An HRRP simulation data correction method based on an adversarial Bayesian residual network

Abstract

This invention discloses a method for correcting HRRP simulation data based on an adversarial Bayesian residual network. The method includes the following steps: constructing an HRRP simulation data correction network; inputting a Bayesian residual network as a network generator; outputting the mean and variance of the HRRP residuals; constructing a confidence function from the variance output by the Bayesian residual network; calculating the confidence weights of each distance unit using an exponential mapping method; adaptively weighting the reconstruction loss and adversarial loss; introducing a discriminator network; inputting the corrected HRRP and the real HRRP into the discriminator for adversarial learning; employing a least-squares adversarial mechanism to improve the convergence ability of the adversarial network; jointly optimizing the Bayesian residual network and the discriminator network; and outputting the corrected HRRP and its corresponding confidence distribution. This invention effectively reduces the distribution bias between simulation data and real data under conditions of scarce experimental samples, while avoiding over-correction in low-confidence regions.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition and simulation data enhancement technology, specifically to a method for correcting high-resolution range profile (HRRP) simulation data based on adversarial Bayesian residual networks. Background Technology

[0002] High-resolution range profiles (HRRPs) are important features in broadband radar target recognition, reflecting the distribution of scattering centers of a target along the radar's line-of-sight. However, due to the influence of complex electromagnetic environments and variations in radar operating parameters, HRRP data still faces challenges such as difficulty in obtaining measured samples and significant discrepancies between simulated and real data, thus hindering the development of HRRP-based automatic target recognition (ATR) technology.

[0003] In recent years, broadband radar technology has developed rapidly, but the development of ATR technology based on HRRP has been relatively slow. The difficulties are mainly reflected in the following three aspects: (1) The cost of obtaining measured HRRP samples is high and the conditions are limited, resulting in a small number of available training samples, which makes it difficult to support the training needs of data-driven models such as deep neural networks. (2) There is a systematic deviation between physical simulation HRRP and measured HRRP. Directly using simulation data to train the recognition model will cause the model's performance on measured data to drop significantly. (3) Changes in target attitude and radar parameters cause the distribution of HRRP data to drift, making it difficult for the recognition model trained for a specific condition to generalize to other conditions.

[0004] To address the aforementioned issues, existing technologies employ generative adversarial networks or residual networks to correct simulation data. However, these methods generally suffer from the following shortcomings: a lack of quantification of the uncertainty of the correction results; a tendency to overfit and deviate from the simulation when measured data is scarce; and the potential generation of false features in areas where the physical simulation is severely distorted. Summary of the Invention

[0005] The problem this invention aims to solve is to overcome the shortcomings of existing HRRP simulation data correction methods, such as the lack of uncertainty quantification, overfitting under small sample conditions leading to simulation bias, and forced correction in low-confidence regions resulting in false patterns. This invention provides an HRRP simulation data correction method based on adversarial Bayesian residual networks. This invention belongs to the interdisciplinary field of radar signal processing, deep learning, generative adversarial networks, and uncertainty quantification techniques, and can be used for training data enhancement and simulation data quality improvement in broadband radar target recognition.

[0006] The technical solution provided by this invention to solve the above problems is: a method for correcting HRRP simulation data based on adversarial Bayesian residual networks, the method comprising the following steps: S1. Construct an HRRP simulation data correction network, using the simulated HRRP data input Bayesian residual network as the network generator, and output the mean and variance of the HRRP residuals. S2 constructs a confidence function from the variance output by the Bayesian residual network, calculates the confidence weight of each distance unit through an exponential mapping method, and uses this weight to adaptively weight the reconstruction loss and adversarial loss. When the confidence is high, the residual mean with higher weights is used to correct the HRRP simulation data, and when the confidence is low, the residual mean with lower weights is used to correct the HRRP simulation data, thus realizing a dynamic weight residual repair optimization mechanism. S3 introduces a discriminator network, which inputs the corrected HRRP and the real HRRP into the discriminator for adversarial learning. The least squares adversarial mechanism is used to improve the convergence ability of the adversarial network. S4 jointly optimizes the Bayesian residual network and the discriminator network, constructs a total loss function that includes weighted reconstruction loss, weighted adversarial loss and uncertainty regularization term, and continuously improves the model performance through an alternating training strategy, finally outputting the corrected HRRP and its corresponding confidence distribution.

[0007] Furthermore, in step S1, the Bayesian residual network simultaneously employs the following two methods to achieve Bayesian inference: (1) By using Monte Carlo Dropout to retain the Dropout layer and perform multiple forward propagations during the training and testing phases, the uncertainty of the model parameters can be obtained; (2) The last layer of the network is designed as an independent Gaussian distribution output layer, which outputs the mean and variance of the residuals of each distance unit to characterize the uncertainty of the observation noise; (3) The mean of the final output residual is the average of the mean values ​​obtained from multiple forward propagations, and the total variance is obtained by combining the variance of the observation noise and the variance of the model parameters.

[0008] Furthermore, in step S2, the formula for calculating the confidence level is: (1); in For the first The variance of each distance cell, This is the scaling factor. (0, 1).

[0009] Furthermore, in step S2, the specific method for adaptively adjusting the weights is to sum the adversarial loss and reconstruction loss by weighting them according to the unit-by-unit confidence level: (2); (3); in To weighted countermeasures against losses, For weighted reconstruction losses, The total number of distance cells. For the revised HRRP Units, For the first time in the real HRRP Units.

[0010] Furthermore, in step S4, the total loss function is: (4); The second term is the adversarial loss, and the third term is the overconservative regularization term, which is used to prevent the network from outputting high variance to all units in order to escape training. This is the balance coefficient.

[0011] Furthermore, during the inference phase, the Bayesian residual network after joint training is decoupled from the discriminator network, retaining only the Bayesian residual network for forward propagation. The simulated HRRP data to be corrected is input, and after one or more Monte Carlo forward propagations (if Monte Carlo Dropout is used), the corrected HRRP data and the variance value corresponding to each distance unit are output. The variance values ​​are used to generate an uncertainty heatmap through pseudo-color mapping, serving as a visual representation of the confidence distribution of the corrected result across each distance unit. In subsequent target recognition tasks, this variance heatmap is used to perform confidence screening on the corrected HRRP data: high-confidence distance units with variances less than a preset low threshold are given high weights to participate in target recognition decisions, while low-confidence distance units with variances greater than a preset high threshold are eliminated or zeroed out, thereby improving the robustness of the target recognition system under noisy or simulation mismatch conditions.

[0012] Furthermore, when the number of real HRRP samples participating in training is less than a preset sample number threshold, the discriminator network adopts a gradient penalty strategy to stabilize the adversarial training process. Specifically, a gradient penalty term is added to the discriminator's loss function. This term constrains the gradient norm of the discriminator output by calculating the linear interpolation points between the real HRRP samples and the corrected HRRP samples, making the gradient norm approach 1. The formula for calculating the gradient penalty term is as follows: (5); in , This is a real HRRP sample. For the corrected HRRP sample, For the discriminator output, The gradient penalty coefficient is used to prevent the discriminator from overfitting or gradient vanishing / exploding on a limited number of real samples, ensuring that the generator obtains a stable and effective adversarial gradient, thereby improving the generalization ability of the correction network.

[0013] Furthermore, the input to the Bayesian residual network includes not only the main sequence of simulated HRRP data, but also an auxiliary physical parameter vector corresponding to the simulated HRRP data. The auxiliary physical parameters include, but are not limited to, one or more of the following: target attitude azimuth angle, target attitude pitch angle, radar operating frequency, target radial distance, and target type identification code. At specific levels of the residual module, the network fuses the auxiliary physical parameters with the depth features extracted from the simulated HRRP through feature splicing or cross-modal attention fusion mechanisms. Specifically, the auxiliary physical parameters are mapped to a conditional vector of the same dimension as the depth features via an independent embedding network, and then injected into the residual block element-wise or through attention weighting. This ensures that the residual mean and variance output by the network not only depend on the simulated HRRP data but are also explicitly constrained by the physical parameters, making the HRRP correction results for different attitudes and frequencies conform to the physical laws of radar scattering, avoiding the generation of range image structures that contradict the physical parameters, and ultimately enhancing the physical consistency and scene generalization ability of the correction method.

[0014] Compared with the prior art, the advantages of the present invention are: (1) Uncertainty quantification: The Bayesian residual network outputs the corrected variance of each distance cell, so that users know the credibility of the correction result and provide a reliable basis for subsequent target recognition.

[0015] (2) Adaptive correction: By using the confidence weight guided by variance, the network is forced to make corrections in the regions where it is confident, and conservatively retains the physical simulation results in the uncertain regions, thus avoiding the false features caused by the traditional method of "forcibly correcting all units".

[0016] (3) Robustness to small samples: Combining adversarial distribution alignment and uncertainty perception, it can still train stably and generate highly reliable corrected data when the measured HRRP data is scarce.

[0017] (4) Physical consistency preservation: The correction intensity is automatically reduced in the low confidence area, which essentially preserves the results of the original physical simulation and avoids the HRRP after correction from violating the basic physical laws of radar scattering.

[0018] This invention effectively reduces the distribution deviation between simulated and real data under conditions of scarce experimental samples, while avoiding over-correction in low-confidence regions. The HRRP data generated by the proposed adversarial Bayesian residual network significantly outperforms the original simulated data and data generated by traditional residual networks or GAN methods in target recognition tasks, exhibiting higher accuracy and robustness. Attached Figure Description

[0019] The accompanying drawings, which are provided to further illustrate the invention and constitute a part of this invention, are illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention.

[0020] Figure 1 Overall framework diagram of the adversarial Bayesian residual network of this invention; Figure 2 Bayesian residual network structure diagram; Figure 3 Discriminator network structure diagram; Figure 4 Schematic diagram of the confidence-adaptive weighting mechanism. Detailed Implementation

[0021] The following will describe in detail the implementation of the present invention with reference to the accompanying drawings and embodiments, so that the process of how the present invention uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0022] Residual networks, as a typical deep learning model, have achieved superior results compared to traditional interpolation methods in simulation data correction and have been successfully applied in the field of radar HRRP data augmentation. However, when residual networks are used for HRRP simulation data correction, problems still exist, such as overfitting and simulation bias under conditions of scarce measured samples, inability to quantify the reliability of correction results, and the generation of false features by forcibly correcting low-confidence regions.

[0023] Adversarial Bayesian residual networks combine Bayesian deep learning with generative adversarial networks, overcoming the limitations of ordinary residual networks in outputting uncertainty and the lack of corrected confidence assessment in ordinary GANs. Furthermore, a variance-guided adaptive weighting mechanism is introduced into the loss function, enabling the network to apply strong corrections in high-confidence regions and conservatively preserve physical simulation results in low-confidence regions. By reducing the correction intensity in unreliable regions, the improved adversarial Bayesian residual network can address the problem of simulation data correction under small sample conditions.

[0024] Therefore, this invention proposes a method for correcting HRRP simulation data based on an adversarial Bayesian residual network. The overall framework diagram of the adversarial Bayesian residual network of this invention is shown below. Figure 1 As shown.

[0025] The inventions include: Figure 1 The diagram shows the overall framework of the adversarial Bayesian residual network of this invention. One network, called the Bayesian residual network, primarily functions by learning the residual mapping relationship between the simulated HRRP and the real HRRP, correcting the simulated HRRP to data closer to the real HRRP, and outputting the variance of each distance cell as a measure of uncertainty. The other network, called the discriminator, determines whether the corrected HRRP is real data. The two networks continuously optimize their network structures in this dynamic "game" until the discriminator can no longer distinguish between the corrected data and the real data. This indicates that the Bayesian residual network has been able to generate convincingly realistic corrected data, and the effectiveness of the correction model is demonstrated.

[0026] Figure 2 The diagram shows the Bayesian residual network structure of this invention. This network is used to simulate HRRP. The input is processed through several one-dimensional convolutional layers and fully connected layers. Each layer is followed by Dropout and ReLU activation functions. The output layer is divided into two branches, one of which outputs the mean residual. One output residual variance The revised HRRP calculation method is as follows: (1); Its uncertainty variance is In this embodiment, the number of convolutional layers is set to 3, with each layer having a kernel size of 3 and channel numbers of 16, 32, and 64 respectively. This is followed by two fully connected layers with a dimension of 128, and a Dropout probability of 0.2. Bayesian inference is implemented using Monte Carlo Dropout. During testing, the same input is subjected to 50 forward propagations. The formulas for calculating the mean and variance of the residuals are as follows: (2); in Table 1 shows the parameter settings for the Bayesian residual network of this invention.

[0027] Table 1 Bayesian Residual Network Parameter Settings

[0028] Figure 3 The diagram shown illustrates the network structure of the discriminator in this invention. Discriminator The input is the corrected HRRP Or a real HRRP The output is a scalar probability. For stable training, a least-squares GAN loss function is used, and the discriminator's optimization objective is to minimize the following function: (3); A Bayesian residual network (acting as a generator) aims to deceive the discriminator; its adversarial loss function is defined as: (4); When the number of measured HRRP samples is small (e.g., less than 100), a spectral normalization layer is used after the convolutional layer of the discriminator to prevent overfitting. The formula for calculating spectral normalization is: (5); in Represents the weight matrix The spectral norm. Table 2 shows the parameter settings for the discriminator of this invention.

[0029] Table 2 Discriminator Parameter Settings

[0030] The diagram shows the adaptive weighting mechanism for confidence levels in this invention. Due to the variance of the Bayesian residual network output... Reflects the first The confidence level of the correction for each distance cell is defined in this invention as follows: (6); Among them The scaling factor, in this embodiment, is taken as... Using an exponential function, a smooth nonlinear mapping from variance to confidence level can be achieved: when the variance is small, the confidence level is close to 1 and decreases slowly; when the variance exceeds a certain threshold, the confidence level rapidly approaches 0. This characteristic meets the design requirement of "high confidence in certain regions and rapid weight reduction in uncertain regions." This confidence function has the following properties: when... hour, ;when hour, Then, the weight coefficients for each distance cell are generated. The adversarial loss and reconstruction loss are weighted according to the unit-by-unit confidence level: (7); (8); Wherein is the total number of distance cells. Through this weighting mechanism, high-confidence cells (such as distance cells corresponding to strong scattering points) occupy a larger weight in the loss function and participate in strong alignment correction; low-confidence cells (such as shadow areas or complex areas with multiple reflections) have weights close to zero, automatically reducing the correction intensity, thereby avoiding the problem of traditional methods forcibly correcting unreliable areas and generating false features.

[0031] The total loss function of this invention consists of a weighted sum of three parts: (9); Among them, the first item To weight the reconstruction loss, the correction result is forced to approximate the true HRRP on distance cells with high confidence; the second term To weight and counteract loss, the simulation data distribution on high-confidence cells is aligned with the real data distribution; the third item To prevent overconservative regularization, its function is: if the network outputs a large variance across all units, then... The first two losses tend to 0, but the third loss tends to 1, thus penalizing this behavior and forcing the network to output low variance where corrections are possible. In this embodiment, the parameter values ​​are... , .

[0032] This embodiment employs an alternating training strategy, where the discriminator is updated first in each iteration, followed by the Bayesian residual network. The discriminator update formula is as follows: (10); The update formula for a Bayesian residual network is: (11); in The learning rate is used. The optimizer employs Adam, and the learning rate is... The mini-batch size is 32, and the total number of training rounds is 200. In each training round, the discriminator is updated... The generator is updated once. Table 3 summarizes the training parameters of the method of this invention.

[0033] Table 3 Training parameter settings

[0034] After training is complete, for any new simulated HRRP During the inference phase, only one forward propagation is required from the Bayesian residual network to output the corrected HRRP. and the variance of each distance cell The modified HRRP can be directly used to train target recognition models (such as CNN, SVM, etc.), while the variance heatmap can be used to remove low-confidence test samples or for weighted fusion decisions, providing a reliable basis for subsequent target recognition tasks.

[0035] This embodiment was validated under conditions of scarce HRRP data (only 30 real samples and 10,000 simulated samples). Experimental results show that the method of this invention outperforms existing methods in both correction accuracy and downstream recognition performance: the recognition accuracy of the original simulated data is 72.3%, that of ordinary residual networks is 79.1%, that of ordinary GAN correction is 81.5%, while the method of this invention reaches 89.8%, and can simultaneously output the confidence distribution of each distance unit. Therefore, the recognition rate of the corrected HRRP dataset generated by this invention through an adversarial Bayesian residual network is superior to the recognition rate of the original simulated dataset and datasets generated by other methods under conditions of scarce measured data.

Claims

1. A method for correcting HRRP simulation data based on adversarial Bayesian residual networks, characterized in that, The method Includes the following steps, S1. Construct an HRRP simulation data correction network, using the simulated HRRP data input Bayesian residual network as the network generator, and output the mean and variance of the HRRP residuals. S2 constructs a confidence function from the variance output by the Bayesian residual network, calculates the confidence weight of each distance unit through an exponential mapping method, and uses this weight to adaptively weight the reconstruction loss and adversarial loss. When the confidence is high, the residual mean with higher weights is used to correct the HRRP simulation data, and when the confidence is low, the residual mean with lower weights is used to correct the HRRP simulation data, thus realizing a dynamic weight residual repair optimization mechanism. S3 introduces a discriminator network, which inputs the corrected HRRP and the real HRRP into the discriminator for adversarial learning. The least squares adversarial mechanism is used to improve the convergence ability of the adversarial network. S4 jointly optimizes the Bayesian residual network and the discriminator network, constructs a total loss function that includes weighted reconstruction loss, weighted adversarial loss and uncertainty regularization term, and continuously improves the model performance through an alternating training strategy, finally outputting the corrected HRRP and its corresponding confidence distribution.

2. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: In step S1, the Bayesian residual network simultaneously employs the following two methods to achieve Bayesian inference: (1) By using Monte Carlo Dropout to retain the Dropout layer and perform multiple forward propagations during the training and testing phases, the uncertainty of the model parameters can be obtained; (2) The last layer of the network is designed as an independent Gaussian distribution output layer, which outputs the mean and variance of the residuals of each distance unit to characterize the uncertainty of the observation noise; (3) The mean of the final output residual is the average of the mean values ​​obtained from multiple forward propagations, and the total variance is obtained by combining the variance of the observation noise and the variance of the model parameters.

3. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: In step S2, the confidence level is calculated using the following formula: (1)； in For the first The variance of each distance cell, This is the scaling adjustment coefficient. (0, 1).

4. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: In step S2, the specific method for adaptively adjusting the weights is to sum the adversarial loss and reconstruction loss by weighting them according to the unit-by-unit confidence level: (2)； (3)； in To weighted countermeasures against losses, For weighted reconstruction losses, The total number of distance cells. For the revised HRRP Units, For the first time in the real HRRP Units.

5. The HRRP simulation data correction method based on an adversarial Bayesian residual network according to claim 1, characterized in that: In step S4, the total loss function is: (4)； The second term is the adversarial loss, and the third term is the overconservative regularization term, which is used to prevent the network from outputting high variance to all units in order to escape training. This is the balance coefficient.

6. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: During the inference phase, the Bayesian residual network, after joint training, is decoupled from the discriminator network, retaining only the Bayesian residual network for forward propagation. The simulated HRRP data to be corrected is input, and after one or more Monte Carlo forward propagations (if Monte Carlo Dropout is used), the corrected HRRP data and the variance value corresponding to each distance cell are output. The variance values ​​are used to generate an uncertainty heatmap through pseudo-color mapping, serving as a visual representation of the confidence distribution of the corrected result across each distance cell. In subsequent target recognition tasks, this variance heatmap is used to perform confidence screening on the corrected HRRP data: high-confidence distance cells with variances less than a preset low threshold are given high weights to participate in target recognition decisions, while low-confidence distance cells with variances greater than a preset high threshold are eliminated or zeroed out, thereby improving the robustness of the target recognition system under noisy or simulation mismatch conditions.

7. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: When the number of real HRRP samples participating in training is less than a preset sample number threshold, the discriminator network adopts a gradient penalty strategy to stabilize the adversarial training process. Specifically, a gradient penalty term is added to the loss function of the discriminator. This term constrains the gradient norm of the discriminator output by calculating the linear interpolation points between the real HRRP samples and the corrected HRRP samples, making the gradient norm approach 1. The formula for calculating the gradient penalty term is as follows: (5)； in , This is a real HRRP sample. For the corrected HRRP sample, For the discriminator output, The gradient penalty coefficient is used to prevent the discriminator from overfitting or gradient vanishing / exploding on a limited number of real samples, ensuring that the generator obtains a stable and effective adversarial gradient, thereby improving the generalization ability of the correction network.

8. The HRRP simulation data correction method based on adversarial Bayesian residual networks according to claim 1, characterized in that: The input to the Bayesian residual network includes not only the main sequence of simulated HRRP data, but also an auxiliary physical parameter vector corresponding to the simulated HRRP data. The auxiliary physical parameters include, but are not limited to, one or more of the following: target attitude azimuth angle, target attitude elevation angle, radar operating frequency, target radial distance, and target type identification code. At specific levels of the residual module, the network fuses the auxiliary physical parameters with the depth features extracted from the simulated HRRP through feature splicing or cross-modal attention fusion mechanisms. Specifically, the auxiliary physical parameters are mapped to a conditional vector of the same dimension as the depth features via an independent embedding network, and then injected into the residual block element-wise or through attention weighting. This ensures that the residual mean and variance output by the network not only depend on the simulated HRRP data but are also explicitly constrained by the physical parameters. This makes the HRRP correction results for different attitudes and frequencies conform to the physical laws of radar scattering, avoiding the generation of range image structures that contradict the physical parameters, and ultimately enhancing the physical consistency and scene generalization ability of the correction method.

Images