Electromagnetic target feature correction method and device based on adversarial network training

Abstract

The invention discloses an electromagnetic target feature correction method and device based on adversarial network training, and the method comprises the steps: generating a high-dimensional feature space based on a measurement learning method of center loss and ternary loss, carrying out the characterization and extraction of feature distribution of a data set, and obtaining the data feature distribution of a source domain and a target domain, inputting the target domain data features into a feature correction network for correction, then inputting the source domain data features and the target domain data features into a feature alignment network to estimate a metric distance between the distribution, and performing adversarial training and iterative optimization on the feature correction network and the feature alignment network by taking minimization of the metric distance as a target to obtain a target domain data feature; and reducing the difference between the data feature distribution of the source domain and the target domain until a convergence state is reached. On the premise that the high-dimensional feature space is not changed and the source domain feature distribution is not changed, the distribution difference between the target domain and the source domain features is learned in an adversarial training mode, and correction and learning of the electromagnetic target features under the background environment change condition are completed.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a method and device for correcting electromagnetic target characteristics based on adversarial network training. Background Art

[0002] The radiation source individual identification technology first relies on the extraction of electromagnetic signal features that can describe the individual fingerprint characteristics. Generally speaking, feature extraction methods can be divided into the following categories, including those based on transient signals, signal parameters and statistics, modulation errors and constellation diagrams, signal analysis, mechanism analysis, and deep learning. The feature extraction methods in actual research are often not limited to a specific category, but multiple categories are intertwined and jointly participate in feature extraction.

[0003] According to the state of the transmitter, the radiation source characteristics are divided into steady-state characteristics and transient characteristics. For transient characteristics, since the communication equipment is still in an unstable state, the transient signal has a more distinct nonlinear and non-stationary characteristic, which makes the subsequent processing and analysis simpler. In addition, since it does not contain load information, it can effectively reduce the interference of external non-fingerprint factors, which is conducive to extracting effective individual fingerprint characteristics of the radiation source. Although transient signals are easier to reflect the characteristics of the radiation source than steady-state signals, it is not easy to accurately locate the starting and ending positions in actual scenarios, and the time available for signal acquisition is extremely short and difficult to capture. It has high requirements for signal integrity and high performance requirements for the receiver.

[0004] In terms of feature extraction and recognition based on signal parameters and statistics, Zhang Min et al. used signal carrier stability as a fingerprint feature and achieved the classification of 5 communication radio stations. Liu Yongyong et al. extracted the second-order moment and fourth-order moment of the signal envelope to construct fingerprint features. Cai Zhongwei et al. proposed to use the signal bispectrum to extract features. The bispectrum is the Fourier transform of the second-order cumulant of the signal. It has certain anti-noise performance and therefore often appears in feature extraction. Since the bispectrum feature has a high dimensionality, it needs to be reduced in dimension. It can be reduced in dimension by PCA feature selection algorithm or by integration. The integration methods mainly include axial integrated bispectrum, radial integrated bispectrum, rectangular integrated bispectrum, etc.

[0005] Deep learning is an algorithm that models highly complex data through multi-layer nonlinear transformations. It improves the network's expressiveness by increasing the depth of the neural network. In theory, deep neural networks can fit arbitrarily complex functions. In recent years, deep learning has made many breakthroughs in the fields of computer vision and natural language processing. Inspired by this, deep neural networks have gradually been applied to radar and communication signal processing. The radiation source feature extraction methods based on deep learning can be divided into two categories. One is to perform simple preprocessing or even no preprocessing on the time domain signal, and the feature extraction is completely completed by the neural network's autonomous learning. The other is to combine the deep neural network with the previous feature extraction methods.

[0006] However, environmental factors pose a huge challenge to the identification of radiation sources: in a new environment, there are new obstacles and a large number of unknown radiation source signal types, and the detection and identification capabilities of systems that have not adapted to the new environment will be significantly reduced. Many studies have proposed open set recognition algorithms for the detection and identification of unknown classes, such as open set recognition methods based on semi-supervised and adversarial autoencoders, prototype networks, zero-shot learning, and adversarial learning.

[0007] At the same time, the signal sample characteristics of the radiation source in different environments have domain differences. The feature differences will make the data distribution of the training set and the test set very different. The independent classifier trained directly based on the original data has poor generalization ability for new scenes, resulting in a low recognition rate. For this type of recognition problem, many existing studies have proposed corresponding domain adaptive methods. A common approach is to use domain discriminators, domain adaptive networks, or combined with generative networks to reduce the difference in feature distribution in different domains, thereby improving the recognition rate in new environments. However, this type of method requires that the distribution of the target domain and the source domain is similar, that is, the difference between the two cannot be too large, and in actual situations, there is a confusion distribution of multiple types of data in the target domain. At the same time, this type of method requires that the source domain data and the target domain data be transformed simultaneously, which takes a lot of time and computational cost and cannot meet the needs of practical applications.

[0008] In actual application scenarios, changes in the background environment will lead to a gap between the data distribution of the model training set and the test set. Feature differences will make the data distribution of the training set and the test set very different. The traditional electromagnetic target recognition model based on finite environmental state parameters and a fixed training data set can only achieve good performance in a relatively closed environment. The independent classifier directly trained based on a large amount of data in its recognition method has poor generalization ability for new scenes and low robustness. Once the electromagnetic environment changes, the electromagnetic target features will be confused and overlapped in the feature space, resulting in a significant decrease in recognition performance. Summary of the invention

[0009] The purpose of this application is to overcome the existing technical defects and provide an electromagnetic target feature correction method and device based on adversarial network training. Under the premise of ensuring that the high-dimensional feature space remains unchanged and the source domain feature distribution remains unchanged, the distribution difference of the target domain and source domain features is learned through adversarial training, and the correction and learning of electromagnetic target features under the conditions of background environment changes are completed, thereby realizing the distortion correction of subtle features of the radiation source.

[0010] The purpose of this application is achieved through the following technical solutions:

[0011] In a first aspect, the present application proposes an electromagnetic target feature correction method based on adversarial network training, the method comprising:

[0012] Step S1, generating a high-dimensional feature space that meets the correction requirements through a metric learning method of center loss and ternary loss;

[0013] Step S2, ensuring that the high-dimensional feature space remains unchanged and characterizing and extracting the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and sending the target domain data features to the feature correction network for feature correction to obtain the target domain correction features;

[0014] Step S3: input source domain data features and target domain correction features into a feature alignment network to estimate the metric distance between source domain features and target domain feature distributions;

[0015] Step S4, with the goal of minimizing the metric distance, adversarial training and parameter optimization are performed on the feature correction network and the feature alignment network at the same time;

[0016] Step S5: Repeat steps S2 to S4 to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

[0017] In one possible implementation, the center loss L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, The yth yth represents the deep feature i Class center of class, is the 2-norm.

[0018] In one possible implementation, the ternary loss l tri The expression is: tri=max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

[0019] In a possible implementation, the feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution difference between the source domain data and the target domain data.

[0020] In a possible implementation, the adversarial training of the feature correction network and the feature alignment network is equivalent to solving an optimization problem, and the joint loss function of the optimization problem is: where θ E is the feature correction network parameter, θ T is the feature alignment network parameter, x s and x t are data samples of the source domain and the target domain respectively, D s and D t They are the datasets of the source domain and the target domain respectively.

[0021] In a second aspect, the present application also proposes an electromagnetic target feature correction device based on adversarial network training, the device comprising:

[0022] The generation module is used to generate a high-dimensional feature space that meets the correction requirements through the metric learning method of center loss and ternary loss;

[0023] The feature correction module is used to ensure that the high-dimensional feature space remains unchanged and to characterize and extract the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and to send the target domain data features to the feature correction network for feature correction to obtain the target domain correction features;

[0024] A feature alignment module is used to input source domain data features and target domain correction features into a feature alignment network to estimate the metric distance between source domain features and target domain feature distributions;

[0025] An adversarial training module is used to perform adversarial training and parameter optimization on the feature correction network and the feature alignment network with the goal of minimizing the metric distance;

[0026] The iterative optimization module is used to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

[0027] In one possible implementation, the center loss function L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, The yth yth represents the deep feature i Class center of class, is the 2-norm.

[0028] In one possible implementation, the ternary loss l tri The expression is: tri =max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

[0029] In a possible implementation, the feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution difference between the source domain data and the target domain data.

[0030] In a possible implementation, the adversarial training of the feature correction network and the feature alignment network is equivalent to solving an optimization problem, and the joint loss function of the optimization problem is: where θ E is the feature correction network parameter, θ T is the feature alignment network parameter, x s and x t are data samples of the source domain and the target domain respectively, D s and D t They are the datasets of the source domain and the target domain respectively.

[0031] The above-mentioned main scheme of the present application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and claimed for protection in the present application; and in the present application, (non-conflicting options) options and other options can also be freely combined. After understanding the scheme of the present application, those skilled in the art can understand that there are multiple combinations based on the prior art and common knowledge, all of which are technical schemes to be protected by the present application, and they are not exhaustively listed here.

[0032] The present application discloses a method and device for correcting electromagnetic target features based on adversarial network training. First, a high-dimensional feature space that meets the correction requirements is generated based on the metric learning method of center loss and ternary loss, and the feature distribution of the data set is characterized and extracted. After the data feature distribution of the source domain and the target domain is obtained, the target domain data features are input into the feature correction network for correction, and then the source domain data features and the target domain data features are input into the feature alignment network to estimate the metric distance between their distributions, with the goal of minimizing the metric distance. At the same time, the feature correction network and the feature alignment network are adversarially trained and iteratively optimized to reduce the difference between the data feature distribution of the source domain and the target domain until a convergence state is reached. Under the premise of ensuring that the high-dimensional feature space remains unchanged and the source domain feature distribution remains unchanged, the distribution difference between the target domain and the source domain features is learned through adversarial training to complete the correction and learning of electromagnetic target features under the condition of background environment changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A flow chart of an electromagnetic target feature correction method based on adversarial network training proposed in an embodiment of the present application is shown.

[0035] Figure 2 A schematic diagram of feature correction and feature alignment adversarial training proposed in an embodiment of the present application is shown.

[0036] Figure 3a The figure shows the effect of migrating from the Gaussian channel to other channels before correction.

[0037] Figure 3b The figure shows the effect of the correction after migrating from the Gaussian channel to other channels.

[0038] Figure 4a The effect diagram before correction is shown when migrating from a scene with a speed of 0 to scenes with other speeds.

[0039] Figure 4b The figure shows the corrected effect of migrating from a scene with a speed of 0 to scenes with other speeds. DETAILED DESCRIPTION

[0040] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0041] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0042] In the prior art, since background environmental changes will lead to a gap between the data distribution of the model training set and the test set, feature differences will make the data distribution of the training set and the test set very different. The traditional electromagnetic target recognition model based on finite environmental state parameters and a fixed training data set can only achieve good performance in a relatively closed environment. The independent classifier directly trained based on a large amount of data in its recognition method has poor generalization ability for new scenes and low robustness. Once the electromagnetic environment changes, the electromagnetic target features will be confused and overlapped in the feature space, resulting in a significant decrease in recognition performance.

[0043] Therefore, in order to solve the problem of confusion and overlap of electromagnetic target features under changing background environment, the embodiment of the present application proposes an electromagnetic target feature correction method and device based on adversarial network training, which performs feature correction learning for two factors that cause distortion to time-frequency signals, namely, changes in spatial environment and changes in speed. Under the premise of ensuring that the high-dimensional feature space and the source domain feature distribution remain unchanged, the distribution difference of target domain and source domain features is learned through adversarial training, so as to achieve distortion correction of subtle features of the radiation source, thereby laying a foundation for accurate identification of electromagnetic targets under changing background environment.

[0044] Please refer to Figure 1 , Figure 1A flow chart of an electromagnetic target feature correction method based on adversarial network training proposed in an embodiment of the present application is shown. The method is mainly divided into feature space metric learning, feature correction network training, and feature alignment network training, wherein feature space metric learning generates a high-dimensional feature space so that in this space, the distance between samples of the same type is as small as possible, and the distance between samples of different types is as large as possible. Feature correction network training constructs a correction network, maps the target domain data distribution again, and aligns the data distribution of the target domain with that of the source domain, thereby improving the generalization ability of the model in the target domain. Feature alignment network training constructs a network to estimate the distribution distance between the source domain features and the corrected target domain features, and minimizes it through optimization. The training and optimization of the feature correction network and the feature alignment network constitute a set of adversarial game tasks. Through adversarial network training of the target domain data, the correction and learning of electromagnetic target features under background environment changes are completed.

[0045] Figure 2 The schematic diagram of feature correction and feature alignment adversarial training proposed in the embodiment of the present application is shown. The feature distribution of source domain data and target domain data is characterized and extracted through feature space metric learning, and the source domain feature vector (Z s ) and the target domain feature vector (Z t ). The feature correction network is used to correct the feature data of the target domain. The input of the network is the target domain feature vector (Z t ), the output is the corrected feature (Z t ′). The feature alignment network is used to estimate the metric distance between the source domain features and the target domain feature distribution. By minimizing the metric distance, adversarial training and parameter optimization are performed on the correction network and the alignment network. After the iterative optimization converges, the final target domain correction features are output.

[0046] Electromagnetic target feature correction methods include:

[0047] Step S1: Generate a high-dimensional feature space that meets the correction requirements through the metric learning method of center loss and ternary loss.

[0048] Although the classic Softmax regression in deep learning can separate heterogeneous data, it does not have an explicit aggregation effect on samples of the same type. The distribution of samples of the same type after the Softmax classifier in the feature space may be scattered, which is not conducive to the determination of abnormal data. Therefore, it is necessary to use the center loss and ternary loss function to constrain the distance between samples of the same type, so as to achieve feature representation and extraction that can highlight subtle features.

[0049] First, prepare the dataset for training, divide the dataset into training set and validation set, and preprocess the image or other forms of data, such as normalization, scaling, etc. Select a basic network structure suitable for feature extraction. This basic network can be any model that can extract effective feature representation from the original input, such as convolutional neural network (CNN). At this stage, there is no need to pay special attention to the design of the final classification layer. The goal is to obtain good feature expression rather than direct classification results.

[0050] Define the center loss and ternary loss. The center loss is to calculate a center vector (the average value of all sample features in the category) for each category, and then for each sample, calculate its distance to the center of the category to which it belongs, and minimize these distances through the center loss, so that samples of the same category are as close as possible in the feature space. Then construct a triplet (anchor point-positive sample-negative sample), where the anchor point is a specified sample, the positive sample is a sample randomly selected from the same category, and the negative sample is another sample from a different category. Then, use the ternary loss function to make the distance between the positive sample and the anchor point smaller than the distance between the negative sample and the anchor point plus a margin value, to ensure that there is sufficient separation between samples of different categories.

[0051] When applied in practice, center loss and ternary loss are usually used in combination to achieve joint optimization by adjusting their respective weight coefficients. This can enhance the ability to distinguish between different categories while ensuring closeness within the same category. The selected basic network is trained using the set loss function. The model parameters and the center positions of each category are continuously updated during training. As the number of iterations increases, it is expected to obtain a feature representation that can distinguish each category well and cluster samples of the same category together. After training, the quality of the learned features is evaluated using performance indicators on the validation set (such as accuracy, recall, etc.). If the effect is not ideal, you need to go back to the previous steps to adjust the parameter configuration or try a different network architecture.

[0052] Central loss function L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, The yth yth represents the deep feature i Class center of class, is the 2-norm.

[0053] Ternary loss tri The expression is: tri=max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

[0054] Margin controls the distinguishability between different categories. The larger the margin, the stronger the distinguishability between different categories after training, but the corresponding training difficulty is greater.

[0055] Step S2, ensure that the high-dimensional feature space remains unchanged and characterize and extract the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and send the target domain data features to the feature correction network for feature correction to obtain the target domain correction features.

[0056] Using the trained feature extractor, the given source domain data set and target domain data set are processed to obtain the corresponding high-dimensional feature vector set. The key to step S2 is to use the existing model or the pre-trained part to fix the feature extraction process to ensure that subsequent operations are performed on the same feature space. It can ensure that the high-dimensional feature representation extracted from the original input remains unchanged, and the feature distribution of data samples from different fields (source domain and target domain) is obtained respectively.

[0057] Then, the target domain data features are sent to the feature correction network for feature correction. The feature correction network is used to transform the data points of the target domain, and the target domain data distribution is mapped again. In the next iteration, the data distribution of the target domain and the source domain are aligned. The difference between it and the domain adaptation method is that it changes alignment to correction, changes the source domain and the target and data features at the same time to complete the distribution alignment, gives priority to the feature transformation method that makes the source domain classification effect the best, and then corrects the feature data of the target domain and migrates it to the classification space of the source domain. The feature correction network can be implemented by different neural networks. The simplest one can be directly implemented by using a convolutional neural network.

[0058] Step S3: input the source domain data features and the target domain correction features into the feature alignment network to estimate the metric distance between the source domain features and the target domain feature distribution.

[0059] The feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution difference between the source domain data and the target domain data. The feature alignment network uses statistical indicators such as relative entropy as a metric, and sends the two sets of feature vectors obtained from the source domain and the target domain into the feature alignment network, and outputs a numerical value representing the degree of difference between the two. The gap in feature distribution between the source domain and the target domain is calculated through a module (feature alignment network) specially designed to measure the difference between two different distributions.

[0060] A distance metric estimation module is trained using the feature alignment network to estimate the distribution distance between the source domain features and the corrected features, characterizing the differences between different distributions. In information theory, relative entropy is equivalent to the difference between the information entropy of two probability distributions. If one of the probability distributions is the true distribution and the other is the theoretical (fitted) distribution, then the relative entropy is equal to the difference between the cross entropy and the information entropy of the true distribution, indicating the information loss generated when the theoretical distribution is used to fit the true distribution: Where p(x) is the probability distribution of the real event, and q(x) is the probability distribution of the event fitted theoretically. The relative entropy is always greater than or equal to 0. The relative entropy is equal to 0 if and only if the two distributions are the same. That is, when the probability distribution of the event fitted theoretically is exactly the same as the real one, the relative entropy is equal to 0. When the fitted distribution is slightly different, the relative entropy is greater than 0. This property is critical because it is exactly the characteristic required by the gradient descent method of deep learning. If the neural network fits perfectly, then it will no longer descend in gradient, but if it is not perfect, it will continue to descend because it is greater than 0.

[0061] Step S4: With the goal of minimizing the metric distance, adversarial training and parameter optimization are performed on the feature correction network and the feature alignment network at the same time.

[0062] According to the metric distance output in step S3, the error information is back-propagated to update the network weights. The whole process follows the basic logic of the gradient descent algorithm and iterates continuously until a local optimal solution is found or the preset stop condition is met. The internal parameters of the feature alignment network can be adjusted so that it can more accurately reflect and reduce the actual cross-domain feature differences.

[0063] The process of adversarial training between the feature correction network and the feature alignment network is equivalent to solving the following optimization problem, where max is to maximize the metric distance between the source domain features and the target domain correction feature distribution, and min is to minimize the metric distance by training the feature alignment network.

[0064]

[0065] Among them, θ E is the feature correction network parameter, θ T is the feature alignment network parameter, xs and x t are data samples of the source domain and the target domain respectively, D s and D t are the datasets of the source domain and the target domain respectively. The above optimization problem is used as the loss function to conduct adversarial training of the feature correction network and the feature alignment network, so as to complete the correction learning of the target domain data features.

[0066] Step S5: Repeat steps S2 to S4 to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

[0067] This is an iterative process. After each complete S2-S4 process, the latest measurement results will be fed back to guide the next adjustment direction. As the number of iterations increases, the feature distributions of the two will theoretically converge until the difference between them can no longer be significantly reduced. Continuously improving the performance of the feature correction network will ultimately achieve two key goals - enhancing the model's generalization ability in new scenarios and improving prediction accuracy.

[0068] In a possible implementation, the electromagnetic signal of the communication network radio station is used as the experimental object. The background environment change focuses on two factors: channel change and speed change. The channel change considers the migration change of the target from the Gaussian channel to other channels, and the speed change considers the migration change of the target from the speed of 0 to other speed scenes.

[0069] Firstly, the collected communication network radio stations are divided into training set and test set according to the background environment changes, and the high-dimensional feature space and data feature distribution are generated by training with the training set. Then, the electromagnetic target feature correction and learning method based on adversarial network training under the background environment change condition proposed by the present invention is used to correct and migrate the data distribution of the test set. Figure 3a The figure shows the effect of migrating from the Gaussian channel to other channels before correction. Figure 3b The figure shows the effect of the correction after migrating from the Gaussian channel to other channels. Figure 4a The figure shows the effect before correction when migrating from a scene with a speed of 0 to scenes with other speeds. Figure 4b The figure shows the corrected effect of migrating from a scene with a speed of 0 to scenes with other speeds.

[0070] It can be seen that when the target migrates from the Gaussian channel to other channels, or from a scene with a speed of 0 to a scene with other speeds, if no correction is performed, multiple target features will be confused, overlapped, and clustered together. After feature correction, multiple target features will be distributed as dispersedly as possible, thereby improving the distinguishability of the data in the high-dimensional feature space after target migration.

[0071] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0072] First, we use the metric learning method of center loss and ternary loss to construct a high-dimensional feature space, which can better reflect the similarities and differences between samples. Center loss encourages samples of the same category to gather in the feature space. Ternary loss emphasizes increasing the distance between samples of different categories to enhance the distinction between them.

[0073] Second, by designing two neural network structures specifically for processing feature correction and feature alignment, and using adversarial training to continuously optimize the functions of these two networks. The role of the feature correction network is to try to convert the data features of the target domain into a form closer to the source domain, that is, to achieve feature "migration" from the target domain to the source domain. The feature alignment network is used to estimate the distance or difference between the corrected target domain features and the source domain features, and to guide how to adjust the parameters of the feature correction network to further narrow this gap. The adversarial training and optimization between the two constitute a set of adversarial game tasks. In the continuous adversarial training of the target domain data, the correction of the electromagnetic target feature distribution is achieved.

[0074] Third, the adjustment is made without changing the adaptability of the original model to the source domain data. There is no need to use a large number of source domain samples to retrain the model. Instead, the feature correction task can be completed with only a small number of target domain samples, which greatly reduces the demand for computing resources and time costs.

[0075] Fourth, even in the face of complex and dynamically changing electromagnetic environment conditions, it is possible to accurately extract the required target signal characteristics and maintain good classification performance. Through the adversarial training mechanism, the target domain features are gradually brought closer to the source domain features, thereby effectively reducing the negative impact caused by environmental changes and enhancing the system's adaptability to different application scenarios. Improve robustness and stability.

[0076] A possible implementation of an electromagnetic target feature correction device based on adversarial network training is given below, which is used to execute the various execution steps and corresponding technical effects of the electromagnetic target feature correction method shown in the above embodiment and possible implementation. The device includes:

[0077] The generation module is used to generate a high-dimensional feature space that meets the correction requirements through the metric learning method of center loss and ternary loss;

[0078] The feature correction module is used to ensure that the high-dimensional feature space remains unchanged and to characterize and extract the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and to send the target domain data features to the feature correction network for feature correction to obtain the target domain correction features;

[0079] A feature alignment module is used to input source domain data features and target domain correction features into a feature alignment network to estimate the metric distance between source domain features and target domain feature distributions;

[0080] An adversarial training module is used to perform adversarial training and parameter optimization on the feature correction network and the feature alignment network with the goal of minimizing the metric distance;

[0081] The iterative optimization module is used to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

[0082] In one possible implementation, the center loss function L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, The yth yth represents the deep feature i Class center of class, is the 2-norm.

[0083] In one possible implementation, the ternary loss l tri The expression is: tri =max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

[0084] In a possible implementation, the feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution difference between the source domain data and the target domain data.

[0085] In a possible implementation, the adversarial training of the feature correction network and the feature alignment network is equivalent to solving an optimization problem, and the joint loss function of the optimization problem is: where θ E is the feature correction network parameter, θ T is the feature alignment network parameter, x s and x t are data samples of the source domain and the target domain respectively, D s and D t They are the datasets of the source domain and the target domain respectively.

[0086] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for correcting electromagnetic target features based on adversarial network training, characterized in that: The method comprises: Step S1, generating a high-dimensional feature space that meets the correction requirements through a metric learning method of center loss and ternary loss; Step S2, ensuring that the high-dimensional feature space remains unchanged and characterizing and extracting the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and sending the target domain data features to the feature correction network for feature correction to obtain the target domain correction features; Step S3: input source domain data features and target domain correction features into a feature alignment network to estimate the metric distance between source domain features and target domain feature distributions; Step S4, with the goal of minimizing the metric distance, adversarial training and parameter optimization are performed on the feature correction network and the feature alignment network at the same time; Step S5: Repeat steps S2 to S4 to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

2. The electromagnetic target feature correction method according to claim 1, characterized in that: Center loss L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, c yi ∈Ρ d The yth yth represents the deep feature i Class center of class, is the 2-norm.

3. The electromagnetic target feature correction method according to claim 1, characterized in that: Ternary loss tri The expression is: tri =max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

4. The electromagnetic target feature correction method according to claim 1, characterized in that: The feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution differences between the source domain data and the target domain data.

5. The electromagnetic target feature correction method according to claim 1, characterized in that: The adversarial training of the feature correction network and the feature alignment network is equivalent to solving the optimization problem. The joint loss function of the optimization problem is: where θ E is the feature correction network parameter, θ T is the feature alignment network parameter, x s and x t are data samples of the source domain and the target domain respectively, s and D t They are the datasets of the source domain and the target domain respectively.

6. An electromagnetic target feature correction device based on adversarial network training, characterized in that: The device comprises: The generation module is used to generate a high-dimensional feature space that meets the correction requirements through the metric learning method of center loss and ternary loss; The feature correction module is used to ensure that the high-dimensional feature space remains unchanged and to characterize and extract the feature distribution of the data set to obtain the data feature distribution of the source domain and the target domain, and to send the target domain data features to the feature correction network for feature correction to obtain the target domain correction features; A feature alignment module is used to input source domain data features and target domain correction features into a feature alignment network to estimate the metric distance between source domain features and target domain feature distributions; An adversarial training module is used to perform adversarial training and parameter optimization on the feature correction network and the feature alignment network with the goal of minimizing the metric distance; The iterative optimization module is used to perform iterative optimization to reduce the difference between the data feature distributions of the source domain and the target domain until a convergence state is reached.

7. The electromagnetic target characteristic correction device according to claim 6, characterized in that: Central loss function L center The expression is: Where m is the total number of samples, x i is the feature vector of the i-th sample, y i is the true category of the i-th sample, c yi ∈Ρ d The yth yth represents the deep feature i Class center of class, is the 2-norm.

8. The electromagnetic target characteristic correction device according to claim 6, characterized in that: Ternary loss tri The expression is: tri =max(d(f(A),f(p))-d(f(A),f(N))+margin,0), where d(·) represents the distance calculation function, f(·) represents the nonlinear mapping function represented by the feature extraction network, f(A) is the feature representation of the anchor sample, f(p) is the feature representation of the positive sample, f(N) is the feature representation of the negative sample, and margin is a hyperparameter.

9. The electromagnetic target characteristic correction device according to claim 6, characterized in that: The feature correction network is implemented using a convolutional neural network, and the feature alignment network is constructed using relative entropy to characterize the distribution differences between the source domain data and the target domain data.

10. The electromagnetic target characteristic correction device according to claim 6, characterized in that: The adversarial training of the feature correction network and the feature alignment network is equivalent to solving the optimization problem. The joint loss function of the optimization problem is: where θ E is the feature correction network parameter, θ T is the feature alignment network parameter, x s and x t are data samples of the source domain and the target domain respectively, s and D t They are the datasets of the source domain and the target domain respectively.

Images