A target recognition method based on an unsupervised dynamic-attribute feature decoupling network

Abstract

The application provides a target recognition method based on an unsupervised dynamic-attribute feature decoupling network, acquires HRRP data of a target in a period of time as an HRRP sequence, processes the target HRRP sequence by using a trained HRRP target recognition model to obtain a target recognition result, and the HRRP target recognition model comprises an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network; the 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. The HRRP target recognition method provided in the application can more accurately and robustly recognize the type of the target under the condition that various dynamic factors such as bearing changes and noise interference are superimposed and the values are unknown.

Description

Technical Field

[0001] This invention relates to the field of target identification technology, specifically to a target identification method based on an unsupervised dynamic-attribute feature decoupling network. Background Technology

[0002] In maritime law enforcement and search and rescue operations, target search and identification using search radar is a primary method in harsh weather conditions. Because high-resolution range profiles (HRRPs) contain discriminative features such as the target's material properties and geometric structure, they are widely used for automatic radar target identification, greatly aiding in rapid target detection. HRRPs consist of the amplitude of the sum of the complex echo vectors of the target scattering body in each range cell; they represent the projection of the complex echo from the target's scattering center onto the radar's line-of-sight (LOS).

[0003] In recent years, due to the development of deep learning technology, significant progress has been made in HRRP target recognition research based on deep neural networks. However, many difficulties and challenges remain, including the azimuth sensitivity of HRRP. Because changes in the relative azimuth of the radar and the target alter the projection distribution of the scattering points, the HRRP waveform of the same target exhibits significant differences under different observation azimuths. Recognition models are prone to misidentification when they have not seen such waveforms before. Many existing HRRP target recognition methods typically build their models under the assumption of complete azimuth samples, assuming the model has seen HRRP data from all azimuths of the target, thus possessing good recognition performance. However, in practical non-cooperative target recognition scenarios, due to the uncontrollable nature of target motion and observation constraints, the available HRRP data often suffers from severe azimuth gaps. Therefore, how to achieve omnidirectional generalized recognition based on partial azimuth samples is of great practical significance.

[0004] Currently, scholars have conducted research on this issue, mainly focusing on data augmentation and model augmentation. Data augmentation primarily involves synthesizing HRRP data with missing azimuth using generative adversarial networks or variational coding structures; however, there is a lack of effective evaluation criteria for the physical plausibility and distribution consistency of the generated samples. Model augmentation includes meta-learning cross-azimuth adaptation methods, transfer feature mapping models, and self-supervised contrastive frameworks. These methods perform well with moderate missing rates, but their performance significantly degrades with high proportions of missing azimuth. Notably, inspired by decoupled representation learning, some scholars have proposed a type-azimuth decoupled network that separates the target's category and azimuth features through dual-branch supervised learning, and then uses the separated category features for target recognition, providing a new approach for HRRP target recognition with missing azimuth. However, this method relies on dual supervision signals, requiring the simultaneous acquisition of target category labels and accurate azimuth information. The target's azimuth information is jointly determined by the radar's LOS direction and the target's attitude. However, in real-world scenarios, environmental factors cause complex and variable target attitudes, making it difficult to accurately obtain the target's attitude data even if the radar's LOS direction relative to the target point can be obtained. Therefore, in such situations, accurate target location information is difficult to provide, posing an applicability challenge to existing HRRP decoupling networks. Furthermore, factors such as constantly changing noise interference, radar hardware failures, and unstable data transmission can also cause variations in the HRRP waveform. When such time-varying, unknown interference factors exist in the environment, the difficulty in obtaining the specific values ​​of these interference factors prevents existing dual-supervision-based decoupling networks from effectively removing these disturbances. This results in the network's inability to extract pure target category features, thus affecting the final recognition performance. Therefore, a more flexible and robust HRRP decoupling strategy is urgently needed to extract the inherent attribute features of the target, thereby achieving more accurate and reliable HRRP target recognition. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a target identification method based on an unsupervised dynamic-attribute feature decoupling network: First, HRRP data of the target over a period of time is acquired as an HRRP sequence; then, a trained HRRP target recognition model is used to process the target HRRP sequence to obtain the target recognition result.

[0006] 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.

[0007] In a further preferred embodiment, the unsupervised dynamic-attribute feature decoupling network includes a dynamic feature encoder, an attribute feature encoder, and a decoder;

[0008] 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 X i = [ X i 1 , ⋯ , X i n , ⋯ , X i N ] 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 [ X i 1 , ⋯ , X i n , ⋯ , X i N ] Corresponding dynamic features ;

[0009] 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 ;

[0010] 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 .

[0011] Further optimization schemes utilize the overall mean. Covariance Matrix Constructing the posterior distribution of dynamic features The covariance matrix Calculate using the following formula:

[0012]

[0013] in The noise was added diagonally. To utilize the covariance parameter The constructed sparse matrix.

[0014] In a further preferred embodiment, the HRRP target recognition network uses attribute features As input, the target category is determined by a classifier consisting of multiple fully connected (FC) layers.

[0015] A further optimized approach involves training an unsupervised dynamic-attribute feature decoupling network using multivariate HRRP sequence sample data in an unsupervised manner. The specific process is as follows:

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

[0017] (2) Training data from multivariate HRRP sequence samples A batch of data is randomly sampled and fed into an unsupervised dynamic-attribute feature decoupling network.

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

[0019]

[0020] in The weights are used for the counterfactual regularization loss; during network training, the goal is to minimize the overall loss.

[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. −  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 REG = 1 I ∑ i = 1 I  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] A further preferred approach involves training the HRRP target recognition network using multivariate HRRP sequence sample data 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) Training data from multivariate HRRP sequence samples The attribute features obtained by processing a batch of randomly sampled data through an unsupervised dynamic-attribute feature decoupling network, along with the corresponding target category labels, are input into the HRRP target recognition network.

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

[0035]

[0036] in, This refers to the batch size of the HRRP sequence samples. It is the number of sample categories. It is the first The true category of each sample This is the category probability distribution output by the HRRP target recognition network; during training, the objective is to minimize this cross-entropy function.

[0037] (4) Update the parameters of the HRRP target recognition network using the backpropagation algorithm. ;

[0038] (5) Repeat (2) to (4) until the set iteration termination condition is met to complete the training of the HRRP target recognition network.

[0039] Furthermore, the present invention also proposes an electronic device and a computer-readable storage medium for implementing the above method:

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

[0041] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described target identification method.

[0042] Beneficial effects

[0043] This invention provides a target identification method based on an unsupervised dynamic-attribute feature decoupling network, which has at least the following beneficial effects:

[0044] (1) The feature decoupling network UDAFD-Net constructed in this invention can achieve decoupling of dynamic features and target attribute features in an unsupervised manner without any label guidance by using asymmetric variational inference and multi-kernel Gaussian process prior modeling. This breaks through the dependence of existing HRRP decoupling methods on supervision signals and solves the feature decoupling problem under the condition that the perturbation factors are unmeasurable.

[0045] (2) This invention adds counterfactual regularization constraints to UDAFD-Net. By constructing counterfactual samples, the false correlation between dynamic features and attribute features is blocked, promoting the complete separation of dynamic interference features and the inherent attribute features of the target, and significantly improving the physical interpretability of decoupled features.

[0046] (3) This invention achieves independent optimization of dynamic factor stripping and target classification by using a phased training strategy to decouple the UDAFD-Net network and the recognition network HRRP-RecNet, ensuring that feature decoupling and target recognition can reach their respective optimal states.

[0047] (4) The HRRP target identification method proposed in this invention can more accurately and robustly identify the type of target when multiple dynamic factors such as orientation change and noise interference are superimposed and the values ​​are unknown.

[0048] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0049] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0050] Figure 1 This diagram illustrates the research process steps of the target identification method based on an unsupervised dynamic-attribute feature decoupling network in an exemplary embodiment of the present invention.

[0051] Figure 2 This diagram illustrates the structure of the HRRP target recognition model in an exemplary embodiment of the present invention.

[0052] Figure 3 A schematic diagram showing the relative positions of the radar and the target using simulated data in an exemplary embodiment of the present invention;

[0053] Figure 4 A schematic diagram showing HRRP simulation data of a first type of target in an exemplary embodiment of the present invention;

[0054] Figure 5 A schematic diagram showing HRRP simulation data of the second type of target in an exemplary embodiment of the present invention;

[0055] Figure 6 This diagram illustrates the multivariate HRRP sequence sample generation process in an exemplary embodiment of the present invention.

[0056] Figure 7 A schematic diagram illustrating the counterfactual sample generation process in an exemplary embodiment of the present invention is shown;

[0057] Figure 8 This diagram illustrates the t-SNE visualization results of the original HRRP samples and attribute features under different perturbation conditions in an exemplary embodiment of the present invention. Detailed Implementation

[0058] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0059] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0060] This embodiment is mainly designed to address the target identification requirements of HRRP under the influence of perturbation factors. It needs to solve the problems of target attribute invariant feature extraction and robust target recognition, as detailed below:

[0061] First, to address the issue that the accuracy of existing HRRP target recognition methods is easily affected by perturbation factors and lacks strong anti-interference and robustness, it is necessary to extract the attribute features that truly determine the target type for target recognition, remove the influence of perturbation factors, and improve the accuracy and robustness of target recognition.

[0062] Second, to address the problem that existing HRRP decoupling methods rely on precise perturbation values ​​(such as azimuth data), making it difficult to extract the truly defining attribute features for target identification when perturbation factors are unpredictable, a novel HRRP feature decoupling network needs to be designed to decouple perturbation features and attribute features in an unsupervised manner, breaking the dependence of existing methods on label signals and enabling the extraction of attribute-invariant features even when perturbation factors are unpredictable.

[0063] Third, to address the difficulty in decoupling and separating perturbation features and attribute features, feature independence constraints need to be added to the decoupling network to ensure that these two types of features are independent and effectively separated. At the same time, visualization methods should be used to illustrate the interpretability of feature decoupling.

[0064] Therefore, this embodiment provides a target identification method based on an unsupervised dynamic-attribute feature decoupling network. First, HRRP data of the target over a period of time is obtained as an HRRP sequence. Then, a trained HRRP target recognition model is used to process the target HRRP sequence to obtain the target identification result. Specifically, the HRRP target recognition model includes an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network. After the obtained target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network, attribute features are obtained. The trained HRRP target recognition network then processes the attribute features to obtain the target identification result.

[0065] The research process of this method is given below, such as Figure 1 As shown, it includes the following steps:

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

[0067] The unsupervised dynamic-attribute feature decoupling network UDAFD-Net includes a dynamic feature encoder, an attribute feature encoder, and a decoder. These three structures are each composed of several Long Short Term Memory (LSTM) layers and fully connected (FC) layers, collectively forming the main architecture of UDAFD-Net.

[0068] UDAFD-Net employs an asymmetric variational inference architecture. The dynamic feature encoder and attribute feature encoder model the posterior distributions of dynamic and attribute features, respectively, while the decoder models the conditional likelihood distribution of HRRP samples. The temporal correlation of dynamic features is modeled through a Gaussian process prior, and the stability of attribute features is maintained by incorporating standard normality constraints. A counterfactual regularization mechanism is constructed to effectively block spurious associations between dynamic and attribute features, ensuring effective decoupling between them.

[0069] The target recognition network HRRP-RecNet consists of three fully connected (FC) layers. The first two FC layers use ReLU as the activation function, and each FC layer is followed by a Dropout layer to alleviate overfitting. The last FC layer uses Softmax as the activation function, transforming the output into a probability distribution.

[0070] Specifically:

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

[0072] Using HRRP data over a period of time as a multivariate HRRP sequence sample ,in Indicates the number of sampling time points. This indicates the number of distance cells in HRRP. , This indicates the number of HRRP sequence samples.

[0073] The process of generating data for input to the dynamic feature encoder is as follows: Divide into equal parts A non-overlapping window, X i = [ X i 1 , ⋯ , X i n , ⋯ , X i N ] For each window Window data ,in Indicates the size of the window.

[0074] The process of generating data for input to the attribute feature encoder is as follows: from Randomly select a segment of length. continuous subsequences .

[0075] The dynamic feature encoder is designed to approximate HRRP samples. dynamic characteristics conditional posterior distribution .like Figure 2 As shown, the dynamic feature encoder consists of LSTM layers and multiple FC layers. Since dynamic features need to reflect the changing patterns and correlation structures of features across different windows in a time series, and are high-dimensional features with complex correlations between their dimensions, the posterior distribution of dynamic features is modeled as a multivariate normal distribution. The multivariate normal distribution can effectively describe the correlations and dependencies between multidimensional variables through the covariance matrix, and its mean and covariance matrix parameters are parameterized by the encoder network. The computation process of the dynamic feature encoder is described below.

[0076] The dynamic feature encoder uses multivariate HRRP sequence samples of Non-overlapping windows [ X i 1 , ⋯ , X i n , ⋯ , X i N ] Data is used as input, with data from each window. The input to the LSTM layer is processed recursively to capture the dynamic dependencies between time steps, and the generated output is represented as follows:

[0077] (1)

[0078] in It's a mask used to filter time steps where the HRRP distance cell data loss rate is too high. It is an activation function. This is the hidden layer dimension of the LSTM. The output of the LSTM. After being mapped to a high-dimensional latent space through two fully connected (FC) layers, the generated output is represented as follows:

[0079] (2)

[0080] in It is an activation function. These represent the weights and biases of the two fully connected (FC) layers following the LSTM layer. Then, the mean of the latent feature distribution is estimated using two independent FC layers. Covariance parameter :

[0081] (3)

[0082] (4)

[0083] in This is the activation function used to ensure that the covariance parameter is positive. These are the weight and bias terms for two independent fully connected (FC) layers. Then... The mean and covariance parameters of the non-overlapping windows are stacked together to obtain the overall mean and covariance parameter estimates:

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

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

[0086] To reduce computational complexity, the covariance parameter is used. To construct sparse matrices This can be expressed as a formula:

[0087] (7)

[0088] in , It is an upper triangular matrix, whose main diagonal and upper adjacent diagonal are... Each element is taken from Covariance matrix The following formula can be used to calculate:

[0089] (8)

[0090] in Small noise is added on the diagonal to enhance the numerical stability of the sparse matrix. Finally, the mean is used. Covariance Matrix To construct the posterior distribution of dynamic features As shown in the following formula:

[0091] (9)

[0092] Posterior distribution of dynamic features Sampling was performed to obtain Non-overlapping windows [ X i 1 , ⋯ , X i n , ⋯ , X i N ] The corresponding dynamic features are denoted as F D i = [ f i 1 , ⋯ , f i n , ⋯ , f i N ] , , Dimensions representing dynamic features.

[0093] In addition, such as Figure 2 As shown, a Gaussian process is used to model the prior distribution of dynamic features. Gaussian process It is defined as the joint distribution of a set of random variables, each corresponding to an input point, and its distribution is given by the mean function. Sum of covariance functions Determined, used to measure input and The correlation between them. Here, a Gaussian process is used to construct a multivariate normal distribution as the prior distribution of dynamic features. The mean of the multivariate normal distribution is 0, and the covariance matrix is ​​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 This method extracts attribute features without considering changes across different time periods. The posterior distribution of the attribute features is modeled as a univariate standard normal distribution, with its mean parameterized by the attribute feature encoder and its standard deviation fixed at 1.

[0104] With the extracted HRRP sequence and the corresponding mask The feature extraction process of the LSTM layer and the first two FC layers is similar to that in a dynamic feature encoder. The input is first processed recursively by the LSTM layer, and the output of the LSTM layer... Mapped to a higher-dimensional latent space through two fully connected (FC) layers, the output of the second FC layer... The mean estimate of the feature distribution is obtained by feeding it into the last fully connected layer. Posterior distribution of attribute features It is a mean A normal distribution with a standard deviation of 1 can be expressed as:

[0105] (15)

[0106] Posterior distribution of attribute features Sampling was performed to obtain the truncated HRRP sequence. Attributes and characteristics , The dimension representing the attribute feature.

[0107] In addition, such as Figure 2 As shown, a standard normal distribution is also used. Prior distribution as attribute features .

[0108] The decoder features dynamic characteristics and attribute features As input, the conditional likelihood distribution of the HRRP sequence is intended to be approximated. The decoder mainly consists of LSTM and several FC layers.

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

[0110] (16)

[0111] (17)

[0112] In the decoder, the input features First, after processing by the first fully connected (FC) layer and the batch normalization (BN) layer, the output is obtained:

[0113] (18)

[0114] in For activation function, It is the output dimension of the first FC layer. These are the weights and biases of the first fully connected (FC) layer in the decoder. Then, after passing through an LSTM layer, the output is obtained:

[0115] (19)

[0116] After processing by the second FC layer, the output is obtained:

[0117] (20)

[0118] in It is the output dimension of the second FC layer. These are the weight and bias terms for the second FC layer in the decoder.

[0119] Finally, the mean of the generated sample distribution is estimated using two independent fully connected (FC) layers. and standard deviation This can be expressed as a formula:

[0120] (twenty one)

[0121] (twenty two)

[0122] in These represent the weights and biases of two independent fully connected (FC) layers in the decoder. The conditional likelihood distribution of the generated HRRP sequence samples. It can be represented as:

[0123] (twenty three)

[0124] The HRRP target recognition network HRRP-RecNet uses attribute features As input, the target category is determined by a classifier consisting of three fully connected (FC) layers, such as... Figure 2 As shown, the first two fully connected (FC) layers use the ReLU function as the activation function, followed by a Dropout layer to alleviate overfitting. The last FC layer uses the Softmax function as the activation function, converting the output into a probability distribution.

[0125] Step S102: Construct an HRRP simulation dataset containing two target models. Use a sliding window method to preprocess the HRRP simulation dataset into a multivariate HRRP sequence sample set. Divide the multivariate HRRP sequence sample set into four different training and test set combinations to simulate four different degrees of orientation loss.

[0126] Simulations were performed in the electromagnetic simulation software CST: two different target aircraft models were established, and the radar simulation band was set to Ku-band, with a center frequency of 35 GHz and a signal bandwidth of 300 MHz. For each type of target, such as Figure 3 As shown, the radar orbits the target in all directions, continuously sampling HRRP data at 3960 time points. At each sampling point, the relative azimuth data between the target and the radar is different. Each HRRP has 101 range cells. Figure 4 and Figure 5 The paper presents HRRP data examples for two types of target aircraft at a certain sampling point.

[0127] A sliding window strategy (with a step size of 1 in this embodiment) is used to segment and preprocess the simulated HRRP dataset to generate a multivariate HRRP sequence sample set. For example... Figure 6 As shown, starting from the first sampling point, HRRP data from 30 consecutive points are used to generate the first multivariate HRRP sequence sample. ,in . Continuing in sequence, for the first... The sampling point is used to generate the HRRP data from the next 30 consecutive points. A multivariate HRRP sequence sample Since there are 3960 sampling points of HRRP data for each target class, according to Figure 6 The classification method generates 3931 multivariate HRRP sequence samples for each target class, which are used for training and testing of the proposed method.

[0128] The multivariate HRRP sequence sample set was divided using different training and testing set partitioning methods to simulate four different degrees of orientation missing (Case 1-Case 4). As shown in Table 1, in this step, the 3931 HRRP sequence samples of each target class were divided into 10 orientation intervals in sequence. Except for the first interval containing 394 samples, the other intervals each contained 393 samples. Taking Case 2 as an example, the network's training set did not include HRRP samples from the 5th and 10th intervals, so the orientation missing rate was 20%. As the missing rate increased, the recognition difficulty of the network gradually increased.

[0129] Table 1. Four distributions of the training and test sets for each class of samples.

[0130]

[0131] Step S103: Using the training sets from the four cases (Case 1-Case 4) above, train the feature decoupling network UTAFD-Net in an unsupervised manner, enabling UTAFD-Net to separate dynamic features and attribute features. This includes the following process:

[0132] Sub-step S1031: Randomly initialize the parameters of UDAFD-Net .

[0133] Sub-step S1032: From each set of training data Randomly sample a batch of data Send to the network, This refers to the batch size.

[0134] Sub-step S1033: The network begins forward propagation and calculates the total loss function value. UTAFD-Net is built on a variational inference framework, and its loss function includes the Evidence Lower Bound (ELBO) loss. The ELBO loss consists of a Kullback-Leibler (KL) divergence term and a Negative Log-Likelihood (NLL) term. To promote the decoupling of dynamic features and attribute features and enable better separation between them, a counterfactual regularization constraint is added to the ELBO loss. Therefore, the overall loss function of UTAFD-Net consists of the ELBO loss and the counterfactual regularization loss. During network training, the goal is to minimize the overall loss. The ELBO loss calculation formula is shown below:

[0135] 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 ) ) ] (twenty four)

[0136] in This refers to the batch size of the HRRP sequence samples. −  F A i , F D i [ l o g ( p ( X i ∣ F A i , F D i ) ) ] The NLL term is used to measure the model's performance in a given context. Generate target samples under the following circumstances The ability to minimize NLL so that the generated HRRP is as close as possible to the input HRRP. The weights control the proportion of KL divergence in the ELBO loss. Two KL divergences... and This is used to measure the difference between the posterior distributions of dynamic features and attribute features and their corresponding prior distributions, encouraging the posterior distributions of features to approximate the priors by minimizing the KL divergence. In equation (24) above, the prior distribution of attribute features... It is a standard normal distribution .

[0137] To promote the decoupling and separation of dynamic features and attribute features, this embodiment adds counterfactual regularization loss to UTAFD-Net to constrain the mutual independence of these two types of features, i.e. .like Figure 2 and Figure 7 As shown, the prior distribution of attribute features Random sampling is used to construct counterfactual attribute features. Then the dynamic features and counterfactual attributes Input decoder to generate counterfactual intervention samples The formula for calculating the counterfactual regularization loss is:

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

[0139] In the above formula This represents attribute characteristics. Counterfactual attributes In counterfactual samples The difference in posterior probabilities (i.e., likelihood ratio) is calculated. Minimizing this loss function forces... , making counterfactual samples It does not contain any attribute information, that is, dynamic features. It contains no attribute information, thus blocking... and The model learns independent dynamic features by using counterfactual regularization constraints to eliminate spurious correlations between features. and attribute features This significantly enhances the feature decoupling capability.

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

[0141] (26)

[0142] in These are the weights for the counterfactual regularization loss. During network training, the goal is to minimize the overall loss.

[0143] Sub-step S1034: Update the parameters of UDAFD-Net using the backpropagation algorithm. In the t-th iteration, the parameter update formula is as follows:

[0144] (27)

[0145] in This represents the learning rate used to train UDAFD-Net. One training iteration is completed once the training set data has been sampled in batches.

[0146] Since the number of iterations of UDAFD-Net is Therefore, it is executed repeatedly. The training of UDAFD-Net can be completed through sub-steps S1032-S1034. After training, the UDAFD-Net network can effectively remove perturbation factors from HRRP sequence samples and extract attribute-invariant features.

[0147] Step S104: Freeze the parameters of UTAFD-Net, use the attribute features decoupled from UTAFD-Net as input to the target recognition network HRRP-RecNet, and train HRRP-RecNet in conjunction with the target category labels to enable HRRP-RecNet to have the ability to classify targets; specifically including the following steps:

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

[0149] Sub-step S1042: From training data Randomly sample a batch of data ,Will The attribute features decoupled by UTAFD-Net and tags Send to the network, This refers to the batch size.

[0150] Sub-step S1043: The network begins forward propagation and calculates the cross-entropy loss value. The formula for calculating the cross-entropy loss is:

[0151] (28)

[0152] in, It is the number of HRRP sequence samples. It is the number of sample categories. It is the first Sample The true category (one-hot encoding) This is the probability distribution output by the HRRP object recognition network. During training, the objective is to minimize this cross-entropy function.

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

[0154] (29)

[0155] in This represents the learning rate used to train HRRP-RecNet. One training iteration is completed once the training set data has been sampled in batches.

[0156] Since the number of iterations of HRRP-RecNet is Therefore, it is executed repeatedly. The training of HRRP-RecNet can be completed through the next steps S1042-S1044.

[0157] After training, HRRP-RecNet has the ability to identify the type of the target.

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

[0159] Sub-step S1051: Transfer the test set Each HRRP sequence sample All data are fed into the trained HRRP target recognition model, which outputs a probability distribution using the Softmax function. Then the sample can be obtained using the following formula. Category prediction results :

[0160] (30)

[0161] Sub-step S1052: The prediction results are... With the true label of the sample The results are compared, and those that match indicate that the identification is correct.

[0162] Sub-step S1053: Divide the number of correctly identified samples by the total number of samples in the test set to obtain the target correct recognition rate for this test set.

[0163] Since this step uses test sets with four different orientation missing rates, the correct target recognition rate of HRRP will be obtained under four different orientation missing cases (Case1-Case4).

[0164] Table 2 compares the accuracy of the proposed method with six different HRRP target recognition algorithms. The methods compared include direct HRRP target recognition methods and those considering aspect gaps. Direct HRRP target recognition methods use state-of-the-art neural networks to extract deep features and identify target types, without considering aspect gaps or potential interference factors. These methods include Target-Aware RecurrentAttentional 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 gaps include Dual Self-Supervised Contrastive Learning method (DSSCL), Scattering Center Neural Network (SCNet), and Type-Aspect Disentanglement Network (TADN). All these methods use a single HRRP dataset as the input sample to the network. During the experiment, TADN was trained and tested in two forms: TADN(√) indicates that the network training had correct orientation and type labels, while TADN(×) indicates that the orientation labels were biased. TADN(×) better reflects actual adversarial situations where it is difficult to obtain the precise orientation of the target. The header "Test set" and "ALL" indicate testing on the test set and testing on all data, respectively.

[0165] As shown in Table 2, in Case 1 where no orientation loss occurred, the recognition performance of each method was not significantly different, with an accuracy difference of less than 3%. However, as the missing value increased, the recognition accuracy of TARAN, VmResNet, SCRAM, and TACNN dropped significantly. This is because they directly extract deep features from the HRRP data for recognition. These deep features include both attribute features affecting target type and pseudo-features related to target orientation factors in the training set. When the target orientation in the test set changes, these pseudo-features cause a shift in the overall deep features, thus affecting the target type recognition result.

[0166] DSSCL and SCNet exhibit some robustness to orientation loss, with overall recognition accuracy exceeding that of the aforementioned methods. When the missing rate increases from 0% to 20% and then to 50%, the accuracy decreases only slightly, by about 5% each time. However, when the missing rate increases from 50% to 70%, the accuracy decreases by about 10%. This indicates that DSSCL and SCNet struggle to effectively handle high proportions of orientation loss. TADN, a type-orientation decoupling network, works ideally when target type and orientation labels are accurately obtained (TADN(√)). The network can remove the interference of orientation factors and separate the attribute features that truly determine the target type for recognition. Therefore, its recognition performance is less affected by orientation loss and can achieve results comparable to the proposed methods. However, in real-world situations where precise target orientation is difficult to obtain (TADN(×)), due to the bias in the orientation labels, the network struggles to accurately separate orientation features from type features. The deep features still contain these two confused features, thus the network's recognition performance is essentially equivalent to direct recognition networks like TARAN.

[0167] In contrast, the method proposed in this embodiment employs an unsupervised feature decoupling approach, separating pure, attribute-invariant features without any label guidance. This breaks TADN's dependence on directional labels, making it more suitable for real-world scenarios where accurate directional data is difficult to obtain. Among all methods compared, the proposed method achieves higher overall recognition accuracy, maintaining 87.28% accuracy on the test set even with a missing rate as high as 70%. The proposed method is more robust to directional missing data; as the missing rate increases from 0 to 70%, the recognition accuracy on the test set decreases by only 9.92% (97.20% → 87.28%), while the best-performing method besides TADN (√) experiences a 20.37% decrease in accuracy (96.06% → 75.69%). The proposed method demonstrates stronger robustness and superiority in HRRP target recognition under directional missing conditions, effectively mitigating the directional sensitivity issue in HRRP target recognition.

[0168] Table 2 Comparison of recognition accuracy of various methods under different orientation 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 cells to simulate more types of perturbation factors. Test the trained HRRP target recognition model on this test set to obtain the HRRP target recognition results under the condition of multiple perturbation factors superimposed. Specifically, the following process is included:

[0171] Sub-step S1061: First, the disturbance factors are simulated. Since the target azimuth in the test set of Case 2 is completely different from that in the training set, meaning there are azimuth variations in the original test set of Case 2, Gaussian white noise with different signal-to-noise ratios is added to the test set to simulate the effects of environmental noise or electromagnetic interference that may change at any time. The signal-to-noise ratio of the added noise ranges from 0dB to 20dB. Additionally, the scenario of radar hardware failure leading to data segment loss or data link instability causing some range cells to fail to transmit is simulated by randomly truncating several range cells in the HRRP. 5 to 20 range cells are randomly truncated at arbitrary locations, and the values ​​of these range cells are set to 0.

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

[0173] Sub-step S1063: Similarly, compare the identification results output by the model with the true labels of the samples, and count the number of correctly identified samples.

[0174] Sub-step S1064: Then, by dividing the number of correctly identified samples by the total number of samples in the test set, the HRRP target correct identification rate under the condition of multiple perturbation factors can be obtained.

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

[0176] Experimental results show that TARAN, VmResNet, SCRAM, and TACNN are most affected by perturbation factors. With the addition of noise and truncation, the recognition accuracy of each method decreases significantly. This is because the confounding effect of perturbation factors causes deep features to deviate from the target's attribute features. Furthermore, the dynamic perturbation factors in the test set may differ from those in the training set. During training, the network may only fit deep features under specific perturbation patterns, failing to learn invariant features that fundamentally affect the target category. This significantly limits its generalization ability when facing unknown and changing perturbations. DSSCL, SCNet, and TADN, which consider orientation loss, do not show significant performance improvements when dealing with multiple perturbation factors. While they do mitigate the impact of orientation changes to some extent, pseudo-features generated by other dynamic factors (such as noise and truncation) are still mixed into the deep features, and the confusion effect persists. As these dynamic factors in the test set change, the deep features dynamically shift, leading to a mismatch between the mapping relationship between deep features and the target category, thus affecting the recognition performance.

[0177] The method proposed in this embodiment can separate the dynamic factors that generate interference from the invariant factors of the target's attributes, extracting the attribute-invariant features as deep features for target recognition, making the recognition results less susceptible to the influence of perturbation factors in the system. Because it employs an unsupervised decoupling approach, it can achieve accurate decoupling even when the composition and value of the dynamic interference factors are unknown, making it more suitable for practical scenarios with complex and unpredictable perturbation factors. The method proposed in this embodiment has the highest recognition accuracy; even under the influence of triple perturbation factors (A+B+C), the accuracy of the method proposed in this embodiment reaches 88.04%, which is 13.49% higher than the TADN(√) of the method second only to the one proposed in this embodiment. The method proposed in this embodiment is less affected by interference and has more robust recognition performance. With a 20% orientation loss, adding random truncation (A+B) only reduces accuracy by 3.43% (94.40%→90.97%). Adding varying noise (A+C) reduces accuracy by only 4.90% (94.40%→89.50%). When all three dynamic factors are combined (A+B+C), accuracy decreases by only 6.36% (94.40%→88.04%). The proposed method can more stably and accurately identify the type of HRRP target even when the perturbation factors are unknown.

[0178] Table 3. Comparison of recognition accuracy of various methods after adding different degrees of perturbation.

[0179]

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

[0181] Figure 8 (a)-(b), (c)-(d), and (e)-(f) respectively show the visualization results of the original HRRP samples and attribute features in the three cases A, A+B, and A+B+C in Table 3. Red and green represent the first type of target and the second type of target, respectively, and the triangle and circle shapes represent the directional interval "5" and interval "10" in the test set of Case 2, respectively.

[0182] In scenario A, only the orientation factor changes in the test set, which can be roughly divided into two intervals: "5" and "10". Since these two intervals are far apart, the orientation differences between samples within each interval will be quite significant. Therefore, from... Figure 8 (a) The original HRRP sample distribution does indeed show differences in the orientational distribution of targets, with the samples represented by the circles on the left and the triangles on the right clustering together. However, the different types of samples represented by red and green are not easily distinguishable. This illustrates, from the perspective of feature space distribution, that HRRP target recognition is affected by orientational changes; the classification boundary of two classes of samples fitted within one orientation interval may not fit another interval. From the perspective of attribute features, Figure 8 The samples in (b) have obvious type distinguishability. The inter-class distance between the two types of samples represented by the two colors is large, and the two types of samples are easy to distinguish. This shows that UTAFD-Net has extracted attribute features that are only related to the target type, and the target recognition effect is no longer affected by the orientation factor.

[0183] In the case of A+B, in addition to the change in orientation, random truncation was also added to the test set. For example... Figure 8 As shown in (c), the HRRP samples still possess a certain degree of directional distinguishability overall, with the circle on the left and the triangle on the right each representing a directional interval. However, due to the addition of random truncation, compared to... Figure 8 (a) The two types of outlier samples are more mixed together and harder to distinguish. However, after feature decoupling, Figure 8 In (d), samples of the same color and type gradually cluster together, exhibiting relatively clear type distinguishability. However, compared to... Figure 8 (b) In comparison, Figure 8 In (d), the inter-class distance between the two classes of samples is significantly reduced, and there are more samples near the classification boundary. This indicates that random truncation interference increases the difficulty of feature decoupling to some extent, but the proposed method can still effectively maintain the core separability of attribute features.

[0184] In the case of A+B+C, the test set includes three dynamic factors: orientation change, random truncation, and noise interference. For example... Figure 8 As shown in (e), the HRRP samples are randomly distributed in space and have no distinguishability, indicating that the coupling of multiple perturbation factors further increases the difficulty of classification. However, after feature decoupling, as... Figure 8 As shown in (f), UTAFD-Net can correctly classify most samples in the image based on attribute features. This indicates that the proposed decoupling method can still extract robust type discrimination features under multiple dynamic disturbances. Even if the original HRRP data loses its separability due to compound disturbances, the decoupled attribute features can still maintain the core type information, significantly improving the reliability and accuracy of HRRP target recognition.

[0185] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A target identification method based on an unsupervised dynamic-attribute feature decoupling network, characterized in that: HRRP data of the target over a period of time is obtained as an HRRP sequence; the trained HRRP target recognition model is used to process the target HRRP sequence to obtain the target recognition result; 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. The unsupervised dynamic-attribute feature decoupling network includes a dynamic feature encoder, an attribute feature encoder, and a decoder. The unsupervised dynamic-attribute feature decoupling network adopts an asymmetric variational inference architecture. The dynamic feature encoder and the attribute feature encoder model the posterior distribution of dynamic features and attribute features, respectively, and the decoder is used to model the conditional likelihood distribution of HRRP samples.

2. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 1, characterized in that: 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 identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: Using the overall mean Covariance Matrix Constructing the posterior distribution of dynamic features The covariance matrix Calculate using the following formula: in The noise was added diagonally. To utilize the covariance parameter The constructed sparse matrix.

4. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 1, characterized in that: The HRRP target recognition network uses attribute features As input, the target category is determined by a classifier consisting of multiple fully connected (FC) layers.

5. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: An unsupervised dynamic-attribute feature decoupling network is trained using multivariate HRRP sequence sample data in an unsupervised manner. The specific process is as follows: (1) Randomly initialize the parameters of the unsupervised dynamic-attribute feature decoupling network. ; (2) Training data from multivariate HRRP sequence samples A batch of data is randomly sampled and fed into an unsupervised dynamic-attribute feature decoupling network. (3) The network performs forward propagation and calculates the total loss function value; total loss function Including loss of the lower bound of evidence and counterfactual regularization loss : in The weights are used for the counterfactual regularization loss; during network training, the goal is to minimize the overall loss. Loss of lower bound of evidence According to the formula The calculation yielded, where This refers to the batch size of the HRRP sequence samples. It is the negative log-likelihood term. 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; Counterfactual regularization loss According to the formula 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 ; (4) Update the parameters of the unsupervised dynamic-attribute feature decoupling network using the backpropagation algorithm. ; (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.

6. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 5, characterized in that: Modeling the prior distribution of dynamic features using Gaussian processes Using a standard normal distribution Prior distribution as attribute features .

7. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 5, characterized in that: The multivariate HRRP sequence sample data includes HRRP data with varying degrees of orientation loss.

8. The target identification method based on an unsupervised dynamic-attribute feature decoupling network according to claim 2, characterized in that: The process of training the HRRP target recognition network using multivariate HRRP sequence sample data is as follows: (1) Freeze the parameters of the unsupervised dynamic-attribute feature decoupling network and randomly initialize the parameters of the HRRP target recognition network. ; (2) Training data from multivariate HRRP sequence samples The attribute features obtained by processing a batch of randomly sampled data through an unsupervised dynamic-attribute feature decoupling network, along with the corresponding target category labels, are input into the HRRP target recognition network. (3) The network performs forward propagation and calculates the cross-entropy loss value: in, This refers to the batch size of the HRRP sequence samples. It is the number of sample categories. It is the first The true category of each sample This is the category probability distribution output by the HRRP target recognition network; during training, the objective is to minimize this cross-entropy function. (4) Update the parameters of the HRRP target recognition network using the backpropagation algorithm. ; (5) Repeat (2) to (4) until the set iteration termination condition is met to complete the training of the HRRP target recognition network.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the target identification method as described in any one of claims 1 to 8.

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

Images