阅读说明：本技术 基于注意力变换网络的雷达高分辨距离像目标识别方法 (Radar high-resolution range profile target identification method based on attention transformation network ) 是由 白雪茹 赵晨 杨敏佳 周峰 于 2021-07-05 设计创作，主要内容包括：本发明公开了一种基于注意力变换网络的雷达高分辨距离像目标识别方法,主要解决现有技术对雷达高分辨距离像识别时,难以关注雷达高分辨距离像的局部细节,难以聚焦于雷达高分辨距离像中更具可分性的目标区域,难以利用其全局时序信息,识别正确率较低,识别性能受限等问题。实现步骤为：(1)生成训练集；(2)构建注意力变换网络；(3)训练注意力变换网络；(4)对待分类的雷达高分辨距离像目标进行识别。本发明同时利用了高分辨距离像的局部细节特征与全局时序信息,对高分辨距离像不同距离单元的重要性进行了区分,使得本发明有效提高了高分辨距离像的识别性能。(The invention discloses a method for identifying a target of a radar high-resolution range profile based on an attention transformation network, which mainly solves the problems that local details of the radar high-resolution range profile are difficult to pay attention to, a target area with higher separability in the radar high-resolution range profile is difficult to focus on, global time sequence information is difficult to utilize, the identification accuracy is low, the identification performance is limited and the like when the radar high-resolution range profile is identified in the prior art. The method comprises the following implementation steps: (1) generating a training set; (2) constructing an attention transformation network; (3) training an attention transformation network; (4) and identifying the radar high-resolution range profile target to be classified. The invention simultaneously utilizes the local detail characteristics and the global time sequence information of the high-resolution range profile to distinguish the importance of different range units of the high-resolution range profile, so that the invention effectively improves the identification performance of the high-resolution range profile.)

1. A radar high-resolution range profile target identification method based on an attention conversion network is characterized in that a convolution attention module consisting of a convolution sub-network and an attention-enhanced convolution sub-network is used for extracting local features of a high-resolution range profile, a position coding module is used for carrying out position coding on the local features, and a multi-head attention conversion encoder module is used for carrying out attention coding on the features subjected to position coding; the radar high-resolution range profile identification method comprises the following steps:

step 1, generating a training set:

(1a) selecting 147950 range images from three types of airplane high-resolution range images acquired by a radar under the conditions of 5520MHz central frequency, 400MHz signal bandwidth and 400Hz pulse repetition frequency to form a sample set;

(1b) sequentially carrying out amplitude normalization and translation alignment pretreatment on each high-resolution range profile in the sample set;

(1c) taking every 30 continuous high-resolution range profiles after the pretreatment as a group, and performing sliding window on the sample set after the pretreatment;

(1d) forming a training set by 9859 groups of sequence samples obtained by the sliding window;

step 2, constructing an attention transformation network;

(2a) building a first convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum value pooling layer and setting sub-network parameters;

(2b) building a second convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum value pooling layer and setting sub-network parameters;

(2c) building a channel attention layer consisting of a global average attention pooling layer, a first convolution layer, a global maximum attention pooling layer, a second convolution layer and a nonlinear activation layer; the output dimensionalities of the global average attention pooling layer and the global maximum attention pooling layer are both 1 x 1, the first convolution layer is provided with 4 convolution kernels with the kernel size of 3 x 3 pixels, the second convolution layer is provided with 32 convolution kernels with the kernel size of 3 x 3 pixels, and the nonlinear activation layer adopts a linear rectification unit activation function;

(2d) building a space attention layer consisting of an average attention pooling layer, a maximum attention pooling layer, a convolution layer and a nonlinear activation layer; the number of output channels of the average attention pooling layer and the maximum attention pooling layer is 1, the convolution layer is provided with 1 convolution kernel with the kernel size of 3 multiplied by 3 pixels, and the nonlinear activation layer adopts a linear rectification unit activation function;

(2e) constructing an attention-enhancing convolution sub-network consisting of a convolution layer, a channel attention layer, a space attention layer, a batch normalization layer, a nonlinear activation layer and a maximum pooling layer; the convolution layer is provided with 32 convolution kernels with the kernel size of 3 x 3 pixels, the number of the convoluted padding is 1, the number of channels of the batch normalization layer is 32, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum pooling layer is 2 x 2 pixels, and the step length is 2 pixels;

(2f) cascading the first convolution sub-network, the second convolution sub-network and the attention-enhancing convolution sub-network into a convolution attention module;

(2g) building a position coding module consisting of a sine encoder and a cosine encoder; the encoding dimensionalities of the sine encoder and the cosine encoder are both 32, the position indexes of the sine encoder are all even numbers selected from [0,96], and the position indexes of the cosine encoder are all odd numbers selected from [0,96 ];

(2h) building a multi-head attention conversion encoder module consisting of a multi-head attention group and a plurality of layers of sensors; the multi-head attention group consists of 8 attention heads connected in parallel, each attention head is obtained by calculating a key value, a query value and a true value through a scaling dot product formula, wherein the lengths of the key value, the query value and the true value are all 97, and the dimensions are all 32 pixels; the multilayer perceptron is formed by cascading a first full-connection layer, a Gaussian error linear unit and a second full-connection layer, wherein the weight dimensions of the first full-connection layer and the second full-connection layer are respectively set to be 32 multiplied by 128 and 128 multiplied by 32;

(2i) the convolution attention module, the position coding module and the multi-head attention transformation coder module are cascaded into an attention transformation network;

step 3, training an attention transformation network:

inputting the training set into an attention transformation network, calculating a cross entropy loss value between the output of the network and a class label of a training image by using a cross entropy loss function, and iteratively updating parameters of the network through a back propagation algorithm until the cross entropy loss function of the network is converged to obtain the trained attention transformation network;

step 4, identifying the radar high-resolution range profile target:

and (3) preprocessing and sliding window processing are carried out on the radar high-resolution range profile to be recognized by adopting the same method as the steps (1b) and (1c) to obtain a sample set to be recognized consisting of 22203 groups of sequence samples, the sample set to be recognized is input into a trained attention transformation network, and the class label is output.

2. The method for radar high-resolution range profile object recognition based on attention transformation network as claimed in claim 1, wherein the parameters of the sub-network in step (2a) are as follows: the convolution layer has 8 convolution kernels with the kernel size of 7 x 7 pixels, the number of filling after convolution is 3, the number of channels of the batch normalization layer is 8, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum pooling layer is 2 x 2 pixels, and the step length is 2 pixels.

3. The method for radar high-resolution range profile object recognition based on attention transformation network as claimed in claim 1, wherein the parameters of the sub-network in step (2b) are as follows: the convolution layer has 16 convolution kernels with kernel size of 5 × 5 pixels, the filling number after convolution is 2, the number of channels of the batch normalization layer is 16, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum pooling layer is 2 × 2 pixels, and the step size is 2 pixels.

4. The method for recognizing radar high-resolution range profile based on attention transformation network as claimed in claim 1, wherein the scaling dot product formula in step (2h) is as follows:

wherein, Head_iDenotes the ith attention head, ∑ denotes the summation operation, M denotes the total number of right fractional ownership scores, M denotes the number of weight scores, exp (-) denotes the exponential operation with the natural constant e as the base, Q_mDenotes a query value used in calculating the mth weight score, T denotes a transposition operation, K_mRepresenting a key value used in calculating the mth weight score, d representing the dimensionality of the key value, N representing the total number of all the union degree scores of the right index operation, N representing the serial number of the union degree scores, Q_nRepresents the query value, K, used in calculating the nth fitness score_nRepresents the key used in calculating the nth degree of association score, and V represents the true value used in calculating the mth weight score.

5. The method for identifying radar high-resolution range profile targets based on attention transformation network as claimed in claim 1, wherein the cross entropy loss function in step 3 is in the form of:

where F represents the cross entropy loss function, j represents the class of the samples in the training set, x_jRepresenting the real label corresponding to each sample in the training set, ln represents the logarithm operation with e as the base, y_jRepresenting the output of the attention transformation network.

Technical Field

The invention belongs to the technical field of radars, and further relates to a radar high-resolution range profile target identification method based on an attention transformation network in the technical field of radar target identification. The invention provides an attention transformation network structure aiming at radar high-resolution range profiles, which can be used for effectively identifying the radar high-resolution range profiles.

Background

The High Resolution Range Profile (HRRP) is the vector sum of the projection of target scattering point sub-echoes acquired by broadband radar signals on a radar sight line, and can provide the distribution condition of the radar scattering cross section area of a target scatterer (such as a nose, a fuselage and the like of an airplane) along the radar sight line under a certain radar viewing angle. Compared with synthetic aperture radar images and inverse synthetic aperture radar images, the method has the advantages of easiness in obtaining and simplicity in processing, and the radar target identification technology based on the high-resolution range profile becomes one of important means for radar real-time target identification. When the high-resolution radar continuously observes a target, a high-resolution range profile sequence of the target can be obtained, and the sequence contains important characteristics such as target shape, structure, scattering intensity and change rule of the target along with the radar visual angle. The existing high-resolution range profile target identification method based on deep learning is mainly improved on the basis of a convolutional neural network or a long-short term memory network, and a complex manual design process is effectively avoided. However, the existing long-short term memory unit for learning the time sequence information of the high-resolution range profile is not enough to pay attention to the local details of the target, and the existing method does not consider the different importance of hidden layers mapped by different range units in the high-resolution range profile to the target recognition, so that the recognition accuracy is low.

In the published paper "analog Network with a conditional Module and a Residual Structure for Radar T area Recognition Based on High-Resolution Range Profile" (Sensors,21January 2020), Zhequan Fu, Li Xiangping Li, Bo Dan, Xukun Wang proposes a method for identifying a target Based on a convolution Residual neural Network with a High Resolution Range. The method comprises the following implementation steps: (1) expanding the radar high-resolution range profile data set after amplitude normalization by using a one-dimensional translation interception technology, and dividing an expanded sample into a training sample and a test sample; (2) taking a deep convolution residual error network formed by cascading a plurality of convolution blocks and residual error blocks as a learner, taking an edge center loss function as a loss function, training the extended training sample, and obtaining a trained learner; (3) and (4) carrying out radar high-resolution range profile automatic target identification by using the trained learner. According to the method, the deep neural network is used for automatically extracting the multilayer features of the target, the edge center loss function is used for improving the feature separability, and the residual error structure is used for effectively improving the identification precision of the deep network. However, the method still has the defects that the network training efficiency is low and the recognition accuracy is low due to the fact that the method ignores the global time sequence information between the high-resolution range profiles.

The patent document "high-resolution range profile target identification method based on attitude adaptive convolutional network" (publication number: CN202110032890.4, application publication number: CN112835008A) applied by the university of sienna electronic technology discloses a high-resolution range profile target identification method based on attitude adaptive convolutional network. The method comprises the following implementation steps: (1) constructing an attitude self-adaptive convolution network; (2) generating a training data set and an auxiliary data set; (3) preprocessing a training data set; (4) generating an adaptive convolution kernel; (5) training a posture self-adaptive convolution network; (6) and (4) target identification. The method can effectively solve the problem of high-resolution range profile echo attitude sensitivity by constructing an attitude self-adaptive convolution network and training the network by utilizing the target high-resolution range profile echo and the target attitude angle information. However, the method still has the disadvantages that hidden layers mapped by different range units in the high-resolution range profile have different importance for target identification, and the method does not distinguish the importance, so that the method cannot focus on a target area with higher separability in the high-resolution range profile during identification, and the identification performance of the algorithm is limited.

Disclosure of Invention

The invention aims to provide a radar high-resolution range profile target identification method based on an attention transformation network aiming at overcoming the defects of the prior art, and aims to solve the problems that in the prior art, the time sequence correlation between high-resolution range profiles is difficult to utilize when radar high-resolution range profile identification is carried out, the network training efficiency is low, and the importance of different areas in the high-resolution range profiles is difficult to distinguish.

The technical idea for realizing the purpose of the invention is as follows: according to the invention, an attention transformation network is constructed, and the trained attention transformation network is used for directly identifying the radar high-resolution range profile, so that the problems that the network training efficiency is low, the target area with higher separability in the high-resolution range profile cannot be focused and the identification accuracy is low when the radar high-resolution range profile target identification is carried out in the prior art are solved. The invention utilizes the convolution attention module formed by cascade connection of the convolution sub-network and the attention-enhancing convolution sub-network to carry out fine feature extraction on the high-resolution range profile, thereby avoiding the problem that the prior art is difficult to pay attention to the local feature details of the high-resolution range profile. The features extracted by the convolution attention module are subjected to position coding through the position coding module, so that the establishment of global time sequence information is realized, and the problem that the prior art is difficult to utilize the global time sequence information of the high-resolution range profile is solved. The multi-head attention transformation encoder module is used for distinguishing and learning the importance degree of the position-encoded features, and the problem that the prior art is difficult to focus on a target area with higher separability in a high-resolution range profile is solved.

The method comprises the following specific steps:

step 1, generating a training set:

(1a) selecting 147950 range images from three types of airplane high-resolution range images acquired by a radar under the conditions of 5520MHz central frequency, 400MHz signal bandwidth and 400Hz pulse repetition frequency to form a sample set;

(1b) sequentially carrying out amplitude normalization and translation alignment pretreatment on each high-resolution range profile in the sample set;

(1c) taking every 30 continuous high-resolution range profiles after the pretreatment as a group, and performing sliding window on the sample set after the pretreatment;

(1d) forming a training set by 9859 groups of sequence samples obtained by the sliding window;

step 2, constructing an attention transformation network;

(2a) building a first convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum value pooling layer and setting sub-network parameters;

(2b) building a second convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum value pooling layer and setting sub-network parameters;

(2c) building a channel attention layer consisting of a global average attention pooling layer, a first convolution layer, a global maximum attention pooling layer, a second convolution layer and a nonlinear activation layer; the output dimensionalities of the global average attention pooling layer and the global maximum attention pooling layer are both 1 x 1, the first convolution layer is provided with 4 convolution kernels with the kernel size of 3 x 3 pixels, the second convolution layer is provided with 32 convolution kernels with the kernel size of 3 x 3 pixels, and the nonlinear activation layer adopts a linear rectification unit activation function;

(2d) building a space attention layer consisting of an average attention pooling layer, a maximum attention pooling layer, a convolution layer and a nonlinear activation layer; the number of output channels of the average attention pooling layer and the maximum attention pooling layer is 1, the convolution layer is provided with 1 convolution kernel with the kernel size of 3 multiplied by 3 pixels, and the nonlinear activation layer adopts a linear rectification unit activation function;

(2e) constructing an attention-enhancing convolution sub-network consisting of a convolution layer, a channel attention layer, a space attention layer, a batch normalization layer, a nonlinear activation layer and a maximum pooling layer; the convolution layer is provided with 32 convolution kernels with the kernel size of 3 x 3 pixels, the number of the convoluted padding is 1, the number of channels of the batch normalization layer is 32, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum pooling layer is 2 x 2 pixels, and the step length is 2 pixels;

(2f) cascading the first convolution sub-network, the second convolution sub-network and the attention-enhancing convolution sub-network into a convolution attention module;

(2g) building a position coding module consisting of a sine encoder and a cosine encoder; the encoding dimensionalities of the sine encoder and the cosine encoder are both 32, the position indexes of the sine encoder are all even numbers selected from [0,96], and the position indexes of the cosine encoder are all odd numbers selected from [0,96 ];

(2h) building a multi-head attention conversion encoder module consisting of a multi-head attention group and a plurality of layers of sensors; the multi-head attention group consists of 8 attention heads connected in parallel, each attention head is obtained by calculating a key value, a query value and a true value through a scaling dot product formula, wherein the lengths of the key value, the query value and the true value are all 97, and the dimensions are all 32 pixels; the multilayer perceptron is formed by cascading a first full-connection layer, a Gaussian error linear unit and a second full-connection layer, wherein the weight dimensions of the first full-connection layer and the second full-connection layer are respectively set to be 32 multiplied by 128 and 128 multiplied by 32;

(2i) the convolution attention module, the position coding module and the multi-head attention transformation coder module are cascaded into an attention transformation network;

step 3, training an attention transformation network:

inputting the training set into an attention transformation network, calculating a cross entropy loss value between the output of the network and a class label of a training image by using a cross entropy loss function, and iteratively updating parameters of the network through a back propagation algorithm until the cross entropy loss function of the network is converged to obtain the trained attention transformation network;

step 4, identifying the radar high-resolution range profile target:

and (3) preprocessing and sliding window processing are carried out on the radar high-resolution range profile to be recognized by the same method as the steps (1b) and (1c), so as to obtain a sample set to be recognized, which is composed of 22203 groups of sequence samples, and the sample set to be recognized is input into a trained attention transformation network to output class labels.

Compared with the prior art, the invention has the following advantages:

firstly, the invention utilizes the convolution attention module formed by cascading the convolution sub-network and the attention enhancement convolution sub-network to extract the precise local features of the radar high-resolution range profile, so that the invention focuses on the more useful local details of the radar high-resolution range profile, thereby overcoming the problem that the local details of the range profile are difficult to focus on when the time sequence information of the radar high-resolution range profile is learned in the prior art, and improving the accuracy of the identification of the radar high-resolution range profile.

Secondly, the invention carries out position coding on the features extracted by the convolution attention module through the position coding module on the basis of fully utilizing the time-varying information of the radar high-resolution range profile, thereby realizing the establishment of the global time sequence information, overcoming the problem that the prior art is difficult to utilize the global time sequence information of the radar high-resolution range profile, and improving the identification precision of the radar high-resolution range profile.

Thirdly, the invention carries out attention coding on the position coded features through a multi-head attention transform encoder module, distinguishes the importance of different areas in the radar high-resolution range profile, overcomes the problem that the prior art is difficult to focus on a target area with more separability in the radar high-resolution range profile, and improves the identification performance of the radar high-resolution range profile.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a schematic structural diagram of a backbone network module according to the present invention;

FIG. 3 is a schematic diagram of the structure of the first and second convolution sub-networks in the backbone network module of the present invention;

fig. 4 is a schematic structural diagram of a channel attention layer and a spatial attention layer in the backbone network module according to the present invention.

Detailed Description

The invention is further described below with reference to the figures and examples.

With reference to fig. 1, the specific steps of the implementation of the present invention will be described in detail.

Step 1, generating a training set.

S1, selecting 147950 range images from three types of airplane high-resolution range images acquired by the radar under the conditions of 5520MHz central frequency, 400MHz signal bandwidth and 400Hz pulse repetition frequency to form a sample set.

And S2, sequentially carrying out amplitude normalization and translation alignment pretreatment on each high-resolution range profile in the sample set.

And S3, taking each 30 continuous high-resolution range profiles after the preprocessing as a group, and performing sliding window on the preprocessed sample set.

The sliding window comprises the following specific steps:

and step 1, arranging all high-resolution range images of the preprocessed sample set into a line to obtain a sample lumped sequence.

And 2, using a rectangular sliding window with the length of 30 high-resolution range profiles and the width of 1 high-resolution range profile to slide on the sample lumped sequence by the step length of 15 high-resolution range profiles, and taking out all high-resolution range profile sequences in the sliding window to form sequence samples behind the sliding window.

And S4, forming a training set by 9859 groups of sequence samples obtained by the sliding window.

And 2, constructing an attention transformation network.

Build a backbone network module of constituteing by 3 modules, its structure does in proper order: a convolution attention module, a position coding module and a multi-head attention transformation coder module. The convolution attention module is composed of a first convolution sub-network, a second convolution sub-network and an attention enhancement convolution sub-network in cascade connection; the position coding module consists of a sine coder and a cosine coder; the multi-head attention transformation encoder module is composed of a multi-head attention group and a cascade of multi-layer sensors.

The backbone network module constructed by the present invention is further described with reference to fig. 2.

The input data of the backbone network module is a radar high-resolution range profile sequence, the convolution attention module extracts local features of the radar high-resolution range profile, the position coding module carries out position coding on the extracted local features, and the multi-head attention transform coder module completes attention coding on the features and outputs a recognition result.

The first and second convolution sub-networks constructed by the present invention are further described with reference to fig. 3.

Building a first convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum pooling layer; the convolutional layer has 8 convolutional kernels with a kernel size of 7 × 7 pixels, the number of padding after convolution is 3, the number of channels of the batch normalization layer is 8, the nonlinear active layer adopts a linear rectification unit active function, the window size of the maximum pooling layer is 2 × 2 pixels, and the step size is 2 pixels, as shown in fig. 3 (a).

Building a first convolution sub-network consisting of a convolution layer, a batch normalization layer, a nonlinear activation layer and a maximum pooling layer; the convolution layer has 16 convolution kernels with kernel size of 5 × 5 pixels, the number of filling after convolution is 2, the number of channels of the batch normalization layer is 16, the nonlinear active layer adopts a linear rectification unit active function, the window size of the maximum pooling layer is 2 × 2 pixels, and the step size is 2 pixels, as shown in fig. 3 (b).

The channel attention layer, the spatial attention layer, built by the present invention is further described with reference to fig. 4.

Building a channel attention layer consisting of a global average attention pooling layer, a first convolution layer, a global maximum attention pooling layer, a second convolution layer and a nonlinear activation layer; the output dimensions of the global average attention pooling layer and the global maximum attention pooling layer are both 1 × 1, the first convolution layer has 4 convolution kernels with a kernel size of 3 × 3 pixels, the second convolution layer has 32 convolution kernels with a kernel size of 3 × 3 pixels, and the nonlinear active layer adopts a linear rectifying unit active function, as shown in fig. 4 (a).

Building a space attention layer consisting of an average attention pooling layer, a maximum attention pooling layer, a convolution layer and a nonlinear activation layer; the output of each of the average attention pooling layer and the maximum attention pooling layer is subjected to channel dimension splicing and then used as the total output of the attention pooling, the number of output channels of each of the average attention pooling layer and the maximum attention pooling layer is 1, each convolution layer has 1 convolution kernel with the kernel size of 3 × 3 pixels, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum attention pooling layer is 2 × 2 pixels, and the step length is 2 pixels, as shown in fig. 4 (b).

Constructing an attention-enhancing convolution sub-network consisting of a convolution layer, a channel attention layer, a space attention layer, a batch normalization layer, a nonlinear activation layer and a maximum pooling layer; the convolution layer has 32 convolution kernels with the kernel size of 3 x 3 pixels, the filling number after convolution is 1, the number of channels of the batch normalization layer is 32, the nonlinear activation layer adopts a linear rectification unit activation function, the window size of the maximum pooling layer is 2 x 2 pixels, and the step length is 2 pixels.

The first convolution sub-network, the second convolution sub-network, and the attention-enhancing convolution sub-network are cascaded into a convolution attention module.

Building a position coding module consisting of a sine encoder and a cosine encoder; the encoding dimensions of the sine encoder and the cosine encoder are both 32, the position indexes of the sine encoder are all even numbers selected from [0,96], and the position indexes of the cosine encoder are all odd numbers selected from [0,96 ].

Building a multi-head attention conversion encoder module consisting of a multi-head attention group and a plurality of layers of sensors; the multi-head attention group comprises 8 attention heads connected in parallel, the characteristics after position coding are respectively copied into three parts as true values, key values and query values, the key values, the query values and the true values are calculated through a scaling dot product formula to obtain each attention head, and the 8 attention heads are added to obtain the output of the multi-head attention group, wherein the lengths of the key values, the query values and the true values are 97, and the dimensions are 32 pixels; the multi-head attention transformation encoder module is obtained by connecting a softmax classifier after the output of a multi-head attention group is connected to the multi-head attention transformation encoder module for feature perception, wherein the weight dimensions of the first full connection layer and the second full connection layer are respectively set to be 32 multiplied by 128 and 128 multiplied by 32.

The scaling dot product formula is as follows:

wherein, Head_iDenotes the ith attention head, ∑ denotes a summing operation, M denotes a total number of right fractional ownership remultirations,m denotes the number of weight fractions, exp (-) denotes the exponential operation based on the natural constant e, Q_mDenotes a query value used in calculating the mth weight score, T denotes a transposition operation, K_mRepresenting a key value used in calculating the mth weight score, d representing the dimensionality of the key value, N representing the total number of all the union degree scores of the right index operation, N representing the serial number of the union degree scores, Q_nRepresents the query value, K, used in calculating the nth fitness score_nRepresents the key used in calculating the nth degree of association score, and V represents the true value used in calculating the mth weight score.

The softmax function is as follows:

wherein p is_tDenotes the probability that the input image belongs to the t-th type, t 1,2, …, M, exp (·) denotes an exponential operation based on a natural constant e, O_vRepresenting the output of the v-th neuron, the value of v is equal to t, L represents the total number of neurons, L represents the serial number of the neuron, O_lRepresenting the output of the l-th neuron.

And 3, training an attention transformation network.

Step 1, initializing weight parameters and bias parameters of each convolution layer, channel attention layer, space attention layer, multi-head attention group and multi-layer perceptron in the attention transformation network.

And 2, inputting the radar high-resolution range profile samples in the training sample set into a convolution attention module, and extracting local features to obtain fine local features.

And 3, inputting the fine local features into a position coding module to obtain the features after position coding.

And 4, inputting the position-coded features into a multi-head attention transformation encoder module to obtain a class label of the attention transformation network.

And 5, calculating the error of the attention transformation network according to the category label of the attention transformation network and the label corresponding to the radar high-resolution distance image sample and the cross entropy loss function.

The cross entropy loss function is of the form:

where F represents the cross entropy loss function, j represents the class of the samples in the training set, x_jRepresenting the real label corresponding to each sample in the training set, ln represents the logarithm operation with e as the base, y_jRepresenting the output of the sequence adjustment network.

And 6, reversely propagating the error of the attention transformation network, and updating the weight parameters and the bias parameters of each convolution layer, each channel attention layer, each space attention layer, each multi-head attention group and each multi-layer sensor of the attention transformation network according to a gradient descent method.

And 7, repeating and iterating the calculation processes from the step 2 to the step 6 by using the updated weight parameters and bias parameters of each convolution layer, the channel attention layer, the spatial attention layer, the multi-head attention group and the multi-layer perceptron of the attention transformation network, and stopping iteration after the error is stably converged to obtain the trained attention transformation network.

And 4, identifying the radar high-resolution range profile target.

And (4) preprocessing and sliding window processing are carried out on the radar high-resolution range profile to be recognized by adopting the same method as the steps S2 and S3 to obtain a sample set to be recognized consisting of 22203 groups of sequence samples, the sample set to be recognized is input into the trained attention transformation network, and the class label is output.

The effect of the present invention will be further described with reference to simulation experiments.

1. And (5) simulating experimental conditions.

The hardware platform of the simulation experiment of the invention is as follows: the processor is an Intel Xeon E5-2650 CPU, the main frequency of the processor is 2.20GHz, the memory is 64GB, and the display card is NVIDIA Geforce GTX 1080 ti.

The simulation experiment software platform of the invention is a Windows 10 operating system, Mtalab2018, Python 3.6 and Pythroch 1.4.

2. And (5) analyzing simulation experiment contents and results.

The simulation experiment of the invention is that under the same data set, the invention and the traditional convolution neural network method are adopted to respectively identify the three types of airplane radar high-resolution range profiles, and the identification result is obtained.

The data set used in the simulation experiment is three types of airplane radar high-resolution range profiles obtained by the radar under the conditions of 5520MHz central frequency, 400MHz signal bandwidth and 400Hz pulse repetition frequency, and the three types of airplane targets are respectively: an-26, prize, Jack-42, where An-26 and prize each have 7 data segments and Jack-42 has 5 data segments. Selecting 5 and 6 segments of An-26, 6 and 7 segments of prize-shape and 2 and 5 segments of Jack-42 as training sample sets, and selecting 1 to 4 segments of An-26, 1 to 5 segments of prize-shape and 1, 3 and 4 segments of Jack-42 as test sample sets.

In the simulation experiment of the invention, the existing radar high-resolution range profile identification method based on the traditional convolutional neural network is taken from a paper 'HRRP feature extraction and reproduction method of radar ground and target use of a capacitive distance network' published by Beicheng Ding, Penghui Chen and the like, the radar high-resolution range profile identification method based on the traditional convolutional neural network firstly carries out fast Fourier transform on the radar high-resolution range profile, then extracts the characteristics of the transformed high-resolution range profile, and inputs the extracted characteristics into the designed convolutional neural network to identify the high-resolution range profile.

Simulation experiment 1: the method is applied to identifying the three types of airplane radar high-resolution range profile targets, firstly, the training sample set is used for training the attention transformation network-based radar high-resolution range profile identification network to obtain the trained attention transformation network-based radar high-resolution range profile identification network, and then, the test sample set is used for testing the trained attention transformation network-based radar high-resolution range profile identification network.

The accuracy rate identified by the simulation experiment 1 is calculated by the following formula:

wherein c represents the identification accuracy of the test sample set, R represents the total number of samples of the test sample set, h (-) represents the category discrimination function, t_rA true class label, y, representing the r-th test sample in the set of test samples_rShowing the output result of the attention transformation network corresponding to the r test sample in the test sample set when t is_rAnd y_rEqual, h (t)_r,y_r) Equal to 1, otherwise, h (t)_r,y_r) Equal to 0.

According to the formula of R-22203,the recognition accuracy of the invention is calculated to be 96.82%.

Simulation experiment 2: the method comprises the steps of identifying three types of airplane radar high-resolution range profiles by applying a traditional convolutional neural network method, firstly training a traditional convolutional neural network-based radar high-resolution range profile identification network by using a training sample set to obtain a trained traditional convolutional neural network-based radar high-resolution range profile identification network, and then testing the trained traditional convolutional neural network-based radar high-resolution range profile identification network by using a test sample set.

The recognition accuracy of the simulation experiment 2 is calculated by the following formula:

wherein c represents the identification accuracy of the test sample set, R represents the total number of samples of the test sample set, h (-) represents the category discrimination function, t_rA true class label, y, representing the r-th test sample in the set of test samples_rShowing the output result of the attention transformation network corresponding to the r test sample in the test sample set when t is_rAnd y_rEqual, h (t)_r,y_r) Equal to 1, otherwise, h (t)_r,y_r) Equal to 0.

According to the formula of R-22203,the recognition accuracy of the traditional convolutional neural network is calculated to be 87.47%.

In conclusion, compared with the existing method, the method for identifying the radar high-resolution range profile based on the attention transformation network can effectively improve the identification performance of the radar high-resolution range profile.