer generation sub-network to generate auxiliary information weights;

[0064] The auxiliary information dataset obtained in step 1 is input into the embedding layer generation sub-network. The auxiliary information dataset passes through the first fully connected layer in the embedding layer generation sub-network to obtain a carrier frequency information feature with a dimension of 1×8. Then it passes through the second fully connected layer to obtain a carrier frequency information feature with a dimension of 1×32. After passing through the third fully connected layer, a carrier frequency information feature with a dimension of 1×2560 is obtained. The dimension of the carrier frequency information feature is converted to 128×20 to obtain the auxiliary information weight.

[0065] Step 3.3: Use the auxiliary information weights obtained in Step 3.2 as the weights of the embedding layer in the ViT sub-network, and input the training data into the ViT sub-network to obtain the predicted classification label.

[0066] Specifically, the auxiliary information weights obtained in step 3.2 are used as the weights of the embedding layer in the ViT sub-network. The training data is input into the embedding layer of the ViT sub-network to obtain the embedding features of the input signal. The embedding features are concatenated with the classification label to obtain the concatenated embedding features. Then, the concatenated embedding features are added to the weights of the position encoding module to obtain the position encoded features. The position encoded features are sequentially input into the first Transformer encoding layer and the second Transformer encoding layer to obtain the Transformer output features. The features in the Transformer output features that correspond to the classification label are input into the classification head. Finally, the predicted classification label of the training data is output.

[0067] It should be noted that in order to train the carrier frequency-assisted attention network more efficiently, the data of different carrier frequencies needs to be segmented. The segmentation process is as follows: data of the same carrier frequency are packaged into a carrier frequency mini-batch, and then the mini-batch data of each carrier frequency is input into the carrier frequency-assisted attention network in sequence.

[0068] After the training data of all carrier frequencies in mini-batch are input into the carrier frequency-assisted attention network, the backpropagation algorithm is used to optimize the total loss function of all carrier frequencies in one mixed training. The parameters of each layer of the carrier frequency-assisted attention network are iteratively updated until the network converges or reaches the maximum set number of training rounds, and finally the trained carrier frequency-assisted attention network is obtained.

[0069] The total loss function for all carrier frequencies in a single mixed training is:

[0070]

[0071] Where K represents the total number of carrier frequencies in the training dataset, B represents the number of samples in each mini-batch after splitting, H represents the cross-entropy loss function, and loss k This represents the loss function for the k-th carrier mini-batch. This represents the label value of the i-th training sample in the k-th carrier mini-batch. Let Σ represent the predicted classification label of the i-th training sample in the k-th carrier mini-batch, and let Σ· represent the summation operation.

[0072] Furthermore, the cross-entropy loss function is as follows:

[0073]

[0074] Where H represents the cross-entropy loss function, M represents the number of target categories in the training dataset, and Y... label Y represents the label value of the data. pre Σ represents the network's predicted classification label for the data, and Σ· represents the summation operation.

[0075] Step 4: Use the pre-trained carrier-assisted attention network to identify high-resolution distance image data.

[0076] In this embodiment, radar high-resolution range image echo data of the target to be identified is first acquired. This echo data is then subjected to the same sliding window processing as the training echo data to generate test data. The carrier frequency information corresponding to the test data is input into the embedding layer generation sub-network. The output of the embedding layer generation sub-network is used as the weight value of the embedding layer of the ViT sub-network. The test data is then input into the ViT sub-network, passing through the embedding layer, the position encoding module, two Transformer encoding layers, and the classification head. Finally, the probability of the target to be identified being identified into each category is calculated through the softmax layer in the classification head, and the category corresponding to the highest probability is selected as the identification result.

[0077] The simulation experiment of this invention uses a hybrid training method combining the method of this invention and traditional methods to identify 10 types of high-resolution range image data from aircraft radar. Compared with the method of this invention, the traditional method's network does not have an embedded layer to generate a sub-network.

[0078] The simulation experiment hardware platform of this invention is: an Intel(R) Core(TM) i7-8700 CPU with a main frequency of 3.2GHz and 16GB of memory. The simulation experiment software platform of this invention is: Windows 10 operating system and Python 3.6. The data used is high-resolution range image data of 10 types of aircraft simulated by CST electromagnetic simulation software, with an attitude angle coverage range of 0°-180°. To verify the performance of the method, the pitch angles corresponding to the training dataset and test dataset for each type of aircraft are 3° and 5°, respectively. A total of 8 carrier frequencies are included, namely 8.8GHz, 8.9GHz, 9.0GHz, 9.1GHz, 9.2GHz, 9.3GHz, 9.4GHz, and 9.5GHz. At each carrier frequency, the training set and test set for each type of aircraft are 900 and 1000 records, respectively, and the total training set and test set for the 10 types of targets are 9000 and 10000 records, respectively. The total training and test sets for all carrier frequencies consist of 72,000 and 80,000 records, respectively.

[0079] The accuracy results of different carrier frequency data identification between the method of this invention and the traditional method are plotted as follows: Figure 3 As shown. Figure 3 The horizontal axis represents the carrier frequency, which is 8.8GHz, 8.9GHz, 9.0GHz, 9.1GHz, 9.2GHz, 9.3GHz, 9.4GHz, and 9.5GHz. These correspond to the eight sets of data with different carrier frequencies in the test dataset. The vertical axis represents the recognition accuracy of the test dataset. Figure 3The solid line represents the relationship curve between the recognition accuracy of the test dataset obtained by the method of the present invention and different carrier frequencies. Figure 3 The dashed line represents the relationship between the recognition accuracy of the test dataset obtained by the traditional method and different carrier frequencies.

[0080] Depend on Figure 3 It can be seen that the recognition accuracy of the method of the present invention is better than that of the traditional method under any test carrier frequency. The recognition accuracy of the traditional method is between 78% and 80% in all carrier frequency test datasets, while the recognition accuracy of the method of the present invention is between 83% and 85% in all carrier frequency test datasets. Compared with the traditional method, the recognition accuracy of the method of the present invention is improved by about 5 percentage points on average per carrier frequency.

[0081] The loss function learning curves of the method of this invention and the traditional method on the training dataset are plotted as follows: Figure 4 As shown. Figure 4 The horizontal axis represents the number of training epochs, which totaled 150 epochs, and the vertical axis represents the loss function. Figure 4 The solid line represents the relationship curve between the loss function of the training dataset obtained by the method of this invention and the number of training epochs. Figure 4 The dashed line represents the relationship between the loss function of the training dataset obtained by the traditional method and the number of training rounds.

[0082] Depend on Figure 4 It can be seen that the network converges faster using the method of this invention. The model of this invention is close to convergence when the training epochs reach 40, while the traditional method is close to convergence when the training epochs reach 140.

[0083] The radar target robustness recognition method based on carrier frequency-assisted attention network in this invention incorporates carrier frequency information into the network training process, which improves the network's ability to recognize data from different carrier frequencies and accelerates network convergence. It is applicable to target recognition tasks in scenarios where radar carrier frequencies change.

[0084] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A robust radar target recognition method based on a carrier-assisted attention network, characterized in that, include: Step 1: Generate a training dataset consisting of radar high-resolution range image data from multiple carrier frequencies and an auxiliary information dataset consisting of corresponding carrier frequencies; Step 2: Construct a carrier-assisted attention network, which includes an embedding layer generation subnetwork and a ViT subnetwork. The embedding layer generation subnetwork encodes the carrier frequency information of the input high-resolution radar range image data to obtain auxiliary weight information. The ViT subnetwork is used to identify the input high-resolution radar range image data. The embedding layer generation subnetwork includes a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence. The ViT subnetwork includes an embedding layer, a position encoding module, two Transformer encoding layers, and a classification head connected in sequence. The embedded layer is used for processing data generated by the radar high-resolution range image sliding window. Encode to obtain embedded features ,in, E These are the weight parameters for the embedding layer; The location encoding module is used to combine the embedded features with the classification label. The features are concatenated to obtain the concatenated embedded features. The concatenated embedded features are then encoded with their positions. Add them together to obtain the position-encoded features. ; The two Transformer encoding layers are used to extract features from the position-encoded features; The classification head is used to classify targets based on the extracted features. Step 3: Train the carrier-assisted attention network using the auxiliary information dataset and the training dataset; Step 4: Use a pre-trained carrier-assisted attention network to perform target recognition on the original high-resolution range image data; Step 3 includes: Step 3.1: Initialize the weights of the carrier-assisted attention network using Gaussian methods; Step 3.2: Input the auxiliary information dataset corresponding to the training dataset into the embedding layer generation sub-network to generate auxiliary information weights; Step 3.3: Use the auxiliary information weights obtained in Step 3.2 as the weights of the embedding layer in the ViT sub-network, and input the training data in the training dataset into the ViT sub-network to obtain the predicted classification label.

2. The robust radar target recognition method based on carrier-assisted attention network according to claim 1, characterized in that, Step 1 includes: Step 1.1: Obtain the contents K Under each carrier frequency M The high-resolution range image echo data of the target class are used to form the training echo dataset. K ≥6, M ≥5; Step 1.2: Perform sliding window processing on the training echo dataset to obtain a sequence with a length of... N The training dataset, wherein the sequence length of the training dataset is... N Determined by the dimension of the high-resolution range image echo data; Step 1.3: For all K Under each carrier frequency M The training dataset for the target class is labeled with class labels, and the label values ​​are: 0, 1, ... M -1; Step 1.4: Record the carrier frequency of each data in the training dataset separately as auxiliary information for the current data, and the carrier frequency information of all samples constitutes the auxiliary information dataset.

3. The robust radar target recognition method based on carrier-assisted attention network according to claim 1, characterized in that, The weight dimensions of the first fully connected layer, the second fully connected layer, and the third fully connected layer are set to 1×8, 8×32, and 32×2560, respectively, and the activation function of each fully connected layer is the ReLU activation function.

4. The robust radar target recognition method based on carrier-assisted attention network according to claim 3, characterized in that, Step 3 also includes: Pack the same carrier frequency data in the auxiliary information dataset corresponding to the training dataset into a carrier frequency mini-batch, and then input the mini-batch training data corresponding to each carrier frequency into the carrier frequency auxiliary attention network in sequence. After the training data of all carrier frequencies in mini-batch are input into the carrier frequency-assisted attention network, the backpropagation algorithm is used to optimize the total loss function of all carrier frequencies in one mixed training, and the parameters of each layer of the carrier frequency-assisted attention network are iteratively updated until the network converges or reaches the maximum set number of training rounds, and finally the trained carrier frequency-assisted attention network is obtained.

5. The robust radar target recognition method based on carrier-assisted attention network according to claim 4, characterized in that, The total loss function for all carrier frequencies in a single mixed training is: in, K This represents the total number of carrier frequencies contained in the training dataset. B This indicates the number of samples in each carrier frequency mini-batch after segmentation. H Represents the cross-entropy loss function. Indicates the first k Loss function for each carrier frequency mini-batch Indicates the first k The first of the carrier frequency small batches i The label values ​​of each training sample. Indicates the first k The first of the carrier frequency small batches i Predicted classification labels for each training sample. This indicates a summation operation.

Images