Target identification method based on unsupervised dynamic-attribute feature decoupling network

Abstract

The invention provides a target identification method based on an unsupervised dynamic-attribute feature decoupling network. The method comprises the following steps: acquiring HRRP data of a target in a period of time as an HRRP sequence; processing the target HRRP sequence by using the trained HRRP target recognition model to obtain a target recognition result; the HRRP target recognition model comprises an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network; an obtained target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network to obtain attribute features, and the trained HRRP target recognition network processes the attribute features to obtain a target recognition result. According to the HRRP target identification method provided by the invention, the type of the target can be identified more accurately and robustly under the condition that various dynamic factors such as azimuth change and noise interference are superposed and numerical values are unknown.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target recognition, in particular to a target recognition method based on an unsupervised dynamic-attribute feature decoupling network. BACKGROUND

[0002] In the scene of maritime law enforcement, maritime search and rescue, etc., the use of search radar for target search and identification is the main means in harsh weather environment. Because the radar high resolution range profile (HRRP) contains the discriminative features of target material properties, geometric structure, etc., it is widely used in radar automatic target recognition and is of great help to quickly find targets. HRRP is composed of the amplitude of the sum of complex echo vectors of scattering bodies in each range cell, which represents the projection of the complex echo of the scattering center on the radar line-of-sight (LOS).

[0003] In recent years, due to the development of deep learning technology, the research on HRRP target recognition based on deep neural network has made significant progress, but there are still many difficulties and challenges, including the problem of azimuth sensitivity of HRRP. Because the relative azimuth change between radar and target will cause the change of scattering point projection distribution, the HRRP waveform of the same target under different observation azimuths has significant difference, and the recognition model is easy to misidentify in the case of not seeing such waveform. Many existing HRRP target recognition methods usually construct recognition models under the assumption of complete azimuth samples, assuming that the model has seen all azimuth HRRP data of the target, so it has good recognition performance. However, in the actual non-cooperative target recognition scene, due to the uncontrollability of target motion and the constraints of observation conditions, the available HRRP data often has serious azimuth missing. Therefore, how to realize the full-azimuth domain generalization recognition based on partial azimuth samples will have very important practical significance.

[0004] Currently, scholars have carried out researches on this issue, which can be mainly divided into two directions of data enhancement and model enhancement. Data enhancement is mainly through the generation of adversarial network or variational encoding structure to synthesize the missing HRRP data, but the physical rationality and distribution consistency of the generated samples lack effective evaluation criteria. The model enhancement direction includes meta-learning cross-orientation adaptation method, transfer feature mapping model, and self-supervised contrast framework, etc. These methods perform well under medium missing rate, but their performance significantly decreases under high proportion of missing orientation. Notably, inspired by decoupled representation learning, some scholars have proposed a type-orientation decoupling network, which separates the class and orientation features of the target through double-branch supervised learning, and then uses the separated class features for target recognition, providing a new idea for HRRP target recognition under orientation missing. However, this method relies on double supervision signals, and needs to obtain the target class label and accurate orientation information at the same time. The orientation information of the target is determined by the LOS direction of the radar and the posture of the target, but in actual scenarios, the target posture is complex and variable due to environmental factors. Even if the LOS direction of the radar relative to the target point can be obtained, the posture data of the target body is still difficult to accurately obtain. Therefore, in such cases, the orientation information of the target is difficult to accurately provide, which poses a challenge to the applicability of the existing HRRP decoupling network. In addition, noise interference, radar hardware failure, unstable data transmission, and other factors that may change at any time can also cause changes in HRRP waveform. When there are such time-varying unknown interference factors in the environment, the existing decoupling network based on double supervision cannot effectively strip these disturbance factors, which affects the final recognition performance. Therefore, there is an urgent need for a more flexible and robust HRRP decoupling strategy to extract the inherent attribute features of the target, so as to achieve more accurate and reliable HRRP target recognition. SUMMARY

[0005] In view of the problems existing in the prior art, the present application provides a target recognition method based on an unsupervised dynamic-attribute feature decoupling network: first, obtain the HRRP data of the target within a period of time as an HRRP sequence; process the target HRRP sequence using a trained HRRP target recognition model to obtain a target recognition result.

[0006] The HRRP target recognition model includes an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network. The obtained target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network to obtain attribute features, and the trained HRRP target recognition network processes the attribute features to obtain a target recognition result.

[0007] Further preferably, the unsupervised dynamic-attribute feature decoupling network comprises a dynamic feature encoder, an attribute feature encoder, and a decoder.

[0008] The dynamic feature encoder comprises an LSTM layer and a multi-layer FC layer; the input HRRP sequence is evenly divided into non-overlapping windows, X i = [ X i 1 , ⋯ , X i n , ⋯ , X i N ] The data of each window and the corresponding mask are input into the LSTM layer for recursive processing to capture the dynamic dependence between time steps, and the output of the LSTM layer is mapped to a high-dimensional latent space by the FC layer to obtain the feature The mean and covariance parameters of the latent feature distribution are estimated by two independent FC layers, respectively. The mean and covariance parameters of the non-overlapping windows are stacked to obtain the overall mean and covariance parameters. The posterior distribution of the dynamic feature is constructed using the overall mean and covariance parameters. The posterior distribution of the dynamic feature is sampled to obtain the dynamic feature of the non-overlapping windows. [ X i 1 , ⋯ , X i n , ⋯ , X i N ] ;

[0009] The attribute feature encoder comprises an LSTM layer and a FC layer; the input HRRP sequence is randomly intercepted to obtain a subsequence of consecutive time steps and the corresponding mask The LSTM layer is input for recursive processing, and the output of the LSTM layer is mapped to a high-dimensional latent space by the FC layer to obtain the feature The mean estimate of the feature distribution is obtained by the FC layer , and a normal distribution with mean and standard deviation 1 is established as the posterior distribution of the attribute feature The posterior distribution of the attribute feature is sampled to obtain the attribute feature of the sequence ; ;

[0010] The input of the decoder is the dynamic feature and the attribute feature ​​​fusion feature , the fusion feature After processing by the FC layer and the LSTM layer, the mean value and the standard deviation of the generated sample distribution are estimated by two independent FC layers respectively ; the conditional likelihood distribution of the HRRP sequence sample is obtained by using the mean value and the standard deviation .

[0011] In a further preferred embodiment, the posterior distribution of the dynamic feature is constructed by using the overall mean value and the covariance matrix , wherein the covariance matrix is calculated by the following formula:

[0012]

[0013] wherein is the noise added on the diagonal line, and the sparse matrix constructed by using the covariance parameter .

[0014] In a further preferred embodiment, the HRRP target recognition network takes the attribute feature as input, and discriminates the target category by a classifier composed of multiple FC layers.

[0015] In a further preferred embodiment, the unsupervised dynamic-attribute feature decoupling network is trained in an unsupervised manner by using the multivariate HRRP sequence sample data, and the specific process is as follows:

[0016] (1) randomly initialize the parameters of the unsupervised dynamic-attribute feature decoupling network ;

[0017] (2) randomly sample a batch of data from the multivariate HRRP sequence sample training data and input them into the unsupervised dynamic-attribute feature decoupling network;

[0018] (3) the network performs forward propagation and calculates the total loss function value; the total loss function includes the evidence lower bound loss and the counterfactual regularization loss :

[0019]

[0020] wherein is the weight of the counterfactual regularization loss; during the network training process, the overall loss is minimized as the goal; ​​​

[0021] Loss of lower bound of evidence According to the formula

[0022] L ELBO = − 1 I ∑ i = 1 I  F A i , F D i [ l o g ( p ( X i ∣ F A i , F D i ) ) ] + β 1 I ∑ i = 1 I [ D K L ( q ( F D i | X i ) ∥ p ( F D ) ) + D K L ( q ( F A i | X i ) ∥ p ( F A ) ) ]

[0023] The calculation yielded, where This refers to the batch size of the HRRP sequence samples. − REG F A i , F D i [ l o g ( p ( X i ∣ F A i , F D i ) ) ] For negative log-likelihood terms, As weight, To measure the posterior distribution of dynamic features Its prior distribution The divergence between the differences To measure the posterior distribution of attribute features Its prior distribution The divergence of the differences between them;

[0024] Counterfactual regularization loss According to the formula

[0025] L  = 1 I ∑ i = 1 I Figure 1 F A i , F D i [ q ( F A i | X i ∗ ) q ( F A i ∗ | X i ∗ ) ]

[0026] Calculations show that Attribute features Counterfactual attributes In counterfactual samples The posterior probability difference; where the prior distribution of attribute features is... Random sampling is used to construct counterfactual attribute features. Dynamic features and counterfactual attributes Input decoder to generate counterfactual intervention samples ;

[0027] (4) Update the parameters of the unsupervised dynamic-attribute feature decoupling network using the backpropagation algorithm. ;

[0028] (5) Repeat (2) to (4) until the set iteration termination condition is met to complete the training of the unsupervised dynamic-attribute feature decoupling network.

[0029] A further optimized approach is to use a Gaussian process to model the prior distribution of dynamic features. Using a standard normal distribution Prior distribution as attribute features .

[0030] In a further preferred embodiment, the multivariate HRRP sequence sample data includes HRRP data with varying degrees of orientation loss.

[0031] Further preferred, the process of training the HRRP target recognition network with the multi-HRRP sequence sample data is as follows:

[0032] (1) freeze the parameters of the unsupervised dynamic-attribute feature decoupling network, and randomly initialize the parameters of the HRRP target recognition network ;

[0033] (2) input the attribute features obtained by processing a batch of data randomly sampled from the multi-HRRP sequence sample training data and the corresponding target class labels into the HRRP target recognition network;

[0034] (3) the network performs forward propagation and calculates the cross-entropy loss value:

[0035]

[0036] wherein, is the batch size of the HRRP sequence sample, is the number of sample categories, is the true category of the i-th sample, is the category probability distribution output by the HRRP target recognition network; in the training process, the goal is to minimize the cross-entropy function;

[0037] (4) update the parameters of the HRRP target recognition network through the back propagation algorithm ;

[0038] (5) repeat (2)~(4), and complete the training of the HRRP target recognition network after reaching the set iteration termination condition.

[0039] In addition, the present application also proposes an electronic device and a computer readable storage medium for implementing the above method:

[0040] An electronic device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the above target recognition method.

[0041] A computer readable storage medium has a computer program stored thereon, wherein the computer program is executed by a processor to implement the above target recognition method.

[0042] Advantages

[0043] The present application provides a target recognition method based on an unsupervised dynamic-attribute feature decoupling network, which has at least the following advantages:

[0044] ​(1) The feature decoupling network UDAFD-Net constructed in the application can realize decoupling of dynamic features and target attribute features in an unsupervised manner without any label guidance by asymmetric variational inference and multi-core Gaussian process prior modeling, break through the dependence of existing HRRP decoupling methods on supervision signals, and solve the feature decoupling problem under the condition that disturbance factors are not measurable.

[0045] (2) The application adds counterfactual regularization constraints in the UDAFD-Net. By constructing counterfactual samples to block the false correlation between dynamic features and attribute features, the complete separation of dynamic interference features and target inherent attribute features is promoted, and the physical interpretability of decoupled features is significantly improved.

[0046] (3) The application realizes independent optimization of dynamic factor stripping and target classification by the phased training strategy of the decoupling network UDAFD-Net and the recognition network HRRP-RecNet, and ensures that feature decoupling and target recognition can both reach their respective optimal states.

[0047] (4) The HRRP target recognition method proposed in the application can more accurately and robustly identify the type of the target under the condition that multiple dynamic factors such as bearing change and noise interference are superimposed and the values are unknown.

[0048] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0049] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, including the appended drawings, of the embodiments.

[0050] Figure 2 A research process step schematic diagram of a target recognition method based on an unsupervised dynamic-attribute feature decoupling network in an exemplary embodiment of the application is shown;

[0051] Figure 3 A structure schematic diagram of an HRRP target recognition model in an exemplary embodiment of the application is shown;

[0052] Figure 4 A schematic diagram of the relative position of a radar and a target in simulation data in an exemplary embodiment of the application is shown;

[0053] Figure 5 A schematic diagram of HRRP simulation data of a first type of target in an exemplary embodiment of the application is shown;

[0054] Figure 6 A schematic diagram of HRRP simulation data of a second type of target in an exemplary embodiment of the application is shown;

[0055] Figure 7 FIG. 3 shows a schematic diagram of a process of generating multi-element HRRP sequence samples in an exemplary embodiment of the present application;

[0056] Figure 8 FIG. 4 shows a schematic diagram of a process of generating counterfactual samples in an exemplary embodiment of the present application;

[0057] Figure 1 FIG. 5 shows a schematic diagram of t-SNE visualization results of original HRRP samples and attribute features under different perturbation conditions in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0058] Example implementations are now described in greater detail with reference to the figures. Like reference numerals can be used to denote like elements throughout the description. However, the example implementations can be implemented in various forms and should not be construed to limited to the examples set forth herein; rather, these provided examples are intended to be full and complete so that one of ordinary skill in the art can fully and completely understand the present application and the scope of the example implementations. The described features, structures, or characteristics can be combined in any suitable manner in one or more implementations.

[0059] Furthermore, the accompanying drawings are only schematic and are non-limiting detailed descriptions of the application, and not necessarily drawn to scale. Identical or similar elements are denoted by the same reference numerals throughout the description, not necessarily referring to the same or similar parts of the drawings, and a repetitive description of these elements can be omitted for the sake of brevity. Some of the blocks in the drawings are functional blocks, and can not necessarily correspond to physical or logical entities that are physically or logically separate. These functional blocks can be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0060] The present embodiment is mainly aimed at the demand for HRRP target recognition under the influence of perturbation factors, and needs to solve the problems of attribute-invariant feature extraction and target robust recognition, which are as follows:

[0061] First, the recognition accuracy of the existing HRRP target recognition method is easily affected by perturbation factors, and the anti-interference and robustness are not strong. Therefore, the attribute features that truly determine the target type need to be extracted for target recognition, the influence of interference factors needs to be removed, and the accuracy and robustness of target recognition need to be improved.

[0062] Second, the existing HRRP decoupling method relies on accurate perturbation values (such as azimuth data), and it is difficult to extract attribute features that truly determine the target type for target recognition in the actual situation where the perturbation factors are not measurable. Therefore, a new HRRP feature decoupling network needs to be designed to realize the decoupling and separation of perturbation features and attribute features in an unsupervised manner, break the dependence of the existing method on label signals, and extract attribute-invariant features in the case where the perturbation factors are not measurable.

[0063] Third, in view of the problem that the disturbance feature and the attribute feature are difficult to decouple and separate, a feature independence constraint needs to be added to the decoupling network to promote the mutual independence and effective separation of the two types of features, and a visualization means is used to illustrate the explainability of feature decoupling.

[0064] To this end, the embodiment provides a target recognition method based on an unsupervised dynamic-attribute feature decoupling network. First, HRRP data of a target in a period of time is acquired as an HRRP sequence, and a trained HRRP target recognition model is used to process the target HRRP sequence to obtain a target recognition result. Specifically, the HRRP target recognition model includes an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network. The acquired target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network to obtain attribute features, and the trained HRRP target recognition network processes the attribute features to obtain the target recognition result.

[0065] The research process of the method is shown below, as shown in Figure 2 , including the following steps:

[0066] Step S101: Construct an HRRP target recognition model, as shown in Figure 2 , the HRRP target recognition model includes an unsupervised dynamic-attribute feature decoupling network UTAFD-Net and an HRRP target recognition network HRRP-RecNet.

[0067] The unsupervised dynamic-attribute feature decoupling network UTAFD-Net includes a dynamic feature encoder, an attribute feature encoder, and a decoder. The dynamic feature encoder, the attribute feature encoder, and the decoder are all composed of a plurality of Long Short Term Memory (LSTM) layers and Fully Connected (FC) layers, which together constitute the main architecture of the UDAFD-Net.

[0068] The UDAFD-Net adopts an asymmetric variational inference architecture. The dynamic feature encoder and the attribute feature encoder model the posterior distribution of the dynamic feature and the attribute feature, respectively. The decoder is used to model the conditional likelihood distribution of the HRRP sample. The time correlation of the dynamic feature is modeled by a Gaussian process prior, and the stability of the attribute feature is maintained by combining a standard normal constraint. A counterfactual regularization mechanism is constructed to effectively block the false association between the dynamic feature and the attribute feature, and to ensure effective decoupling of the dynamic feature and the attribute feature.

[0069] The target recognition network HRRP-RecNet is composed of three FC layers. The first two FC layers take the ReLU function as the activation function, and each FC is followed by a Dropout layer to alleviate the overfitting of the network. The last FC layer takes the Softmax function as the activation function to convert the output into the form of probability distribution.

[0070] In particular:

[0071] Before inputting the HRRP data into the unsupervised dynamic-attribute feature decoupling network UTAFD-Net, the input HRRP data is segmented, including generating data for input to the dynamic feature encoder and data for input to the attribute feature encoder.

[0072] The HRRP data in a period of time is taken as a multivariate HRRP sequence sample , wherein represents the number of sampling time points, represents the number of distance units of the HRRP, , represents the number of HRRP sequence samples.

[0073] The process of generating data for input to the dynamic feature encoder is to divide into non-overlapping windows, X i = [ X i 1 , ⋯ , X i n , ⋯ , X i N ] For each window , the data of the window , wherein represents the size of the window.

[0074] The process of generating data for input to the attribute feature encoder is to randomly cut a continuous subsequence from with a length of .

[0075] The dynamic feature encoder aims to approximate the conditional posterior distribution of the dynamic feature of the HRRP sample . As shown in μ , the dynamic feature encoder is composed of an LSTM layer and a multi-layer FC layer. Since the dynamic feature needs to reflect the change pattern and association structure of the features between different windows of the time series, it is a high-dimensional feature with complex correlation between dimensions. Therefore, the posterior distribution of the dynamic feature is modeled as a multivariate normal distribution. The multivariate normal distribution can effectively describe the correlation and dependence between multi-dimensional variables through the covariance matrix, and the parameters of the mean and covariance matrix are parameterized by the encoder network. Next, the calculation process of the dynamic feature encoder is introduced.

[0076] The dynamic feature encoder takes the multivariate HRRP sequence samples of non-overlapping windows [ X i 1 , ⋯ , X i n , ⋯ , X i N ] Data as input, and the data of each window is input into the LSTM layer for recursive processing to capture the dynamic dependencies between time steps, and the generated output is represented as:

[0077] (1)

[0078] where is a mask used to filter out time steps with high HRRP distance cell data loss rates, is an activation function, is the hidden layer dimension of the LSTM. The output of the LSTM is mapped to a high-dimensional latent space through two FC layers, and the generated output is represented as:

[0079] (2)

[0080] where is an activation function, are the weight and bias terms of the two FC layers after the LSTM layer, respectively. Then, the mean and covariance parameters of the latent feature distribution are estimated through two independent FC layers, respectively:

[0081] (3)

[0082] (4)

[0083] where is an activation function used to ensure that the covariance parameter is positive, are the weight and bias terms of the two independent FC layers, respectively. Then, the mean and covariance parameters of the non-overlapping windows are stacked to obtain the overall mean and covariance parameter estimates:

[0084] μ F D i = [ μ i 1 , ⋯ , μ i n , ⋯ , μ i N ] , ℝ F D i ∈ γ d D × N (5)

[0085] γ F D i = [ γ i 1 , ⋯ , γ i n , … , γ i N ] , ℝ F D i ∈ Figure 2 D × 2 N (6)

[0086] To reduce computational complexity, the covariance parameter is used to construct a sparse matrix , which is represented by the formula: ​

[0087] (7)

[0088] where , is an upper triangular band matrix whose main diagonal and upper adjacent diagonal elements are taken from . The covariance matrix can be computed as follows:

[0089] (8)

[0090] where is a small noise added on the diagonal to enhance the numerical stability of the sparse matrix. Finally, the posterior distribution of the dynamic features is constructed using the mean and the covariance matrix as follows:

[0091] (9)

[0092] The posterior distribution of the dynamic features is sampled to obtain non-overlapping windows [ X i 1 , ⋯ , X i n , ⋯ , X i N ] of the corresponding dynamic features, denoted as F D i = [ f i 1 , ⋯ , f i n , ⋯ , f i N ] , , denotes the dimension of the dynamic features.

[0093] In addition, the prior distribution of the dynamic features Figure 2 is modeled using a Gaussian process, as shown in . A Gaussian process is defined as the joint distribution of a set of random variables, each corresponding to an input point, whose distribution is determined by the mean function and the covariance function measuring the correlation between inputs and . Here, the Gaussian process is used to construct a multivariate normal distribution as the prior distribution of the dynamic features, with mean 0 and covariance matrix generated by a multi-kernel function and its multi-scale properties.

[0094] ​Specifically, the RBF kernel and the Cauchy kernel are used to model the prior distribution of dynamic features through a Gaussian process. The RBF kernel is suitable for capturing smooth features with small local variations, while the Cauchy kernel is suitable for capturing features with strong jumps and long-range dependencies. By combining the characteristics of both, the complex dynamic patterns in the HRRP sequence can be captured more accurately. The calculation formulas for the RBF kernel and the Cauchy kernel are as follows:

[0095] (10)

[0096] (11)

[0097] in and Indicates a time window. , The length scale represents the time series and controls the smoothness of the kernel function. To capture the multi-scale characteristics of time series, the length scale of each kernel function is scaled according to the following rules:

[0098] (12)

[0099] in It refers to the number of scales, which is taken in this implementation scheme. The multi-scale kernel matrices of the two kernel functions are concatenated to form the covariance matrix. This can be expressed as a formula:

[0100] (13)

[0101] Finally, the prior distribution of the dynamic features is obtained. It can be represented as:

[0102] (14)

[0103] The attribute feature encoder aims to approximate HRRP samples. Attributes and characteristics conditional posterior distribution .like Figure 2 As shown, its structure is similar to that of the dynamic feature encoder, consisting of LSTM layers and FC layers. The difference lies in the fact that attribute features possess globally invariant properties. Theoretically, attribute features extracted from HRRP sequences across different time periods should remain consistent. Therefore, the attribute feature encoder extracts globally invariant properties from the original samples. Randomly select continuous Subsequence at each time step The attribute feature is extracted without considering the changing relationship between different time periods. The posterior distribution of the attribute feature is modeled as a standard normal distribution with a single variable, whose mean is parameterized by the attribute feature encoder, and the standard deviation is fixed at 1.

[0104] The intercepted HRRP sequence and the corresponding mask are input, and the feature extraction process of the LSTM layer and the first two FC layers is similar to the processing in the dynamic feature encoder. First, the LSTM layer is input for recursive processing, and the output of the LSTM layer is mapped to a high-dimensional latent space through two FC layers, and the output of the second FC layer is sent to the last FC layer to obtain the mean estimate of the feature distribution . The posterior distribution of the attribute feature is a normal distribution with a mean of and a standard deviation of 1, which can be represented as:

[0105] (15)

[0106] The posterior distribution of the attribute feature is sampled to obtain the attribute feature of the intercepted HRRP sequence , where d represents the dimension of the attribute feature.

[0107] In addition, as shown in Figure 2 , a standard normal distribution is also used as the prior distribution of the attribute feature .

[0108] The decoder takes the dynamic feature and the attribute feature as input, aiming to approximate the conditional likelihood distribution of the HRRP sequence. The decoder mainly consists of an LSTM and several FC layers.

[0109] Before entering the decoder, the dynamic feature and the attribute feature sampled from the dynamic feature posterior distribution and the attribute feature posterior distribution are first fused to form a new feature . To facilitate fusion, the attribute feature is expanded and copied to the same shape as the dynamic feature, and then spliced with the dynamic feature. This process can be represented as:

[0110] (16)

[0111] (17)

[0112] In the decoder, the input feature First, the input feature is processed by the first FC layer and batch normalization layer (BN), and the output is obtained as follows:

[0113] (18)

[0114] wherein is an activation function, is the output dimension of the first FC layer, is the weight term and bias term of the first FC layer in the decoder. Then, the output is obtained by an LSTM layer as follows:

[0115] (19)

[0116] Then, the output is obtained by the second FC layer as follows:

[0117] (20)

[0118] wherein is the output dimension of the second FC layer, is the weight term and bias term of the second FC layer in the decoder.

[0119] Finally, the mean and the standard deviation of the generated sample distribution are estimated by two independent FC layers, respectively, and are expressed by the following formula:

[0120] (21)

[0121] (22)

[0122] wherein are the weight term and bias term of the two independent FC layers in the decoder, respectively. The conditional likelihood distribution of the generated HRRP sequence sample can be expressed as follows:

[0123] (23)

[0124] The HRRP target recognition network HRRP-RecNet takes the attribute feature as input, and discriminates the target class through a classifier composed of three FC layers, as shown in Figure 3 . The first two FC layers take the ReLU function as the activation function, and each FC is followed by a Dropout layer to alleviate the overfitting of the network, and the last FC layer takes the Softmax function as the activation function to convert the output into the form of probability distribution.

[0125] Step S102: Constructing a HRRP simulation dataset containing two target models, pre-processing the HRRP simulation dataset into a multivariate HRRP sequence sample set in a sliding window manner, and dividing the multivariate HRRP sequence sample set into four different combinations of training and test sets to simulate four different degrees of azimuth missing;

[0126] Simulation in electromagnetic simulation software CST: Two different target aircraft models are established respectively, the radar simulation band is set to Ku band, the center frequency is 35 GHz, and the signal bandwidth is 300 MHz. For each target, as shown in Figure 4 , the radar surrounds the target for a whole circle, and samples the HRRP data of 3960 time points continuously. At each sampling point, the relative azimuth data between the target and the radar is different. The number of distance units of each HRRP is 101. Figure 5 and Figure 6 show the HRRP data samples of two types of target aircraft at a sampling point.

[0127] The simulated HRRP data set is segmented and pre-processed by using a sliding window (step size is 1 in this embodiment) strategy to generate a multivariate HRRP sequence sample set. As shown in Figure 6 , starting from the 1st sampling point, the HRRP data of the next 30 points is taken to generate the first multivariate HRRP sequence sample , where . In turn, for the th sampling point, the HRRP data of the next 30 points is taken to generate the th multivariate HRRP sequence sample . Since each type of target has 3960 HRRP data points, according to the division method, each type of target will generate 3931 multivariate HRRP sequence samples for training and testing of the proposed method. ELBO

[0128] The multivariate HRRP sequence sample set is divided in different training and test set division ways to simulate four different degrees of azimuth missing (Case1-Case4). As shown in Table 1, the 3931 HRRP sequence samples of each target are evenly divided into 10 azimuth intervals in order. Except that the first interval contains 394 samples, the remaining intervals contain 393 samples. Taking Case2 as an example, the training set of the network does not contain the HRRP samples of the 5th and 10th intervals, so the azimuth missing rate is 20%. As the missing rate increases, the difficulty of network recognition gradually increases.

[0129] Table 1 Four distribution situations of training set and test set of each sample

[0130] ​

[0131] Step S103: training the feature decoupling network UTAFD-Net in an unsupervised manner using the training sets in the above four cases (Case1-Case4) to enable the feature decoupling network UTAFD-Net to separate the dynamic features and the attribute features. The process includes the following steps:

[0132] Sub-step S1031: randomly initializing the parameters of the UDAFD-Net .

[0133] Sub-step S1032: randomly sampling a batch of data from each set of training data and feeding the data into the network, where the batch size is

[0134] Sub-step S1033: the network starts forward propagation and calculates the total loss function value. The UTAFD-Net is constructed based on the variational inference framework, and its loss function includes the Evidence Lower Bound (ELBO) loss. The ELBO loss is composed of a Kullback-Leibler (KL) divergence term and a Negative Log-Likelihood (NLL) term. To promote the mutual decoupling of the dynamic features and the attribute features and enable them to be better separated, an counterfactual regularization constraint is added to the ELBO loss. Therefore, the total loss function of the UDAFD-Net is composed of the ELBO loss and the counterfactual regularization loss. During the network training process, the goal is to minimize the total loss. The ELBO loss calculation formula is as follows:

[0135] L  = − 1 I ∑ i = 1 I Figure 2 F A i , F D i [ l o g ( p ( X i ∣ F A i , F D i ) ) ] + β 1 I ∑ i = 1 I [ D K L ( q ( F D i | X i ) ∥ p ( F D ) ) + D K L ( q ( F A i | X i ) ∥ p ( F A ) ) ] (24)

[0136] where is the batch size of the HRRP sequence samples, − Figure 7 F A i , F D i [ l o g ( p ( X i ∣ F A i , F D i ) ) ] is the NLL term, which is used to measure the ability of the model to generate the target sample under the given , and the HRRP generated by minimizing the NLL is as close as possible to the input HRRP. is the weight, which controls the proportion of the KL divergence in the ELBO loss. The two KL divergences and are used to measure the difference between the posterior distribution of the dynamic features and the attribute features and their corresponding prior distribution, and by minimizing the KL divergence, the posterior distribution of the features is encouraged to be close to the prior. In the above formula (24), the prior distribution of the attribute features is a standard normal distribution .

[0137] To promote the decoupling of dynamic features and attribute features, the embodiment adds counterfactual regularization loss to the UTAFD-Net to constrain the mutual independence of the two types of features, that is . As shown in REG and  , the counterfactual attribute feature is constructed by randomly sampling from the prior distribution of the attribute feature. Then the dynamic feature and the counterfactual attribute feature are input into the decoder to generate the counterfactual intervention sample . The calculation formula of the counterfactual regularization loss is:

[0138] L Figure 8 = 1 I ∑ i = 1 I Figure 8 F A i , F D i [ q ( F A i | X i ∗ ) q ( F A i ∗ | X i ∗ ) ] (25)

[0139] In the above formula, represents the posterior probability difference (i.e. likelihood ratio) of the attribute feature and the counterfactual attribute feature under the counterfactual sample . Minimizing this loss function forces so that the counterfactual sample does not contain any attribute information, that is, the dynamic feature does not contain any attribute information, that is, the spurious association between and is blocked. Through the counterfactual regularization constraint, the model can learn the mutually independent dynamic feature and attribute feature , significantly enhancing the feature decoupling capability.

[0140] The overall loss function of the UDAFD-Net is represented as:

[0141] (26)

[0142] where is the weight of the counterfactual regularization loss. During network training, the overall loss is minimized as the goal.

[0143] Sub-step S1034: update the parameters of the UDAFD-Net through the back propagation algorithm . In the t-th iteration, the update formula of the parameters is as shown in the following formula:

[0144] (27)

[0145] where denotes the learning rate of training UDAFD-Net. When the data of the training set is sampled in batches, one training iteration is completed.

[0146] Since the number of iterations of UDAFD-Net is , the steps S1032-S1034 are repeated times, and the training of UDAFD-Net is completed. After the training is completed, the UDAFD-Net network can effectively strip the disturbance factors in the HRRP sequence sample and extract the attribute-invariant features.

[0147] Step S104: Freeze the parameters of UTAFD-Net, and use the attribute features decoupled by UTAFD-Net as the input of the target recognition network HRRP-RecNet, and train HRRP-RecNet jointly with the target class label, so that HRRP-RecNet has the ability of target classification; specifically including the following steps:

[0148] Sub-step S1041: Randomly initialize the parameters of HRRP-RecNet .

[0149] Sub-step S1042: Randomly sample a batch of data from the training data , and send the attribute features decoupled by UDAFD-Net and the label of the batch of data to the network, and the batch size is .

[0150] Sub-step S1043: The network starts forward propagation and calculates the cross-entropy loss value, and the calculation formula of the cross-entropy loss is:

[0151] (28)

[0152] wherein, is the number of HRRP sequence samples, is the number of sample categories, is the real category (one-hot encoding) of the i-th sample , and is the probability distribution output by the HRRP target recognition network. In the training process, the minimum cross-entropy function is taken as the target.

[0153] Sub-step S1044: Update the parameters of HRRP-RecNet by the back propagation algorithm . In the t-th iteration, the update formula of the parameters is as follows:

[0154] ​ (29)

[0155] wherein is the learning rate of training HRRP-RecNet. When the data of the training set is sampled in batches, it means that one training iteration is completed.

[0156] Since the number of iterations of HRRP-RecNet is , the sub-steps S1042-S1044 are repeated times, that is, the training of HRRP-RecNet is completed.

[0157] After the training is completed, HRRP-RecNet has the target type recognition ability.

[0158] Step S105: input the test set with four different azimuth missing rates into the trained HRRP target recognition model to obtain the HRRP target recognition results under four different azimuth missing conditions; specifically including the following steps:

[0159] Sub-step S1051: input each HRRP sequence sample in the test set into the trained HRRP target recognition model, and the model outputs the probability distribution through the Softmax function, and then the class prediction result of the sample can be obtained by the following formula:

[0160] (30)

[0161] Sub-step S1052: compare the prediction result with the real label of the sample , and the result consistent with the real label means that the recognition is correct.

[0162] Sub-step S1053: divide the number of samples correctly recognized by the total number of samples in the test set to obtain the target correct recognition rate under this group of test set.

[0163] Since this step is tested by four different azimuth missing rate test sets, the HRRP target correct recognition rate under four different azimuth missing conditions (Case1-Case4) can be obtained.

[0164] ​Table 2 compares the correct recognition rate of the proposed method with six different HRRP recognition algorithms. The compared methods include direct HRRP target recognition methods and HRRP target recognition methods considering aspect missing. Direct HRRP target recognition methods use advanced neural networks to extract deep features and recognize target types without considering aspect missing and some possible interference factors. The methods participating in the comparison include Target-Aware Recurrent Attentional Network (TARAN), Voting Mechanism 1-D ResNet (VmResNet), Stacked CNN-bi-RNN with Attention Mechanism (SCRAM), and Target-Attentional Convolutional Neural Network (TACNN). HRRP target recognition methods considering aspect missing include Dual Self-Supervised Contrastive Learning method (DSSCL), Scattering Center neural Network (SCNet), and Type-Aspect Disentanglement Network (TADN). These compared methods all take single HRRP data as sample input to the network. In the experiment, TADN is trained and tested in two forms, TADN(√) indicates that the network is trained with correct aspect and type labels, and TADN(×) indicates that the aspect label is biased. TADN(×) is more suitable for the actual confrontation of difficult to obtain accurate target aspect. The table headers "Test set" and "ALL" represent testing on the test set and testing on all data, respectively.

[0165] From the results of Table 2, it can be seen that in Case 1 where no aspect missing occurs, the recognition performance of each method is not much different, with an accuracy difference of within 3%. As the missing rate increases, the recognition accuracy of TARAN, VmResNet, SCRAM, and TACNN methods decreases significantly, because they directly extract deep features from HRRP data for recognition. Deep features contain both attribute features that affect target types and pseudo-features related to target aspect factors in the training set. When the target aspect in the test set changes, the pseudo-features will cause the overall deep features to shift, affecting the target type recognition result.

[0166] While DSSCL and SCNet are robust to the azimuth missing case, the overall recognition accuracy is higher than the aforementioned several methods. When the missing rate increases from 0 to 20% and then to 50%, the recognition accuracy decreases by about 5% each time, and when the missing rate increases from 50% to 70%, the recognition accuracy decreases by about 10%. This shows that DSSCL and SCNet are difficult to effectively deal with high proportion of azimuth missing phenomenon. TADN is a type-azimuth decoupling network. In the ideal case that the target type and azimuth label can be accurately obtained, i.e. TADN(√), the network can remove the interference of azimuth factor and separate the attribute features that truly determine the target type for recognition, so its recognition performance is not easily affected by the azimuth missing case, and can achieve an effect comparable to the proposed method. But in the real case that it is difficult to provide accurate azimuth of the target, i.e. TADN(x), since the azimuth label is biased, the network is difficult to accurately separate the azimuth features and type features, and the deep features still contain the two confused features, so the recognition performance of the network is basically equivalent to that of TARAN, which is a direct recognition network.

[0167] In contrast, the method proposed in the embodiment adopts an unsupervised feature decoupling manner, which can separate pure attribute invariant features without any label guidance, breaking the dependence of TADN on the azimuth label, and being more suitable for real scenarios where the azimuth data is difficult to accurately obtain. Among all the methods compared, the overall recognition accuracy of the proposed method is higher, and even when the missing rate is as high as 70%, the recognition accuracy on the test set can reach 87.28%. The method proposed in the embodiment is more robust to azimuth missing, and when the missing rate increases from 0 to 70%, the recognition accuracy on the test set only decreases by 9.92% (97.20%→87.28%), while the best SCNet except TADN(√) decreases by 20.37% (96.06%→75.69%). The method proposed in the embodiment has stronger robustness and excellence for HRRP target recognition under the azimuth missing case, and can effectively alleviate the azimuth sensitivity problem in HRRP target recognition.

[0168] Table 2 Recognition accuracy of each method under different azimuth missing conditions

[0169]

[0170] Step S106: Add noise with different signal-to-noise ratios to one of the test sets, and randomly remove a certain number of HRRP distance units to simulate more types of disturbance factors. Test the trained HRRP target recognition model on this test set to obtain the HRRP target recognition result under the superposition of multiple disturbance factors. The specific process includes the following steps:

[0171] Sub-step S1061: First, the simulation of disturbance factors. Since the target azimuth in the test set of Case 2 and the target azimuth in the training set are completely different, that is, there is an azimuth change in the original test set of Case 2. Now continue to add Gaussian white noise with different signal-to-noise ratios to the test set to simulate the influence of environmental noise or electromagnetic interference that may change at any time, and the signal-to-noise ratio of the added noise ranges from 0 dB to 20 dB. In addition, the radar hardware failure causes the data segment to be lost or the data link to be unstable, causing part of the distance units not to be transmitted. By randomly truncating several distance units in the HRRP, the situation is simulated. Randomly truncate 5-20 distance units at any position, and set the values of these distance units to 0.

[0172] Sub-step S1062: Combine the azimuth change, noise interference and random truncation of these disturbance factors in the test set, input each sample in the test set into the trained HRRP target recognition model for testing, and obtain the HRRP target recognition result under the superposition of multiple disturbance factors.

[0173] Sub-step S1063: Similarly, compare the recognition result output by the model with the real label of the sample, and count the number of samples correctly recognized.

[0174] Sub-step S1064: Then divide the number of samples correctly recognized by the total number of samples in the test set, and the HRRP target correct recognition rate under the superposition of multiple disturbance factors can be obtained.

[0175] The recognition performance of each method under the influence of these dynamic factors is tested on the test set of Case 2, and the experimental results are shown in Table 3. The letters A, B and C respectively represent the three disturbance situations of azimuth change, random truncation and noise interference in the test set.

[0176] From the experimental results, it can be seen that the four methods of TARAN, VmResNet, SCRAM and TACNN are most affected by the disturbance factors. With the addition of noise and truncation factors, the recognition accuracy of each method decreases significantly. This is because the mixed effects of disturbance factors cause the deep features to deviate from the attribute features of the target. The dynamic disturbance factors of the test set may differ from those of the training set. The network may only fit the deep features under a specific disturbance pattern during the training phase, but it fails to learn the invariance features that have a fundamental impact on the target class. This limits its generalization ability when facing unknown and changing interference. DSSCL, SCNet and TADN, which consider the case of azimuth absence, do not show obvious performance improvement when dealing with multiple disturbance factors. They can indeed reduce the impact of azimuth changes to some extent, but the pseudo-features generated by other dynamic factors (such as noise and truncation) are still mixed in the deep features, and the confusion effect still exists. With the changes in these dynamic factors in the test set, the deep features will dynamically shift, causing the mapping relationship between the deep features and the target class to mismatch, thereby affecting the recognition effect.

[0177] The method proposed in this embodiment can separate the dynamic factors that cause interference from the attribute invariance factors of the target, extract the attribute invariance features as deep features for target recognition, and make the recognition result less affected by the disturbance factors in the system. Since it uses an unsupervised decoupling method, it can achieve accurate decoupling even when the components and values of the dynamic interference factors are unknown, which is more suitable for practical scenarios where disturbance factors are complex and unpredictable. The method proposed in this embodiment has the highest recognition accuracy. Even under the influence of three disturbance factors (A+B+C), the accuracy of the method proposed in this embodiment can reach 88.04%, which is 13.49% higher than that of TADN(√), which is the second best method. The method proposed in this embodiment is less affected by interference and has more robust recognition performance. After adding random truncation on the basis of 20% azimuth absence (A+B), the accuracy only decreases by 3.43% (94.40%→90.97%). After adding changing noise (A+C), the accuracy only decreases by 4.90% (94.40%→89.50%). When the three dynamic factors are superimposed together (A+B+C), the accuracy only decreases by 6.36% (94.40%→88.04%). The proposed method can more stably and accurately identify the type of HRRP target under unknown disturbance factors.

[0178] Table 3 Comparison of recognition accuracy of each method after adding different levels of disturbance

[0179]

[0180] Step S107: Visualize the HRRP samples under the condition of multiple disturbance factors superposition in step S106 and the attribute features decoupled by UTAFD-Net using t-SNE technology, and verify the feature decoupling effect of UTAFD-Net from the feature level.

[0181] Figure 8 (a)-(b), (c)-(d), (e)-(f) respectively show the visualization results of the original HRRP samples and attribute features in A, A+B, A+B+C in Table 3. Red and green represent the first and second types of targets respectively, and triangles and circles represent the azimuth intervals "5" and "10" in the test set of Case2 respectively.

[0182] In the A case, only the azimuth factor changes in the test set, which can be divided into two intervals "5" and "10" according to the approximate azimuth interval. Since the two intervals are far apart, the azimuth difference of the samples in the interval will be more obvious, so from the perspective of the distribution of the original HRRP samples, Figure 8 The azimuth distribution difference of the targets can indeed be seen from the original HRRP sample distribution of (a), and the samples represented by the circles on the left and the triangles on the right are clustered together. However, the samples represented by red and green are not easy to distinguish. This illustrates from the feature space distribution level that the HRRP target recognition will be affected by the change of azimuth, and the classification boundary of the two types of samples fitted in one azimuth interval may not be suitable for another interval. From the perspective of attribute features, Figure 8 The samples in (b) have obvious type distinguishability, and the inter-class distance between the two types of samples represented by the two colors is large, so the two types of samples are easy to distinguish, which shows that the UTAFD-Net extracts attribute features related only to the type of target, and the target recognition effect is no longer affected by the azimuth factor.

[0183] In the A+B case, random truncation is added to the test set in addition to the change of azimuth. As shown in Figure 8 (c), the HRRP samples still have certain azimuth distinguishability as a whole, and the circles on the left and the triangles on the right each represent an azimuth interval. However, due to the addition of random truncation, the samples of different types represented by the two colors are more mixed together and are more difficult to distinguish than Figure 8 (a). After feature decoupling, Figure 8 the same type of samples represented by the same color gradually gather together and have obvious type distinguishability in (d). However, compared with Figure 8 (b), Figure 8 the inter-class distance between the two types of samples in (d) is significantly reduced, and the samples near the classification boundary become more, which shows that the random truncation interference increases the difficulty of feature decoupling to some extent, but the method can still effectively maintain the core separability of the attribute features.

[0184] In the case of A+B+C, the three dynamic factors of azimuth variation, random truncation and noise interference are contained in the test set. As shown in Fig. ​ (e), the HRRP samples are randomly distributed in space and have no distinguishability, which shows that the coupling of multiple disturbance factors further increases the difficulty of classification. After feature decoupling, as shown in Fig. ​ (f), the UTAFD-Net can correctly classify most samples in the figure according to the attribute features. This shows that the proposed decoupling method can still extract robust type discrimination features under multiple dynamic interferences. Even if the original HRRP data loses distinguishability due to compound interference, the attribute features after decoupling can still maintain the core type information, significantly improving the reliability and accuracy of HRRP target recognition.

[0185] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those of ordinary skill in the art can make changes, modifications, replacements and variations to the above embodiments without departing from the principles and purposes of the present application within the scope of the present application.

Claims

1. A target recognition method based on an unsupervised dynamic-attribute feature decoupling network, characterized in that: The HRRP data of the target in a period of time is acquired as an HRRP sequence; the HRRP sequence of the target is processed by using a trained HRRP target identification model to obtain a target identification result; The HRRP target identification model comprises an unsupervised dynamic-attribute feature decoupling network and an HRRP target identification network; the acquired HRRP sequence of the target is processed by using the trained unsupervised dynamic-attribute feature decoupling network to obtain attribute features, and the trained HRRP target identification network is used to process the attribute features to obtain the target identification result. 2.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 1, characterized in that: The unsupervised dynamic-attribute feature decoupling network comprises a dynamic feature encoder, an attribute feature encoder and a decoder. The dynamic feature encoder consists of an LSTM layer and multiple FC layers; the input HRRP sequence Divide into equal parts After a non-overlapping window Data for each window and the corresponding mask First, the input is processed recursively by an LSTM layer to capture the dynamic dependencies between time steps. The output of the LSTM layer... Features obtained by mapping from the FC layer to a high-dimensional latent space The mean of the latent feature distribution is then estimated separately through two independent FC layers. Covariance parameter ,Will The mean and covariance parameters of each non-overlapping window are stacked together to obtain the overall mean. Covariance parameter ; using the overall mean Covariance parameter Constructing the posterior distribution of dynamic features posterior distribution of dynamic features Sampling was performed to obtain Non-overlapping windows Corresponding dynamic features ; The attribute feature encoder consists of an LSTM layer and an FC layer; the input HRRP sequence Randomly select continuous Subsequence at each time step and the corresponding mask First, the input is processed recursively by the LSTM layer, and then the output of the LSTM layer... Features obtained by mapping from the FC layer to a high-dimensional latent space Then, the mean estimate of the feature distribution is obtained through the FC layer. Establish the mean as A normal distribution with a standard deviation of 1 is used as the posterior distribution of attribute features. ; Posterior distribution of attribute features Sampling is performed to obtain the sequence. Attributes and characteristics ; The input to the decoder is dynamic features. and attribute features Fusion characteristics The fusion feature After processing with FC and LSTM layers, the mean of the generated sample distribution is estimated using two separate FC layers. and standard deviation Using the mean and standard deviation Obtain the conditional likelihood distribution of HRRP sequence samples . 3.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: Utilizing the overall mean and covariance matrix Constructing the posterior distribution of the dynamic features where the covariance matrix is computed with the following formula: wherein is the noise added on the diagonal, is the covariance parameter is the sparse matrix constructed. 4.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 1, characterized in that: The HRRP target recognition network identifies the target with attribute features As input, discrimination of the target class is performed by a classifier consisting of multiple FC layers. 5.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: The unsupervised dynamic-attribute feature decoupling network is trained in an unsupervised manner by using multi-element HRRP sequence sample data, and the specific process is as follows: (1) randomly initializing parameters of the unsupervised dynamic-attribute feature decoupling network ; (2) Randomly sample a batch of data from the multi-element HRRP sequence sample training data into the unsupervised dynamic-attribute feature decoupling network; (3) The network performs forward propagation and calculates the total loss function value; the total loss function including the evidence lower bound loss and the counterfactual regularization loss : wherein is a weight for the counterfactual regularization loss; during network training, the objective is to minimize the overall loss; Evidence lower bound loss According to the formula is computed as is the batch size of HRRP sequence samples, is the negative log-likelihood term, is the weight, is the divergence of the difference between the metric dynamic feature posterior distribution and its prior distribution , is the divergence of the difference between the metric attribute feature posterior distribution and its prior distribution . Counterfactual regularization loss According to the formula The calculation is as follows, The attribute feature is The counterfactual attribute feature is The posterior probability difference under the counterfactual sample ; wherein the prior distribution of the attribute feature is randomly sampled, and the counterfactual attribute feature is constructed ; the dynamic feature and the counterfactual attribute feature are input into the decoder to generate a counterfactual intervention sample ; (4) updating parameters of the unsupervised dynamic-attribute feature decoupling network through a back propagation algorithm ; (5), repeating (2)-(4), and completing the training of the unsupervised dynamic-attribute feature decoupling network after reaching the set iteration termination condition. 6.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 5, characterized in that: Modeling the prior distribution of dynamic features using Gaussian processes ; adopting a standard normal distribution as the prior distribution of attribute features . 7.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 5, characterized in that: The multi-element HRRP sequence sample data comprises HRRP data with different degrees of azimuth missing. 8.The target recognition method based on the unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: The process of training the HRRP target identification network by using the multi-element HRRP sequence sample data is as follows: (1) Parameters of the frozen unsupervised dynamic-attribute feature decoupling network, randomly initialize the parameters of the HRRP target recognition network ; (2), the attribute features obtained after a batch of data randomly sampled from the multi-element HRRP sequence sample training data and the corresponding target class labels are input into the HRRP target recognition network after being processed by the unsupervised dynamic-attribute feature decoupling network. (3), the network is forward propagated and the cross-entropy loss value is calculated: wherein, is a batch size of HRRP sequence samples, is a number of sample classes, is a true class of the th sample, is a class probability distribution output by the HRRP target recognition network; during the training process, a minimization of this cross-entropy function is targeted; (4) updating the parameters of the HRRP target recognition network through a back propagation algorithm ; (5), repeating (2)-(4), and completing the training of the HRRP target identification network after reaching the set iteration termination condition.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the target identification method according to any one of claims 1-8 when executing the program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the target identification method according to any one of claims 1-8 when executed by the processor.

Images