Radar target recognition method based on Transformer and temporal convolutional network

Abstract

The present invention discloses a radar target recognition method based on Transformer and temporal convolutional network. First, the original HRRP sample set is preprocessed; the sample is feature extracted and the data format is changed through a convolution module; then, a VIT-based MoCo contrastive learning module is used to learn to distinguish different samples in the feature space; then, the features obtained by the MoCo contrastive learning module are input into the temporal convolutional module to further enhance the representation ability of the model and obtain more separable features; finally, more effective features are retained through a fully connected layer, and finally, the network output is classified using softmax. The present invention applies the introduction of a temporal convolutional network, which enables the model to obtain more comprehensive information in the HRRP, further improves the feature extraction ability of the model, and obtains more separable features for classification.

Description

Technical Field

[0001] The present invention belongs to the field of radar target recognition, and specifically relates to a radar target recognition method based on Transformer and temporal convolutional network Background Art

[0002] Radar, originally meaning "radio detection and ranging," originated in the early 20th century. It is an electronic device that uses electromagnetic waves to detect targets. Radar transmits electromagnetic waves with specified parameters toward a target and receives the backscattered echo to obtain information such as the target's range, direction, and altitude. Radar enables all-day, all-weather, and long-range target detection. It is widely used in target reconnaissance, imaging, and missile guidance. It holds a high position in the military and plays a vital role in national defense.

[0003] Modern warfare is centered around high-tech information warfare and electronic warfare. In military scenarios, the ability to detect incoming targets in advance, effectively acquire and monitor real-time information, and efficiently monitor the battlefield situation in real time has become a crucial factor in victory or defeat. For this reason, radar, as a crucial device capable of reliably acquiring information, holds a crucial position in battlefield intelligence acquisition.

[0004] RATA can be categorized into two different radar target recognition tasks, low-resolution narrowband and high-resolution wideband, depending on the radar's range resolution. Low-resolution narrowband signals have poor range resolution, meaning the radar target is perceived as a "point" in the echo. Consequently, radar echoes contain less target information, typically enabling only simple recognition or coarse classification based on echo sequence fluctuations, micro-Doppler, and modulation characteristics. This clearly fails to meet the demands of RATA applications in the modern information society.

[0005] High-resolution wideband radars have excellent range resolution and are much smaller than the target size. Their echoes are also known as the target's one-dimensional high-resolution range profile (HRRP). The HRRP signal provides the distribution of the target's scattering centers along the radar's line of sight, reflecting important information such as the target's size and structure. Due to its small data size and ease of acquisition and processing, it is widely used in radar automatic target recognition (RATR). Traditional radar target recognition methods based on HRRP are mostly unsupervised and lossy. Separable features are partially lost during the modeling process, affecting the model's recognition accuracy and generalization. Some deep learning-based HRRP recognition methods, however, experience a sharp drop in recognition performance when trained on a small amount of training data. Summary of the Invention

[0006] To solve the above problems, we propose a new deep learning radar target recognition model.

[0007] The model consists of a data preprocessing module, a convolution module, a Transformer-based MoCo contrastive learning module, a convolutional temporal network, and a classification module. The MoCo contrastive learning module consists of data augmentation, a Transformer encoder module, and a contrastive loss module. The preprocessed HRRP data is first passed through the convolution module and transformed into a three-channel image format. The MoCo contrastive learning module then learns to distinguish different samples in the feature space, narrowing the distance between samples of the same category and increasing the distance between samples of different categories. The representation obtained by the contrastive learning module is then input into the temporal convolutional network to obtain more separable features for classification, completing HRRP recognition.

[0008] The radar target recognition method based on Transformer and temporal convolutional network includes the following steps:

[0009] S1: Preprocess the original HRRP sample set.

[0010] L2 norm normalization is used to process the raw HRRP echo data, dividing the amplitude into a unified scale to eliminate the amplitude sensitivity of HRRP. The center of gravity alignment method is used to improve the translation sensitivity of HRRP.

[0011] S2: Extract features from samples and change the data format through the convolution module;

[0012] The preprocessed HRRP data is processed by the convolution module CNN. By controlling and adjusting the size and step size of the convolution kernel of the convolution module, the one-dimensional distance image is convolved into the data format required by the subsequent network.

[0013] S3: Through the VIT-based MoCo contrastive learning module, we learn to distinguish different samples in the feature space, shorten the distance between samples of the same category in the feature space, and increase the distance between samples of different categories.

[0014] S4: The features obtained by the MoCo contrastive learning module are input into the temporal convolution module to further enhance the representation ability of the model and obtain more separable features.

[0015] S5: The more effective features are retained through the fully connected layer, and finally the softmax is used to classify the network output;

[0016] Preferably, the specific steps of step S1 are as follows:

[0017] S1.1: L2 norm normalization. Divide the amplitude into a unified scale. The original radar HRRP data is represented as X = [x1, x2, ..., x M ], then the X obtained after L2 norm normalization is norm as follows:

[0018]

[0019] Among them, X represents the original HRRP data, M represents the number of distance units contained in HRRP, and x m Represents the amplitude within the mth distance unit. After normalization by norm, we get

[0020] S1.2: Center of gravity alignment method. The center of gravity alignment method consists of two steps: first, the center of gravity of the HRRP needs to be calculated, and then the center of gravity is translated to the center of the HRRP range unit. The radar HRRP data is normalized by the norm to obtain The calculation of the center of gravity G is as follows:

[0021]

[0022] Get data by translation

[0023] Preferably, the detailed steps of S2 are:

[0024] The HRRP sample X after data preprocessing G , after CH convolution kernels of size 1×5 are used for convolution operation, the convolution output is obtained Where X contains L distance units, CH represents the number of channels, and F cnn (ch) represents the convolution output of the chth channel. F cnn The calculation is as follows:

[0025]

[0026] in, is the convolution operation, W ch Represents the weight matrix parameters of the chth convolution kernel, b ch is the corresponding bias, and L represents the sequence length.

[0027] Then, the features of the convolution output are batch normalized and activated by ReLU to obtain the output F br , to avoid the impact of data distribution differences on recognition performance. Finally, F br Transform (Reshape) the data format required for contrast learning to obtain feature F in ,in Among them, CH×1×L=C×H×W, C, H, W are the channel, height and width of the image respectively. The features are obtained through the convolution module. Feature F in The data format is consistent with the picture, so F in This is called HRRP image format data after convolution processing.

[0028] Preferably, the detailed steps of S3 are:

[0029] The MoCo contrastive learning module consists of data augmentation, encoder module and contrastive loss module.

[0030] S3.1: Data enhancement module.

[0031] A data augmentation method combining random cropping, horizontal flipping with probability p, and vertical flipping with probability p is designed. The HRRP image format data after convolution processing in a small batch is Enhance this batch of data:

[0032]

[0033] For one of the picture format data Its enhanced The same image format data is obtained after two data enhancements. and Enter the gradient encoder f respectively q and momentum encoder f k

[0034] S3.2: Encoder module.

[0035] The encoder module includes a gradient encoder f q and momentum encoder f k , constructed as a gradient encoder f through the VIT module q and momentum encoder f k The difference between the two is the gradient encoder f q The gradient method updates the parameters, while the momentum encoder f k The momentum method is used to update parameters, and the module structures of the two are the same.

[0036] The VIT module obtains a deep bidirectional representation of image-format data. It needs to segment and stretch the data into vector form, that is, embed the blocks and add position encoding. It then calculates the attention distribution through the multi-head attention module and finally obtains the output features through the MLP layer.

[0037] The VIT modules are as follows:

[0038] S3.2.1: Block embedding and position encoding. First, the image format data is divided into several blocks (patches), and then linearly compressed into a feature vector structure. For the input feature Given a block size P, let F″′ in Divide into N P×P×C blocks, as follows:

[0039] C×H×W→N×(P 2 C)

[0040] After obtaining N blocks, they are linearly transformed into N feature vectors of length E through a fully connected network, and a flag vector x is added at the starting position of the vector class The calculation process is as follows:

[0041]

[0042] in, represents the i-th block, W i is the weight parameter, and Concat[·] is a concatenation operation. The final F0 is the output result of the block embedding.

[0043] Similarly, the features obtained by block embedding lack the relative position information between blocks, so it is necessary to add position encoding information to F0. Using a trainable one-dimensional position encoding, initialize a trainable random variable and add it to the embedded feature F0:

[0044] F1=F0+E pos

[0045] Among them, F1 is the block embedding feature after adding position information,

[0046] S3.2.2: Layer normalization. For the features obtained by block embedding Use layer normalization to process it, the calculation process is:

[0047] F LN =LayerNorm(F1)

[0048]

[0049] Where μ is the mean, is the variance, ε is a small value to prevent the denominator from being 0, γ and β represent the parameter vectors of scaling and translation, which are consistent with the dimension of the input features.

[0050] S3.2.3: Multi-head attention module. First, we introduce the self-attention mechanism, the layer-normalized embedding representation F LN Multiplying by three different parameter matrices, mapping to query matrix Q, queried key-value matrix K and value matrix V, the formula can be expressed as:

[0051]

[0052] in,

[0053] Next, use the Q matrix to query the K matrix, establish the connection between each dimension feature and the global feature, and obtain the dot product matrix QK T The matrix obtained by clicking usually has a large variance. The result after normalization is not very stable, which is not conducive to model convergence. Therefore, the "scaling factor" is introduced. ". QK T Divide by Then normalize it through the Softmax function to obtain the attention weight matrix of V.

[0054] Finally, the output F of the self-attention mechanism self-att as follows:

[0055]

[0056] The principle of the multi-head attention mechanism is similar to the principle of the self-attention mechanism mentioned above. The difference is that this process is repeated h times, which is called multi-head. Since the parameter matrix of each projection and click operation is different, the h self-attention outputs obtained are also different. These h self-attention results are concatenated and then subjected to a linear transformation to finally obtain the result of the multi-head attention mechanism F MH The calculation process of the multi-head attention mechanism is as follows:

[0057]

[0058] head i =Attention(F LN W q i ,F LN W k i ,F LN W v i )

[0059] FMH =Concat[head1;head2;...;head h ]W MH

[0060] in, head i Represents the output of the i-th attention head.

[0061] S3.2.4: MLP layer.

[0062] MLP is actually two fully connected layers. The first fully connected layer normalizes the residual multi-head attention output feature F lr Mapped to a high-dimensional hidden space, the second fully connected layer maps it to a low-dimensional space to further retain useful feature information. The process is as follows:

[0063] F fc1 =F lr W mlp_dim +b mlp_dim

[0064] F fc2 =ReLU(F fc1 W embed_dim +b embed_dim )(

[0065] Among them, F fc1 Represents the output of the first fully connected layer, F fc2 Represents the output of the second fully connected layer. W mlp_dim and W embed_dim is the parameter matrix of the two linear projections, b mlp_dim and b embed_dim is the corresponding bias.

[0066] Finally, take the flag vector of the MLP layer As the output vector of the encoder, this vector integrates the information of all patches.

[0067] S3.3: Contrast loss module. The loss function used in the contrast module is as follows:

[0068]

[0069] Where q is the result q obtained by the HRRP sample x through the above modules. + is the result obtained by the above modules for positive examples similar to x, k - is the result of the negative sample that is not similar to x after the above module. This loss function makes q and the positive sample k + The similarity with other negative examples is as large as possible, and the similarity with other negative examples is as small as possible.

[0070] Since the overall architecture input of the MoCo contrastive learning module is different from that of the general contrastive learning module, the encoder module of the MoCo contrastive learning module has two inputs, so the loss function is also different. For the same HRRP data, two enhanced versions are obtained due to the randomness of the data enhancement module after two data enhancement modules. and Through the gradient encoder f q and momentum encoder f k Get feature q1, and q2, Let q1, The loss ctr1 is obtained by the above formula, and q2, The loss ctr2 is obtained by the above formula, and then the contrast loss function of the MoCo contrast learning module is obtained as:

[0071] L moco =ctr1+ctr2

[0072] Encoder f q The parameters are optimized by gradient descent of the loss function, and the encoder f k Instead of performing gradient updates, momentum updates are performed using the following formula:

[0073] θ k ←mθ k +(1-m)θ q (5.16)

[0074] Among them, θ q is the encoder f q The parameter θ k is the encoder f k Parameters. m is the momentum parameter, m = 0.999, so the encoder f k Although the update is extremely slow, it can still update the parameters and solve the consistency problem.

[0075] Preferably, the detailed steps of S4 are:

[0076] The temporal convolution module consists of causal dilated convolution, batch normalization, ReLU function, Dropout and residual connection.

[0077] When the convolution kernel f: The HRRP features output by the MoCo contrastive learning module are: Where t is the characteristic length, Then, the output of the HRRP feature after causal dilation convolution is defined as:

[0078]

[0079] Where d represents the expansion coefficient, which increases with the exponential power of 2 as the convolution layer deepens, and f(j) represents the value of the convolution kernel at position j, 0≤j≤ker-1, is the feature output after causal dilation convolution. The following formula represents the size of the receptive field:

[0080] field=(k-1)·d

[0081] Then, batch normalization is performed on the features obtained by causal dilation convolution. Each batch (mini-batch) contains num data: Define the output of the batch data after the causal expansion convolution operation as F i represents the causal expansion convolution output corresponding to the i-th data in the mini-batch. Then, Batch normalization can be defined as:

[0082]

[0083] Among them, F o (k,l), are the lth value of the kth channel of the oth data before and after the batch normalization operation, γ k , β k It is a trainable parameter, and ε is defined as a very small value to prevent the denominator from being zero. BN 、 are the mean and variance respectively.

[0084] Next, Input activation function to get

[0085]

[0086] Repeat the above calculation process for each data in the mini-batch to obtain the mapping output feature F(q1), and then obtain shallow information through residual connection to prevent overfitting:

[0087] F res =ReLU(q1+F(q1))

[0088] Preferably, the detailed steps of S5 are:

[0089] Finally, a fully connected layer and Softmax function are used to predict the features. The process is to first pass the output feature F of the temporal convolutional network through a fully connected layer. res Perform dimensionality reduction to obtain the prediction vector:

[0090] Ff =F res W f +b f

[0091] in, C is the total number of target categories of HRRP samples. The prediction vector is then mapped to a probability distribution P using the Softmax function:

[0092]

[0093] Among them, F f (c) represents the cth element, P(c|F f ) represents feature F f The probability of belonging to the c-th category target. Finally, the classification result is obtained by the following formula:

[0094]

[0095] Where argmax(·) indicates the value of the variable c when the function in the brackets reaches its maximum value.

[0096] The beneficial effects of the present invention are:

[0097] 1. This paper introduces a Transformer-based MoCo contrastive learning module. Contrastive learning uses the Transformer encoder to obtain a deep bidirectional representation of HRRP and uses contrastive loss to narrow the distance between samples of the same category in the feature space.

[0098] 2. The present invention introduces a temporal convolutional network to enable the model to obtain more comprehensive information from HRRP, further improve the model's feature extraction capability, and obtain more separable features for classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 :Flowchart of the steps of radar target recognition method based on improved temporal convolutional network. DETAILED DESCRIPTION

[0100] Reference Figure 1 , is a flow chart of a radar target recognition method based on an improved temporal convolutional network of the present invention, and the specific implementation steps are as follows:

[0101] Training phase:

[0102] S1: Collect data sets. The HRRP data sets collected by the radar are merged according to the types of targets. For each type of sample, training samples and test samples are selected from different data segments. In the process of selecting the training set and test set, it is ensured that the posture formed by the selected training set samples and the radar covers the posture formed by the test set samples and the radar. The ratio of the number of training set and test set samples for each type of target is 8:2. The selected data set is denoted as T = {(x n ,y c )} n∈[1,N],c∈[1,C] , where x n represents the nth sample, y c Indicates that the sample belongs to the cth category, a total of C category targets were collected, and N represents the total number of samples.

[0103] S2: Preprocess the samples in the dataset extracted by S1. The specific steps are as follows:

[0104] L2 norm normalization is used to process the raw HRRP echo data, dividing the amplitude into a unified scale to eliminate the amplitude sensitivity of HRRP. The center of gravity alignment method is used to improve the translation sensitivity of HRRP.

[0105] S2.1: L2 norm normalization. Divide the amplitude into a unified scale. The original radar HRRP data is represented as X = [x1, x2, ..., x M ], then the X obtained after L2 norm normalization is norm as follows:

[0106]

[0107] Among them, X represents the original HRRP data, M represents the number of distance units contained in HRRP, and x m Represents the amplitude within the mth distance unit. After normalization by norm, we get

[0108] S2.2: Center of gravity alignment method. The center of gravity alignment method consists of two steps: first, the center of gravity of the HRRP needs to be calculated, and then the center of gravity is translated to the center of the HRRP range unit. The radar HRRP data is normalized by the norm to obtain The calculation of the center of gravity G is as follows:

[0109]

[0110] Get data by translation

[0111] S3: Use the convolution module to extract features from the sample and change the data format. The specific steps are as follows:

[0112] The preprocessed HRRP data is processed by the convolution module CNN. By controlling and adjusting the size and step size of the convolution kernel of the convolution module, the one-dimensional distance image is convolved into the data format required by the subsequent network.

[0113] S3: HRRP sample X after data preprocessing G , after CH convolution kernels of size 1×5 are used for convolution operation, the convolution output is obtained Where X contains L distance units, CH represents the number of channels, and F cnn (ch) represents the convolution output of the chth channel. F cnn The calculation is as follows:

[0114]

[0115] in, is the convolution operation, W ch Represents the weight matrix parameters of the chth convolution kernel, b ch is the corresponding bias, and L represents the sequence length.

[0116] Then, the features of the convolution output are batch normalized and activated by ReLU to obtain the output F br , to avoid the impact of data distribution differences on recognition performance. Finally, F br Transform (Reshape) the data format required for contrast learning to obtain feature F in ,in Among them, CH×1×L=C×H×W, C, H, W are the channel, height and width of the image respectively. The features are obtained through the convolution module. Feature F in The data format is consistent with the picture, so F in This is called HRRP image format data after convolution processing.

[0117] S4: Through the VIT-based MoCo contrastive learning module, we learn to distinguish different samples in the feature space, narrowing the distance between samples of the same category in the feature space while widening the distance between samples of different categories. The specific steps are as follows:

[0118] S4.1: Data enhancement module.

[0119] A data augmentation method combining random cropping, horizontal flipping with probability p, and vertical flipping with probability p is designed. The HRRP image format data after convolution processing in a small batch is Enhance this batch of data:

[0120]

[0121] For one of the picture format data Its enhanced The same image format data is obtained after two data enhancements. and Enter the gradient encoder f respectively q and momentum encoder f k

[0122] S4.2: Encoder module.

[0123] The encoder module includes a gradient encoder f q and momentum encoder f k , constructed as a gradient encoder f through the VIT module q and momentum encoder f k The difference between the two is the gradient encoder f q The gradient method updates the parameters, while the momentum encoder f k The momentum method is used to update parameters, and the module structures of the two are the same.

[0124] The VIT module obtains a deep bidirectional representation of image-format data. It needs to segment and stretch the data into vector form, that is, embed the blocks and add position encoding. It then calculates the attention distribution through the multi-head attention module and finally obtains the output features through the MLP layer.

[0125] The detailed steps of VIT are:

[0126] S4.2.1: Block embedding and position encoding. First, the image format data is divided into several blocks (patches), and then linearly compressed into a feature vector structure. For the input feature Given a block size P, let F″′ in Divide into N P×P×C blocks, as follows:

[0127] C×H×W→N×(P 2 C)

[0128] After obtaining N blocks, they are linearly transformed into N feature vectors of length E through a fully connected network, and a flag vector x is added at the starting position of the vector class , this vector integrates global semantic information, thus better representing the information of the entire feature sequence, which is beneficial to the subsequent network recognition work. The calculation process is as follows:

[0129]

[0130] in, represents the i-th block, W i is the weight parameter, and Concat[·] is a concatenation operation. The final F0 is the output result of the block embedding.

[0131] Similarly, the features obtained by block embedding lack the relative position information between blocks, so it is necessary to add position encoding information to F0. Using a trainable one-dimensional position encoding, initialize a trainable random variable and add it to the embedded feature F0:

[0132] F1=F0+E pos

[0133] Among them, F1 is the block embedding feature after adding position information,

[0134] S4.2.2: Layer normalization. For the features obtained by block embedding Use layer normalization to process it, the calculation process is:

[0135] F LN =LayerNorm(F1)

[0136]

[0137] Where μ is the mean, is the variance, ε is a small value to prevent the denominator from being 0, γ and β represent the parameter vectors of scaling and translation, which are consistent with the dimension of the input features.

[0138] S4.2.3: Multi-head attention module. First, we introduce the self-attention mechanism, the layer-normalized embedding representation F LN Multiplying by three different parameter matrices, mapping to query matrix Q, queried key-value matrix K and value matrix V, the formula can be expressed as:

[0139]

[0140] in,

[0141] Next, use the Q matrix to query the K matrix, establish the connection between each dimension feature and the global feature, and obtain the dot product matrix QK T The matrix obtained by clicking usually has a large variance. The result after normalization is not very stable, which is not conducive to model convergence. Therefore, the "scaling factor" is introduced. ". QK T Divide by Then normalize it through the Softmax function to obtain the attention weight matrix of V.

[0142] Finally, the output F of the self-attention mechanism self-att as follows:

[0143]

[0144] The principle of the multi-head attention mechanism is similar to the principle of the self-attention mechanism mentioned above. The difference is that this process is repeated h times, which is called multi-head. Since the parameter matrix of each projection and click operation is different, the h self-attention outputs obtained are also different. These h self-attention results are concatenated and then subjected to a linear transformation to finally obtain the result of the multi-head attention mechanism F MH The calculation process of the multi-head attention mechanism is as follows:

[0145]

[0146] head i =Attention(F LN W q i ,F LN W k i ,F LN W v i )

[0147] F MH =Concat[head1;head2;...;head h ]W MH

[0148] in, head i Represents the output of the i-th attention head.

[0149] S4.2.4: MLP layer.

[0150] MLP is actually two fully connected layers. The first fully connected layer normalizes the residual multi-head attention output feature F lr Mapped to a high-dimensional hidden space, the second fully connected layer maps it to a low-dimensional space to further retain useful feature information. The process is as follows:

[0151] F fc1 =F lr W mlp_dim +b mlp_dim

[0152] F fc2 =ReLU(Ffc1 W embed_dim +b embed_dim )(

[0153] Among them, F fc1 Represents the output of the first fully connected layer, F fc2 Represents the output of the second fully connected layer. W mlp_dim and W embed_dim is the parameter matrix of the two linear projections, b mlp_dim and b embed_dim is the corresponding bias.

[0154] Finally, take the flag vector of the MLP layer As the output vector of the encoder, this vector integrates the information of all patches.

[0155] S4.3: Contrast loss module. The loss function used in the contrast module is as follows:

[0156]

[0157] Where q is the result q obtained by the HRRP sample x through the above modules. + is the result obtained by the above modules for positive examples similar to x, k - is the result of the negative sample that is not similar to x after the above module. This loss function makes q and the positive sample k + The similarity with other negative examples is as large as possible, and the similarity with other negative examples is as small as possible.

[0158] Since the overall architecture input of the MoCo contrastive learning module is different from that of the general contrastive learning module, the encoder module of the MoCo contrastive learning module has two inputs, so the loss function is also different. For the same HRRP data, two enhanced versions are obtained due to the randomness of the data enhancement module after two data enhancement modules. and Through the gradient encoder f q and momentum encoder f k Get feature q1, and q2, Let q1, The loss ctr1 is obtained by the above formula, and q2, The loss ctr2 is obtained by the above formula, and then the contrast loss function of the MoCo contrast learning module is obtained as:

[0159] L moco =ctr1+ctr2

[0160] Encoder f q The parameters are optimized by gradient descent of the loss function, and the encoder fk Instead of performing gradient updates, momentum updates are performed using the following formula:

[0161] θ k ←mθ k +(1-m)θ q (5.16)

[0162] Among them, θ q is the encoder f q The parameter θ k is the encoder f k Parameters. m is the momentum parameter, m = 0.999, so the encoder f k Although the update is extremely slow, it can still update the parameters and solve the consistency problem.

[0163] S5: The features obtained by the MoCo contrastive learning module are input into the temporal convolution module to further enhance the representation ability of the model and obtain more separable features. The specific steps are as follows:

[0164] The temporal convolution module consists of causal dilated convolution, batch normalization, ReLU function, Dropout and residual connection.

[0165] When the convolution kernel f: The HRRP features output by the MoCo contrastive learning module are: Where t is the characteristic length, Then, the output of the HRRP feature after causal dilation convolution is defined as:

[0166]

[0167] Where d represents the expansion coefficient, which increases with the exponential power of 2 as the convolution layer deepens, and f(j) represents the value of the convolution kernel at position j, 0≤j≤ker-1, is the feature output after causal dilation convolution. The following formula represents the size of the receptive field:

[0168] field=(k-1)·d

[0169] Then, batch normalization is performed on the features obtained by causal dilation convolution. Each batch (mini-batch) contains num data: Define the output of the batch data after the causal expansion convolution operation as F i represents the causal expansion convolution output corresponding to the i-th data in the mini-batch. Then, Batch normalization can be defined as:

[0170]

[0171] Among them, F o (k,l), are the lth value of the kth channel of the oth data before and after the batch normalization operation, γ k , β k It is a trainable parameter, and ε is defined as a very small value to prevent the denominator from being zero. BN 、 are the mean and variance respectively.

[0172] Next, Input activation function to get

[0173]

[0174] Repeat the above calculation process for each data in the mini-batch to obtain the mapping output feature F(q1), and then obtain shallow information through residual connection to prevent overfitting:

[0175] F res =ReLU(q1+F(q1))

[0176] S6: The more effective features are retained through the fully connected layer, and finally the softmax is used to classify the network output:

[0177] Finally, a fully connected layer and Softmax function are used to predict the features. The process is to first pass the output feature F of the temporal convolutional network through a fully connected layer. res Perform dimensionality reduction to obtain the prediction vector:

[0178] F f =F res W f +b f

[0179] in, C is the total number of target categories of HRRP samples. The prediction vector is then mapped to a probability distribution P using the Softmax function:

[0180]

[0181] Among them, F f (c) represents the cth element, P(c|F f ) represents feature F f The probability of belonging to the c-th category target. Finally, the classification result is obtained by the following formula:

[0182]

[0183] Where argmax(·) indicates the value of the variable c when the function in the brackets reaches its maximum value.

[0184] S7: Send the HRRP sample test set processed by S2 to the radar recognition model based on transformer and temporal convolutional network that has been trained in steps S3, S4, S5, and S6 for training and testing.

[0185] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the present invention have been described in conjunction with the accompanying drawings, it should be understood by those skilled in the art that various changes in form and details may be made without departing from the spirit and scope of the present invention as defined by the appended claims.

Claims

1. Radar target recognition method based on Transformer and temporal convolutional network, characterized by: The following steps are involved: S1: Preprocess the original HRRP sample set; L2 norm normalization is used to process the HRRP raw echo data, dividing the amplitude into a unified scale to eliminate the amplitude sensitivity of HRRP; the center of gravity alignment method is used to improve the translation sensitivity of HRRP; S2: Extract features from samples and change the data format through the convolution module; The pre-processed HRRP data is processed through the convolution module CNN. By controlling and adjusting the size and step size of the convolution kernel of the convolution module, the one-dimensional distance image is convolved into the data format required by the subsequent network; S3: Through the VIT-based MoCo contrastive learning module, we learn to distinguish different samples in the feature space, narrowing the distance between samples of the same category in the feature space, while widening the distance between samples of different categories; S4: The features obtained by the MoCo contrastive learning module are input into the temporal convolution module to further enhance the representation ability of the model and obtain more separable features; S5: The more effective features are retained through the fully connected layer, and finally the softmax is used to classify the network output.

2. The radar target recognition method based on Transformer and temporal convolutional network according to claim 1 is characterized in that The specific steps of step S1 are as follows: S1.1: L2 norm normalization; divide the amplitude into a unified scale; express the original radar HRRP data as X = [x1, x2, ..., x M ], then the X obtained after L2 norm normalization is norm as follows: Among them, X represents the original HRRP data, M represents the number of distance units contained in HRRP, and x m Represents the amplitude within the mth distance unit; after norm normalization, it is obtained S1.2: Center of gravity alignment method; The center of gravity alignment method consists of two steps: first, the center of gravity position of the HRRP needs to be calculated, and then the center of gravity is located at the center of the HRRP range unit by translation; the radar HRRP data is obtained by norm normalization The calculation of the center of gravity G is as follows: Get data by translation 3. The radar target recognition method based on Transformer and temporal convolutional network according to claim 2 is characterized in that The detailed steps of S2 are: The HRRP sample X after data preprocessing G , after CH convolution kernels of size 1×5 are used for convolution operation, the convolution output is obtained Where X contains L distance units, CH represents the number of channels, and F cnn (ch) represents the convolution output of the chth channel; F cnn The calculation is as follows: in, is the convolution operation, W ch Represents the weight matrix parameters of the chth convolution kernel, b ch is the corresponding bias, L represents the sequence length; Then, the features of the convolution output are batch normalized and activated by ReLU to obtain the output F br , to avoid the impact of data distribution differences on recognition performance; finally, F br Transform into the data format required for Reshape contrast learning to obtain feature F in ,in Among them, CH×1×L=C×H×W, C, H, W are the channel, height and width of the image respectively; the features are obtained through the convolution module Feature F in The data format is consistent with the picture, so F in This is called HRRP image format data after convolution processing.

4. The radar target recognition method based on Transformer and temporal convolutional network according to claim 3 is characterized in that The detailed steps of S3 are: The MoCo contrastive learning module consists of a data enhancement module, an encoder module, and a contrastive loss module; S3.1: data enhancement module; A data enhancement method combining random cropping, horizontal flipping with probability p, and vertical flipping with probability p is designed; the HRRP image format data after convolution processing of a small batch is Enhance this batch of data: For one of the picture format data Its enhanced is its positive sample, and the rest are negative samples; the same image format data is obtained after two data enhancements and Enter the gradient encoder f respectively q and momentum encoder f k S3.2: encoder module; The encoder module includes a gradient encoder f q and momentum encoder f k , constructed as a gradient encoder f through the VIT module q and momentum encoder f k The difference between the two is the gradient encoder f q The gradient method updates the parameters, while the momentum encoder f k The momentum method is used to update parameters, and the module structure of the two is the same; The VIT module obtains a deep bidirectional representation of image-formatted data. It needs to segment and stretch the data into vector form, i.e., embed the blocks and add position encoding. Then, it calculates the attention distribution through the multi-head attention module, and finally obtains the output features through the MLP layer. The VIT modules are as follows: S3.2.1: Block embedding and position encoding; first divide the image format data into several patches, and then linearly compress them into a feature vector structure; for the input feature Given a block size P, Divide into N P×P×C blocks, as follows: C×H×W→N×(P 2 C) After obtaining N blocks, they are linearly transformed into N feature vectors of length E through a fully connected network, and a flag vector x is added at the starting position of the vector class ; The calculation process is as follows: in, represents the i-th block, W i is the weight parameter, and Concat[·] is a concatenation operation; the final F0 is the output result of the block embedding. Similarly, the features obtained by block embedding lack the relative position information between blocks, so it is necessary to add position encoding information to F0; use a trainable one-dimensional position encoding to initialize a trainable random variable and add it to the embedded feature F0: F1=F0+E pos Among them, F1 is the block embedding feature after adding position information, S3.2.2: Layer normalization; for features obtained by block embedding Use layer normalization to process it, the calculation process is: F LN =LayerNorm(F1) Among them, μ is the mean, σ 2 is the variance, ε is a very small value to prevent the denominator from being 0, γ and β represent the parameter vectors of scaling and translation, which are consistent with the dimensions of the input features; S3.2.3: Multi-head attention module; first introduce the self-attention mechanism, the layer-normalized embedding representation F LN Multiplying by three different parameter matrices, mapping to query matrix Q, queried key value matrix K and value matrix V, the formula is expressed as: in, d q =d k ; Next, use the Q matrix to query the K matrix, establish the connection between each dimension feature and the global feature, and obtain the dot product matrix QK T ; QK T Divide by the scaling factor Then normalize it through the Softmax function to obtain the attention weight matrix of V; Finally, the output F of the self-attention mechanism self-att as follows: The principle of the multi-head attention mechanism is similar to the principle of the self-attention mechanism mentioned above. The difference is that this process is repeated h times, which is called multi-head. Since the parameter matrix of each projection and click operation is different, the h self-attention outputs obtained are also different. These h self-attention results are concatenated and then subjected to a linear transformation to finally obtain the result F of the multi-head attention mechanism. MH ; The calculation process of the multi-head attention mechanism is as follows: head i =Attention(F LN W q i ,F LN W k i ,F LN W v i ) F MH =Concat[head1;head2;...;head h ]W MH in, head i Represents the output result of the i-th attention head; S3.2.4: MLP layer; MLP is actually two fully connected layers. The first fully connected layer normalizes the residual multi-head attention output feature F lr Mapped to a high-dimensional hidden space, the second fully connected layer maps it to a low-dimensional space to further retain useful feature information. The process is as follows: F fc1 =F lr W mlp_dim +b mlp_dim F fc2 =ReLU(F fc1 W embed_dim +b embed_dim ) Among them, F fc1 Represents the output of the first fully connected layer, F fc2 Represents the output of the second fully connected layer; W mlp_dim and W embed_dim is the parameter matrix of the two linear projections, b mlp_dim and b embed_dim is the corresponding bias; Finally, take the flag vector of the MLP layer As the output vector of the encoder, this vector integrates the information of all block patches; S3.3: Contrastive loss module, the loss function used is as follows: Where q is the result obtained by the encoder module of HRRP sample x; k + is the result obtained by the encoder module for positive samples similar to x, k - is the result of the encoder module for negative samples that are dissimilar to x; the loss function makes q and positive samples k + The similarity with is as large as possible, and the similarity with other negative examples is as small as possible; Since the overall architecture input of the MoCo contrastive learning module is different from that of the general contrastive learning module, the encoder module of the MoCo contrastive learning module has two inputs, so the loss function is also different; for the same HRRP data, two enhanced versions are obtained due to the randomness of the data enhancement module after two data enhancement modules. and Through the gradient encoder f q and momentum encoder f k Get feature q1, and q2, Let q1, The loss ctr1 is obtained by the above formula, and q2, The loss ctr2 is obtained by the above formula, and then the contrast loss function of the MoCo contrast learning module is obtained as: L moco =ctr1+ctr2 Encoder f q The parameters are optimized by gradient descent of the loss function, and the encoder f k Instead of performing gradient updates, momentum updates are performed using the following formula: i k ←mθ k +(1-m)θ q Among them, θ q is the encoder f q The parameter θ k is the encoder f k Parameters; m is the momentum parameter, m = 0.999, so the encoder f k Although the update is extremely slow, it can still update the parameters and solve the consistency problem.

5. The radar target recognition method based on Transformer and temporal convolutional network according to claim 4 is characterized in that: The detailed steps of S4 are: The temporal convolution module consists of causal dilation convolution, batch normalization, ReLU function, Dropout and residual connection; When the convolution kernel The HRRP features output by the MoCo contrastive learning module are: Where t is the characteristic length, Then, the output of the HRRP feature after causal dilation convolution is defined as: Where d represents the expansion coefficient, which increases with the exponential power of 2 as the convolution layer deepens, and f(j) represents the value of the convolution kernel at position j, 0≤j≤ker-1, It is the feature output after causal expansion convolution; the following formula represents the size of the receptive field: field=(ker-1)·d Then, batch normalization is performed on the features obtained by causal expansion convolution; each mini-batch contains num data: Define the output of the batch data after the causal expansion convolution operation as F i represents the causal expansion convolution output corresponding to the i-th data in the mini-batch; then, Batch normalization can be defined as: Among them, F o (k,l), are the lth value of the kth channel of the oth data before and after the batch normalization operation, γ k , β k is a trainable parameter, ε is defined as a very small value to prevent the denominator from being 0; μ BN 、 are the mean and variance respectively; Next, Input activation function to get Repeat the above calculation process for each data in the mini-batch to obtain the mapping output feature F(q1), and then obtain shallow information through residual connection to prevent overfitting: F res =ReLU(q1+F(q1)).

6. The radar target recognition method based on Transformer and temporal convolutional network according to claim 5, characterized in that: The detailed steps of S5 are: Use a fully connected layer and Softmax function to predict the features. The process is to first pass the output feature F of the temporal convolutional network through a fully connected layer. res Perform dimensionality reduction to obtain the prediction vector: F f =F res W f +b f in, C is the total number of HRRP sample target categories; the prediction vector is then mapped to the probability distribution P through the Softmax function: Among them, F f (c) represents the cth element, P(c|F f ) represents feature F f The probability of belonging to the c-th category target; finally, the classification result is obtained by the following formula: Where argmax(·) indicates the value of the variable c when the function in the brackets reaches its maximum value.

Images