Unsupervised pre-training-fine-tuning radar target recognition method based on contrastive learning

Abstract

The present invention discloses a radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning, first pre-processing the original sample and performing data expansion; then using reshape and data enhancement to increase the structural diversity of HRRP samples; then inputting the SimSiam module; fine-tuning the output of the SimSiam module encoder network on the downstream classification module, and then inputting it into the downstream classification module after shape operation, and finally realizing HRRP recognition and classification; the present invention adopts data enhancement to allow the model to learn more effective feature information, which helps to improve the recognition performance of the downstream classification task. By defining different downstream classification modules, various general robust models applied to HRRP target recognition can be obtained, which provides a new idea and method for the field of radar HRRP target recognition.

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 contrastive learning and unsupervised pre-training-fine-tuning. Background Art

[0002] The range resolution of high-resolution broadband radar is much smaller than the target size, and its echo is also called the one-dimensional high-resolution range profile (HRRP) of the target. HRRP contains valuable structural information for classification and recognition, such as the radial size of the target, the distribution of scattering points, etc., and has broad prospects for engineering applications. Therefore, radar automatic target recognition methods based on HRRP have gradually become a hot topic in the field of radar automatic target recognition.

[0003] 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.

[0004] Traditional HRRP feature extraction methods are mainly divided into two parts: (1) Transformer-based feature extraction methods, such as spectrograms. 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 selection 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. Learning and extracting HRRP target features is a crucial link in the radar target recognition process. In view of the shortcomings of the above traditional methods and the characteristics of HRRP data, this paper uses an improved deep learning 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. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning.

[0006] The present invention proposes a deep learning model using contrastive learning method to learn the universal feature representation of radar HRRP data, and use it for downstream tasks, so that this unsupervised contrastive learning method enables the network to achieve deeper training, and uses the unsupervised pre-trained model to learn the effective feature representation in the HRRP sample. After that, a large number of labeled HRRP training samples are no longer needed, and it can be applied to various downstream tasks and fine-tuned, thereby obtaining a variety of different high-performance robust models suitable for HRRP target recognition, providing a new idea and general method for radar HRRP recognition. The biggest advantage of contrastive learning is that it can better extract the common features of similarity contained between HRRP samples in different angular domains, and use them for identification and classification of appropriate downstream tasks. It can still maintain a certain recognition ability in a harsh small sample target environment, providing a new general idea and method for radar target recognition.

[0007] The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning includes the following steps:

[0008] S1: Preprocess the original HRRP sample set.

[0009] The original HRRP echo is processed by the l2 intensity normalization method to improve the intensity sensitivity of HRRP. HRRP is intercepted from the radar echo data through the range window. During the interception process, the position of the recorded range image in the range gate is not fixed, which leads to the translation sensitivity of HRRP. In order to make the training and testing have a unified standard, the center of gravity alignment method is used to eliminate the translation sensitivity.

[0010] S2: Perform translation processing on the processed HRRP samples to achieve data expansion.

[0011] S3: Reshaping and data enhancement are used to increase the structural diversity of HRRP samples so that the model can learn more effective feature information.

[0012] S4: Input the HRRP samples obtained by different data enhancement methods into the SimSiam module.

[0013] The SimSiam module consists of four parts: random data augmentation, backbone network, projector, and predictor, where the backbone network and projector form the encoder network. The expanded HRRP samples are sampled through different data augmentation methods and then enter the encoder network for encoding. Feature matching is performed by maximizing the consistency between feature vectors from different views of the same HRRP sample. The projector then extracts high-quality main feature representations of the HRRP samples, which helps to obtain a universal consistent representation for comparative prediction tasks. The output features are passed into the predictor to estimate the overall expectation of the SimSiam module, that is, the network parameters are optimized through back-propagation, and some invalid feature information at high levels is filtered out, so that the feature information in the HRRP sample is fully retained for comparative prediction optimization.

[0014] S5: After unsupervised pre-training of the SimSiam module using contrastive learning on the expanded radar HRRP samples, the output of the encoder network is made to have the same general feature information as the output obtained by supervised training. The output of the encoder network is then fine-tuned on the downstream classification module and input into the downstream classification module after shape operation, finally achieving HRRP recognition and classification.

[0015] S6: The predicted capsule features of the HRRP samples that are correctly classified are reconstructed into target data through the reconstruction module, restored to the initial input, and participate in training.

[0016] Furthermore, the detailed steps of S1 are:

[0017] S1.1: 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] S1.2: Sample alignment. Translate HRRP so that its center of gravity g1 moves to Nearby, so that the distance cells containing information in HRRP will be distributed near the center. The calculation method of HRRP center of gravity g1 is as follows:

[0020]

[0021] Furthermore, the detailed steps of S2 are:

[0022] In order to avoid overfitting during the pre-training process and obtain important semantic information in the HRRP data, data expansion is performed by shifting the center of gravity of each sensitivity-processed HRRP sample to the left and right by 1 to 4 distance units respectively. This increases the amount of data available for unsupervised pre-training by 8 times based on the original training set, thereby improving the network's generalization ability for new samples to a certain extent.

[0023] Furthermore, the detailed steps of S3 are:

[0024] S3.1: First, the radar HRRP samples after data sensitivity preprocessing are still 1×256 one-dimensional vector data. In order to match the data dimension, the unsqueeze() function is used to increase the data dimension and transform the shape of the HRRP sample into a four-dimensional tensor. The HRRP sample is then reshaped, and the encoder network entering the SimSiam module is in the form of Secondly, data augmentation operations are performed on 16×16 HRRP samples, including random cropping, random Gaussian blur, horizontal flipping and scaling to increase the structural diversity of HRRP. Finally, the data augmented HRRP samples are normalized.

[0025] Furthermore, the detailed steps of S4 are:

[0026] S4.1: Random Data Augmentation:

[0027] The random data augmentation operation is to generate two different sets of views for each HRRP sample for the self-supervised contrastive learning task, that is, the views x′ and x″ used for the contrastive learning task are obtained by performing two random data augmentations on the HRRP samples respectively.

[0028] S4.2: Backbone network part:

[0029] Resnet-50 is selected as the backbone network. ResNet-50 consists of 1 convolution input layer, 4 residual convolution blocks and 1 fully connected output layer. The residual convolution block includes 16 convolution blocks, and each convolution block consists of the same number of convolution layers. The setting of the convolution input layer is to retain the semantic information and global feature information of the enhanced HRRP view adjacent features and build a reasonable spatial relationship, which is helpful to extract the effective features of HRRP. As the main component of the network, the residual block first uses the convolution layer with a 1×1 convolution kernel to convert the dimension of the feature (up or down). Then the convolution layer with a 3×3 convolution kernel is used to extract effective features. Finally, the identity mapping is introduced to accelerate the flow of correlation information between deep network layers and make full use of the features of each layer. Finally, an effective and meaningful HRRP feature representation is extracted to improve the recognition performance. Each residual convolution block uses the ReLU activation function and batch normalization unit to improve the generalization ability of the model.

[0030] S4.3: Projector part:

[0031] The core function of the projector is to filter out redundant information or irrelevant semantic information in the HRRP view feature representation, retain the most important nonlinear features and perform l2 regularization on the feature vector, and then map it to the unit hypersphere space for comparative matching prediction, maximizing the similarity structural information between the two HRRP view representations, that is, constraining the consistency between the vectors in the hypersphere space and the network training process, so as to obtain a general HRRP sample similarity feature representation. The deep projector can improve the quality of HRRP feature expression and the performance of contrastive learning. Therefore, three linear fully connected layers (FC) are constructed for the projector, and each layer is followed by a BN layer. The output dimension of the fully connected layer is 2048. The feature output by the backbone network is denoted as y i , then the output feature z of the encoder network i The calculation expression is as follows:

[0032] z i =g(y i )=W (3) σ(W (2) σ(W (1) y i ))

[0033] Where σ represents the ReLU activation function, g represents the projection function, and W represents the weight matrix of each fully connected layer. For the convenience of analysis, the influence of the bias term on the network is ignored. The existence of the BN layer in the projector readjusts the spatial distribution of the output of each layer, removes some common features between HRRP samples in each batch, and retains the features with differences for the comparative prediction task.

[0034] S4.4: Predictor part:

[0035] The predictor consists of two fully connected layers, a BN layer, and an activation layer. The input and output dimensions of the first fully connected layer are both 2048, and the output dimension of the second fully connected layer is 512. The feature vector p output by the predictor is i The calculation expression is as follows:

[0036] p i =h(z i )=W (2) σ(W (1) z i )

[0037] Where h represents the prediction function, σ represents the ReLU activation function, and W (1) , W (2) Represent the weight matrices of the first fully connected layer and the second fully connected layer respectively.

[0038] S4.5: SimSiam working mechanism:

[0039] The loss of SimSiam is defined using the EM algorithm, and the expression is as follows:

[0040]

[0041] in represents the encoder network for feature extraction, θ is a learnable parameter, x is a HRRP sample, Represents the random data enhancement function before HRPP data input, and the expectation is Represents the HRRP sample x and random data enhancement method In other words, It is equivalent to the expected sum of the losses of all HRRP samples and random data augmentation. x is the feature expression of HRRP sample x, i.e., the feature vector z output by the encoder network i . Use Mean Squared Error (MSE) to calculate similarity. At this time, SimSiam works similarly to the K-means clustering algorithm, fixing one variable to solve another variable, which is the EM iterative algorithm. Convert it into the following two sub-problems:

[0042]

[0043]

[0044] Where ← represents the assignment operation, and r represents the number of iterations of the algorithm. The Stochastic Gradient Descent (SGD) algorithm is used to calculate θ in the first subproblem. r The solution is to stop the back propagation of the gradient to η through formula (4-5-2). r-1 , then η r-1 In formula (4-5-2), it is a constant. If the back propagation of the gradient is not stopped, there are two variables in the formula, which makes it impossible to solve.

[0045] Find θ r After solving the problem, substitute it into the second sub-problem. At this time, there is only one variable η in formula (4-5-3), and the expectation of each HRRP sample x needs to be minimized. Substituting formula (4-5-1) into formula (4-5-3), the solution to the second sub-problem is converted to:

[0046]

[0047] According to the expected formula transformation, we get:

[0048]

[0049] This means that at the rth iteration update, the feature representation of a HRRP sample x is obtained by random data enhancement expectation of the sample x.

[0050] According to formula (4-5-5), perform random data enhancement on the transformed second sub-problem The formula is as follows:

[0051]

[0052] Substituting it into formula (4-5-2), we get:

[0053]

[0054] where θ r is the solution of equation (4-5-2), and Represents two different data enhancement methods acting on a certain HRRP sample, then the formula can be regarded as a twin tower architecture.

[0055] Add a predictor on one side of the SimSiam module branch and define it as h1. z1 is the feature representation of the HRRP sample. According to the expected formula, convert formula (4-5-4) into:

[0056] h1(z1)=E z [z1]=ET [f(T(x))] (4-5-8)

[0057] Since the random enhancement is directly calculated Expected value It is difficult to analyze the enhanced expectation, so the enhanced expectation is equivalent to the original expectation for the convenience of analysis.

[0058] Furthermore, the detailed steps of S5 are:

[0059] The capsule network is selected as the downstream classification module. During training, only the parameters in the capsule network are fine-tuned and updated. The weight parameters of the encoder in SimSiam are obtained and fixed by unsupervised pre-training learning. The encoder is used as a feature extraction network to extract meaningful key features from HRRP samples. The feature dimension extracted by the encoder network is 2048, that is, 1×2048 vector features. Then, the dimension is reduced through a 2048×256 fully connected layer to obtain a feature with a dimension of 256. After the shape operation, it is input into the capsule network to finally achieve HRRP recognition and classification.

[0060] Furthermore, the detailed steps of S6 are:

[0061] S6.1: The reconstruction module can be regarded as a decoder, which is used to restore the predicted capsule features of the HRRP sample correctly classified into the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire model to assist in comparing the results of HRRP classification. Specifically, the high-level capsule representation with the largest unit length in the capsule layer predicted by the capsule network is used as the reconstruction target, and after three fully connected layers with activation functions, the reconstructed target data is output.

[0062] S6.2: The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning includes a data preprocessing module, a SiamSam module, a downstream classification module, and a reconstruction module. Initialize all weights and biases to be trained in the SimSiam module and the downstream classification module, set training parameters, including learning rate, batch_size, and training batches, and train the SimSiam module and capsule network.

[0063] Furthermore, the total loss of the network constructed by the contrastive learning unsupervised pre-training-fine-tuning radar target recognition method 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:

[0064] Loss 总 =L d +L rec

[0065] The capsule form enables multiple classifications to exist simultaneously, so the training process of the capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as:

[0066] L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m - ) 2

[0067] 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.

[0068] 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 Euclidean distance between the two is used as the reconstruction loss, and the loss of the entire capsule network is added to train the network parameters. The expression is as follows:

[0069]

[0070] 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.

[0071] Therefore, the total loss function of the capsule network can be defined as follows:

[0072]

[0073] The beneficial effects of the present invention are:

[0074] Inspired by the significant progress made in learning representation by large-scale unsupervised pre-training models, the present invention proposes a method for obtaining consistent universal features contained between HRRP samples in different angular domains based on the contrastive learning SimSiam module, and applies the features to different downstream tasks for classification, thus becoming a set of universal HRRP recognition method standards. At the same time, additional preprocessing is performed on the data in the pre-training stage to increase the diversity of HRRP samples. The preprocessed model will no longer be limited to a large number of training samples, and through fine-tuning, it also has good recognition performance on a small sample size, which overcomes the small sample sensitivity of HRRP to a certain extent and has practical significance in practical applications.

[0075] 1. The present invention adopts data enhancement to increase the structural diversity of HRRP samples, providing a large number of rich and diverse training samples for unsupervised pre-training of contrastive learning, so as to obtain better pre-training effect on the large Resnet-50, allowing the model to learn more effective feature information, which is helpful to improve the recognition performance of migration to downstream classification tasks.

[0076] 2. The present invention adopts a self-supervised learning method, and uses unlabeled samples for self-supervised representation learning. The contrastive learning method fully mines the abstract semantic information in the data set, helps the model learn unlabeled data to obtain an encoder, extracts effective general features and applies them to different downstream tasks. The performance of contrastive learning also depends on the depth of the network. The deeper the depth, the stronger the representation learning ability. In addition, the longer the training time of contrastive learning, the better the performance. Applying appropriate deep networks, training methods and optimization methods to contrastive learning methods may not be weaker than end-to-end supervised learning. As a relatively new research direction at present, contrastive learning has immeasurable potential for its development in practical engineering applications in different fields in the future. From the perspective of people's goal of using scientific and technological means, unsupervised learning is the general trend of the future development of artificial intelligence. Compared with supervised learning for large-scale applications, contrastive learning methods reduce the investment of time, human resources and other costs, and the performance is not inferior to supervised learning under the premise of improving efficiency. Therefore, radar HRRP target recognition based on contrastive learning methods has high research prospects and value.

[0077] 3. The present invention uses SimSiam representation learning network, which is a representation learning network with an asymmetric twin structure with a stop-grad mechanism. A predictor is introduced in one branch of the simple twin network, and a stop-grad mechanism is introduced in the other branch to construct an asymmetric two-branch network to avoid mode collapse. This module retains the meaningful feature information in the original HRRP sample to the greatest extent and obtains a feature extractor that extracts effective consistency information of HRRP through pre-training learning, and then connects to the downstream classifier to fine-tune the parameters of the classifier part. This approach helps to better learn the similarity structure information between different HRRP samples, and fully extract and utilize it to achieve a good recognition effect.

[0078] 4. The present invention applies a downstream classification module, and after unsupervised pre-training with contrastive learning on an expanded radar HRRP data set, the output from the encoder and the output obtained by supervised training have the same general feature information, and then fine-tunes the downstream classification module using these features. By defining different downstream classification modules, we can obtain various general robust models for HRRP target recognition, providing a new idea and method for the field of radar HRRP target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 : Flowchart of the steps of radar target recognition method based on unsupervised pre-training-fine-tuning paradigm of contrastive learning.

[0080] Figure 2 :SimSiam module workflow diagram. DETAILED DESCRIPTION

[0081] Reference Figure 1 , is a flow chart of a radar high-resolution range profile recognition technology based on unsupervised pre-training-fine-tuning paradigm of contrastive learning of the present invention, and the specific implementation steps are as follows:

[0082] S1: Preprocess the original HRRP sample set.

[0083] Since the intensity of HRRP is jointly determined by factors such as radar transmission power, target distance, radar antenna gain, and radar receiver gain, before using HRRP for target recognition, we process the original HRRP echo through the l2 intensity normalization method to improve the intensity sensitivity of HRRP. HRRP is intercepted from the radar echo data through the range window. During the interception process, the position of the recorded range image in the range gate is not fixed, which leads to the translation sensitivity of HRRP. In order to make training and testing have a unified standard, the center of gravity alignment method is used to eliminate the translation sensitivity.

[0084] S1.1: 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:

[0085]

[0086] S1.2: Sample alignment. Translate HRRP so that its center of gravity g1 moves to Nearby, so that the distance cells containing information in HRRP will be distributed near the center. The calculation method of HRRP center of gravity g1 is as follows:

[0087]

[0088] 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.

[0089] S2: The processed HRRP samples are translated to achieve data expansion, providing a large number of rich and diverse training samples for unsupervised pre-training of contrastive learning, thereby improving the network's generalization ability for new samples to a certain extent.

[0090] In order to avoid overfitting during the pre-training process and obtain important semantic information in the HRRP data, data expansion is performed by shifting the center of gravity of each sensitivity-processed HRRP sample to the left and right by 1 to 4 distance units respectively. This increases the amount of data available for unsupervised pre-training by 8 times based on the original training set, thereby improving the network's generalization ability for new samples to a certain extent.

[0091] S3: Reshaping and data enhancement are used to increase the structural diversity of HRRP samples, so that the model can learn more effective feature information, which helps to improve the recognition performance of downstream classification tasks.

[0092] S3.1: First, the radar HRRP samples after data sensitivity preprocessing are still 1×256 one-dimensional vector data. In order to match the data dimension, the unsqueeze() function is used to increase the data dimension and transform the shape of the HRRP sample into a four-dimensional tensor. The HRRP sample is then reshaped, and the encoder network entering the SimSiam module is in the form of Secondly, data augmentation operations are performed on 16×16 HRRP samples, including random cropping, random Gaussian blur, horizontal flipping and scaling to increase the structural diversity of HRRP. Finally, the data-augmented HRRP samples are normalized to remove some common features by subtracting the mean from each HRRP sample, highlighting the differences in features between different samples, which helps the model learn and identify the main features in HRRP samples, reduces the gradient during back propagation, speeds up the convergence of contrastive learning pre-training and the learning speed of the network, and increases generalization ability.

[0093] S4: Input the HRRP samples obtained by different data enhancement methods into the SimSiam module.

[0094] The SimSiam module consists of four parts: random data augmentation, backbone network, projector, and predictor, where the backbone network and projector form the encoder network. The expanded HRRP samples are sampled through different data augmentation methods and then enter the encoder network for encoding. Feature matching is performed by maximizing the consistency between feature vectors from different views of the same HRRP sample. The projector then extracts high-quality main feature representations of the HRRP samples, which helps to obtain a universal consistent representation for comparative prediction tasks. The output features are passed into the predictor to estimate the overall expectation of the SimSiam module, that is, the network parameters are optimized through back-propagation, and some invalid feature information at high levels is filtered out, so that the feature information in the HRRP sample is fully retained for comparative prediction optimization.

[0095] S4.1: Random Data Augmentation:

[0096] The random data augmentation operation is to generate two different sets of views for each HRRP sample for the self-supervised contrastive learning task, that is, the views x′ and x″ used for the contrastive learning task are obtained by respectively undergoing two random data augmentations on the HRRP samples. Unlike supervised learning, which obtains output from labeled input, self-supervised learning performs self-supervision in a reasonable way when the data is unlabeled. First, the training sample x is divided into two parts, x′ and x″. x′ is input into the model to obtain the output y, and then y and x″ are compared. The smaller the distance between the two in a certain feature space, the better. Part of the data x′ is used as input, and the other part x″ is used as a supervision signal. This is self-supervised learning.

[0097] S4.2: Backbone network part:

[0098] The present invention compares the influence of the parameter amount, calculation amount and network structure in different depth networks on the feature representation quality and recognition performance of HRRP data, and finally selects Resnet-50 with strong representation ability, high cost performance and three-dimensional network structure as the backbone network. The network not only shows good recognition performance on HRRP samples, but also has strong real-time performance, which is very suitable for the recognition task of HRRP targets in actual radar detection environment. ResNet-50 consists of 1 convolution input layer, 4 residual convolution blocks and 1 fully connected output layer, wherein the residual convolution block includes 16 convolution blocks, and each convolution block is composed of the same number of convolution layers. The setting of the convolution input layer is to retain the semantic information and global feature information of the adjacent features of the enhanced HRRP view, and to construct a reasonable spatial relationship, which is helpful to extract the effective features of HRRP. As the main component of the network, the residual block first uses the convolution layer of 1×1 convolution kernel to convert the dimension of the feature (upgrading or reducing the dimension). The number of convolution kernels affects the output dimension of each layer, and at the same time reduces the number of output parameters, which plays a role in reducing the overall network parameter amount. Then, a convolution layer with a 3×3 convolution kernel is used to extract effective features. Finally, the identity mapping is introduced to speed up the flow of related information between deep network layers and make full use of the features of each layer. Finally, effective and meaningful HRRP feature representation is extracted to improve recognition performance. Each residual convolution block uses the ReLU activation function and batch normalization unit to improve the generalization ability of the model.

[0099] S4.3: Projector part:

[0100] The core function of the projector is to filter out redundant information or irrelevant semantic information in the HRRP view feature representation, retain the most important nonlinear features and perform l2 regularization on the feature vector, and then map it to the unit hypersphere space for comparative matching prediction, maximizing the similarity structural information between the two HRRP view representations, that is, constraining the consistency between the vectors in the hypersphere space and the network training process, so as to obtain a general HRRP sample similarity feature representation. The deep projector can improve the quality of HRRP feature expression and the performance of contrastive learning. Therefore, three linear fully connected layers (FC) are constructed for the projector, and each layer is followed by a BN layer. The output dimension of the fully connected layer is 2048. The feature output by the backbone network is denoted as y i , then the output feature z of the encoder network i The calculation expression is as follows:

[0101] z i =g(y i )=W (3) σ(W (2) σ(W (1)y i ))

[0102] Where σ represents the ReLU activation function, g represents the projection function, and W represents the weight matrix of each fully connected layer. For the convenience of analysis, the influence of the bias term on the network is ignored. The existence of the BN layer in the projector readjusts the spatial distribution of the output of each layer, removes some common features between HRRP samples in each batch, and retains the features with differences for the comparison prediction task, which also makes an implicit comparison to a certain extent.

[0103] S4.4: Predictor part:

[0104] The predictor consists of two fully connected layers, a BN layer, and an activation layer. The input and output dimensions of the first fully connected layer are both 2048, and the output dimension of the second fully connected layer is 512. The feature vector p output by the predictor is i The calculation expression is as follows:

[0105] p i =h(z i )=W (2) σ(W (1) z i )

[0106] Where h represents the prediction function, σ represents the ReLU activation function, and W (1) , W (2) Represent the weight matrices of the first fully connected layer and the second fully connected layer respectively.

[0107] S4.5: SimSiam working mechanism:

[0108] The loss of SimSiam is defined using the EM algorithm, and the expression is as follows:

[0109]

[0110] in represents the encoder network for feature extraction, θ is a learnable parameter, x is a HRRP sample, Represents the random data enhancement function before HRPP data input, and the expectation is Represents the HRRP sample x and random data enhancement method In other words, It is equivalent to the expected sum of the losses of all HRRP samples and random data augmentation. x is the feature expression of HRRP sample x, i.e., the feature vector z output by the encoder network iThe mean square error (MSE) is used to calculate the similarity. This form of optimization is very similar to the EM algorithm or K-means algorithm. The variable θ is a learnable parameter in the encoder and can be regarded as a cluster center. The variable η x This is similar to the corresponding vector of HRRP sample x (such as the one-hot vector of K-means), which is the characteristic expression of HRRP sample. At this time, SimSiam works in a similar way to the K-means clustering algorithm, fixing one variable to solve another variable, which is the EM iterative algorithm. Convert it into the following two sub-problems:

[0111]

[0112]

[0113] Where ← represents the assignment operation, and r represents the number of iterations of the algorithm. The Stochastic Gradient Descent (SGD) algorithm is used to calculate θ in the first subproblem. r The solution is to stop the back propagation of the gradient to η through formula (4-5-2). r-1 , then η r-1 In formula (4-5-2), it is a constant. If the back propagation of the gradient is not stopped, there are two variables in the formula, which makes it impossible to solve. Therefore, the stop-grad operation in SimSiam is reasonably explained.

[0114] Find θ r After solving the problem, substitute it into the second sub-problem. At this time, there is only one variable η in formula (4-5-3), and the expectation of each HRRP sample x needs to be minimized. Substituting formula (4-5-1) into formula (4-5-3), the solution to the second sub-problem is converted to:

[0115]

[0116] According to the expected formula transformation, we get:

[0117]

[0118] This means that at the rth iteration update, the feature representation of a HRRP sample x is obtained by random data enhancement expectation of the sample x.

[0119] The process of performing an SGD optimization solution on these two sub-problems can be approximated to the working process of SimSiam. According to formula (4-5-5), a random data enhancement is performed on the transformed second sub-problem The formula is as follows:

[0120]

[0121] Substituting it into formula (4-5-2), we get:

[0122]

[0123] where θ r is the solution of equation (4-5-2), and Represents two different data enhancement methods acting on a certain HRRP sample, then the formula can be regarded as a twin tower architecture.

[0124] We use SGD optimization once in one iteration to reduce the loss. This calculation process is similar to the optimization process of SimSiam with the twin tower architecture that introduces the stop-grad mechanism. In addition, by increasing the number of SGD in iterative optimization, the generalization ability and performance of SimSiam can be improved to a certain extent, while this article only uses SGD once.

[0125] Add a predictor on one side of the SimSiam module branch and define it as h1. z1 is the feature representation of the HRRP sample. According to the expected formula, convert formula (4-5-4) into:

[0126] h1(z1)=E z [z1]=E T [f(T(x))] (4-5-8)

[0127] Since the random enhancement is directly calculated Expected value It is relatively difficult. For the convenience of analysis, the enhanced expectation is equivalent to its own expectation. At this time, the existence of the predictor can compensate for the gap between the feature representation z1 and the expectation that is difficult to directly associate, because the sampling using the random enhancement method in several different epochs conforms to a stable uniform distribution. Such a distribution is easier for the network to learn and remember, thereby predicting the expected value by representing z1.

[0128] S5: After unsupervised pre-training of the SimSiam module using contrastive learning on the expanded radar HRRP samples, the output of the encoder network is made to have the same general feature information as the output obtained by supervised training. The output of the encoder network is then fine-tuned on the downstream classification module and input into the downstream classification module after shape operation, finally achieving HRRP recognition and classification.

[0129] The capsule network is selected as the downstream classification module. During training, only the parameters in the capsule network are fine-tuned and updated. The weight parameters of the encoder network in SimSiam are obtained and fixed by unsupervised pre-training learning. The encoder network, as a feature extraction network, extracts meaningful key features from HRRP samples. This universal feature expression can be well migrated to different downstream tasks and has extremely strong performance. The feature dimension extracted by the encoder network is 2048, that is, 1×2048 vector features. Then, the dimension is reduced through a 2048×256 fully connected layer to obtain a feature with a dimension of 256. After the shape operation, it is input into the capsule network to finally achieve HRRP recognition and classification.

[0130] S6: The predicted capsule features of the HRRP samples that are correctly classified are reconstructed into target data through the reconstruction module, restored to the initial input, and participate in training.

[0131] S6.1: The reconstruction module can be regarded as a decoder, which is used to restore the predicted capsule features of the HRRP sample correctly classified into the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire model to assist in comparing the results of HRRP classification. Specifically, the high-level capsule representation with the largest unit length in the capsule layer predicted by the capsule network is used as the reconstruction target, and after three fully connected layers with activation functions, the reconstructed target data is output.

[0132] S6.2: The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning includes a data preprocessing module, a SiamSam module, a downstream classification module, and a reconstruction module. Initialize all weights and biases to be trained in the SimSiam module and the downstream classification module, set training parameters, including learning rate, batch_size, and training batches, and train the SimSiam module and capsule network.

[0133] Example

[0134] Training phase:

[0135] S1: Collect data sets, merge HRRP samples collected by radar according to target types, select training samples and test samples for each type of sample in different data segments, and ensure that the postures formed by the selected training set samples and radar cover the postures formed by the test set samples and radar during the selection of training and test sets. The ratio of the number of samples in the training set and test set for each type of target is 7:3.

[0136] S2: Preprocess the samples in the data set extracted by S1. The specific steps are as follows:

[0137] 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:

[0138]

[0139] 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 are distributed near the center. The calculation method of HRRP center of gravity is as follows:

[0140]

[0141] S3: Reshape and enhance the data of the samples. The specific steps are as follows:

[0142] S3.1: After data sensitivity preprocessing, the radar HRRP samples are still 1×256 one-dimensional vector data. In order to match the data dimension, we use the unsqueeze() function to increase the data dimension at the beginning of the experiment. At this time, the shape of the HRRP data is transformed into a four-dimensional tensor. Then reshape the HRRP sample, and the encoder network entering the contrastive learning module is in the form of

[0143] S3.2: Secondly, we perform data augmentation operations on the 16×16 HRRP training samples, including random cropping, random Gaussian blur, horizontal flipping and scaling to increase the structural diversity of HRRP. Finally, the cropped and scaled HRRP samples are normalized by subtracting the mean from each HRRP sample to remove some common features and highlight the differences in features between different samples.

[0144] S4: Pass the processed samples to the SimSiam module. The specific steps are as follows:

[0145] S4.1: The HRRP samples obtained by different data enhancement methods are input into the SimSiam module consisting of random data enhancement, backbone network, projector and predictor. After random data enhancement, they are encoded into the encoder, and feature matching is performed by maximizing the consistency between feature vectors from different views of the same HRRP sample. Then, high-quality main feature representations of the HRRP samples are extracted by entering the projector. The output features are passed into the predictor for back-propagation to optimize the network parameters.

[0146] S5: The output features are passed to the downstream classification module after shape operation. The specific steps are as follows:

[0147] By using contrastive learning for unsupervised pre-training on the expanded radar HRRP samples, the output from the encoder and the output obtained by supervised training have the same common feature information. These features are then used to fine-tune the downstream classification module and input into the capsule network after shape operation to finally achieve HRRP recognition and classification.

[0148] S6: Build the reconstruction module, design the loss function, and start training. The specific steps are as follows:

[0149] S6.1: The high-level capsule representation with the largest unit length in the capsule layer predicted by the capsule network is taken as the reconstruction target, and after passing through three fully connected layers with activation functions, the reconstructed target data is output.

[0150] S6.2: Loss function design. The total loss of the network constructed by the contrastive learning unsupervised pre-training-fine-tuning radar target recognition method 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:

[0151] Loss 总 =L d +L rec

[0152] The capsule form enables multiple classifications to exist simultaneously, so the training process of the capsule network uses MarginLoss as the loss function L d , which is obtained by adding the losses of each prediction capsule, is expressed as:

[0153] L d =Y d max(0,m + -||v d ||) 2 +λ(1-Y d )max(0,||v d ||-m - ) 2

[0154] 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.

[0155] 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 Euclidean distance between the two is used as the reconstruction loss, and the loss of the entire capsule network is added to train the network parameters. The expression is as follows:

[0156]

[0157] 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.

[0158] Therefore, the total loss function of the capsule network can be defined as follows:

[0159]

[0160] S6.3: Initialize the SimSiam module and all weights and biases to be trained in the capsule network, set the training parameters, including learning rate, batch_size, training batch, and start model training.

[0161] Testing phase:

[0162] S7: Perform preprocessing operations of steps S2 and S3 of the training phase on the test data collected by S1;

[0163] S8: Send the sample processed by S7 to the trained SimSiam module and capsule network for testing to obtain the result. That is, the output of SimSiam will be classified by the downstream classification module. HRRP test sample x test The probability corresponding to the k-th radar target in the target set can be calculated as:

[0164]

[0165] Where exp(·) represents exponential operation, and c represents the number of categories.

[0166] The test HRRP sample x is tested by the maximum a posteriori probability test Classify into k0 with maximum target probability:

[0167]

[0168] After the above 8 steps, the radar target recognition method based on contrastive learning and unsupervised pre-training-fine-tuning proposed by the present invention can be obtained.

Claims

1. A radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning, characterized in that: The following steps are involved: S1: Preprocess the original HRRP sample set; The original HRRP echo is processed by the l2 intensity normalization method to improve the intensity sensitivity of HRRP. HRRP is intercepted from the radar echo data through the range window. During the interception process, the position of the acquired range image in the range gate is not fixed, which leads to the translation sensitivity of HRRP. In order to make the training and testing have a unified standard, the center of gravity alignment method is used to eliminate the translation sensitivity. S2: Perform translation processing on the processed HRRP samples to achieve data expansion; S3: Use reshape and data enhancement to increase the structural diversity of HRRP samples so that the model can learn more effective feature information; S4: Input the HRRP samples obtained by different data enhancement methods into the SimSiam module; The SimSiam module includes four parts: random data enhancement, backbone network, projector and predictor, among which the backbone network and projector form the encoder network; the expanded HRRP samples are sampled through different data enhancement methods, and then enter the encoder network for encoding, and feature matching is performed by maximizing the consistency between feature vectors from different views of the same HRRP sample. The projector is then used to extract the high-quality main feature representation of the HRRP sample, which helps to obtain a universal consistent representation for comparative prediction tasks; the output features are passed into the predictor to estimate the overall expectation of the SimSiam module, that is, the network parameters are optimized through back propagation, and some invalid feature information at the high level is filtered out, so that the feature information in the HRRP sample is fully retained for comparative prediction optimization; S5: After unsupervised pre-training of the SimSiam module using contrastive learning on the expanded radar HRRP samples, the output of the encoder network and the output obtained by supervised training have the same general feature information, and then the output of the encoder network is fine-tuned on the downstream classification module, and then input into the downstream classification module after shape operation, finally realizing HRRP recognition and classification; S6: The predicted capsule features of the HRRP samples that are correctly classified are reconstructed into target data through the reconstruction module, restored to the initial input, and participate in training; The detailed steps of S3 are: S3.1: First, the radar HRRP samples after data sensitivity preprocessing are still 1×256 one-dimensional vector data. In order to match the data dimension, the unsqueeze() function is used to increase the data dimension and transform the shape of the HRRP sample into a four-dimensional tensor. The HRRP sample is then reshaped, and the encoder network entering the SimSiam module is in the form of Secondly, for 16 × The 16 forms of HRRP samples are subjected to data augmentation operations, including random cropping, random Gaussian blur, horizontal flipping and scaling to increase the structural diversity of HRRP; finally, the data-augmented HRRP samples are normalized; The detailed steps of S4 are: S4.1: Random Data Augmentation: The random data augmentation operation is to generate two different sets of views ″′ of each HRRP sample for the self-supervised contrastive learning task, that is, the views x and x used for the contrastive learning task are obtained by performing two random data augmentations on the HRRP samples respectively; S4.2: Backbone network part: Resnet-50 is selected as the backbone network. ResNet-50 consists of 1 convolution input layer, 4 residual convolution blocks and 1 fully connected output layer. The residual convolution block includes 16 convolution blocks, and each convolution block is composed of the same number of convolution layers. The convolution input layer is set to retain the semantic information and global feature information of the enhanced HRRP view adjacent features and build a reasonable spatial relationship, which is helpful to extract the effective features of HRRP. As the main component of the network, the residual block first uses the convolution layer with a 1×1 convolution kernel to transform the dimension of the feature. Then, a convolution layer with a 3×3 convolution kernel is used to extract effective features. Finally, an identity mapping is introduced to speed up the flow of correlation information between deep network layers and make full use of the features of each layer. Finally, an effective and meaningful HRRP feature representation is extracted to improve recognition performance. Each residual convolution block uses the ReLU activation function and batch normalization unit to improve the generalization ability of the model. S4.3: Projector part: The core function of the projector is to filter out redundant information or irrelevant semantic information in the HRRP view feature representation, retain the most important nonlinear features and perform l2 regularization on the feature vector, and then map it to the unit hypersphere space for comparison and matching prediction, maximizing the similarity structure information between the two HRRP view representations, that is, constraining the consistency between the vectors in the hypersphere space and the network training process, so as to obtain a universal HRRP sample similarity feature representation; The deep projector can improve the quality of HRRP feature expression and the performance of contrastive learning. Therefore, three linear fully connected layers are constructed for the projector, each of which is followed by a BN layer. The output dimension of the fully connected layer is 2048. The feature output by the backbone network is recorded as y i , then the output feature z of the encoder network i The calculation expression is as follows: With i =g(y i )=In (3) σ(W (2) σ(W (1) and i )) Where σ represents the ReLU activation function, g represents the projection function, and W represents the weight matrix of each fully connected layer. For the convenience of analysis, the influence of the bias term on the network is ignored. The existence of the BN layer in the projector readjusts the spatial distribution of the output of each layer, removes some common features between HRRP samples in each batch, and retains the features with differences for the task of comparative prediction. S4.4: Predictor part: The predictor consists of two fully connected layers, a BN layer, and an activation layer. The input and output dimensions of the first fully connected layer are both 2048, and the output dimension of the second fully connected layer is 512. The feature vector p output by the predictor is i The calculation expression is as follows: p i =h(z i )=In (2) σ(W (1) With i ) Where h represents the prediction function, σ represents the ReLU activation function, and W (1) , W (2) Represent the weight matrices of the first fully connected layer and the second fully connected layer respectively; S4.5: SimSiam working mechanism: The loss of SimSiam is defined using the EM algorithm, and the expression is as follows: in represents the encoder network for feature extraction, θ is a learnable parameter, x is a HRRP sample, Represents the random data enhancement function before HRPP data input, and the expectation is Represents the HRRP sample x and random data enhancement method In other words, Equivalent to the expected sum of the losses of all HRRP samples and random data augmentation; η x is the feature expression of HRRP sample x, i.e., the feature vector z output by the encoder network i ; Use mean square error MSE to calculate similarity; At this time, SimSiam works similarly to the K-means clustering algorithm, fixing one variable to solve another variable, which is the EM iterative algorithm; Convert it into the following two sub-problems: Where ← represents the assignment operation, r represents the number of algorithm iterations; the stochastic gradient descent SGD algorithm is used to calculate θ in the first subproblem r The solution is to stop the back propagation of the gradient to η through formula (4-5-2). r-1 , then η r-1 In formula (4-5-2), it is a constant. If the back propagation of the gradient is not stopped, there are two variables in the formula, which makes it impossible to solve; Find θ r After solving the problem, substitute it into the second sub-problem. At this time, there is only one variable η in formula (4-5-3), and the expectation of each HRRP sample x needs to be minimized. Substituting formula (4-5-1) into formula (4-5-3), the solution to the second sub-problem is converted to: According to the expected formula transformation, we get: This means that at the rth iteration update, the feature representation of a HRRP sample x is obtained by random data enhancement expectation of sample x; According to formula (4-5-5), perform random data enhancement on the transformed second sub-problem The formula is as follows: Substituting it into formula (4-5-2), we get: where θ r is the solution of equation (4-5-2), and Represents two different data enhancement methods acting on a HRRP sample, then the formula can be regarded as a twin tower architecture; Add a predictor on one side of the SimSiam module branch and define it as h1. z1 is the feature representation of the HRRP sample. According to the expected formula, convert formula (4-5-4) into: h1(z1)=E z [z1]=E T [f(T(x))](4-5-8) Since the random enhancement is directly calculated Expected value It is difficult, so for the convenience of analysis, the enhanced expectation is equivalent to its own expectation; The detailed steps of S5 are: The capsule network is selected as the downstream classification module. During training, only the parameters in the capsule network are fine-tuned and updated. The weight parameters of the encoder in SimSiam are obtained and fixed by unsupervised pre-training learning. The encoder is used as a feature extraction network to extract meaningful key features from HRRP samples. The feature dimension extracted by the encoder network is 2048, that is, 1×2048 vector features. Then, the dimension is reduced through a 2048×256 fully connected layer to obtain features with a dimension of 256. After the shape operation, it is input into the capsule network to finally realize HRRP recognition and classification. The detailed steps of S6 are: S6.1: The reconstruction module can be regarded as a decoder, which is used to restore the predicted capsule features of the HRRP sample correctly classified into the initial input, that is, the output dimension is 256 vectors, and participate in the training of the entire model to assist in comparing the results of HRRP classification; specifically, the high-level capsule representation with the largest unit length in the capsule layer predicted by the capsule network is used as the reconstruction target, and after three fully connected layers with activation functions, the reconstructed target data is output; S6.2: The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning includes a data preprocessing module, a SiamSam module, a downstream classification module, and a reconstruction module; initialize all weights and biases to be trained in the SimSiam module and the downstream classification module, set training parameters, including learning rate, batch_size, and training batches, and train the SimSiam module and the capsule network; The total loss of the network constructed by the contrastive learning unsupervised pre-training-fine-tuning radar target recognition method 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: Loss 总 =L d +L rec The capsule form enables multiple classifications to exist simultaneously, so the training process of the capsule network uses Margin Loss 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 Euclidean distance between the two is used as the reconstruction loss, and the loss of the entire capsule network is added 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 capsule network can be defined as follows:

2. The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning according to claim 1 is characterized in that: The detailed steps of S1 are: S1.1: Intensity normalization; denote the original HRRP 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: S1.2: Sample alignment; translate HRRP so that its center of gravity g1 moves to Nearby, so that the distance cells containing information in HRRP will be distributed near the center; the calculation method of HRRP center of gravity g1 is as follows:

3. The radar target recognition method based on contrastive learning unsupervised pre-training-fine-tuning according to claim 2 is characterized in that: The detailed steps of S2 are: In order to avoid overfitting during the pre-training process and obtain important semantic information in the HRRP data, data expansion is performed by shifting the center of gravity of each sensitivity-processed HRRP sample to the left and right by 1 to 4 distance units respectively. This increases the amount of data available for unsupervised pre-training by 8 times based on the original training set, thereby improving the network's generalization ability for new samples to a certain extent.

Images