A radar target recognition method based on multi-scale convolutional capsule network

Abstract

The present invention discloses a radar target recognition method based on a multi-scale convolutional capsule network, which uses an improved deep learning network, namely a multi-scale convolutional capsule network, to identify and classify target HRRP, integrates a network with strong generalization ability to fully capture effective structural information in HRRP, and optimizes the parameters in the network while building a complex model, so as to achieve faster convergence and higher recognition performance. Based on the ability of CNN to obtain local structural information and the Inception structure, a multi-scale CNN structure is proposed, which captures the separable information of different levels of HRRP targets in detail from multiple angles, and increases the sparsity of the network. At the same time, a more robust capsule network is used as a supplement to CNN, which makes up for the deficiency of CNN in reflecting the target HRRP sequence structural information, and exerts greater advantages in the field of radar HRRP recognition with stronger fitting ability.

Description

Technical Field

[0001] The invention belongs to the field of radar target recognition, and specifically relates to a radar target recognition method based on multi-scale convolution and capsule network. Background Art

[0002] With the continuous maturity of radar technology, most of the traditional HRRP recognition methods based on statistical models, popular learning and kernel methods have been able to obtain the distribution of strong scattering points of targets and perform recognition and classification. However, most traditional methods are based on fully connected structures for frame modeling, missing inter-frame correlation information, and unable to capture structural information reflecting HRRP characteristics. At the same time, they also have high requirements for the completeness of target data. However, in actual environments, the objects to be recognized are usually non-cooperative targets, the angular domain information of the collected samples is incomplete, and the distribution of data cannot reach the ideal state of the experiment, which increases the dependence on the experience of researchers in the feature extraction process. The rise of deep learning algorithms in recent years has changed the traditional fully connected structure and can automatically obtain the deep features contained in HRRP data.

[0003] Learning and extracting HRRP target features is an important part of radar target recognition. Traditional HRRP feature extraction methods are mainly divided into two parts: (1) Transformer-based feature extraction methods, such as spectrograms, spectrum amplitude features, etc. These methods project the HRRP signal into the frequency domain, and then model and identify its frequency domain features. (2) Feature extraction methods based on reduced-dimensional data. Although traditional feature extraction methods have good recognition performance, most of these methods are unsupervised and lossy, and the choice of feature extraction methods is highly dependent on the researchers' understanding and experience of the data. Therefore, in many cases, it is difficult to achieve good results. In recent years, with the development of deep learning, methods based on deep learning have gradually been introduced into radar target recognition. Summary of the invention

[0004] In view of the shortcomings of the above traditional methods and the characteristics of HRRP data, the present invention provides a radar target recognition method based on a multi-scale convolutional capsule network.

[0005] This paper uses an improved deep learning network, namely a multi-scale convolutional capsule network, to identify and classify target HRRP. It integrates a network with strong generalization ability to fully capture the effective structural information in HRRP. While building a complex model, it also optimizes the parameters in the network to achieve faster convergence and higher recognition performance. Based on CNN's ability to obtain local structural information and the Inception structure, a multi-scale CNN structure is proposed to capture the separable information of different levels of HRRP targets in detail from multiple angles, increasing the sparsity of the network. At the same time, a more robust capsule network is used as a supplement to CNN to make up for the insufficiency of CNN in reflecting the target HRRP sequence structural information, and to play a greater advantage in the field of radar HRRP recognition with a stronger fitting ability.

[0006] The present invention uses a multi-scale convolutional capsule network to give the model a certain depth. The model organized in this way can better rely on the data to gradually abstract high-level structural features, and construct a multi-scale convolution module to replace the stacked deep convolution layer to increase the sparsity of the network, which not only broadens the network layer, but also enhances the applicability to structures of different scales, and controls the computational amount and parameter amount of the network to a certain extent while ensuring higher feature extraction performance. The basic capsule layer constructs a vector capsule for expressing the spatial relationship between different features of HRRP, and passes the features to high-level prediction capsules for classification through dynamic routing. Since the pooling operation in CNN will lose the relevant position information in the HRRP sample, this module uses the fully connected layer of the dynamic routing mechanism to transmit the vector features after convolution to obtain the capsule features, further explores the relationship between different attributes in the HRRP sample, and achieves better classification in the form of vector output.

[0007] A radar target recognition method based on a multi-scale convolutional capsule network comprises the following steps:

[0008] S1: Dataset collection.

[0009] S2: Preprocess the samples in the dataset.

[0010] S3: Input the preprocessed HRRP samples into the multi-scale convolution module for feature extraction;

[0011] S4: The features obtained in S3 are passed to a convolution layer and a ReLU activation layer for processing, and then the obtained high-dimensional features are passed to the capsule network module. After the capsule features are obtained by convolution, the HRRP features stored in the basic capsule are converted into more comprehensive overall features through the dynamic routing mechanism, and passed to the prediction capsule layer representing the HRRP target category for classification. In order to avoid overfitting caused by the fully connected routing mechanism, the dropout method is introduced to discard the transmission of some invalid vector neurons, further explore the relationship between different attributes in the HRRP sample, and achieve better classification in the form of vector output;

[0012] S5: Build a reconstruction module, take the high-level capsule representation with the largest unit length in the prediction capsule layer as the reconstruction target, and pass it through three fully connected network layers with activation functions to output the reconstructed target data;

[0013] S6: Train the multi-scale convolutional capsule network constructed by the multi-scale convolutional module and the capsule network module using the data set processed by S2.

[0014] Furthermore, the detailed steps of S1 are:

[0015] The HRRP data set collected by the radar is merged according to the type of target. For each type of sample, training samples and test samples are selected in different data segments. In the process of selecting the training set and the 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 samples in the training set and the test set of each type of target is 7:3.

[0016] Furthermore, the detailed steps of S2 are:

[0017] S2.1: Intensity normalization. The original HRRP is expressed as Where L1 represents the total number of distance units contained in HRRP, then the HRRP after intensity normalization can be expressed as:

[0018]

[0019] S2.2: Align the center of gravity. Translate HRRP so that its center of gravity moves g1 to Nearby, so that the distance units containing information in HRRP will be distributed near the center. The calculation method of HRRP center of gravity g1 is as follows:

[0020]

[0021] Preferably, the detailed steps of S3 are:

[0022] S3.1: The preprocessed HRRP samples are input into a multi-scale convolution module, which includes three stacked feature extraction sub-blocks and one fully connected layer, wherein each feature extraction sub-block is composed of the same network layer, including a convolution layer, a batch normalization layer, and an activation layer, and then passes through a fully connected layer for feature mapping, and the target category of the input HRRP sample is determined by the maximum value of the output of the fully connected layer.

[0023] The feature extraction sub-block in the multi-scale convolution module contains a total of K convolution kernels processed synchronously, with the step size set to 2. The zero-filling layer is filled with 0 to increase sparsity. The multi-level position feature representation of the sample output by the k1th convolution kernel of different scales is: Where D represents the number of channels and W represents the width of the feature.

[0024] S3.2: The data after the convolution layer needs further processing. In order to make the model easier to converge and the network training process more stable, batch normalization is added after the convolution. By calculating the mean and variance of the data in each mini_batch, assuming that there are N in a mini batch m HRRP samples, then the output is defined as Among them, F n Represents the convolution output corresponding to the nth HRRP sample. In each small batch, The HRRP data in is batch normalized to obtain It is expressed as:

[0025]

[0026] Among them, F n (k,l) represents the lth element in the kth channel of the convolutional layer output corresponding to the HRRP sample before batch normalization. That is, the HRRP data after batch normalization, α k and β k is the trainable parameter corresponding to the kth channel, ε is a very small number, in order to prevent the divisor from being 0, E(.) is the mean operation, Var(.) represents the variance operation;

[0027] S3.3: Then use the activation function ReLU Each element in is nonlinearly activated to obtain If the input is Then the corresponding output after ReLU is It is expressed as:

[0028]

[0029] The multi-level features F(k1) are respectively adjusted by batch normalization distribution and nonlinear mapping of the activation layer, and then fused into features of multiple different levels to obtain the final output feature map F m , which can be expressed as:

[0030] F m =[F(1),F(2),...,F(k)]

[0031] Where [·] represents the features of each layer that are spliced ​​together. The value of D×K represents the height of the feature map, which is equal to the width W of the feature map. In other words, the feature map extracted by the multi-scale convolution module can be regarded as a square image, which can be used in the subsequent network module to further extract the envelope feature information of the HRRP sample.

[0032] Preferably, the detailed steps of S4 are:

[0033] S4.1: The output of the multi-scale convolution module is processed by a convolution layer and a ReLU activation layer. While obtaining the local hierarchical information in the position features extracted by the multi-scale convolution module, it is mapped to a higher-dimensional space. The number of channels is converted from 1 to 256 to facilitate the transmission of capsule features in vector form.

[0034] S4.2: The capsule network module consists of a primary capsule layer and a prediction capsule layer. First, the previous step output is passed into the primary capsule layer, and every eight convolution feature nodes in the previous step output are stacked into a basic capsule feature, recorded as That is, a vector neuron in the form of 1×8 obtains 32 channels at the same time, and each channel outputs 64 (8×8) basic capsules of length 8. Therefore, the primary capsule layer contains a total of 2048 (8×8×32) basic capsules.

[0035] S4.3: The primary capsule layer is denoted as the l1th layer of the capsule network module, and the n1th capsule after the l1th layer is stacked is denoted as Where n1 = 1, 2, ..., 2048, the prediction capsule layer in the routing process is denoted as layer l1+1, and the mth (m = 0, 1, 2) prediction capsule output by this layer is denoted as v m . First, through the transformation matrix W nm Get vector It can be expressed as:

[0036]

[0037] Secondly, the coupling coefficient c can be adjusted dynamically nm From primary capsule and prediction capsule v mThe similarity between them is determined by the softmax function:

[0038]

[0039] Where d represents the category of the target aircraft, b nm Representation vector With v m The prior probability of the connection is initialized to 0 in the forward propagation. Then the transformed vector With parameter c nm The weighted summation is used to obtain the output vector s of the intermediate layer. m , which can be expressed as:

[0040]

[0041] In order to make the final capsule contain rich and accurate global feature representation, the vector s is squashed using a nonlinear activation function. m The modulus value is distributed between 0 and 1, and finally the prediction capsule v for classification is output m ,and Output v m The length of ||v m || represents the probability of a certain type of HRRP target existing, v m The direction of represents the different feature information in the sample component. The prediction capsule v output by the l+1 layer m It is calculated by the following formula:

[0042]

[0043] By continuously iteratively adjusting b nm To update the network weight parameter c nm . b nm The update method is expressed by the following formula:

[0044]

[0045] Preferably, the detailed steps of S5 are:

[0046] The reconstruction module can be regarded as a decoder, which is used to restore the output results of the correct classification of HRRP samples to the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire multi-scale convolutional capsule network to assist in comparing the results of HRRP classification. Specifically, the capsule representation with the largest unit length in the prediction capsule layer is used as the reconstruction target, and after three fully connected network layers with activation functions, the reconstructed target data is output.

[0047] Preferably, the detailed steps of S6 are:

[0048] S6.1: The total loss of the multi-scale convolutional capsule network is given by the capsule classification loss L d and the reconstruction loss L rec It consists of two parts and loses L d Plays a leading role, the expression is as follows:

[0049] Loss 总 =L d +L rec

[0050] The capsule form enables multiple classifications to exist simultaneously, so the training process of the multi-scale convolutional capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as:

[0051] L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m - ) 2

[0052] Where d represents the HRRP target category, Y d is the classification label in training, m + 、m - are fixed hyperparameters, set to 0.9 and 0.1 respectively, and λ is a coefficient to prevent the network from solving local optimal solutions, set to 0.5. When the prediction capsule matches the correct HRRP sample d during training, Y d =1, and when the probability predicted by this prediction capsule exceeds 0.9, the loss is set to zero, and when the probability is less than 0.9, the corresponding loss value is calculated for optimization; similarly, when the prediction capsule matches the wrong HRRP sample, that is, the class d does not exist in the sample, then Y d = 0. The above process actually performs a binary classification operation on each prediction capsule, and the parameter λ is introduced to ensure the stability of the network training process.

[0053] The reconstruction loss L of the decoding process in the reconstruction module rec It represents the gap between the input HRRP sample and the reconstructed data. The present invention uses the Euclidean distance between the two as the reconstruction loss and adds the loss of the entire multi-scale convolutional capsule network to train the network parameters. The expression is as follows:

[0054]

[0055] where hic represents the initial HRRP sample, h ir represents the reconstructed data, α is the scaling factor and is set to 0.0005 so that the capsule classification loss dominates the total loss.

[0056] Therefore, the total loss function of the multi-scale convolutional capsule network can be defined as follows:

[0057]

[0058] S6.2: Initialize all weights and biases to be trained, set training parameters, including learning rate, batch_size, training batches, and train the model.

[0059] The beneficial effects of the present invention are:

[0060] 1. The present invention constructs a multi-scale convolution block to replace the stacked deep convolution layer to increase the sparsity of the network, which not only broadens the network layer, but also enhances the applicability to structures of different scales, controls the computational complexity and parameter quantity of the network to a certain extent, and ensures high feature extraction performance. The convolution layer with convolution kernels of different sizes is used as a sliding window to extract HRRP position features at different levels, and the position features of different receptive fields are merged into a whole feature map for use by subsequent modules. The different convolution kernel parameters are obtained by network training and applied to HRRP feature extraction in the test phase, which makes up for the shortcoming that the traditional capsule network uses only one convolution layer to extract incomplete sample features, and improves the quality of subsequent basic capsule construction.

[0061] 2. The basic capsule layer in the capsule network module is essentially a convolution algorithm-capsule layer. Its main function is to convert the features of the convolution layer into capsule features, that is, to adjust the neurons from scalar form to vector form that can be used for capsule feature transmission, rather than scalar input and scalar output like convolutional neural networks. Each basic capsule will be presented in vector form to represent an instance feature information of the HRRP sample. Similar to the structural components of the feature representation obtained by the convolutional neural network, we use basic capsules as structural components for extracting features. The instantiation parameters of different types of aircraft represented by these structural components are described by the scalar neurons input into the basic capsule layer. The vector capsule used to describe a certain instantiation information (cabin, wing, hangar or tail, etc.) in the HRRP sample is composed of eight scalars, that is, the length is 8. In this way, each basic capsule covers eight different feature attributes in the sample, such as the direction, speed, size and edge attributes of the target aircraft. This vector representation method can describe and summarize a certain instantiation information of different types of aircraft, enrich the description information of the feature, and can not only indicate whether there is a certain instance feature, but also indicate similar feature attributes between HRRP samples.

[0062] 3. The prediction capsule layer can be regarded as a linear layer, which mainly transforms the HRRP features stored in the basic capsule into more comprehensive overall features through a dynamic routing mechanism, and passes them to the high-level prediction capsule representing the HRRP target category for classification. Each prediction capsule is connected to the basic capsule of the previous layer in a fully connected form, and the output is in vector form, corresponding to the classification result of an HRRP target. In addition, in order to avoid overfitting caused by the fully connected routing mechanism, we introduced the dropout method to discard the transmission of some invalid vector neurons, eliminate the weakened joint adaptability between neuron nodes, and enhance the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A flowchart of the steps of an embodiment of the present invention. DETAILED DESCRIPTION

[0064] Reference Figure 1 , is a technical flow chart of a radar high-resolution range profile recognition method based on multi-scale convolution and capsule network of the present invention, and the specific implementation steps are as follows:

[0065] S1: Dataset collection.

[0066] The HRRP data set collected by the radar is merged according to the type of target. For each type of sample, training samples and test samples are selected in different data segments. In the process of selecting the training set and the 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 samples in the training set and the test set of each type of target is 7:3.

[0067] S2: Preprocess the samples in the dataset.

[0068] S2.1: Intensity normalization. The original HRRP is expressed as Where L1 represents the total number of distance units contained in HRRP, then the HRRP after intensity normalization can be expressed as:

[0069]

[0070] S2.2: Align the center of gravity. Translate HRRP so that its center of gravity moves g1 to Nearby, so that the distance units containing information in HRRP will be distributed near the center. The calculation method of HRRP center of gravity g1 is as follows:

[0071]

[0072] After intensity normalization and center of gravity alignment, the amplitude of the original HRRP samples has been limited to between 0 and 1, which not only unifies the scale, but also the value between 0 and 1 is very beneficial for subsequent neural network processing; HRRP echo signals that are biased to the right or left are adjusted to near the center point.

[0073] S3: Input the preprocessed HRRP samples into the multi-scale convolution module for feature extraction;

[0074] S3.1: The preprocessed HRRP samples are input into a multi-scale convolution module, which includes three stacked feature extraction sub-blocks and one fully connected layer, wherein each feature extraction sub-block is composed of the same network layer, including a convolution layer, a batch normalization layer, and an activation layer, and then passes through a fully connected layer for feature mapping, and the target category of the input HRRP sample is determined by the maximum value of the output of the fully connected layer.

[0075] The feature extraction sub-block in the multi-scale convolution module contains a total of K convolution kernels processed synchronously, with the step size set to 2. The zero-filling layer is filled with 0 to increase sparsity. The multi-level position feature representation of the sample output by the k1th convolution kernel of different scales is: Where D represents the number of channels and W represents the width of the feature.

[0076] S3.2: The data after the convolution layer needs further processing. In order to make the model easier to converge and the network training process more stable, batch normalization is added after the convolution. By calculating the mean and variance of the data in each mini_batch, assuming that there are N in a mini batch m HRRP samples, then the output is defined as Among them, F n Represents the convolution output corresponding to the nth HRRP sample. In each small batch, The HRRP data in is batch normalized to obtain It is expressed as:

[0077]

[0078] Among them, F n (k,l) represents the lth element in the kth channel of the convolutional layer output corresponding to the HRRP sample before batch normalization. That is, the HRRP data after batch normalization, α k and β k is the trainable parameter corresponding to the kth channel, ε is a very small number, in order to prevent the divisor from being 0, E(.) is the mean operation, Var(.) represents the variance operation;

[0079] S3.3: Then use the activation function ReLU Each element in is nonlinearly activated to obtain If the input is Then the corresponding output after ReLU is It is expressed as:

[0080]

[0081] The multi-level features F(k1) are respectively adjusted by batch normalization distribution and nonlinear mapping of the activation layer, and then fused into features of multiple different levels to obtain the final output feature map F m , which can be expressed as:

[0082] F m =[F(1),F(2),...,F(k)]

[0083] Where [·] represents the features of each layer that are spliced ​​together. The value of D×K represents the height of the feature map, which is equal to the width W of the feature map. In other words, the feature map extracted by the multi-scale convolution module can be regarded as a square image, which can be used in the subsequent network module to further extract the envelope feature information of the HRRP sample.

[0084] S4: The features obtained in S3 are passed to a convolution layer and a ReLU activation layer for processing, and then the obtained high-dimensional features are passed to the capsule network module. After the capsule features are obtained by convolution, the HRRP features stored in the basic capsule are converted into more comprehensive overall features through the dynamic routing mechanism, and passed to the prediction capsule layer representing the HRRP target category for classification. In order to avoid overfitting caused by the fully connected routing mechanism, the dropout method is introduced to discard the transmission of some invalid vector neurons, further explore the relationship between different attributes in the HRRP sample, and achieve better classification in the form of vector output;

[0085] S4.1: The output of the multi-scale convolution module is processed by a convolution layer and a ReLU activation layer. While obtaining the local hierarchical information in the position features extracted by the multi-scale convolution module, it is mapped to a higher-dimensional space. The number of channels is converted from 1 to 256 to facilitate the transmission of capsule features in vector form.

[0086] S4.2: The capsule network module consists of a primary capsule layer and a prediction capsule layer. First, the previous step output is passed into the primary capsule layer, and every eight convolution feature nodes in the previous step output are stacked into a basic capsule feature, recorded as That is, a vector neuron in the form of 1×8 obtains 32 channels at the same time, and each channel outputs 64 (8×8) basic capsules of length 8. Therefore, the primary capsule layer contains a total of 2048 (8×8×32) basic capsules.

[0087] S4.3: The primary capsule layer is denoted as the l1th layer of the capsule network module, and the n1th capsule after the l1th layer is stacked is denoted as Where n1 = 1, 2, ..., 2048, the prediction capsule layer in the routing process is denoted as layer l1+1, and the mth (m = 0, 1, 2) prediction capsule output by this layer is denoted as v m . First, through the transformation matrix W nm Get vector It can be expressed as:

[0088]

[0089] Secondly, the coupling coefficient c can be adjusted dynamically nm From primary capsule and prediction capsule v m The similarity between them is determined by the softmax function:

[0090]

[0091] Where d represents the category of the target aircraft, b nm Representation vector With v m The prior probability of the connection is initialized to 0 in the forward propagation. Then the transformed vector With parameter c nm The weighted summation is used to obtain the output vector s of the intermediate layer. m , which can be expressed as:

[0092]

[0093] In order to make the final capsule contain rich and accurate global feature representation, the vector s is squashed using a nonlinear activation function. m The modulus value is distributed between 0 and 1, and finally the prediction capsule v for classification is output m ,and Output v m The length of ||v m || represents the probability of a certain type of HRRP target existing, v m The direction of represents the different feature information in the sample component. The prediction capsule v output by the l+1 layer m It is calculated by the following formula:

[0094]

[0095] By continuously iteratively adjusting b nm To update the network weight parameter c nm . bnm The update method is expressed by the following formula:

[0096]

[0097] S5: Build a reconstruction module, take the high-level capsule representation with the largest unit length in the prediction capsule layer as the reconstruction target, and pass it through three fully connected network layers with activation functions to output the reconstructed target data;

[0098] The reconstruction module can be regarded as a decoder, which is used to restore the output results of the correct classification of HRRP samples to the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire multi-scale convolutional capsule network to assist in comparing the results of HRRP classification. Specifically, the capsule representation with the largest unit length in the prediction capsule layer is used as the reconstruction target, and after three fully connected network layers with activation functions, the reconstructed target data is output.

[0099] S6: Train the multi-scale convolutional capsule network constructed by the multi-scale convolutional module and the capsule network module using the data set processed by S2.

[0100] S6.1: The total loss of the multi-scale convolutional capsule network is given by the capsule classification loss L d and the reconstruction loss L rec It consists of two parts and loses L d Plays a leading role, the expression is as follows:

[0101] Loss 总 =L d +L rec

[0102] The capsule form enables multiple classifications to exist simultaneously, so the training process of the multi-scale convolutional capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as:

[0103] L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m-) 2

[0104] Where d represents the HRRP target category, Y d is the classification label in training, m + 、m -are fixed hyperparameters, set to 0.9 and 0.1 respectively, and λ is a coefficient to prevent the network from solving local optimal solutions, set to 0.5. When the prediction capsule matches the correct HRRP sample d during training, Y d =1, and when the probability predicted by this prediction capsule exceeds 0.9, the loss is set to zero, and when the probability is less than 0.9, the corresponding loss value is calculated for optimization; similarly, when the prediction capsule matches the wrong HRRP sample, that is, the class d does not exist in the sample, then Y d = 0. The above process actually performs a binary classification operation on each prediction capsule, and the parameter λ is introduced to ensure the stability of the network training process.

[0105] The reconstruction loss L of the decoding process in the reconstruction module rec It represents the gap between the input HRRP sample and the reconstructed data. The present invention uses the Euclidean distance between the two as the reconstruction loss and adds the loss of the entire multi-scale convolutional capsule network to train the network parameters. The expression is as follows:

[0106]

[0107] where h ic represents the initial HRRP sample, h ir represents the reconstructed data, α is the scaling factor and is set to 0.0005 so that the capsule classification loss dominates the total loss.

[0108] Therefore, the total loss function of the multi-scale convolutional capsule network can be defined as follows:

[0109]

[0110] S6.2: Initialize all weights and biases to be trained, set training parameters, including learning rate, batch_size, training batches, and train the model.

[0111] Example

[0112] Training phase:

[0113] 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 in different data segments. In the process of selecting training sets and test sets, 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 samples in the training set and the test set of each type of target is 7:3.

[0114] S2: Preprocess the samples in the data set extracted by S1. The specific steps are as follows:

[0115] S2.1: Intensity normalization. The original HRRP is represented as X raw =[x1,x2,...,x L1 ], where L1 represents the total number of distance units contained in HRRP, then the HRRP after intensity normalization can be expressed as:

[0116]

[0117] S2.2: Sample alignment. Translate HRRP so that its center of gravity g1 moves to Nearby, so that the distance units containing information in HRRP will be distributed near the center. The calculation method of HRRP center of gravity g1 is as follows:

[0118]

[0119] S3: Send the preprocessed data set to the multi-scale convolution module. The specific steps are as follows:

[0120] S3.1: The feature extraction sub-block in the multi-scale convolution module contains a total of K convolution kernels processed simultaneously, with the step size set to 2. The zero-filling layer is filled with 0 to increase sparsity. The multi-level position features of the sample output by the k1th convolution kernel of different scales can be expressed as Where D represents the number of channels and W represents the width of the feature.

[0121] S3.2: Add batch normalization after convolution, by calculating the mean and variance of the data in each mini_batch, assuming that there are N in a mini batch. m HRRP samples, then the output is defined as Among them, F n Represents the convolution output corresponding to the nth HRRP sample. In each small batch, The HRRP data in is batch normalized to obtain It is expressed as:

[0122]

[0123] Among them, F n (k,l) represents the lth element in the kth channel of the convolutional layer output corresponding to the HRRP sample before batch normalization. That is, the HRRP data after batch normalization, α k and β k is the trainable parameter corresponding to the kth channel, ε is a very small number, in order to prevent the divisor from being 0, E(.) is the mean operation, Var(.) represents the variance operation;

[0124] S3.3: Then use the activation function ReLU Each element in is nonlinearly activated to obtain If the input is Then the corresponding output after ReLU is It is expressed as:

[0125]

[0126] After batch normalization to adjust the distribution of the multi-level features F(k1) and the nonlinear mapping of the activation layer, the multi-level features F(k1) are fused into features of multiple different levels to obtain the final output feature map F m , which can be expressed as

[0127] F m =[F(1),F(2),...,F(k)]

[0128] Where [·] represents the process of each layer feature. The value of D×K represents the height of the feature map, which is equal to the width W of the feature map.

[0129] S4: Pass the data from the multi-scale convolution module to the capsule network module. The specific steps are as follows:

[0130] S4.1: After a convolution layer and a ReLU activation layer, the local hierarchical information in the position features extracted by multi-scale convolution is obtained and mapped to a higher-dimensional space. The number of channels is converted from 1 to 256 to facilitate the transmission of capsule features in vector form.

[0131] S4.2: The capsule network module consists of a primary capsule layer and a prediction capsule layer. First, the output of the previous step is passed to the primary capsule layer, and every eight convolution feature nodes in the feature map obtained by the previous step are stacked into a basic capsule feature, recorded as That is, a vector neuron in the form of 1×8 obtains 32 channels at the same time, and each channel outputs 64 (8×8) basic capsules of length 8. Therefore, the primary capsule layer contains a total of 2048 (8×8×32) basic capsules.

[0132] S4.3: The primary capsule layer is denoted as the l1th layer of the network, and the n1th capsule after the l1th layer is stacked is denoted as Where n1 = 1, 2, ..., 2048, the prediction capsule layer in the routing process is denoted as layer l1+1, and the mth (m = 0, 1, 2) prediction capsule output by this layer is denoted as v m . First, through the transformation matrix W nm Get vector It can be expressed as:

[0133]

[0134] Secondly, the coupling coefficient c can be adjusted dynamically nm From primary capsule and prediction capsule v m The higher the similarity between them, the higher the c nm The larger the value of , the expression can be calculated by the softmax function:

[0135]

[0136] Where d represents the category of the target aircraft, b nm Representation vector With v m The prior probability of the connection is initialized to 0 in the forward propagation. Then the transformed vector With parameter c nm The weighted summation is used to obtain the output vector s of the intermediate layer. m , which can be expressed as:

[0137]

[0138] In order to make the final capsule contain rich and accurate global feature representation, we use the nonlinear activation function squash to transform the vector s m The modulus value is distributed between 0 and 1, and finally the prediction capsule v for classification is output m ,and Output v m The length of ||v m || represents the probability of the existence of a certain type of HRRP target. The longer the modulus, the greater the possibility of belonging to this type of target aircraft, and vice versa. m The direction of represents the different feature information in the sample component. The prediction capsule v output by the l+1 layer m It can be calculated by the following formula:

[0139]

[0140] By continuously iteratively adjusting b nm To update the network weight parameter c nm . b nm The update method can be expressed by the following formula:

[0141]

[0142] S5: Construct the reconstruction module. The specific steps are as follows:

[0143] S5.1: The capsule representation with the largest unit length in the predicted capsule layer is taken as the reconstruction target, and after passing through three fully connected network layers with activation functions, the reconstructed target data is output.

[0144] S6: Design a multi-scale convolutional capsule network loss function and start training. The specific steps are as follows:

[0145] S6.1: The loss function is designed as cross entropy. The total loss of the multi-scale convolutional capsule network is composed of the capsule classification loss L d and the reconstruction loss L rec It consists of two parts and loses L d Plays a leading role, the expression is as follows:

[0146] Loss 总 =L d +L rec

[0147] The capsule form enables multiple classifications to exist simultaneously, so the training process of the convolutional capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as:

[0148] L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m - ) 2

[0149] Where d represents the HRRP target category, Y d is the classification label in training, m + 、m - are fixed hyperparameters, set to 0.9 and 0.1 respectively, and λ is a coefficient to prevent the network from solving local optimal solutions, set to 0.5. When the prediction capsule matches the correct HRRP sample d during training, Y d =1, and when the probability predicted by this prediction capsule exceeds 0.9, the loss is set to zero, and when the probability is less than 0.9, the corresponding loss value is calculated for optimization; similarly, when the prediction capsule matches the wrong HRRP sample, that is, the class d does not exist in the sample, then Y d = 0. The above process actually performs a binary classification operation on each prediction capsule, and the parameter λ is introduced to ensure the stability of the network training process.

[0150] The reconstruction loss L of the decoding process in the reconstruction module recIt represents the gap between the input HRRP sample and the reconstructed data. The Euclidean distance between the two is used as the reconstruction loss, and the loss of the entire multi-scale convolutional capsule network is added to train the network parameters. The expression is as follows:

[0151]

[0152] where h ic represents the initial HRRP sample, h ir represents the reconstructed data, α is the scaling factor and is set to 0.0005 so that the capsule classification loss dominates the total loss.

[0153] Therefore, the total loss function of the multi-scale convolutional capsule network can be defined as follows:

[0154]

[0155] S6.2: Initialize all weights and biases to be trained in the above model, set training parameters, including learning rate, batch_size, training batch, and start model training.

[0156] Testing phase:

[0157] S7: Perform the preprocessing operation of step S2 in the training phase on the test data collected by S1

[0158] S8: Send the sample processed by S7 to the multi-scale convolutional capsule network for testing and obtain the result. The i2th HRRP test sample The probability corresponding to the k-th radar target in the target set can be calculated as:

[0159]

[0160] Where exp(·) represents exponential operation, and c represents the number of categories.

[0161] We test the HRRP sample x by the maximum a posteriori probability test Classify into k0 with maximum target probability:

[0162]

[0163] After the above 8 steps, a radar high-resolution range profile recognition method based on multi-scale convolution and capsule network proposed in the present invention can be obtained.

Claims

1. A radar target recognition method based on a multi-scale convolutional capsule network, characterized in that: The following steps are involved: S1: Dataset collection; S2: Preprocess the samples in the dataset; S3: Input the preprocessed HRRP samples into the multi-scale convolution module for feature extraction; S4: The features obtained in S3 are passed to a convolution layer and a ReLU activation layer for processing, and then the obtained high-dimensional features are passed to the capsule network module. After the capsule features are obtained by convolution, the HRRP features stored in the basic capsule are converted into more comprehensive overall features through the dynamic routing mechanism, and passed to the prediction capsule layer representing the HRRP target category for classification; in order to avoid overfitting caused by the fully connected routing mechanism, the dropout method is introduced to discard the transmission of some invalid vector neurons, further explore the relationship between different attributes in the HRRP sample, and achieve better classification in the form of vector output; S5: Build a reconstruction module, take the high-level capsule representation with the largest unit length in the prediction capsule layer as the reconstruction target, and pass it through three fully connected network layers with activation functions to output the reconstructed target data; S6: training the multi-scale convolution capsule network constructed by the multi-scale convolution module and the capsule network module through the data set processed by S2; The detailed steps of S1 are: The HRRP data set collected by the radar is merged according to the types of targets. For each type of sample, training samples and test samples are selected in different data segments. In the process of selecting the training set and the 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 samples in the training set and the test set of each type of target is 7:

3. The detailed steps of S2 are: S2.1: Intensity normalization; the original HRRP is expressed as Where L1 represents the total number of distance units contained in HRRP, then the HRRP after intensity normalization can be expressed as: S2.2: Center of gravity alignment; translate HRRP so that its center of gravity moves g1 to Nearby, so that the distance units containing information in HRRP will be distributed near the center; the calculation method of HRRP center of gravity g1 is as follows: The detailed steps of S3 are: S3.1: Input the preprocessed HRRP sample into a multi-scale convolution module, which includes three stacked feature extraction sub-blocks and one fully connected layer, wherein each feature extraction sub-block is composed of the same network layer, including a convolution layer, a batch normalization layer, and an activation layer, and then passes through a fully connected layer for feature mapping, and the target category of the input HRRP sample is determined by the maximum value of the output of the fully connected layer; The feature extraction sub-block in the multi-scale convolution module contains a total of K convolution kernels processed synchronously, with the step size set to 2. The zero-filling layer is filled with 0 to increase sparsity. The multi-level position feature representation of the sample output by the k1th convolution kernel of different scales is: Where D represents the number of channels and W represents the width of the feature; S3.2: The data after the convolution layer needs further processing. In order to make the model easier to converge and the network training process more stable, batch normalization is added after the convolution. By calculating the mean and variance of the data in each mini_batch, assuming that there are N in a mini batch m HRRP samples, then the output is defined as where F n Represents the convolution output corresponding to the nth HRRP sample. In each small batch, The HRRP data in is batch normalized to obtain It is expressed as: Among them, F n (k,l) represents the lth element in the kth channel of the convolutional layer output corresponding to the HRRP sample before batch normalization. That is the HRRP data after batch normalization, α k and β k is the trainable parameter corresponding to the kth channel, ε is a very small number, in order to prevent the divisor from being 0, E(.) is the mean operation, Var(.) represents the variance operation; S3.3: Then use the activation function ReLU Each element in is nonlinearly activated to obtain If the input is Then the corresponding output after ReLU is It is expressed as: The multi-level features F(k1) are respectively adjusted by batch normalization distribution and nonlinear mapping of the activation layer, and then fused into features of multiple different levels to obtain the final output feature map F m , which can be expressed as: F m =[F(1),F(2),...,F(k)] Where [·] represents the features of each layer that are spliced ​​together. The value of D×K represents the height of the feature map, which is equal to the width W of the feature map; that is, the feature map extracted by the multi-scale convolution module can be regarded as a square picture, which can be used in the subsequent network module to further extract the envelope feature information of the HRRP sample; The detailed steps of S4 are: S4.1: The output of the multi-scale convolution module is processed by a convolution layer and a ReLU activation layer. While obtaining the local hierarchical information in the position features extracted by the multi-scale convolution module, it is mapped to a higher-dimensional space. The number of channels is converted from 1 to 256 to facilitate the transmission of capsule features in vector form. S4.2: The capsule network module consists of a primary capsule layer and a prediction capsule layer. First, the previous step output is passed into the primary capsule layer, and every eight convolution feature nodes in the previous step output are stacked into a basic capsule feature, recorded as That is, a vector neuron in the form of 1×8 gets 32 channels at the same time, and each channel outputs 64 (8×8) basic capsules of length 8, so the primary capsule layer contains a total of 2048 (8×8×32) basic capsules; S4.3: The primary capsule layer is denoted as the l1th layer of the capsule network module, and the n1th capsule after the l1th layer is stacked is denoted as Where n1 = 1, 2, ..., 2048, the prediction capsule layer in the routing process is denoted as layer l1+1, and the mth prediction capsule output by this layer is denoted as v m , m=0,1,2; First, through the transformation matrix W nm Get vector It can be expressed as: Secondly, the coupling coefficient c can be adjusted dynamically nm From primary capsule and prediction capsule v m The similarity between them is determined by the softmax function: Where d represents the category of the target aircraft, b nm Representation vector With v m The prior probability connected to it is initialized to 0 in the forward propagation; then the transformed vector With parameter c nm The weighted summation is used to obtain the output vector s of the intermediate layer. m , which can be expressed as: In order to make the final capsule contain rich and accurate global feature representation, the vector s is squashed using a nonlinear activation function. m The modulus value is distributed between 0 and 1, and finally the prediction capsule v for classification is output m ,and Output v m The length of ||v m || represents the probability of a certain type of HRRP target existing, v m The direction represents the different feature information in the sample component; the prediction capsule v output by the l+1 layer m It is calculated by the following formula: By iteratively adjusting b nm To update the network weight parameter c nm ; b nm The update method is expressed by the following formula: The detailed steps of S5 are: The reconstruction module can be regarded as a decoder, which is used to restore the output results of the correct classification of HRRP samples to the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire multi-scale convolutional capsule network to assist in comparing the results of HRRP classification; specifically, the capsule representation with the largest unit length in the prediction capsule layer is used as the reconstruction target, and after three fully connected network layers with activation functions, the reconstructed target data is output; The detailed steps of S6 are: S6.1: The total loss of the multi-scale convolutional capsule network is given by the capsule classification loss L d and the reconstruction loss L rec It consists of two parts and loses L d Plays a leading role, the expression is as follows: Loss 总 =L d +L rec The capsule form enables multiple classifications to exist simultaneously, so the training process of the multi-scale convolutional capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as: L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m-) 2 Where d represents the HRRP target category, Y d is the classification label in training, m + 、m - are fixed hyperparameters, set to 0.9 and 0.1 respectively, and λ is a coefficient to prevent the network from solving local optimal solutions, set to 0.5; when the prediction capsule matches the correct HRRP sample d during training, Y d =1, and when the probability predicted by this prediction capsule exceeds 0.9, the loss is set to zero, and when the probability is less than 0.9, the corresponding loss value is calculated for optimization; similarly, when the prediction capsule matches the wrong HRRP sample, that is, the class d does not exist in the sample, then Y d =0; The above process actually performs a binary classification operation on each prediction capsule, and the parameter λ is introduced to ensure the stability of the network training process; The reconstruction loss L of the decoding process in the reconstruction module rec It represents the gap between the input HRRP sample and the reconstructed data. The present invention uses the Euclidean distance between the two as the reconstruction loss and adds the loss of the entire multi-scale convolutional capsule network to train the network parameters. The expression is as follows: where h ic represents the initial HRRP sample, h ir represents the reconstructed data, α is the scaling factor and is set to 0.0005, so that the capsule classification loss dominates the total loss; Therefore, the total loss function of the multi-scale convolutional capsule network can be defined as follows: S6.2: Initialize all weights and biases to be trained, set training parameters, including learning rate, batch_size, training batches, and train the model.

Images