Simulation data assisted radar HRRP small sample target identification method

Abstract

The invention relates to a simulation data-assisted radar HRRP small sample target recognition method, which comprises the steps of inputting to-be-detected radar HRRP data into a trained radar HRRP small sample target recognition model to execute feature extraction and distance measurement, and generating a classification result of the to-be-detected radar HRRP data; wherein the trained radar HRRP small sample target recognition model generates a plurality of first intermediate domain features by fusing M pieces of simulation data including target full attitude information and N pieces of actually measured HRRP data including part of target attitude information in iterative training, and the M pieces of simulation data, the N pieces of actually measured HRRP data and the plurality of first intermediate domain features are utilized to recognize the target full attitude information. And respectively calculating to obtain classification loss, domain-level alignment loss and class-level alignment loss in a manner of jointly optimizing an overall objective function formed by a plurality of losses. The method can solve the problem that the performance of a traditional radar target recognition method is reduced under the condition of insufficient training data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a simulation data assisted radar HRRP small sample target recognition method. BACKGROUND

[0002] Radar high-resolution range profile (HRRP) target recognition technology realizes the discrimination of target attributes by analyzing the structural characteristics in the target backscattering echo. This technology has important application value in the fields of security monitoring, environmental perception, etc. due to the advantages of convenient data acquisition, high processing efficiency and good real-time performance. Compared with synthetic aperture radar (SAR), HRRP has lower hardware requirements and faster processing speed, and thus becomes a research hotspot in the field of radar target recognition.

[0003] Current HRRP-based target recognition methods mainly include two categories: traditional machine learning-based methods and deep learning-based methods. Traditional methods rely on manually designed features and prior knowledge, and use classifiers to discriminate after extracting target characteristics through feature engineering; deep learning methods automatically learn the mapping relationship from raw data to class labels through end-to-end network structure, realizing the joint optimization of feature extraction and classification. Both of these two methods have achieved good recognition results in laboratory environment.

[0004] However, the existing technology faces significant challenges in practical application: first, most methods assume that the training data has complete attitude coverage and sufficient sample quantity, which is often difficult to meet in actual complex scenarios; second, traditional methods rely on manual feature design, and the discriminative ability of features decreases significantly when the target attitude changes; in addition, deep learning methods are prone to overfitting when the training sample is insufficient, which significantly reduces the model generalization ability. Especially in the case of high cost of obtaining a large number of labeled samples in practical application, how to utilize limited samples to achieve robust recognition becomes a key problem to be solved. SUMMARY

[0005] In order to solve the above problems existing in the prior art, the application provides a simulation data assisted radar HRRP small sample target recognition method. The technical problem to be solved by the application is realized by the following technical scheme: The application provides a simulation data assisted radar HRRP small sample target recognition method, which comprises: acquiring radar HRRP data to be detected; inputting the radar HRRP data to be detected into the trained radar HRRP small sample target recognition model to perform feature extraction, obtaining target features to be detected, filtering out the minimum Euclidean distance between the target features to be detected and the prototypes corresponding to each target class, and generating a classification result of the radar HRRP data to be detected, wherein the classification result includes classifying the radar HRRP data to be detected into a target class corresponding to the minimum Euclidean distance; The trained radar HRRP small sample target recognition model is obtained by iteratively training, fusing M simulation data containing target full attitude information and N measured HRRP data containing partial target attitude information to generate a plurality of first intermediate domain features, M is greater than N, and using the M simulation data, the N measured HRRP data and the plurality of first intermediate domain features to respectively calculate a classification loss, a domain-level alignment loss and a class-level alignment loss, and a total target function composed of the classification loss, the domain-level alignment loss and the class-level alignment loss is jointly optimized.

[0006] Compared with the prior art, the beneficial effects of the present application are: (1) Significantly improve the target recognition performance under small sample conditions, effectively utilize simulation data to make up for the deficiency of measured data, and overcome the problem of sharp decline in recognition performance of traditional methods when the training sample is scarce.

[0007] (2) Enhance the adaptability and generalization ability of the model to attitude changes, and solve the problem of limited recognition effect of traditional methods under incomplete target attitude by domain-invariant feature learning and intermediate domain feature generation.

[0008] (3) Realize effective transfer of cross-domain knowledge, establish feature alignment between simulation domain and measured domain by jointly optimizing classification loss, domain-level alignment loss and class-level alignment loss, and break through the technical bottleneck of difficult knowledge transfer between different domains of traditional methods.

[0009] (4) Improve the discriminability and robustness of feature representation, learn more discriminative feature representation through prototype contrast learning and multi-domain feature alignment, and overcome the defect of insufficient feature discriminability of traditional methods in complex scenes.

[0010] (5) Reduce the dependence on a large amount of labeled measured data in actual application, significantly reduce the demand for measured samples while ensuring recognition performance through a simulation data assisted training mechanism, and solve the practical problem of high labeling cost in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is an application scenario example diagram of the simulation data assisted radar HRRP small sample target recognition method provided by the embodiments of the present application. Figure 2 is a model training example diagram of a radar HRRP small sample target recognition model provided by an embodiment of the present application. Figure 3 is a structural schematic diagram of a feature extraction module provided by an embodiment of the present application. Figure 4 is a schematic diagram of generating intermediate domain features provided by an embodiment of the present application. Figure 5 is a confusion matrix diagram generated by different methods in the case of 1 / 2 angle domain missing and 10 measured training samples per class provided by an embodiment of the present application. Figure 6 is a confusion matrix diagram generated by the ben method in the case of 1 / 2 angle domain missing and different numbers of measured training samples provided by an embodiment of the present application. Figure 7 is a t-SNE visualization diagram of features extracted by different methods in the method provided by an embodiment of the present application. Figure 5 Figure 8 is a t-SNE visualization diagram of features extracted by the method in the method provided by an embodiment of the present application. Figure 5 is a t-SNE visualization diagram of features extracted by the method in the method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0012] The present application will be further described in detail below in combination with specific embodiments, but the embodiments of the present application are not limited thereto.

[0013] Now, in combination with the drawings, a radar HRRP small sample target recognition method assisted by simulation data proposed by the present application will be described in detail.

[0014] Figure 1 is an application scenario example diagram of the radar HRRP small sample target recognition method assisted by simulation data provided by an embodiment of the present application. As shown in Figure 1 ​As shown, the method comprises: obtaining radar HRRP data to be detected; inputting the radar HRRP data to be detected into a trained radar HRRP small sample target recognition model to perform feature extraction, obtaining target features to be detected, generating a classification result of the radar HRRP data to be detected by calculating the Euclidean distance between the target features to be detected and prototypes corresponding to each target class, and the classification result comprises classifying the radar HRRP data to be detected into a target class corresponding to the minimum Euclidean distance; wherein the trained radar HRRP small sample target recognition model is obtained by iteratively training, generating a plurality of first intermediate domain features by fusing M simulation data containing target full attitude information and N measured HRRP data containing partial target attitude information, M is greater than N, using the M simulation data, the N measured HRRP data and the plurality of first intermediate domain features to respectively calculate a classification loss, a domain-level alignment loss and a class-level alignment loss, and obtaining a total target function composed of the classification loss, the domain-level alignment loss and the class-level alignment loss in a joint optimization manner.

[0015] The process of training the radar HRRP small sample target recognition model will be described. Figure 2 is a model training example diagram of the radar HRRP small sample target recognition model provided by the embodiment of the present application. As shown in the figure, Figure 2 The radar HRRP small sample target recognition model comprises: a feature extraction module, a domain aligner, an intermediate domain data generation module, a domain discriminator, a prototype network and a classification module. Here, a multi-target scene is constructed in a special radar test field, containing different types of ground vehicle targets, and an X-band pulse Doppler radar system is used to collect data in this multi-target environment to obtain a measured data set; and an electromagnetic calculation software (such as FEKO, CST) is used to establish a target electromagnetic model to generate HRRP data covering full attitude angles (i.e. a simulation data set).

[0016] Exemplarily, a simulation data set of a completely labeled source domain is defined , which contains annotated samples. In addition, a measured data set of a completely labeled target domain is defined , which contains annotated samples. The source domain and the target domain data both contain classes, but the data distribution is different.

[0017] After obtaining the measured data set and the simulation data set, iterative training is started. It should be noted that in the training process, the specific structure and connection mode of the model do not change, only the corresponding model parameters change. Specifically: In the i-th training process, the radar HRRP small sample target recognition model obtained from the (i-1)-th training is used as the i-th radar HRRP small sample target recognition model. That is, the i-th radar HRRP small sample target recognition model includes: the i-th feature extraction module, the i-th domain aligner, the i-th intermediate domain data generation module, the i-th domain discriminator, and the i-th prototype network. Specifically: S110: Obtain P simulation data points from the simulation dataset and Q measured HRRP data points from the measured dataset, where P is greater than Q; all simulation data points and all measured HRRP data points can be classified into C target categories.

[0018] It should be noted that the number of categories in the P simulated data points and the Q measured HRRP data points is the same. Assuming C is 5, P is 1000, and Q is 50, then in the 1000 simulated data points, there are an average of 200 samples per category, and in the 50 measured HRRP data points, there are an average of 10 samples per category.

[0019] It should be noted that, due to the practical constraints of data acquisition in real-world application scenarios, the data samples in the actual test dataset only contain echo information of the target within a limited range of attitude angles.

[0020] S120: Using the i-th feature extraction module, feature extraction is performed on P simulation data and Q measured HRRP data respectively to obtain P simulation domain feature vectors and Q measured domain feature vectors that correspond one-to-one.

[0021] Figure 3 This is a schematic diagram of the feature extraction module provided in an embodiment of the present invention. (Combined with...) Figures 2-3 As shown, the feature extraction module in the current model includes multiple sequentially connected one-dimensional convolutional neural network blocks and at least one fully connected layer. Each one-dimensional convolutional neural network block includes a one-dimensional convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. Each one-dimensional convolutional neural network block is used to extract features from P simulated data and Q measured HRRP data respectively. The fully connected layer is used to integrate and transform the features extracted by the one-dimensional convolutional neural network block, outputting P simulated domain feature vectors and Q measured domain feature vectors.

[0022] For example, the number of one-dimensional convolutional neural network blocks is 3, and the number of fully connected layers is 2. The one-dimensional convolutional neural network blocks can accurately capture the local fluctuation features related to the target physical structure in HRRP, and achieve layer-by-layer mining from the bottom waveform to the deep semantic features through multi-layer convolution stacking.

[0023] Here, it is assumed that each one-dimensional convolutional neural network contains The nth convolutional layer, the nth layer( ) comprises convolution kernels. For the first layer, the first ( ) convolution kernel is denoted as , which has a channel number of ( ). It is assumed that each channel is a vector of size , i.e. . Let denote the input of the first convolution layer (i.e. the output of the first layer, is an HRRP sample), and its channel number, the vector dimension of each channel. and , denoted as , can be represented as: ; where ∗ denotes the convolution operator, denotes the bias, is a nonlinear activation function (ReLU is used in this method). The final output of the first layer, i.e. , is the sum of all channels, i.e. , which has a size of . For multi-channel and , their convolution, i.e. , is the sum of all channels. In order to increase the receptive field of the convolution kernel and the sparsity of the hidden features, a subsampling layer is usually embedded after the above convolution calculation and nonlinear transformation. In this method, a max-pooling operation is used to remove all elements in the pooling region except the maximum element for each feature vector of each convolution layer. This operation enhances the tolerance to small attitude fluctuations by local feature selection, while reducing the risk of overfitting by dimensionality reduction.

[0024] Batch normalization is used after each convolution operation. For a layer with an input of dimensions, the batch normalization operation first normalizes each dimension by: ; where the mean and variance are calculated over a small batch from the training dataset, and a constant is added to the small batch variance to maintain numerical stability. After normalization, a pair of parameters , are introduced to scale and shift the normalized value: ; Learning parameters on the basis of original model parameters and By normalizing each input unit to have zero mean and unit variance, the batch normalization layer helps to deal with the problem of poor initialization in the training phase and helps to alleviate the problem of internal covariate shift caused by the change of target posture of HRRP data, accelerating learning and convergence.

[0025] Here, the order of feature extraction of the simulation data and the measured HRRP data is not limited. Feature extraction is performed on each simulation data to obtain a one-to-one corresponding simulation domain feature vector, and feature extraction is performed on each measured HRRP data to obtain a one-to-one corresponding measured domain feature vector.

[0026] S130: Using the ith domain aligner, map the Q measured domain feature vectors to the domain-invariant space to obtain Q domain-invariant feature vectors.

[0027] It should be noted that mapping the measured domain features to the domain-invariant space can transform the measured domain features to a space similar to the simulation domain features, which can reduce the distribution difference between the simulation domain and the measured domain, so that the features of the two domains are aligned in the same space. In addition, in the domain-invariant space, the features of the measured domain and the simulation domain are closer, so it is more reasonable and effective to generate intermediate domain features (for example, by interpolation) in this space, and the generated intermediate domain features can better connect the two domains.

[0028] S140: Input the Q domain-invariant feature vectors and the P simulation domain feature vectors into the intermediate domain data generation module in the current model to generate a plurality of second intermediate domain features.

[0029] Figure 4 is a schematic diagram of generating intermediate domain features provided by an embodiment of the present application. As shown in Figure 4 , the Q domain-invariant feature vectors and the P simulation domain feature vectors contain the same number of target categories; the Q domain-invariant feature vectors and the P simulation domain feature vectors are used to generate a plurality of intermediate domain features, including: calculating the feature mean of the plurality of domain-invariant feature vectors under each target category as a plurality of target domain anchor points, and calculating the feature mean of the plurality of simulation domain feature vectors under each target category in the plurality of target categories as a plurality of source domain anchor points; performing linear interpolation between the source domain anchor point and the target domain anchor point corresponding to any one target category to generate a first intermediate domain anchor point and a second intermediate domain anchor point; randomly sampling a semantic transformation direction vector corresponding to the target category from a zero-mean multivariate normal distribution; adding the semantic transformation direction vector corresponding to the target category to the first intermediate domain anchor point and the second intermediate domain anchor point, respectively, to obtain the first intermediate domain feature and the second intermediate domain feature corresponding to the target category, and then obtaining all the intermediate domain features.

[0030] Exemplarily, in the i-th intermediate domain data generation module, first, the feature mean of each class in the source domain and the target domain is estimated respectively The feature mean is taken as the source domain / target domain anchor point to integrate the semantic knowledge of each class in the source domain and the target domain. Linear interpolation is performed between the source domain and the target domain anchor points of the same class to generate two intermediate domain anchor points: ; ; wherein, is the anchor point of the class in the feature space on the source domain / target domain; is a function about the current iteration number . With the progress of the training process, gradually approaches from 0. .

[0031] In the feature space, some transformation directions represent meaningful semantic transformation. Therefore, the generated intermediate domain anchor points , ( ) are approximated to simulate this semantic transformation, so that meaningful cross-domain semantic knowledge is transferred from the source domain to the target domain. A vector is randomly sampled from a zero-mean multivariate normal distribution , which is taken as the semantic transformation direction of the intermediate domain anchor points , , and then the enhanced intermediate domain features , are obtained. Wherein, represents the covariance of the source domain feature .

[0032]

[0033]

[0034] wherein the generated intermediate domain features , obeys the normal distribution.

[0035] S150: The i-th domain discriminator is used to judge the Q domain-invariant feature vectors, the P simulated domain feature vectors and the plurality of second intermediate domain features, and the i-th domain-level alignment loss function is generated according to the judgment result to measure the domain alignment degree of the Q domain-invariant feature vectors and the P simulated domain feature vectors, and the confusion degree of the plurality of second intermediate domain features.

[0036] ​In the iterative training process of the model, the i-th domain discriminator is constructed as a binary classifier, and the design goal is to distinguish whether the input feature comes from the simulation domain or the real measurement domain. Exemplarily, the score corresponding to the simulation domain data is 1, and the score corresponding to the real measurement domain is 0. In an ideal state, when the input is the domain-invariant feature vector obtained by the domain aligner, the domain discriminator should be able to accurately identify its simulation domain source; and when the input is the second intermediate domain feature, the output probability of the domain discriminator should tend to 0.5, indicating that it cannot effectively distinguish the domain of the feature, thereby achieving the purpose of assisting the generator to confuse the discriminator.

[0037] However, in the initial stage of training, the parameters of the domain aligner have not been optimized, and the domain-invariant features generated by the domain aligner often still retain obvious real measurement domain statistical characteristics, causing the domain discriminator to be able to easily identify the source of the features. At the same time, the intermediate domain data generation module generates the second intermediate domain feature distribution that may deviate from the ideal intermediate position due to unadjusted parameters, and cannot effectively play a confusing role. Therefore, the i-th domain-level alignment loss function constructed using the judgment result output by the i-th domain discriminator is used to update the model parameters of the domain aligner and the domain discriminator in real time.

[0038] In a possible implementation, the calculation expression of the domain-level alignment loss function in each training process is as follows: ; Wherein, is used to calculate the mathematical expectation of the simulation domain feature vector in the process of feature extraction, domain alignment and intermediate domain data generation, is used to calculate the mathematical expectation of the real measurement domain feature vector in the process of feature extraction, domain alignment and intermediate domain data generation; refers to the feature transformation of the feature extraction module in the current model, refers to a set of real measurement HRRP data, refers to a set of simulation data, refers to the discrimination result of the domain discriminator pair, refers to a set of second intermediate domain features.

[0039] In a possible implementation, the domain aligner in each training process updates the parameters by the following formula: .

[0040] S160: Using the i-th prototype network, performing class alignment processing on the Q domain-invariant feature vectors, the P simulation domain feature vectors and the plurality of second intermediate domain features simultaneously to obtain a class-aligned data set, and generating an i-th class-level alignment loss function.

[0041] Here, the i-th class-level alignment loss function includes the i-th cross-domain prototype contrast loss function and the i-th cross-domain sample contrast loss function; the class alignment processing includes: respectively assigning learnable prototype vectors to the Q domain-invariant feature vectors, the P simulated domain feature vectors and the plurality of second intermediate domain features, and by minimizing the i-th cross-domain prototype contrast loss function and the i-th cross-domain sample contrast loss function, to constrain the distribution relationship of all data in the feature space; wherein the constraint target of the distribution relationship is to: realize the intra-class compactness of the cross-domain features under the same class, and the inter-class separation of the features between different classes.

[0042] Specifically, the i-th cross-domain prototype contrast loss function is obtained by the following way: S1: assigning a prototype vector to each data in the i-th domain alignment result; S2: for each target class in the C target classes, calculating the similarity between the prototype vectors of each target class in different domains to construct the i-th cross-domain prototype contrast loss function covering all domain combinations.

[0043] The i-th cross-domain sample contrast loss function is obtained by the following way: S1: for any one data in the i-th domain alignment result, calculating the similarity between the data and all prototype vectors of all domains to obtain a plurality of similarity values; S2: based on the feature-level scaling coefficient, normalizing the plurality of similarity values by the Softmax function to obtain a standardized similarity probability distribution; S3: based on the standardized similarity probability distribution, calculating the cross-entropy loss corresponding to the data according to the real class of any one data in the i-th domain alignment result; S4: repeating S1-S3 for all data in the i-th domain alignment result to obtain all cross-entropy losses, and taking the average of all cross-entropy losses to obtain the i-th cross-domain sample contrast loss function.

[0044] In one possible implementation, the calculation expression of the class-level alignment loss function in each training process is: ; ; ; wherein, is the cross-domain prototype contrast loss function, is the cross-domain sample contrast loss function, is the prototype vector of the c-th target class in the measured domain, including the prototype vector corresponding to the feature vector belonging to the c-th target class in the Q domain alignment processed domain-invariant feature vector, It is the prototype vector of the c-th target category in the simulation domain, including the prototype vectors corresponding to the feature vectors of the c-th target category among the P domain-aligned simulation domain feature vectors. It is the prototype vector of the c-th target category in the intermediate domain, including the prototype vectors corresponding to the feature vectors of the c-th target category in the second intermediate domain features after multiple domain alignment processing. This represents calculating the similarity between the prototype vectors of the c-th target category in the simulation domain and the measured domain. This represents calculating the similarity between the prototype vectors of the c-th target category in the simulation domain and the intermediate domain. This represents calculating the similarity between the prototype vectors of the c-th target category in the measured domain and the intermediate domain. All prototype vectors representing the simulation domain and the measured domain are bidirectionally aligned. All prototype vectors representing the simulation domain are aligned with all prototype vectors representing the measured domain. This means that all prototype vectors in the measured domain are aligned with all prototype vectors in the simulation domain. All prototype vectors representing the simulation domain and intermediate domain are bidirectionally aligned. All prototype vectors representing the measured domain and intermediate domain are bidirectionally aligned.

[0045] It should be noted that the calculation expressions are the same for one-way calculations. That is... , , , , and The calculation expressions are the same. In specific calculations, simply replacing the subscripts will yield the corresponding calculation formula. and Let's take the calculation expression as an example to illustrate. Specifically: ; .

[0046] in, Represents the feature-level scaling factor. Represents the inner domain of the feature space The Middle The feature vector of each sample Indicates the first The category corresponding to each sample.

[0047] S170: Construct the i-th classification loss function using the i-th classification result, and then combine the i-th domain-level alignment loss function and the i-th class-level alignment loss function to obtain the i-th overall objective function.

[0048] Here, the i-th classification loss function includes an i-th simulation domain sample classification loss function, an i-th measured domain sample classification loss function, and an i-th intermediate domain sample classification loss function; the expression of the i-th classification loss function is: ; ; ; ; wherein, is the i-th simulation domain sample classification loss function, is the i-th measured domain sample classification loss function, is the i-th intermediate domain sample classification loss function, is a domain-invariant feature vector, is a prototype vector of an n-th class in a simulation domain, is a prototype vector of a k-th class in the simulation domain, is an Euclidean distance function, is a prototype vector of an n-th class in a measured domain, is a prototype vector of a k-th class in the measured domain, is a prototype vector of an n-th class in an intermediate domain, is a prototype vector of a k-th class in the intermediate domain, is a feature vector of an intermediate domain sample, is a feature vector of a measured domain sample, is a feature vector of a simulation domain sample, is a real class label of a sample.

[0049] Further, the calculation expression of the overall objective function in each training process is: ; wherein, is a classification loss function, is a domain-level alignment loss function, is a class-level alignment loss function, all are preset weights.

[0050] S180: judging whether the i-th overall objective function converges or not; S190: if yes, taking the current model as a trained radar HRRP small sample target recognition model; if no, updating the model parameters by using the i-th overall objective function, performing the i+1-th training, until the model converges, and obtaining the trained radar HRRP small sample target recognition model.

[0051] Specifically, the class-level alignment loss function is used to adjust the model parameters of the prototype network, the domain-level alignment loss function is used to adjust the parameters of the domain aligner and the domain discriminator, and the classification loss function is used to adjust the parameters of the classification module.

[0052] Compared with the prior art, the present application has the following beneficial effects: (1) Significantly improve the target recognition performance under the condition of small sample, effectively utilize the simulation data to make up for the shortage of measured data, overcome the problem of sharp decline in recognition performance of traditional method when training sample is scarce.

[0053] (2) Enhance the adaptability and generalization ability of the model to the posture change, solve the problem of limited recognition effect of traditional method under the condition of incomplete target posture through domain invariant feature learning and intermediate domain feature generation.

[0054] (3) Realize effective migration of cross-domain knowledge, establish feature alignment between simulation domain and measured domain through joint optimization of classification loss, domain-level alignment loss and class-level alignment loss, break through the technical bottleneck of difficult knowledge migration between different domains of traditional method.

[0055] (4) Improve the discriminability and robustness of feature representation, learn more discriminative feature representation through prototype contrast learning and multi-domain feature alignment, overcome the defect of insufficient feature discriminability of traditional method in complex scene.

[0056] (5) Reduce the dependence on a large amount of labeled measured data in actual application, through the training mechanism assisted by simulation data, significantly reduce the demand for measured samples while ensuring the recognition performance, solve the practical problem of high labeling cost in actual application.

[0057] In order to prove the effectiveness of the present application, simulation experiment is carried out on the SAMPLE data set to evaluate the present application. Table 1 shows the number of simulation and measured HRRP data of each large class category, the azimuth angle of the data is from 10 degrees to 80 degrees, and the pitch angle is from 14 degrees to 17 degrees. In the experimental setting, the test data has a pitch angle of 17 degrees, and the training data has a pitch angle of 14 degrees, 15 degrees and 16 degrees.

[0058] Table 1

[0059] In the experimental setting, the source domain is a simulation HRRP data set, and its category label can be obtained. The target domain is a measured HRRP data set, and its category label can also be obtained. Follow the process of inductive learning, in which the target domain is divided into training set and test set. The training set and the source domain data participate in the learning of the cross-domain HRRP target recognition model together. The test set is used to evaluate the performance of the model.

[0060] For the network used in the present application, all weight matrices are randomly initialized, all bias terms are initialized to 0, the small batch size is set to 64, the hyper-loss weight parameters are , the temperature coefficient hyperparameters are , . Optimization is performed using the Adam optimizer with an initial learning rate of 0.001, and all experiments are trained under the condition of GPU 4060 8GB, RAM 16GB. Unless otherwise specified, all experiments use this setting.

[0061] In order to verify the performance of the present application in the radar HRRP small sample target recognition task, the proposed method is compared with the traditional deep radar HRRP target recognition method and the traditional domain adaptation method, as shown in Table 2. Among them, PN is the traditional deep radar HRRP recognition method; DAN and DANN are traditional domain adaptation methods, wherein DAN is a method of aligning the global distribution of simulation domain and measured domain features using MMD constraint, and DANN is a method of aligning the global distribution of simulation domain and measured domain features using domain adversarial learning. Both of these domain adaptation methods use simulation data as the source domain and a small amount of target domain measured data as the target domain for domain adaptation transfer learning in the experiment.

[0062] Table 2 gives the recognition performance of the above various typical radar HRRP recognition methods in different angular domain ranges, with the number of target domain training samples being 5shot, 10shot, 20shot and all samples (with the average result of 20 experiments), wherein the entire azimuth angle angular domain range is 10 degrees to 80 degrees, the azimuth angle is 10 degrees to 45 degrees when the angular domain range is 1 / 2, and the azimuth angle is 10 degrees to 28 degrees when the angular domain range is 1 / 4. The number of measured training samples 5shot, 10shot and 20shot all represent that 5, 10 and 20 samples are randomly sampled in the measured training set. As can be seen from the table, the best recognition performance is achieved by the present application. The traditional deep network (PN) is pre-trained using a large amount of electromagnetic simulation data, and is fine-tuned using a small amount of measured data in the training process. The fine-tuning method considers that the network shallow knowledge is domain-invariant information that is commonly shared, and through freezing the shallow parameters and fine-tuning the deep parameters, knowledge transfer is achieved while learning the domain-specific information of the measured domain. However, for this simple and direct transfer learning method, the transferable knowledge is highly sensitive to the difference between the source domain and the target domain data and the number of frozen shallow layers, so the overall small sample recognition performance is weaker than the domain adaptation method. The traditional domain adaptation method (DAN, DANN) relies on the accurate alignment of the simulation domain (source domain) features and the measured domain (target domain) features, so as to transfer the knowledge learned by the source domain to the target domain to assist the target domain recognition task. However, in the case of only a small amount of training data in the target domain, the traditional domain adaptation method cannot accurately align the simulation domain and the measured domain feature distribution, and the limited target domain training data affects the feature alignment effect of the source domain and the target domain, so that the recognition model only aligns the features of the source domain and the limited target domain training samples, resulting in that the performance of the traditional domain adaptation method is still limited under the small sample condition.

[0063] Compared with the traditional deep network method PN and the traditional domain adaptation method DAN and DANN in Table 2, the performance of the present application is improved by 4%-13%, which shows that the intermediate domain feature expansion strategy and the cross-domain knowledge transfer learning strategy used in the present application effectively alleviate the small sample problem, thereby achieving the optimal performance.

[0064] Table 2

[0065] Figure 5 is a confusion matrix diagram generated by the different methods provided by the embodiments of the present application under the condition that 1 / 2 of the angular domain is missing and there are 10 measured training samples per class. Figure 5 (a) in (a) is a confusion matrix diagram corresponding to the PN method, Figure 5 (b) in (b) is a confusion matrix diagram corresponding to the DAN method, Figure 5 (c) in (c) is a confusion matrix diagram corresponding to the DANN method, Figure 5 (d) in (d) is a confusion matrix diagram corresponding to the present application. AsFigure 5 It can be seen that the model learned by the traditional deep network method PN is over-fitted, and the recognition performance is low, which verifies that the model learned under the condition of small sample cannot be well generalized to the whole measured data set, and the traditional domain adaptation method DAN and DANN align the feature distribution of the simulation domain and the measured domain to a certain extent through the domain adaptation strategy, but the feature alignment effect is not ideal due to the limitation of small sample measured training data, and the performance improvement is small.

[0066] Figure 6 is a confusion matrix diagram generated by the ben method provided by the embodiment of the present application under the condition of 1 / 2 different measured training sample quantities in the angular domain. Figure 6 (a) in is a confusion matrix diagram generated under the condition of 5 measured training samples per class, Figure 6 (b) in is a confusion matrix diagram generated under the condition of 10 measured training samples per class, Figure 6 (c) in is a confusion matrix diagram generated under the condition of 20 measured training samples per class, Figure 6 (d) in is a confusion matrix diagram generated under the condition of corresponding all measured training samples per class. Figure 6 It can be seen that the present application obtains a higher recognition accuracy in all 6 classes of targets under different conditions, and the confusion degree between classes is greatly reduced, which shows that the features learned by the present application are more discriminative on the measured HPPR data set.

[0067] In order to more intuitively illustrate the recognition effect of the present application and the separability of the features proposed, Figure 7 is a t-SNE visualization diagram of the features proposed by different methods in Figure 5 Figure 7 (a) in is a t-SNE visualization diagram corresponding to the PN method, Figure 7 (b) in is a t-SNE visualization diagram corresponding to the DAN method, Figure 7 (c) in is a t-SNE visualization diagram corresponding to the DANN method, Figure 7 (d) in is a t-SNE visualization diagram corresponding to the present application. In the diagram, the diamond symbol refers to the source domain training feature, and the circular symbol refers to the target domain training feature. Figure 7 It can be seen that the features proposed by the traditional deep network PN have a large amount of confusion between classes, the model is over-fitted, and the generalization is poor, while the measured domain features extracted by the traditional domain adaptation method DAN and DANN still have limited separability, resulting in poor recognition performance, and the confusion between classes of the features extracted by the present application is significantly reduced, and the feature separability is significantly improved. It shows that the present application has stronger generalization under the condition of small measured training samples, and further has better recognition performance.

[0068] In order to further show the role of the intermediate domain generation module, Figure 8 ​is provided by the embodiment of the present application Figure 5 Fig. 6 is a t-SNE visualization diagram of the HRRP data of a type of ground vehicle target in different iterations of the method provided by the embodiment of the present application, wherein, Figure 7 Fig. 6(a) is a t-SNE visualization diagram in the first iteration, Figure 7 Fig. 6(b) is a t-SNE visualization diagram in the fifth iteration, Figure 7 Fig. 6(c) is a t-SNE visualization diagram in the thirtieth iteration, Figure 8 Fig. 6(d) is a t-SNE visualization diagram in the one hundredth iteration. ​ As can be seen from Fig. 6, under the action of the intermediate domain generation module, the test data and the training data of the target domain are gradually drawn closer, and finally alignment is achieved. The effect of the intermediate domain generation module can be seen from the visualization experiment. By generating intermediate domains near the source domain and the target domain, these intermediate domains can be regarded as being generated by mixing the source domain and the target domain in a larger range, so that the target domain samples are indirectly expanded, the small sample problem caused by the angle domain missing of the measured target is solved, the overfitting problem of the model in the training process is alleviated, and the generalization ability of the model is greatly improved. At the same time, the intermediate domain generation module also solves the problem that the distribution distance between the source domain and the target domain is far. In the training process, the source domain and the target domain are drawn closer to each other with the intermediate domain. Specifically, the source domain is drawn closer to the intermediate domain close to the source domain, and the target domain is drawn closer to the intermediate domain close to the target domain (this is because in the domain alignment module, the classification scores of the intermediate domain input to the domain discriminator are 0.5 and 0.5, so the source domain / target domain will also be close to the intermediate domain close to the source domain / target domain, and the classification scores input to the domain discriminator will also be 0.5 and 0.5, achieving the purpose of confusing the domain discriminator), and the two intermediate domains also gradually approach each other due to the gradual setting and the action of the domain alignment and class alignment, so with the increase of the number of iterations, the source domain and the target domain slowly approach each other, and finally achieve the effect of alignment of the source domain, the target domain and the intermediate domain.

[0069] In summary, the intermediate domain feature expansion strategy combining the source domain simulation data and the target domain measured data and the simulation data assisted cross-domain knowledge transfer learning strategy promote the generalization learning ability of the target domain separable feature in the small sample situation, and are beneficial to the generalization description ability of the prototype to the corresponding class feature distribution, thereby improving the robustness of the target recognition in the case of a small amount of training samples in the measured domain.

[0070] In order to more intuitively analyze the influence of different modules of the present application on the recognition performance, Table 3 shows the ablation experiment results of 10 measured training samples of each class under the condition of 1 / 2 angle domain missing, wherein S2R refers to a method of directly testing the measured domain test set by using the model trained by the simulation domain, which can be used as a baseline comparison method of training by using a large amount of simulation data and a small amount of target domain measured data at the same time.

[0071] The ablation experiment results of Table 3 show that each module designed can improve performance, and the optimal performance is achieved by the synergistic combination of each module. Experiment (a) refers to the result of training using the extended measured domain data and simulation data together. Compared with S2R, the data augmentation strategy can effectively expand the measured domain HRRP sample, and alleviate the problem of lack of information caused by small samples (the accuracy is improved from 51.72% to 68.36%). Experiments (b) and (c) show the effectiveness of the domain alignment and class alignment modules. Compared with only using the data augmentation strategy, the performance improvement of adding the domain alignment and class alignment modules is 10.06% and 4.49% respectively, and the effect of domain alignment is more significant. This is because under the condition of not performing domain alignment, the feature domain difference between the simulation domain and the measured domain is too large, at this time, the difficulty of direct class alignment is too large, the alignment degree is limited, and the effect is not significant; the domain alignment only aligns the overall distribution of the simulation domain and the measured domain features, which is less difficult and easy to align accurately, and due to the inherent characteristics of automatic aggregation of features of the same class, the effect of cross-domain feature class alignment is formed to some extent even without class alignment constraint in the domain alignment. Therefore, overall, the domain alignment and class alignment are more superior in small sample recognition. However, under the premise of using domain alignment, class alignment can further improve the target domain recognition performance, which shows the synergistic effect of domain alignment and class alignment, and the combination of the two is more effective, as shown in experiment (g). Experiments (d), (e) and (f) further illustrate the effectiveness of the data augmentation and domain alignment module and the class alignment module, and it can be seen that without data augmentation, the effects of domain alignment and class alignment are similar to the conclusions of experiments (b) and (c) as described above. In addition, by comparing (b) and (d), (c) and (e), (f) and (g), the experiments show that in the case of data scarcity, the alignment effect of the domain alignment and class alignment modules will decrease significantly, which shows the basic role of the data augmentation strategy for the domain alignment module and the class alignment module, and the modules cooperate with each other to better improve the recognition performance. Experiment (g) is the method proposed in this chapter, and compared with the results of other experiments, it can be concluded that a single module cannot achieve good target domain recognition, and the modules cooperate with each other to achieve the maximum performance improvement, effectively solve the problem of feature alignment between the simulation domain and the measured domain under the condition of small samples, thereby realizing the simulation data assisted knowledge transfer learning, and achieving the best recognition performance.

[0072] Table 3

[0073] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the protection scope of the present application.

Claims

1. A radar HRRP small sample target recognition method assisted by simulation data, characterized in that, The method comprises the following steps: acquiring radar HRRP data to be detected; inputting the radar HRRP data to be detected into a trained radar HRRP small sample target recognition model to perform feature extraction, obtaining target features to be detected, calculating the Euclidean distance between the target features to be detected and prototypes corresponding to each target class, and screening out the minimum Euclidean distance to generate a classification result of the radar HRRP data to be detected, wherein the classification result comprises classifying the radar HRRP data to be detected into a target class corresponding to the minimum Euclidean distance; wherein the trained radar HRRP small sample target recognition model is obtained by iteratively training, generating a plurality of first intermediate domain features by fusing M simulation data containing target full attitude information and N measured HRRP data containing partial target attitude information, M being greater than N, respectively calculating a classification loss, a domain-level alignment loss and a class-level alignment loss by using the M simulation data, the N measured HRRP data and the plurality of first intermediate domain features, and jointly optimizing a total objective function composed of the classification loss, the domain-level alignment loss and the class-level alignment loss.

2. The method of claim 1, wherein, The radar HRRP small sample target recognition model comprises a feature extraction module, a domain aligner, an intermediate domain data generation module, a domain discriminator, a prototype network and a classification module; the process of training the trained radar HRRP small sample target recognition model comprises: in the i-th training process, taking the radar HRRP small sample target recognition model obtained in the (i-1)-th training as the i-th radar HRRP small sample target recognition model: acquiring P simulation data in a simulation data set and Q measured HRRP data in a measured data set, P being greater than Q; all simulation data and all measured HRRP data can be divided into C target classes; using the i-th feature extraction module to perform feature extraction on the P simulation data and the Q measured HRRP data respectively, obtaining one-to-one corresponding P simulation domain feature vectors and Q measured domain feature vectors; using the i-th domain aligner to map the Q measured domain feature vectors to a domain-invariant space to obtain Q domain-invariant feature vectors; inputting the Q domain-invariant feature vectors and the P simulation domain feature vectors into the i-th intermediate domain data generation module to generate a plurality of second intermediate domain features; using the i-th domain discriminator to judge the Q domain-invariant feature vectors, the P simulation domain feature vectors and the plurality of second intermediate domain features, and generating an i-th domain-level alignment loss function according to the judgment result to measure the domain alignment degree of the Q domain-invariant feature vectors and the P simulation domain feature vectors, and the confusion degree of the plurality of second intermediate domain features; using the i-th prototype network to perform class alignment processing on the Q domain-invariant feature vectors, the P simulation domain feature vectors and the plurality of second intermediate domain features synchronously to obtain a class-aligned data set, and generating an i-th class-level alignment loss function. The Euclidean distance between each data in the class-aligned data set and the prototype corresponding to each target class is calculated using the i-th classification module, the minimum Euclidean distance is screened out, the target class corresponding to the minimum Euclidean distance is taken as the classification result, and the i-th group of classification results is obtained by summarizing; and the i-th classification result is used to construct the i-th classification loss function; The i-th classification result constructs the i-th classification loss function, and the i-th domain-level alignment loss function and the i-th class-level alignment loss function are combined to obtain the i-th overall objective function by weighting. It is judged whether the i-th overall objective function converges or not. If yes, the current model is taken as the trained radar HRRP small sample target recognition model; if no, the model parameters are updated using the i-th overall objective function, the i+1-th training is performed, and the trained radar HRRP small sample target recognition model is obtained until the model converges.

3. The method of claim 2, wherein, The target class number contained in the Q domain-invariant feature vectors and the P simulation domain feature vectors is the same; the Q domain-invariant feature vectors and the P simulation domain feature vectors are input into the i-th intermediate domain data generation module to generate a plurality of second intermediate domain features, including: The feature mean of the plurality of domain-invariant feature vectors under each target class is calculated to serve as a plurality of target domain anchor points, and the feature mean of the plurality of simulation domain feature vectors under each target class in the plurality of target classes is calculated to serve as a plurality of source domain anchor points; Linear interpolation is performed between the source domain anchor point and the target domain anchor point corresponding to any one target class to generate a first intermediate domain anchor point and a second intermediate domain anchor point; A semantic transformation direction vector corresponding to the target class is randomly sampled from a zero-mean multivariate normal distribution; The semantic transformation direction vector corresponding to the target class is added to the first intermediate domain anchor point and the second intermediate domain anchor point respectively to obtain the first intermediate domain feature and the second intermediate domain feature corresponding to the target class, and then all intermediate domain features are obtained.

4. The method of claim 2, wherein, The i-th class-level alignment loss function includes an i-th cross-domain prototype contrast loss function and an i-th cross-domain sample contrast loss function; the class alignment processing includes: Learnable prototype vectors are respectively assigned to the Q domain-invariant feature vectors, the P simulation domain feature vectors and the plurality of second intermediate domain features, and the distribution relationship of all data in the feature space is constrained by minimizing the i-th cross-domain prototype contrast loss function and the i-th cross-domain sample contrast loss function; wherein the constraint target of the distribution relationship is to realize the intra-class compactness of cross-domain features under the same class and the inter-class separation of features between different classes.

5. The method of claim 4, wherein, The i-th cross-domain prototype contrast loss function is obtained by the following way: S1: a prototype vector is assigned to each data in the i-th domain alignment result; S2: for each target class in the C target classes, the similarity between the prototype vectors of the target class in different domains is calculated to construct the i-th cross-domain prototype contrast loss function covering all domain combinations.

6. The method of claim 5, wherein, The i-th cross-domain sample contrast loss function is obtained by the following way: S1: For any one data in the i-th domain alignment result, the similarity between the data and all prototype vectors of all domains is calculated to obtain a plurality of similarity values; S2: Based on the feature-level scaling coefficient, the plurality of similarity values are normalized by a Softmax function to obtain a standardized similarity probability distribution; S3: According to the true class of any one data in the i-th domain alignment result, the cross-entropy loss corresponding to the data is calculated based on the standardized similarity probability distribution; S4: Repeat S1-S3 for all data in the i-th domain alignment result to obtain all cross-entropy losses, and take the average of the all cross-entropy losses to obtain the i-th cross-domain sample comparison loss function.

7. The method of claim 2, wherein the method is a simulated data aided radar HRRP small sample target recognition method. The calculation expression of the domain-level alignment loss function in each training process is: ； wherein, for calculating the mathematical expectation of the simulated domain feature vectors in the process of feature extraction, domain alignment and intermediate domain data generation, for calculating the mathematical expectation of the measured domain feature vectors in the process of feature extraction, domain alignment and intermediate domain data generation; refers to the feature transformation of the feature extraction module in the current model, refers to the set of measured HRRP data, refers to the set of simulated data, refers to the discrimination result of the domain discriminator pair, refers to the set of the plurality of second intermediate domain features.

8. The method of claim 6, wherein the method is a simulated data aided radar HRRP small sample target recognition method. The calculation expression of the class-level alignment loss function in each training process is: ； ； ； wherein, is a cross-domain prototype contrast loss function, is a cross-domain instance contrast loss function, is a prototype vector of the cth target class in the measured domain, including a prototype vector corresponding to a feature vector belonging to the cth target class in the Q domain invariant feature vectors after domain alignment processing, is a prototype vector of the cth target class in the simulation domain, including a prototype vector corresponding to a feature vector belonging to the cth target class in the P simulation domain feature vectors after domain alignment processing, is a prototype vector of the cth target class in the intermediate domain, including a prototype vector corresponding to a feature vector belonging to the cth target class in the plurality of second intermediate domain features after domain alignment processing, represents calculating the similarity between the prototype vectors of the cth target class in the simulation domain and the measured domain, represents calculating the similarity between the prototype vectors of the cth target class in the simulation domain and the intermediate domain, represents calculating the similarity between the prototype vectors of the cth target class in the measured domain and the intermediate domain, represents bidirectional alignment of all prototype vectors of the simulation domain and the measured domain, represents that all prototype vectors of the simulation domain are aligned with all prototype vectors of the measured domain, represents that all prototype vectors of the measured domain are aligned with all prototype vectors of the simulation domain, represents bidirectional alignment of all prototype vectors of the simulation domain and the intermediate domain, represents bidirectional alignment of all prototype vectors of the measured domain and the intermediate domain.

9. The method of claim 2, wherein, The i-th classification loss function includes an i-th simulated domain sample classification loss function, an i-th measured domain sample classification loss function, and an i-th intermediate domain sample classification loss function; and the expression of the i-th classification loss function is: ； ； ； ； wherein, is the ith simulated domain sample classification loss function, is the ith measured domain sample classification loss function, is the ith intermediate domain sample classification loss function, is the domain-invariant feature vector, is the prototype vector of the nth class in the simulated domain, is the prototype vector of the kth class in the simulated domain, is the Euclidean distance function, is the prototype vector of the nth class in the measured domain, is the prototype vector of the kth class in the measured domain, is the prototype vector of the nth class in the intermediate domain, is the prototype vector of the kth class in the intermediate domain, is the feature vector of the intermediate domain sample, is the feature vector of the measured domain sample, is the feature vector of the simulated domain sample, is the true class label of the sample.

10. The method of claim 2, wherein, The calculation expression of the overall objective function in each training process is: ； wherein, is a classification loss function, is a domain-level alignment loss function, is a class-level alignment loss function, are all preset weights.

Images