HRRP sample generation method and system based on category-guided DDPM

Abstract

The present invention discloses a HRRP sample generation method and system based on category-guided DDPM, the method comprising: normalizing and centroid-aligning preprocessing the HRRP raw data to generate a small sample data set; constructing a one-dimensional conditional convolutional network, the input of which is a sample to be denoised, and the output is a predicted noise with the same dimension as the input, the one-dimensional conditional convolutional network comprising a downsampling module and an upsampling module integrated into the time step and the angular domain; training the one-dimensional conditional convolutional network using the small sample data set to obtain a trained category-guided denoising diffusion probability model; and generating an extended sample data set using the trained category-guided denoising diffusion probability model. The present invention solves the problems in the field of HRRP sample generation, such as unstable model training, poor quality of samples generated under small sample conditions, uncontrollable angular domain of samples generated by conventional diffusion models, and easy generation of out-of-distribution samples.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar echo sample generation, and in particular relates to a HRRP sample generation method and system based on category-guided DDPM. Background Art

[0002] The range resolution of high-resolution radar is much smaller than the size of the illuminated targets such as ships, aircraft, and vehicles. For radar, the target is a distributed target and can be regarded as a collection of statistically uniform scattering units. Each scattering unit will generate a sub-echo for the radar signal. For a certain radar line of sight, the superposition of the sub-echo vectors is the echo signal received by the radar, that is, the projection of the sub-echo on the radar perspective in the form of vector sum. This is the high-resolution range profile (HRRP), which reflects the distribution of the target scattering center along the radar line of sight and contains a lot of target structural information. Compared with synthetic aperture radar (SAR) images, HRRP has many advantages such as easy acquisition, easy storage, and easy processing, and has received continuous attention in the field of radar automatic target recognition.

[0003] The existing methods of using HRRP for radar target recognition tasks can be roughly divided into two categories: those driven by physical mechanisms and those driven by data. Among them, data-driven methods, especially deep learning methods, have developed rapidly in recent years due to their excellent results. However, deep learning methods require a large number of high-quality samples for training to ensure good accuracy. However, it is extremely difficult to obtain HRRP data of non-cooperative targets under actual radar illumination, and the existing samples are very limited. Therefore, how to expand the samples for recognition tasks under small sample conditions has become a key issue in the field of radar target recognition.

[0004] Compared with traditional methods, there are problems such as low quality of generated samples and limited range of posture angles. In recent years, deep learning and data science have developed rapidly. Probabilistic generative models, such as variational autoencoders (VAE), generative adversarial networks (GAN), denoising diffusion probability models (DDPM) and other methods, have shown the ability to generate realistic images on natural images, and in the case of small samples, probabilistic modeling methods can often perform better than deterministic methods. In the existing work of using probabilistic generative models to generate HRRP samples, researchers mainly use VAE and GAN and have achieved good results. The application of DDPM in the field of HRRP sample generation is still relatively lacking.

[0005] VAE often performs poorly when faced with data with complex probability distributions, and is prone to problems such as missing details of generated samples and limited diversity. First, VAE defines the latent space as a Gaussian distribution with independent components, but the distribution of the latent space of real data is not necessarily a Gaussian distribution. In other words, when faced with real data, no matter how the hyperparameters of the neural network are adjusted, the probability distribution of generated samples cannot be perfectly approximated to the probability distribution of real samples. Therefore, VAE cannot fit complex probability distributions and suffers from the problem of missing sample details. Secondly, the existence of the reconstruction loss term in the loss function of VAE causes values ​​near the same mean in the latent space to be decoded into close samples, and The divergence term makes the distribution of the latent space close to the given prior distribution, which causes the decoder to tend to generate an average sample, resulting in limited sample diversity.

[0006] GAN can learn complex probability distributions and generate high-quality samples with better details, but the confrontation between the discriminator and the generator network causes problems such as unstable training and mode collapse. First, during the confrontation between the generator and the discriminator, if the performance of either party is far ahead, it will become difficult to optimize the other party, and it will be difficult to control the degree of training of both parties. This is the problem of vanishing gradients in GAN. Secondly, during the training process, a situation may arise where the generator finds a pattern that can easily deceive the discriminator, so the probability of generating samples of this pattern will increase, the generated samples will tend to be consistent, and the diversity will decrease, leading to the problem of mode collapse. Therefore, the training of GAN is very unstable.

[0007] See also Figure 1 , Figure 1 is a schematic diagram of the structure of a denoising diffusion probability model (DDPM) provided by an embodiment of the present invention. DDPM determines a forward diffusion process with Markov properties. Map the input to a probability space consistent with its dimension, and then obtain a reverse denoising process through learning The input is restored to complete the generation task. The input of each layer of DDPM is samples with different noise levels, which can learn different levels of details of the samples, thus showing a strong ability to generate details and a stable training process.

[0008] DDPM has many advantages such as strong detail generation capability and stable training process, and has achieved excellent results in the field of image generation. However, the technology for using it in the field of high-quality HRRP sample generation is still lacking. First, HRRP is a one-dimensional signal, and the diffusion model in the image field cannot be directly applied. It is necessary to design the network structure in a targeted manner. Secondly, compared with natural images, HRRP samples have strong angle sensitivity. It only maintains the correlation of sample data in a small angular domain, and the probability distribution of the data is quite different from that of the image samples. Finally, it is difficult for conventional methods to learn the true distribution of samples under small sample conditions. Existing HRRP sample generation methods have problems such as poor performance under small sample conditions and inability to integrate angle information for sample generation, resulting in uncontrollable angular domains of generated samples and generation of out-of-distribution samples that will not appear in actual scenes. They cannot be used to train and improve classifier performance. Summary of the invention

[0009] In order to solve the above problems existing in the prior art, the present invention provides a method and system for generating HRRP samples based on category-guided DDPM. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0010] One aspect of the present invention provides a method for generating HRRP samples based on category-guided DDPM, comprising:

[0011] The acquired HRRP raw data are normalized and preprocessed by centroid alignment to generate a small sample data set;

[0012] Constructing a one-dimensional conditional convolutional network, wherein the input of the one-dimensional conditional convolutional network is the sample to be denoised, and the output is the predicted noise with the same dimension as the input, and the one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that incorporate the time step and the angular domain;

[0013] Using the small sample data set to train the one-dimensional conditional convolutional network to obtain a trained category-guided denoising diffusion probability model;

[0014] The trained category-guided denoising diffusion probability model is used to generate an expanded sample data set.

[0015] Another aspect of the present invention further provides a HRRP sample generation system based on category-guided DDPM, which is used to execute the HRRP sample generation method described in the above embodiment. The HRRP sample generation system includes a data preprocessing module and a sample generation module, wherein:

[0016] The data preprocessing module is used to perform normalization and centroid alignment preprocessing on the acquired HRRP raw data to generate a small sample data set;

[0017] The sample generation module is used to generate an expanded sample data set using the trained category-guided denoising diffusion probability model, wherein the trained category-guided denoising diffusion probability model is obtained by training a one-dimensional conditional convolutional network constructed using the small sample data set, the input of the one-dimensional conditional convolutional network is the sample to be denoised, and the output is the predicted noise with the same dimension as the input, and the one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that incorporate time steps and angular domains.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. The present invention proposes a HRRP sample generation method based on category-guided DDPM. Starting from the angular sensitivity of the HRRP signal, a DDPM model is constructed using a one-dimensional conditional convolutional network according to the characteristics of HRRP data. The reverse process of the DDPM model is conditionally probability modeled, and the angular domain is used as a condition to control the probability distribution of the generated samples. The angular domain is integrated to achieve high-quality HRRP sample generation. The hierarchical probability modeling method of the DDPM model is more suitable for learning the advantages of data distribution under small sample conditions, and solves the problems in the field of HRRP sample generation, such as unstable model training, poor quality of samples generated under small sample conditions, uncontrollable angular domain of samples generated by conventional diffusion models, and easy generation of samples outside the distribution. Compared with the existing methods, the HRRP samples generated by the present invention under small sample conditions can be used for classifier training and significantly improve the classification accuracy.

[0020] 2. The present invention models the reverse process of the DDPM model as conditional probability, uses the HRRP angular domain as a condition to guide model learning, and combines the detail generation ability of the probabilistic statistical diffusion model under small sample conditions. A probabilistic statistical diffusion model based on angular domain guidance is proposed and applied to the radar HRRP sample generation task, thereby improving the HRRP target recognition accuracy under small sample conditions.

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a structural schematic diagram of a denoising diffusion probability model provided by an embodiment of the present invention;

[0023] Figure 2 is a flow chart of a method for generating HRRP samples based on category-guided DDPM provided by an embodiment of the present invention;

[0024] Figure 3 It is a schematic diagram of the structure of a one-dimensional conditional convolutional network provided by an embodiment of the present invention;

[0025] Figure 4It is a schematic diagram of the change of feature dimension after each layer of calculation of a one-dimensional conditional convolutional network provided by an embodiment of the present invention;

[0026] Figure 5 is a schematic diagram of the structure of a residual block provided by an embodiment of the present invention;

[0027] Figure 6 It is a structural diagram of a category-guided denoising diffusion probability model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a HRRP sample generation method and system based on category-guided DDPM proposed in accordance with the present invention in combination with the accompanying drawings and specific implementation methods.

[0029] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0030] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0031] Embodiment 1

[0032] See also Figure 2 , Figure 2 : is a flow chart of a method for generating HRRP samples based on category-guided DDPM provided by an embodiment of the present invention, the method comprising the following steps:

[0033] S1: The acquired HRRP raw data is normalized and preprocessed by centroid alignment to generate a small sample data set.

[0034] The amplitude sensitivity and translation sensitivity of HRRP signals in actual scenarios will seriously affect the performance of the model trained by it. Therefore, after obtaining the HRRP data from the high-resolution radar, data preprocessing is required first.

[0035] Amplitude sensitivity means that the amplitude of the HRRP signal will be affected by various factors such as radar parameters and target distance. The amplitude difference between each sample is large, which is not conducive to the stable learning of the network. Translation sensitivity means that since the HRRP sample is obtained from the radar echo according to a distance window of a given length, when the target is in high-speed motion, its echo may appear at different positions within the distance window, causing the HRRP vector of the same target in the same posture to be misaligned, resulting in a reduction in recognition rate.

[0036] The preprocessing of this step includes normalizing and center-of-gravity alignment of the HRRP raw data to eliminate the influence of amplitude sensitivity and translation sensitivity, which specifically includes the following steps:

[0037] S1.1: Normalize the acquired HRRP raw data to obtain normalized HRRP data :

[0038]

[0039] in, Indicates the acquired HRRP raw data. Indicates the number of distance units contained in the HRRP raw data. Indicates The amplitude within a distance unit, , Indicates The normalized amplitude within a distance unit.

[0040] S1.2: Based on the mass points of each HRRP raw data, the normalized HRRP data are Translate to make the center position coincide with the center of mass position, and obtain the HRRP data after the center of mass is aligned:

[0041]

[0042] in, represents the HRRP data after centroid alignment, represents the Fourier transform operation, represents the inverse Fourier transform operation, represents the phase of the centroid position of the HRRP data, represents the centroid position of the HRRP data, represents the phase of the HRRP data center location, Indicates the center position of HRRP data, Indicates the distance that needs to be moved to align the center position with the center of mass position. Denotes a complex unit. The expression is:

[0043] .

[0044] S2: Construct a one-dimensional conditional convolutional network, the input of which is the sample to be denoised, and the output is the predicted noise with the same dimension as the input. The one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that incorporate the time step and angular domain.

[0045] See also Figure 3 and Figure 4 , Figure 3 is a schematic diagram of the structure of a one-dimensional conditional convolutional network provided by an embodiment of the present invention. Figure 4 This is a schematic diagram of the change of feature dimensions after each layer of calculation of a one-dimensional conditional convolutional network provided by an embodiment of the present invention. The network uses U-Net, and based on the residual block, a one-dimensional U-Net that simultaneously incorporates time step and control condition information (i.e., angular domain) is implemented. Figure 4 It shows the change of feature dimension after each layer calculation.

[0046] Specifically, the input data of the one-dimensional conditional convolutional network is , indicating the The samples that need to be denoised in time steps are output as predicted noise with the same dimension as the input. ,in, Indicates The number of samples that need to be denoised at each time step.

[0047] As shown above, the one-dimensional conditional convolution network includes a downsampling module and an upsampling module. The downsampling module of this embodiment includes a first convolution unit, a first downsampling unit, a second downsampling unit and a third downsampling unit connected in sequence, wherein the input end of the first convolution unit is used to input the data to be processed, the output end of the first convolution unit is connected to the input end of the first downsampling unit, and the first convolution unit can output more channels while keeping the data size unchanged to extract more features of the sample; the first downsampling unit includes a first residual block, a second residual block and a first downsampling convolution layer connected in sequence, and the output end of the first downsampling convolution layer is connected to the input end of the second downsampling unit; the second downsampling unit includes a third residual block, a fourth residual block and a second downsampling convolution layer connected in sequence, and the output end of the second downsampling convolution layer is connected to the input end of the third downsampling unit; the third downsampling unit includes a fifth residual block, a sixth residual block, a third downsampling convolution layer, a seventh residual block, an eighth residual block and a ninth residual block connected in sequence.

[0048] Further, the upsampling module of the present embodiment includes a first upsampling unit, a second upsampling unit, a third upsampling unit and a second convolution unit. The first upsampling unit includes a tenth residual block, an eleventh residual block, a twelfth residual block and a first upsampling transposed convolution layer in sequence, the input of the tenth residual block is the data of the output of the ninth residual block and the output of the eighth residual block spliced ​​in the channel dimension, the input of the eleventh residual block is the data of the output of the tenth residual block and the output of the eighth residual block spliced ​​in the channel dimension, and the input of the twelfth residual block is the data of the output of the eleventh residual block and the output of the seventh residual block spliced ​​in the channel dimension;

[0049] The second upsampling unit includes a thirteenth residual block, a fourteenth residual block, a fifteenth residual block, and a second upsampling transposed convolution layer in sequence, wherein the input of the thirteenth residual block is the data of the output of the first upsampling unit and the output of the fifth residual block after splicing in the channel dimension, the input of the fourteenth residual block is the data of the output of the thirteenth residual block and the output of the fourth residual block after splicing in the channel dimension, and the input of the fifteenth residual block is the data of the output of the fourteenth residual block and the output of the third residual block after splicing in the channel dimension; the third upsampling unit includes a sixteenth residual block, a seventeenth residual block, an eighteenth residual block in sequence. and the third upsampling transposed convolution layer, wherein the input of the sixteenth residual block is the data of the output of the second upsampling unit and the output of the second residual block in the channel dimension, the input of the seventeenth residual block is the data of the output of the sixteenth residual block and the output of the first residual block in the channel dimension, and the input of the eighteenth residual block is the data of the output of the seventeenth residual block and the output of the first convolution unit in the channel dimension; the input end of the second convolution unit is connected to the output end of the third upsampling unit, which is used to convert the output of the third upsampling unit into a dimension consistent with the input data to be processed of the first convolution unit. The jump connection represented by the dotted line in the figure ensures that the network will not have the problem of gradient disappearance during training, making the training more stable and enhancing the network's ability to learn features.

[0050] Further, the first residual block, the second residual block, the third residual block, the fourth residual block, the fifth residual block, the sixth residual block, the seventh residual block, the eighth residual block, the ninth residual block, the tenth residual block, the eleventh residual block, the twelfth residual block, the thirteenth residual block, the fourteenth residual block, the fifteenth residual block, the sixteenth residual block, the seventeenth residual block and the eighteenth residual block are residual blocks with the same structure, which are used to fuse the position-encoded time step and angular domain into the input feature data.

[0051] See also Figure 5 , Figure 5 is a schematic diagram of the structure of a first residual block provided by an embodiment of the present invention, including a first convolutional layer Conv1, a second convolutional layer Conv2, a third convolutional layer Conv3, a linear layer Linear, a first fusion unit (in Figure 5 " ”) and the second fusion unit (in Figure 5 " ” represents), wherein the first convolution layer Conv1 is used to perform a convolution operation on the input data, and the output end of the first convolution layer Conv1 is connected to the first input end of the first fusion unit; the linear layer Linear is used to perform a linear transformation on the time step after the position coding process and input it to the second input end of the first fusion unit, and the first fusion unit is used to fuse the data inputted from its first input end and the data inputted from its second input end (i.e., add them one by one) and output them; the input of the second convolution layer Conv2 is the data after the splicing of the first fusion unit and the data after the angular domain after the position coding process is multiplied, and the output end of the second convolution layer Conv2 is connected to the first output end of the second fusion unit; the third convolution layer Conv3 is used to perform a convolution operation on the input data to convert the input data into the same number of channels as the output data of the second convolution layer Conv2, and the output end of the third convolution layer Conv3 is connected to the input end of the second fusion unit; the second fusion unit is used to fuse the data inputted from its first input end and the data inputted from the second input end and output them to obtain the data features of the fused angular domain and time step.

[0052] As mentioned above, the structures of other residual blocks (the second residual block to the eighteenth residual block) in the one-dimensional conditional convolutional network are the same as the structure of the first residual block, and will not be repeated here.

[0053] like Figure 5 As shown, Indicates the number of channels of the input data of the current residual block, Represents the number of channels of the output data of the second convolutional layer Conv2, Indicates the dimension of the current feature data. The dotted line means that if , then the third convolutional layer Conv3 is used to perform a convolution operation on the input data so that the number of channels of the output data of the third convolutional layer Conv3 becomes , so that the number of channels of the data inputted at the first input end and the data inputted at the second input end of the second fusion unit are equal, and then the two data are added at the second fusion unit; if , then the number of channels of the input data and output data of the third convolution layer Conv3 remains unchanged, and the data input at the first input end and the data input at the second input end are fused at the second fusion unit. The first convolution layer, the second convolution layer, and the third convolution layer are all one-dimensional convolutions, and the data size before and after the convolution remains unchanged. represents the size of the convolution kernel, represents the step size of the convolution, Indicates the length of the data to be filled with 0 before and after. and Respectively represent the number of channels of the data before and after the convolution operation, then The choice of parameters in the first, second, and third convolutional layers is not unique and can satisfy Usually, a smaller convolution kernel is selected and the network depth is increased to achieve a larger receptive field while reducing the number of parameters. The residual block is the core module in the convolutional network used in this technology. It is responsible for adding the time step after position encoding to the sample features and introducing residual connections to achieve better gradient transmission when the network is deeper.

[0054] In addition, in this one-dimensional conditional convolutional network, when convolution operations are involved, the activation functions after the corresponding convolutional layers are all ReLU functions. The residual blocks (including the first residual block to the eighteenth residual block) integrate the conditional information (i.e., the angular domain) and the time step into the network while learning the sample features through convolution operations. Each upsampling layer enables the network to learn more spatial information and reduce the size of the feature map, thereby reducing memory usage and computation. In addition, the input and output sizes of the U-Net network are the same, which does not lose the details in the extracted features while meeting the requirement of DDPM that the input and output sizes must be the same during reverse reasoning.

[0055] Furthermore, according to the description of the DDPM model, the reverse denoising process is divided into Step by step, and It is usually large. Using the same network (the one-dimensional conditional convolutional network mentioned above in this embodiment) to process all time steps requires position encoding of the time steps to let the network know the current time step. The position encoding is responsible for transforming the time step and embedding it into the one-dimensional conditional convolutional network.

[0056] The specific method is to set the time step ( ) is encoded into a vector with the same dimension as the number of channels of the current sample, and then superimposed with the data sample for input. Each time step has a unique code and can represent the absolute or relative position relationship of the sample. The encoding method used in this embodiment is trigonometric function position encoding, which uses sine and cosine waves to represent the position of the model input, and distinguishes different positions by different frequencies. The calculation formula is:

[0057]

[0058] in, In the feature dimension On elements, is the number of elements in the dimension direction. The dimension of the position encoding is equal to the dimension of the HRRP data input sequence, and the time step can be integrated into the HRRP data input sequence by adding the two.

[0059] Similarly, the angular domain is encoded in the same way into a vector with a dimension equal to the number of channels of the current sample.

[0060] S3: Train the one-dimensional conditional convolutional network based on the small sample dataset to obtain the trained category-guided denoising diffusion probability model.

[0061] In this step, the loss function is used to train the one-dimensional conditional convolutional network built above. The first step is the forward noise addition process. Assume that the input sample is , the corresponding azimuth (i.e. angular domain) is , in the forward noise addition process The mean coefficient is continuously added to the input sample. , the variance is Gaussian noise, is a hyperparameter given in advance that is consistent with the data dimension, usually less than 1, that is, Adding noise gives ,exist Adding noise gives , and so on, we get the noise level increasing , this process does not require the participation of a one-dimensional conditional convolutional network, and Usually take a larger number, then you can consider is standard Gaussian noise.

[0062] Then, using the noisy training samples Train the above one-dimensional conditional convolutional network, the network input is , indicating the Samples that need to be denoised in time steps, output prediction noise , and input the position-encoded time step and angle domain To tell the network the current time step And integrate into the corner area.

[0063] See also Figure 6 , Figure 6 1 is a schematic diagram of the structure of a category-guided denoising diffusion probability model provided by an embodiment of the present invention. The difference between the category-guided denoising diffusion probability model (category-guided DDPM) and the general denoising diffusion probability model is that the category-guided denoising diffusion probability model adds category information when predicting noise. , that is, the angular domain. The conditional probability method is used to guide the probability distribution of the generation process using the angular domain. The sampling process of the improved category-guided denoising diffusion probability model is defined as ,in, , is a hyperparameter given in advance that is consistent with the data dimension, , Indicates The coefficient of the variance of the noise added at time steps, represents the identity matrix, the loss function Can be used express, It means to find the second norm.

[0064] Use this loss function The constructed one-dimensional conditional convolutional network is trained, and finally the one-dimensional conditional convolutional network can be used on the input and When completed Denoising restores This task results in a trained category-guided denoising diffusion probability model.

[0065] S4: Generate an augmented sample dataset using the trained category-guided denoising diffusion probability model.

[0066] Specifically, with standard Gaussian noise As input, determine the angular domain as , input the trained category-guided denoising diffusion probability model to obtain the time step Output data at ;

[0067] Will Input the trained category-guided denoising diffusion probability model to obtain the time step Output data at , and so on The denoised output of the step is is the predicted sample, i.e. the generated HRRP sample. Then, and different angular domains as input, and repeat the above steps to obtain the required expanded sample data set.

[0068] Embodiment 2

[0069] On the basis of Example 1, this embodiment also provides an HRRP sample generation system based on category-guided DDPM, the HRRP sample generation system includes a data preprocessing module and a sample generation module, wherein the data preprocessing module is used to normalize and perform centroid alignment preprocessing on the acquired HRRP original data to generate a small sample data set; the sample generation module is used to generate an expanded sample data set using the trained category-guided denoising diffusion probability model, wherein the trained category-guided denoising diffusion probability model is obtained by training a one-dimensional conditional convolutional network constructed using a small sample data set, the input of the one-dimensional conditional convolutional network is the sample to be denoised, and the output is the predicted noise with the same dimension as the input, and the one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that incorporate time steps and angular domains.

[0070] For the specific structure of the downsampling module and the upsampling module of the one-dimensional conditional convolutional network and the specific training process, please refer to Example 1, which will not be repeated here.

[0071] The effect of the HRRP sample generation method based on the angle-domain guided DDPM of the present invention is further illustrated below through simulation experiments.

[0072] (1) Simulation generation of training and test data

[0073] The HRRP data used in the present invention is simulation data, and the simulation radar parameters are center frequency 5520Mhz, bandwidth 400Mhz, horizontal polarization, elevation angle 84°, azimuth angle range 0°~180°, including four types of aircraft: CorsairⅡ, ZancudoT38, HawkerHT7, and MirageⅢ, with a total of 7680 samples, and the sample dimension is 256 dimensions. It is randomly and evenly divided into training sets and test sets. The number of samples in each category before division is 1920, with a total of 7680 samples. After division, the number of test sets is 3 times the number of training sets, that is, the number of training sets after division is 1920, and the number of training samples of four types of targets is 480. The number of test sets after division is 5760, and the number of test samples of four types of targets is 1440.

[0074] (2) Training models for comparative experiments

[0075] A comparative experiment is adopted, and the criterion for judging the generation effect is to train the CNN classifier by mixing real HRRP data samples with HRRP data samples generated by the model. The higher the accuracy of the classifier, the better the quality of the generated samples. At the same time, the existing variational autoencoder (VAE) and the category-guided denoising diffusion probability model (i.e., the trained one-dimensional conditional convolutional network) of this embodiment are trained using the same HRRP data samples, and the classifiers are trained with the generated samples to compare the accuracy. The convolution kernel size in each residual block is set as follows: the convolution kernel length of the first convolution layer Conv1 is 5, the convolution kernel length of the second convolution layer Conv2 is 3, the convolution kernel length of the third convolution layer Conv3 is 1, the convolution kernel length in the downsampling module is 5, the convolution kernel length in the upsampling module is 6, the convolution kernel length in the channel transformation convolution is 5, and the total number of diffusion steps (time steps) is =1000, , the disturbance term selection in the sampling process ,BatchSize=32,Epochs=100,Adam optimizer,and initial learning rate of 1e-5 were selected based on the device performance and data size.,The experimental results are shown in Table 1.

[0076] Table 1 Comparison of HRRP sample generation quality based on different methods

[0077]

[0078] As can be seen from Table 1, the introduction of generated samples improves the performance of the classifier, but the improvement is small. Among them, the accuracy of the samples generated by the category-guided denoising diffusion probability model is as high as 97.93%, which is 2.7% higher than the accuracy of 95.23% obtained by training only on real data, and higher than the 95.84% after injecting the generated samples of the variational autoencoder. In order to further simulate the extreme small sample situation, only 96 real samples are used to train the classifier and different generated models and retest them. The experimental results are shown in Table 2.

[0079] It can be seen from Table 2 that in the case of very few samples, the injection of samples generated by the variational autoencoder did not improve the performance of the classifier, even though these generated samples looked "almost" real. The accuracy of the samples generated by the category-guided denoising diffusion probability model of the present invention reached a maximum of 90.24%, which was 14.56% higher than the accuracy of 75.67% obtained by training only on real data.

[0080] Table 2 Comparison of HRRP sample generation quality based on different methods under very few samples

[0081]

[0082] The present invention proposes a HRRP sample generation method based on category-guided DDPM. Starting from the angle sensitivity of the HRRP signal, a DDPM model is constructed using a one-dimensional conditional convolutional network according to the characteristics of HRRP data. The reverse process of the DDPM model is conditionally modeled with probability, and the probability distribution of the generated samples is controlled by the angular domain as a condition. The angular domain is integrated to achieve high-quality HRRP sample generation. The hierarchical probability modeling method of the DDPM model is more suitable for learning the advantages of data distribution under small sample conditions, and solves the problems of unstable model training, poor quality of generated samples under small sample conditions, uncontrollable angular domain of samples generated by conventional diffusion models, and easy generation of samples outside the distribution in the field of HRRP sample generation. Compared with the existing methods, the HRRP samples generated by the present invention under small sample conditions can be used for classifier training and significantly improve the classification accuracy. The present invention models the reverse process of DDPM as conditional probability, uses the HRRP angular domain as a condition to guide model learning, and combines the detail generation ability of the probability statistical diffusion model under small sample conditions. A probability statistical diffusion model based on angular domain guidance is proposed, and it is applied to the radar HRRP sample generation task, thereby improving the HRRP target recognition accuracy under small sample conditions.

[0083] Another embodiment of the present invention provides a storage medium, in which a computer program is stored, and the computer program is used to execute the steps of the HRRP sample generation method based on category-guided DDPM described in the above embodiment. Another aspect of the present invention provides an electronic device, including a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus; the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform the operation corresponding to the HRRP sample generation method based on category-guided DDPM as described in any one of the above embodiments. Specifically, the above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including several instructions for enabling an electronic device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0084] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A HRRP sample generation method based on category-guided DDPM, characterized in that: include: The acquired HRRP raw data are normalized and preprocessed by centroid alignment to generate a small sample data set; A one-dimensional conditional convolutional network is constructed, wherein the input of the one-dimensional conditional convolutional network is a sample to be denoised, and the output is a predicted noise with the same dimension as the input. The one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that integrate the time step and the angular domain. The one-dimensional conditional convolutional network uses a residual block that integrates the time step information and the angle information as a basis to realize the simultaneous integration of the time step and the angular domain. The residual block is used to integrate the angular domain and the time step into the network while learning the sample features through the convolution operation. The residual block includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a linear layer, a first fusion unit and a second fusion unit, wherein, The first convolution layer is used to perform a convolution operation on the input data, and the output end of the first convolution layer is connected to the first input end of the first fusion unit; The linear layer is used to perform a linear transformation on the encoded time step and input it to the second input end of the first fusion unit, the first fusion unit is used to fuse the data input from the first input end and the data input from the second input end and output it, and the encoded time step has the same dimension as the input data; The input of the second convolution layer is the data obtained by multiplying the fused data of the first fusion unit and the angular domain after encoding, and the output end of the second convolution layer is connected to the first output end of the second fusion unit, wherein the angular domain after position encoding has the same dimension as the input data; The third convolutional layer is used to perform a convolution operation on the input data to convert the input data into the same number of channels as the output data of the second convolutional layer, and the output end of the third convolutional layer is connected to the input end of the second fusion unit; The second fusion unit is used to fuse and output the data input from the first input terminal and the data input from the second input terminal to obtain data features of the fused angle domain and time step; Using the small sample data set to train the one-dimensional conditional convolutional network to obtain a trained category-guided denoising diffusion probability model; The trained category-guided denoising diffusion probability model is used to generate an expanded sample data set.

2. The HRRP sample generation method based on category-guided DDPM according to claim 1, characterized in that: The acquired HRRP raw data is normalized and centroid aligned to generate a small sample data set, including: The acquired HRRP raw data is normalized to obtain the normalized HRRP data. : in, represents the HRRP original data, represents the number of distance units contained in the HRRP raw data, Indicates The amplitude within a distance unit; Based on the mass points of each HRRP original data, the normalized HRRP data Translate to make the center position coincide with the centroid position, and obtain the HRRP data after centroid alignment: in, represents the HRRP data after centroid alignment, represents the Fourier transform operation, represents the inverse Fourier transform operation, represents the phase of the centroid position of the HRRP data, represents the centroid position of the HRRP data, represents the phase of the HRRP data center location, Indicates the center position of HRRP data, Indicates the distance needed to move to align the center position with the center of mass position. Indicates plural units.

3. The HRRP sample generation method based on category-guided DDPM according to claim 1, characterized in that: The downsampling module includes a first convolution unit, a first downsampling unit, a second downsampling unit and a third downsampling unit connected in sequence, wherein: The input end of the first convolution unit is used to input the data to be processed, the output end of the first convolution unit is connected to the input end of the first downsampling unit, and the first convolution unit can output more channels to extract more features while keeping the data size unchanged; The first downsampling unit includes a first residual block, a second residual block and a first downsampling convolution layer connected in sequence, and the output end of the first downsampling convolution layer is connected to the input end of the second downsampling unit; the second downsampling unit includes a third residual block, a fourth residual block and a second downsampling convolution layer connected in sequence, and the output end of the second downsampling convolution layer is connected to the input end of the third downsampling unit; the third downsampling unit includes a fifth residual block, a sixth residual block, a third downsampling convolution layer, a seventh residual block, an eighth residual block and a ninth residual block connected in sequence.

4. The HRRP sample generation method based on category-guided DDPM according to claim 3, characterized in that: The up-sampling module includes a first up-sampling unit, a second up-sampling unit, a third up-sampling unit and a second convolution unit, wherein: The first upsampling unit sequentially includes a tenth residual block, an eleventh residual block, a twelfth residual block and a first upsampling transposed convolution layer, the input of the tenth residual block is the data of the output of the ninth residual block and the output of the eighth residual block after splicing in the channel dimension, the input of the eleventh residual block is the data of the output of the tenth residual block and the output of the eighth residual block after splicing in the channel dimension, and the input of the twelfth residual block is the data of the output of the eleventh residual block and the output of the seventh residual block after splicing in the channel dimension; The second upsampling unit includes a thirteenth residual block, a fourteenth residual block, a fifteenth residual block and a second upsampling transposed convolution layer in sequence, wherein the input of the thirteenth residual block is the data of the output of the first upsampling unit and the output of the fifth residual block after splicing in the channel dimension, the input of the fourteenth residual block is the data of the output of the thirteenth residual block and the output of the fourth residual block after splicing in the channel dimension, and the input of the fifteenth residual block is the data of the output of the fourteenth residual block and the output of the third residual block after splicing in the channel dimension; The third upsampling unit includes a sixteenth residual block, a seventeenth residual block, an eighteenth residual block and a third upsampling transposed convolution layer in sequence, wherein the input of the sixteenth residual block is the data obtained by splicing the output of the second upsampling unit and the output of the second residual block in the channel dimension, the input of the seventeenth residual block is the data obtained by splicing the output of the sixteenth residual block and the output of the first residual block in the channel dimension, and the input of the eighteenth residual block is the data obtained by splicing the output of the seventeenth residual block and the output of the first convolution unit in the channel dimension; The input end of the second convolution unit is connected to the output end of the third up-sampling unit, and is used to convert the output of the third up-sampling unit into a dimension consistent with the to-be-processed data input by the first convolution unit.

5. The HRRP sample generation method based on category-guided DDPM according to claim 4, characterized in that: The first residual block, the second residual block, the third residual block, the fourth residual block, the fifth residual block, the sixth residual block, the seventh residual block, the eighth residual block, the ninth residual block, the tenth residual block, the eleventh residual block, the twelfth residual block, the thirteenth residual block, the fourteenth residual block, the fifteenth residual block, the sixteenth residual block, the seventeenth residual block and the eighteenth residual block are residual blocks with the same structure, which are used to fuse the time step and angle domain after position encoding into the input feature data.

6. The HRRP sample generation method based on category-guided DDPM according to claim 5, characterized in that: The one-dimensional conditional convolutional network is trained using the small sample data set to obtain a trained category-guided denoising diffusion probability model, including: Perform the forward noise addition process, assuming that the input sample is , the corresponding angular domain is ,point time steps to the input sample Gaussian noise is continuously added to obtain training samples with increasing noise levels. ,in, Indicates The training samples after adding noise in time steps, is considered as standard Gaussian noise, represents the total time step, ≥ ≥1; Using the noisy training samples Train the one-dimensional conditional convolutional network, the input of the one-dimensional conditional convolutional network is and the current time step after position encoding and angle domain , output prediction noise ; The loss function is used for training to obtain the trained category-guided denoising diffusion probability model.

7. The HRRP sample generation method based on category-guided DDPM according to claim 6, characterized in that: The trained category-guided denoising diffusion probability model is used to generate an expanded sample data set, including: The standard Gaussian noise and its angular domain after position encoding and time step , input the trained category-guided denoising diffusion probability model to obtain the time step Output data at ; Will The angular domain and time step after position encoding are input into the trained category-guided denoising diffusion probability model to obtain the time step Output data at ; Repeatedly output the output data of the previous time step to the trained category-guided denoising diffusion probability model, after The HRRP samples generated by the denoising output of the step.

8. A HRRP sample generation system based on category-guided DDPM, characterized in that: The HRRP sample generation system is used to execute the HRRP sample generation method according to any one of claims 1 to 7, wherein the HRRP sample generation system comprises a data preprocessing module and a sample generation module, wherein: The data preprocessing module is used to perform normalization and centroid alignment preprocessing on the acquired HRRP raw data to generate a small sample data set; The sample generation module is used to generate an expanded sample data set using the trained category-guided denoising diffusion probability model, wherein the trained category-guided denoising diffusion probability model is obtained by training a one-dimensional conditional convolutional network constructed using the small sample data set, the input of the one-dimensional conditional convolutional network is the sample to be denoised, and the output is the predicted noise with the same dimension as the input, and the one-dimensional conditional convolutional network includes a downsampling module and an upsampling module that incorporate time steps and angular domains.

Images