HRRP Sample Generation Method Based on Conditional Denoising Diffusion Probability Model

Abstract

The present invention discloses an HRRP sample generation method based on a conditional denoising diffusion probability model, which relates to the technical field of radar signal processing and includes: generating a small sample training set; constructing a conditional denoising diffusion probability model and setting parameters; wherein, the conditional denoising diffusion probability model includes a U-Net network module and an output module, and the U-Net network module includes a downsampling layer, a skip connection layer and an upsampling layer; training the conditional denoising diffusion probability model according to the small sample training set to obtain a trained conditional denoising diffusion probability model; generating an augmented sample training set according to the trained conditional denoising diffusion probability model. The present invention can make full use of the azimuth information of radar HRRP samples, can generate HRRP samples at a specified azimuth, and the generated HRRP samples have high quality.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and particularly relates to a method for generating HRRP samples based on a conditional denoising diffusion probability model. Background Art

[0002] The High Resolution Range Profile (HRRP) of a radar refers to the fact that in a broadband radar, the echo signal of the target to be measured can be regarded as the vector sum of all scattered echoes within each resolved range cell. HRRP is one-dimensional information, containing characteristic information such as the geometric structure of the target and the energy distribution of scattering points. Compared with two-dimensional echo signals (SAR, ISAR), etc., it has the advantages of being easy to obtain, store, and process, and thus is very valuable for radar target recognition and classification.

[0003] In the prior art, when establishing an HRRP recognition database for non-cooperative targets of the enemy, it is difficult for the radar to detect and continuously track the target. Therefore, it is difficult to obtain sufficient HRRP samples covering all azimuth angles. Further, when using the samples of the incomplete HRRP recognition database as a training set to train the recognition system, due to the incomplete pose of the input HRRP samples, the recognition performance is poor, and the features extracted by the recognition system cannot represent the essential characteristics of the target, affecting the recognition performance and generalization ability of the classification system.

[0004] Therefore, it is urgent to improve the defects existing in the prior art. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a method for generating HRRP samples based on a conditional denoising diffusion probability model. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a method for generating HRRP samples based on a conditional denoising diffusion probability model, including:

[0007] Generating a small sample training set;

[0008] Constructing a conditional denoising diffusion probability model and setting parameters; wherein, the conditional denoising diffusion probability model includes a U-Net network module and an output module, and the U-Net network module includes a downsampling layer, a skip connection layer, and an upsampling layer;

[0009] Training the conditional denoising diffusion probability model according to the small sample training set to obtain a trained conditional denoising diffusion probability model;

[0010] Generating an augmented sample training set according to the trained conditional denoising diffusion probability model.

[0011] Advantages of the present invention:

[0012] A method for generating HRRP samples based on a conditional denoising diffusion probability model provided by the present invention. On the one hand, compared with the generative adversarial network (GAN) model that trains the network through the confrontation between the generator and the discriminator and requires training two networks, the network model proposed in the present invention is a conditional denoising diffusion probability model. This model is mainly divided into a forward diffusion process and a reverse generation process, and only one network needs to be trained. Compared with the GAN model, the diffusion model is easier to converge, has better stability, and higher diversity of generated samples. On the other hand, the present invention takes each sample in the training set together with the corresponding class label and azimuth angle label information as the input of the model network, overcoming the problem in the prior art that when generating HRRP samples, the azimuth angle is not considered, resulting in incomplete azimuth angles of the generated HRRP samples and affecting the recognition performance of the subsequent classification system. The present invention can make full use of the azimuth angle information of radar HRRP samples, can generate HRRP samples at a specified azimuth angle, and the generated HRRP samples have high quality.

[0013] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0014] Figure 1 is a flowchart of a method for generating HRRP samples based on a conditional denoising diffusion probability model provided by an embodiment of the present invention;

[0015] Figure 2 is a schematic structural diagram of a U-Net network module provided by an embodiment of the present invention;

[0016] Figure 3 is a schematic structural diagram of a CNN classifier recognition system provided by an embodiment of the present invention. Specific Embodiments

[0017] The following further describes the present invention in detail with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0018] In the prior art, Nie Jianghua from Nanchang Hangkong University disclosed a method for expanding the HRRP recognition database samples based on the Least Squares Generative Adversarial Networks (LSGAN) and Convolutional Neural Network (CNN) in the patent document "A HRRP Radar Target Recognition Method Based on Improved LSGAN and CNN" (Patent Application No.: 202010940775.2, Publication No.: 112230210A). This method first receives noisy and clean HRRP data through a data denoising module. Then, in the network structure design stage, a penalty term is added to the loss functions of both the discriminative network and the generative network to obtain the LSGAN network composed of the discriminative network and the generative network. Next, the LSGAN network is used to generate HRRP data with high signal-to-noise ratio for expanding the recognition database samples. Finally, the target recognition module receives the HRRP data, and the CNN is used for target recognition to obtain the recognition result. The drawback of this method is that when the number of samples in the HRRP recognition database is small, the deviation of the LSGAN network in feature extraction and feature selection of data will be relatively large, resulting in poor quality of the HRRP data generated by the LSGAN network. The recognition performance of the classification system trained after expanding the recognition database samples with the generated HRRP data is relatively low.

[0019] Ma Peiwen from Xidian University disclosed a method for expanding the HRRP recognition database samples based on the Conditional Auxiliary Classification Generative Adversarial Network (CACGAN) in the patent document "A Method for Expanding HRRP Recognition Database Samples Based on CACGAN" (Patent Application No.: 202110283773.5, Publication No.: 112784930A). This method first concatenates the samples with the corresponding class labels as the input of the Conditional Auxiliary Classification Generative Adversarial Network (CACGAN). Then, in the network structure design stage, a conditional auxiliary classification generative network composed of a generator, a discriminator, and an auxiliary classifier is constructed. A gradient penalty term is added to the loss function of the discriminator, and the cross-entropy loss function is used to calculate the loss value of the auxiliary classifier. Finally, the CACGAN network is used to generate HRRP samples of different classes for expanding the recognition database samples. The drawback of this method is that for each class of HRRP samples, the azimuth angle information is not considered as a label to generate samples at a specified azimuth angle, resulting in incomplete azimuth angles and poor quality of the HRRP samples generated by the CACGAN network. The recognition performance of the classification system trained after expanding the recognition database samples with the generated HRRP data is relatively low.

[0020] In view of this, the present invention provides a method for generating HRRP samples based on a conditional denoising diffusion probability model, which is used to solve the problem that when using LSGAN and CACGAN networks to expand the HRRP recognition database samples, the azimuth angles of the HRRP samples are incomplete and the azimuth angle information of the HRRP samples is not fully utilized, resulting in incomplete azimuth angles and poor quality of the generated HRRP data, and the recognition performance of the classifier trained after expanding the recognition database samples is low.

[0021] Please refer to Figures 1 to 2 as shown in Figure 1 FIG. is a flowchart of a method for generating HRRP samples based on a conditional denoising diffusion probability model provided by an embodiment of the present invention. Figure 2 FIG. is a schematic structural diagram of a U-Net network module provided by an embodiment of the present invention. A method for generating HRRP samples based on a conditional denoising diffusion probability model provided by the present invention includes:

[0022] S101. Generate a small sample training set.

[0023] Specifically, in this embodiment, the small sample training set is obtained through the following process.

[0024] S1011. Extract HRRP samples of multiple categories covering multiple azimuth angle domains from the radar echo along the range dimension on the radar line of sight.

[0025] S1012. Use the average frame division method to divide the HRRP samples of each category into multiple azimuth frames, and assign two labels to the category and the azimuth frame.

[0026] It should be noted that the average frame division method includes:

[0027] Obtain the maximum angle at which the scatter points do not undergo range cell migration, and equally divide the azimuth angle of the HRRP samples according to the maximum angle; wherein, the maximum angle The expression of is:

[0028]

[0029] wherein, L is the lateral size of the target relative to the radar, and ΔR is the range resolution cell length.

[0030] S1013. Combine all the HRRP samples, as well as the category labels and azimuth frame labels corresponding to the HRRP samples, to form the small sample training set.

[0031] It should be noted that the process of assigning category labels and azimuth frame labels to the HRRP samples includes:

[0032] Denote the class label of each HRRP sample with class serial number 1 in the small sample training set as y1, the class label of each HRRP sample with class serial number 2 as y2, and so on. Denote the class label of each HRRP sample with class serial number U as y U , where y1 takes the value of 1, y2 takes the value of 2, and so on. y U takes the value of U, and U represents the total number of class labels in the small sample training set. Denote the azimuth frame label of each HRRP sample with azimuth frame serial number 1 in the small sample training set as z1, the azimuth frame label of each HRRP sample with azimuth frame serial number 2 as z2, and so on. Denote the azimuth frame label of each HRRP sample with azimuth frame serial number V as z V , where z1 takes the value of 1, z2 takes the value of 2, and so on. z V takes the value of V, and V represents the total number of azimuth frame labels.

[0033] S102. Construct a conditional denoising diffusion probability model and set parameters. Among them, the conditional denoising diffusion probability model includes a U-Net network module and an output module, and the U-Net network module includes a downsampling layer, a skip connection layer, and an upsampling layer.

[0034] Specifically, please continue to refer to Figure 2 As shown, in this embodiment, it is necessary to first build the main body part of the conditional denoising diffusion probability model, the U-Net network module. Among them, the U-Net network module includes a downsampling layer, a skip connection layer, and an upsampling layer. The downsampling layer includes multiple layers, namely the first downsampling layer, the second downsampling layer, the third downsampling layer, the fourth downsampling layer, and the fifth downsampling layer. The output end of the first downsampling layer is connected to the input end of the second downsampling layer, the output end of the second downsampling layer is connected to the input end of the third downsampling layer, the output end of the third downsampling layer is connected to the input end of the fourth downsampling layer, and the output end of the fourth downsampling layer is connected to the input end of the fifth downsampling layer. Among them, the first downsampling layer sequentially includes a first convolutional module, a first residual module, a second residual module, and a second convolutional module. The second downsampling layer sequentially includes a third residual module, a fourth residual module, and a third convolutional module. The third downsampling layer sequentially includes a fifth residual module, a first self-attention module, a sixth residual module, a second self-attention module, and a fourth convolutional module. The fourth downsampling layer sequentially includes a seventh residual module and an eighth residual module. The fifth downsampling layer includes a ninth residual module;

[0035] The upsampling layer includes a multi-layer structure, namely the first upsampling layer, the second upsampling layer, the third upsampling layer, the fourth upsampling layer and the fifth upsampling layer. The input end of the first upsampling layer is connected to the output end of the second upsampling layer and the output end of the first downsampling layer. The input end of the second upsampling layer is connected to the output end of the third upsampling layer and the output end of the second downsampling layer. The input end of the third upsampling layer is connected to the output end of the fourth upsampling layer and the output end of the third downsampling layer. The input end of the fourth upsampling layer is connected to the output end of the fifth upsampling layer and the output end of the fourth downsampling layer. Among them, the structure of the first upsampling layer is the same as that of the first downsampling layer, the structure of the second upsampling layer is the same as that of the second downsampling layer, the structure of the third upsampling layer is the same as that of the third downsampling layer, the structure of the fourth upsampling layer is the same as that of the fourth downsampling layer, and the structure of the fifth upsampling layer is the same as that of the fifth downsampling layer.

[0036] The input end of the skip connection layer is connected to the output end of the fifth downsampling layer, and the output end of the skip connection layer is connected to the input end of the fifth upsampling layer. The skip connection layer is used to connect the feature map output by the downsampling layer to the upsampling layer.

[0037] In an optional embodiment of the present invention, the first residual module, the second residual module, the third residual module, the fourth residual module, the fifth residual module, the sixth residual module, the seventh residual module, the eighth residual module and the ninth residual module all include a convolution module one, a convolution module two and a position encoding module. The convolution module one includes a group normalization layer, a SiLU function and a 1×3 convolution layer arranged in sequence. The convolution module two includes a group normalization layer, a SiLU function, a Dropout and a 1×3 convolution layer arranged in sequence. The position encoding module includes a SiLU function and a fully connected layer arranged in sequence.

[0038] The first convolution module, the second convolution module, the third convolution module and the fourth convolution module are all composed of 1×3 convolution layers; the convolution kernel sizes used in the first self-attention module and the second self-attention module are both 1×1.

[0039] Among them, the expression of the SiLU function is:

[0040] f(w) = w * sigmoid(w);

[0041] Among them, w is the network parameter before passing through the activation layer, f(w) is the network parameter after passing through the activation layer, and sigmoid(·) is the activation function ReLU function.

[0042] In this embodiment, an output module of the conditional denoising diffusion probability model also needs to be built. The output module includes a group normalization layer, a SiLU function and a 1×3 convolution layer.

[0043] In this embodiment, it is also necessary to set the parameters of the conditional denoising diffusion probability model, which are divided into two parts. One part is the parameter setting of the U-Net network module, including the number of convolutional modules, the number of residual modules, the number of self-attention modules, the optimizer, the loss function, and the activation function. The other part is the training parameters of the conditional diffusion probability model Its expression is:

[0044]

[0045] where s is the offset, T is the total number of time steps, initialized to 500, t is a certain moment, and both t and T are integers

[0046] In this embodiment, please continue to refer to Figure 2 As shown, the feature map is input into the constructed conditional denoising diffusion probability model. After being processed by the first convolutional module, the first residual module, the second residual module, and the second convolutional module in the first downsampling layer in sequence, the processed feature map is output. This feature map is respectively transmitted to the first upsampling layer and the second downsampling layer. After being processed by the third residual module, the fourth residual module, and the third convolutional module in the second downsampling layer, the processed feature map is output until it is processed by the fifth downsampling layer and the processed feature map is output. This processed feature map is transmitted to the skip connection layer. After being processed by the skip connection layer, it is transmitted to the fifth upsampling layer. The fifth upsampling layer transmits the processed feature map to the fourth upsampling layer. This processed feature map is combined with the feature map processed by the fourth downsampling layer. The fourth upsampling layer processes the combined feature map until it is transmitted to the input end of the first upsampling layer. The first upsampling layer processes the combined feature map and processes it again by the group normalization layer, the SiLU function, and the convolutional layer in the output module in sequence to obtain the output structure. In this way, it can be ensured that the input of the U-Net network module is the HRRP sample, and the output is the noise of the same length as the HRRP sample

[0047] S103. Train the conditional denoising diffusion probability model to obtain a trained conditional denoising diffusion probability model

[0048] Specifically, in this embodiment, the conditional denoising diffusion probability model is trained through the following process

[0049] S1031. Use the conditional denoising diffusion probability model to process the small sample training set to obtain the predicted noise

[0050] S1032. Use the mean square error loss function to calculate the loss between the actual noise and the predicted noise; then use the backpropagation algorithm to iteratively update the parameters of the conditional denoising diffusion probability model until convergence, obtain the trained conditional denoising diffusion probability model, and save the parameters of the trained conditional denoising diffusion probability model

[0051] It should be noted that during the training process of the conditional denoising diffusion probability model, the loss function of the model is as follows:

[0052]

[0053] where θ is the trainable parameter of the U-Net network module, ∈ θ (·) is the predicted noise output by the U-Net network module, x0 is a set of HRRP samples sampled from the small sample training set, y is the category label vector corresponding to this set of HRRP samples, z is the azimuth frame label vector corresponding to this set of HRRP samples, t is the time, and ∈ is Gaussian noise.

[0054] S104. Generate an augmented sample training set according to the trained conditional denoising diffusion probability model.

[0055] Specifically, in this implementation, the augmented sample training set is generated through the following process.

[0056] S1041. Initialize the conditional denoising diffusion probability model according to the parameters of the trained conditional denoising diffusion probability model;

[0057] S1042. Randomly generate M noise samples from the normal distribution, and randomly generate the category label and azimuth frame label corresponding to the noise samples;

[0058] S1043. Use the initialized conditional denoising diffusion probability model to process the M noise samples, their category labels and azimuth frame labels, and generate HRRP samples at a specified azimuth angle to form a generated sample set;

[0059] S1044. Combine the generated sample set with the small sample training set to form an augmented sample training set.

[0060] In summary, the HRRP sample generation method based on the conditional denoising diffusion probability model provided by the present invention, on the one hand, compared with the generative adversarial network GAN model that trains the network through the confrontation between the generator and the discriminator and requires training two networks, the network model proposed by the present invention is a conditional denoising diffusion probability model. This model is mainly divided into a forward diffusion process and a reverse generation process, and this model only needs to train one network. Compared with the generative adversarial network GAN model, the diffusion model is easier to converge, has better stability, and has higher sample diversity. On the other hand, the present invention takes each sample in the training set together with the corresponding class label and azimuth angle label information as the input of the model network, overcoming the problem in the prior art that when generating HRRP samples, the azimuth angle is not considered, resulting in incomplete azimuth angles of the generated HRRP samples and affecting the recognition performance of the subsequent classification system. The present invention can make full use of the azimuth angle information of radar HRRP samples, can generate HRRP samples at a specified azimuth angle, and the generated HRRP samples have high quality.

[0061] In an optional embodiment of the present invention, the performance of the sample generation method proposed by the present invention is verified through simulation experiments.

[0062] I. Simulation Conditions

[0063] The hardware platform for the simulation experiment in this embodiment is: Intel i7-10700 2.9GHz, with a memory of 16GB, the operating system is Windows10, and the Python version is 3.9.

[0064] II. Simulation Content and Result Analysis

[0065] Simulation experiment 1 in this embodiment is to use the present invention to generate HRRP data at a specified azimuth angle, use the generated HRRP data to expand the samples in the small sample set, and obtain an expanded training set after expansion. Expand the small sample training set and the training set after expansion by using the method provided in the above embodiment, input the samples of the expanded training set into the CNN classifier recognition system to obtain a trained CNN classifier; input the samples of the test set generated in the simulation experiment of this embodiment into the trained CNN classifier respectively, and output the predicted class of each sample in the test set.

[0066] In the simulation experiment of this embodiment, both the recognition database and the training set used are HRRP electromagnetic simulation data of 5 types of aircraft. Among them, the small sample training set contains 1200 pieces of HRRP data of the first type, 1200 pieces of HRRP data of the second type, 1200 pieces of HRRP data of the third type, 1200 pieces of HRRP data of the fourth type, and 1200 pieces of HRRP data of the fifth type; the test sample set contains 1600 pieces of HRRP data of the first type, 1600 pieces of HRRP data of the second type, 1600 pieces of HRRP data of the third type, 1600 pieces of HRRP data of the fourth type, and 1600 pieces of HRRP data of the fifth type. Each HRRP sample contains 256 range cells.

[0067] The simulation experiment 1 of this embodiment is to generate the HRRP data by using the sample expansion method of the present invention to obtain the generated data set of the present invention. The generated data set contains 400 pieces of HRRP data of the first type, 400 pieces of HRRP data of the second type, 400 pieces of HRRP data of the third type, 400 pieces of HRRP data of the fourth type, and 400 pieces of HRRP data of the fifth type; use the method provided in the above embodiment to generate the data set to complete the sample expansion of the small sample training set, and obtain the expanded training set of the present invention after expansion.

[0068] Build a five-layer CNN classifier recognition system, the structure of which is successively the first convolutional layer, the second convolutional layer, the third convolutional layer, the first fully connected layer, and the second fully connected layer. Please refer to Figure 3 as shown Figure 3 is a schematic structural diagram of the CNN classifier recognition system provided by the embodiment of the present invention. The number of feature maps of the first to third convolutional layers are respectively set to 32, 64, and 128, the convolutional kernel size is set to 1×9, the convolutional kernel sliding step size is set to 1, the pooling downsampling kernel size is set to 1×2, the downsampling kernel sliding step is set to 2, and the input dimensions of the first and second fully connected layers are 4096 and 128 respectively, and the output dimensions are 128 and 5 respectively.

[0069] Input the training sets before and after expansion into the CNN classifier respectively. After 300 iterations of training, two trained CNN classifiers are obtained; use the two CNN classifiers to predict the category of each sample in the test set respectively, and then calculate the ratio of the number of test samples whose predicted category by the two CNN classifiers for each sample in the test set matches the category of the sample to the total number of test samples respectively, to obtain two target recognition accuracies; the higher the target recognition accuracy, the higher the recognition performance of the CNN classifier, and the more complete the azimuth angle of the HRRP samples in the expanded training set.

[0070] The results of the above two target recognition accuracies are shown in Table 1.

[0071] Table 1 Comparison table of target recognition accuracies

[0072]

[0073]

[0074] As can be seen from Table 1, the recognition performance of the CNN trained after sample augmentation using the present invention is better than that of the CNN trained with a small sample training set, indicating that the present invention generates HRRP samples with a specified azimuth angle, and the generated HRRP has a high recognition performance for the classification system of the CNN trained after sample augmentation of the small sample training set.

[0075] It should be noted that in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant are intended to cover non-exclusive inclusion, so that an article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the article or device including the said element. Terms such as "connected" or "coupled" do not necessarily refer to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0076] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0077] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, which should all be regarded as falling within the protection scope of the present invention.

Claims

1. A method for generating HRRP samples based on a conditional denoising diffusion probability model, characterized in that, Including: Generating a small sample training set; Constructing a conditional denoising diffusion probability model and setting parameters; wherein, the conditional denoising diffusion probability model includes a U-Net network module and an output module, and the U-Net network module includes a downsampling layer, a skip connection layer and an upsampling layer; Training the conditional denoising diffusion probability model according to the small sample training set to obtain a trained conditional denoising diffusion probability model; Generating an augmented sample training set according to the trained conditional denoising diffusion probability model, including: Initializing the conditional denoising diffusion probability model according to the parameters of the trained conditional denoising diffusion probability model; Randomly generate from a normal distribution noise samples, and randomly generate class labels and azimuth frame labels corresponding to the noise samples; Using the initialized conditional denoising diffusion probability model, process the noise samples, their class labels, and azimuth frame labels to generate HRRP samples at a specified azimuth angle, and form a generated sample set; Combining the generated sample set with the small sample training set to form an augmented sample training set.

2. The HRRP sample generation method based on the conditional denoising diffusion probability model according to claim 1, wherein The process of generating the small sample training set includes: Extracting HRRP samples of multiple categories covering multiple azimuth domains from the radar echo along the range dimension on the radar line of sight; Dividing the HRRP samples of each category into multiple azimuth frames using the average frame division method, and assigning two labels to the category and the azimuth frame; Composing the small sample training set with all the HRRP samples, as well as the category labels and azimuth frame labels corresponding to the HRRP samples.

3. The HRRP sample generation method based on the conditional denoising diffusion probability model according to claim 2, wherein, The average frame division method includes: Obtain the maximum angle at which the scatter points do not undergo range cell migration, and equally divide the azimuth angles of the HRRP samples according to this maximum angle; among them, the maximum angle has the following expression: ； Among them, is the lateral dimension of the target relative to the radar, is the length of the range resolution cell.

4. The HRRP sample generation method based on the conditional denoising diffusion probability model according to claim 1, wherein The downsampling layer includes multiple layers, namely the first downsampling layer, the second downsampling layer, the third downsampling layer, the fourth downsampling layer and the fifth downsampling layer. The output end of the first downsampling layer is connected to the input end of the second downsampling layer, the output end of the second downsampling layer is connected to the input end of the third downsampling layer, the output end of the third downsampling layer is connected to the input end of the fourth downsampling layer, and the output end of the fourth downsampling layer is connected to the input end of the fifth downsampling layer; wherein, the first downsampling layer sequentially includes a first convolution module, a first residual module, a second residual module and a second convolution module, the second downsampling layer sequentially includes a third residual module, a fourth residual module and a third convolution module, the third downsampling layer sequentially includes a fifth residual module, a first self-attention module, a sixth residual module, a second self-attention module and a fourth convolution module, the fourth downsampling layer sequentially includes a seventh residual module and an eighth residual module, and the fifth downsampling layer includes a ninth residual module; The upsampling layer includes a multi-layer structure, namely the first upsampling layer, the second upsampling layer, the third upsampling layer, the fourth upsampling layer, and the fifth upsampling layer. The input end of the first upsampling layer is connected to the output end of the second upsampling layer and the output end of the first downsampling layer. The input end of the second upsampling layer is connected to the output end of the third upsampling layer and the output end of the second downsampling layer. The input end of the third upsampling layer is connected to the output end of the fourth upsampling layer and the output end of the third downsampling layer. The input end of the fourth upsampling layer is connected to the output end of the fifth upsampling layer and the output end of the fourth downsampling layer. Among them, the structure of the first upsampling layer is the same as that of the first downsampling layer, the structure of the second upsampling layer is the same as that of the second downsampling layer, the structure of the third upsampling layer is the same as that of the third downsampling layer, the structure of the fourth upsampling layer is the same as that of the fourth downsampling layer, and the structure of the fifth upsampling layer is the same as that of the fifth downsampling layer; The input end of the skip connection layer is connected to the output end of the fifth downsampling layer, and the output end of the skip connection layer is connected to the input end of the fifth upsampling layer. The skip connection layer is used to connect the feature map output by the downsampling layer to the upsampling layer.

5. The HRRP sample generation method based on the conditional denoising diffusion probability model according to claim 4, wherein The first residual module, the second residual module, the third residual module, the fourth residual module, the fifth residual module, the sixth residual module, the seventh residual module, the eighth residual module, and the ninth residual module all include a convolution module one, a convolution module two, and a position encoding module. The convolution module one includes a group normalization layer, a SiLU function, and a convolutional layer arranged in sequence. The convolution module two includes a group normalization layer, a SiLU function, a Dropout, and a convolutional layer arranged in sequence. The position encoding module includes a SiLU function and a fully connected layer arranged in sequence.

6. The HRRP sample generation method based on conditional denoising diffusion probability model according to claim 5, wherein The expression of the SiLU function is: ； Among them, are the network parameters before passing through the activation layer, are the network parameters after passing through the activation layer, is the ReLU activation function.

7. The HRRP sample generation method based on conditional denoising diffusion probability model according to claim 1, characterized in that, The process of training the conditional denoising diffusion probability model according to the small sample training set to obtain a trained conditional denoising diffusion probability model includes: Using the conditional denoising diffusion probability model to process the small sample training set to obtain predicted noise; Using the mean squared error loss function to calculate the loss between the actual noise and the predicted noise; then using the backpropagation algorithm to iteratively update the parameters of the conditional denoising diffusion probability model until convergence, obtaining a trained conditional denoising diffusion probability model, and saving the parameters of the trained conditional denoising diffusion probability model.

8. The HRRP sample generation method based on conditional denoising diffusion probabilistic model according to claim 7, characterized in that, The loss function of the U-Net network module is: ； ； Among them, are the trainable parameters of the U-Net network module, is the predicted noise output by the U-Net network module, is a set of HRRP samples sampled from the small-sample training set, is the class label vector corresponding to this set of HRRP samples, is the azimuth frame label vector corresponding to this set of HRRP samples, is Gaussian noise, are the initialized parameters, is the offset, is the total number of time steps, is a certain moment, and The values are all integers.

Images