Method for training and method for identifying target recognition model of HRRP noise label based on meta learning

Abstract

The application provides a training method and a recognition method of a target recognition model of HRRP noise labels based on meta learning, and the method comprises the following steps: constructing a meta dataset based on HRRP and a training dataset based on HRRP and with label noise; training a preset diffusion model by using the meta dataset and supplementing the training dataset obtained by training into the meta dataset to obtain a new meta dataset; pre-training an initial recognition model by using the training dataset for multiple rounds to obtain a pre-trained recognition model; updating the labels of each second sample according to the meta model A and the meta model B, the new meta dataset and the original labels of each second sample to generate a new training dataset, and training the pre-trained recognition model based on the new training dataset until the model converges to obtain a radar target recognition model based on HRRP; and the method can reduce the interference of label noise of the training sample and improve the model training effect by updating the labels of the training sample.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and specifically to a training method and recognition method for a target recognition model based on meta-learning and HRRP noise labels. Background Technology

[0002] High-resolution range profiles (HRRPs) reflect the distribution of target scattering centers along the radar line of sight, containing a wealth of structural information about the target. They are easy to acquire, store, and process, and have thus attracted continuous attention in the field of automatic radar target identification. Existing HRRP target identification methods typically assume that the labels on the HRRP data are completely accurate; however, this assumption is difficult to meet in real-world battlefield environments for non-cooperative targets.

[0003] In actual combat, electronic jamming and deception, the complexity of the battlefield environment, enemy camouflage and deception strategies, and human mislabeling all contribute to label noise during the HRRP data collection process for non-cooperative targets. In noisy label scenarios, the performance of traditional HRRP target identification methods deteriorates sharply. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a training method and a recognition method for a target recognition model based on meta-learning for HRRP noise labels. Specifically, it includes:

[0005] In a first aspect, the present invention provides a training method for a target recognition model based on meta-learning for HRRP noise labels, comprising:

[0006] Construct a metadata dataset and a training dataset based on HRRP. The metadata dataset includes multiple first samples, each with the correct original label. The training dataset includes multiple second samples, each with label noise in its original label. The number of samples in the metadata dataset is much smaller than the number of samples in the training dataset.

[0007] The pre-defined diffusion model is trained using the meta-dataset, and the trained dataset is added to the meta-dataset to obtain a new meta-dataset.

[0008] The initial recognition model is pre-trained multiple times using the training dataset to obtain a pre-trained recognition model.

[0009] Based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample, the labels of each second sample are updated to generate a new training dataset. Based on the new training dataset, the pre-trained recognition model is trained until the model converges, thereby obtaining the radar target recognition model based on HRRP.

[0010] The step of updating the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset, and training the pre-trained recognition model based on the new training dataset until the model converges, includes:

[0011] The first predicted label corresponding to each second sample in the training dataset is obtained by predicting using the pre-trained recognition model;

[0012] The first predicted label corresponding to each first sample in the new metadata set is obtained by predicting the pre-trained recognition model, and the parameters of the meta-model A and the meta-model B are updated according to the original label and the first predicted label of each first sample to obtain the first meta-model A and the first meta-model B.

[0013] Based on the original label and the first predicted label of each second sample, and the first meta-model A, the first weight corresponding to each second sample is obtained;

[0014] Based on the original label and the first predicted label of each second sample, and the first meta-model B, the second weight corresponding to each second sample is obtained;

[0015] Based on the first weight, second weight, original label, and first predicted label corresponding to each second sample, the first soft label corresponding to each second sample is determined;

[0016] Replace the original label of each second sample with the corresponding first soft label to obtain the first new training dataset;

[0017] The pre-trained recognition model is trained multiple times using the new training dataset until the model converges; wherein, the step of training the pre-trained recognition model multiple times using the new training dataset until the model converges includes:

[0018] In the t-th round of training, t is a positive integer greater than or equal to 2.

[0019] The recognition model obtained through the (t-1)th round of training predicts the t-th predicted label corresponding to each of the first samples in the new metadata set, and updates the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B according to the original label and the t-th predicted label of each of the first samples, to obtain the t-th meta-model A and the t-th meta-model B. The parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are generated during the (t-1)th round of training.

[0020] Based on the original label and the t-th predicted label corresponding to each first sample, and the t-th meta-model A, the t-th first weight corresponding to each first sample is obtained;

[0021] Based on the (t-1)th soft label and the tth predicted label corresponding to each of the first samples, and the tth meta-model B, the tth second weight corresponding to each of the first samples is obtained;

[0022] Based on the first weight, second weight, original label, t-th test label, and t-1-th soft label corresponding to each first sample, the t-th soft label corresponding to each first sample is determined;

[0023] Replace the original label of each of the first samples with the corresponding t-th soft label to obtain the t-th new training dataset;

[0024] The recognition model obtained from the (t-1)th training round is trained based on the t-th new training dataset to update the model parameters; wherein, the t-th soft label corresponding to each first sample is determined based on the first weight, second weight, original label, first predicted label, and (t-1)-th soft label, as expressed as:

[0025]

[0026] in, This represents the t-th soft tag. Indicates the first weight. Indicates the original label. Indicates the second weight. This represents the (t-1)th soft tag. This represents the t-th test label. The parameters of metamodel A are represented. This represents the parameters of meta-model B.

[0027] Secondly, the present invention also provides a target recognition method for HRRP noise labels based on meta-learning, comprising:

[0028] The first HRRP corresponding to the radar target to be identified;

[0029] The first HRRP is input into the HRRP-based radar target recognition model trained according to any of the methods provided in the first aspect, in order to identify the type of radar target.

[0030] Thirdly, the present invention provides a training apparatus for a target recognition model based on meta-learning and HRRP noise labels, comprising:

[0031] The building module is used to build the HRRP-based metadata dataset and the HRRP-based training dataset. The metadata dataset includes multiple first samples, each with the correct original label. The training dataset includes multiple second samples, each with label noise in its original label. The number of samples in the metadata dataset is much smaller than the number of samples in the training dataset.

[0032] The training module is used to train the pre-defined diffusion model using the meta-dataset and then add the trained dataset to the meta-dataset to obtain a new meta-dataset.

[0033] The training module is also used to perform multiple rounds of pre-training on the initial recognition model using the training dataset to obtain a pre-trained recognition model;

[0034] The training module is also used to train the pre-trained recognition model based on meta-model A and meta-model B, as well as the new meta-dataset and training dataset, until the model converges, thereby obtaining a radar target recognition model based on HRRP.

[0035] The step of updating the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset, and training the pre-trained recognition model based on the new training dataset until the model converges, includes:

[0036] The first predicted label corresponding to each second sample in the training dataset is obtained by predicting using the pre-trained recognition model;

[0037] The first predicted label corresponding to each first sample in the new metadata set is obtained by predicting the pre-trained recognition model, and the parameters of the meta-model A and the meta-model B are updated according to the original label and the first predicted label of each first sample to obtain the first meta-model A and the first meta-model B.

[0038] Based on the original label and the first predicted label of each second sample, and the first meta-model A, the first weight corresponding to each second sample is obtained;

[0039] Based on the original label and the first predicted label of each second sample, and the first meta-model B, the second weight corresponding to each second sample is obtained;

[0040] Based on the first weight, second weight, original label, and first predicted label corresponding to each second sample, the first soft label corresponding to each second sample is determined;

[0041] Replace the original label of each second sample with the corresponding first soft label to obtain the first new training dataset;

[0042] The pre-trained recognition model is trained multiple times using the new training dataset until the model converges; wherein, the step of training the pre-trained recognition model multiple times using the new training dataset until the model converges includes:

[0043] In the t-th round of training, t is a positive integer greater than or equal to 2.

[0044] The recognition model obtained through the (t-1)th round of training predicts the t-th predicted label corresponding to each of the first samples in the new metadata set, and updates the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B according to the original label and the t-th predicted label of each of the first samples, to obtain the t-th meta-model A and the t-th meta-model B. The parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are generated during the (t-1)th round of training.

[0045] Based on the original label and the t-th predicted label corresponding to each first sample, and the t-th meta-model A, the t-th first weight corresponding to each first sample is obtained;

[0046] Based on the (t-1)th soft label and the tth predicted label corresponding to each of the first samples, and the tth meta-model B, the tth second weight corresponding to each of the first samples is obtained;

[0047] Based on the first weight, second weight, original label, t-th test label, and t-1-th soft label corresponding to each first sample, the t-th soft label corresponding to each first sample is determined;

[0048] Replace the original label of each of the first samples with the corresponding t-th soft label to obtain the t-th new training dataset;

[0049] The recognition model obtained from the (t-1)th training round is trained based on the t-th new training dataset to update the model parameters; wherein, the t-th soft label corresponding to each first sample is determined based on the first weight, second weight, original label, first predicted label, and (t-1)-th soft label, as expressed as:

[0050]

[0051] in, This represents the t-th soft tag. Indicates the first weight. Indicates the original label. Indicates the second weight. This represents the (t-1)th soft tag. This represents the t-th test label. The parameters of metamodel A are represented. This represents the parameters of meta-model B.

[0052] Fourthly, the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0053] Memory, used to store computer programs;

[0054] The processor, when executing a program stored in memory, implements any of the methods provided in the first aspect.

[0055] Fifthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods provided in the first aspect.

[0056] The beneficial effects of this invention are:

[0057] This invention provides a training and recognition method for a target recognition model based on meta-learning and HRRP noise labels. The method involves constructing an HRRP-based meta-dataset and an HRRP-based training dataset. The meta-dataset includes multiple first samples, each with a correct original label. The training dataset includes multiple second samples, each with label noise in its original label. The number of samples in the meta-dataset is significantly smaller than the number of samples in the training dataset. A pre-defined diffusion model is trained using the meta-dataset, and the trained dataset is then added to the meta-dataset to obtain a new meta-dataset. Finally, the initial recognition model is pre-trained multiple times using the training dataset to obtain a pre-trained model. The recognition model is constructed by updating the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset. The pre-trained recognition model is then trained on the new training dataset until the model converges, thus obtaining a radar target recognition model based on HRRP. This model can generate a small-scale clean dataset using a diffusion model, which can be used to expand the constructed meta-dataset, increasing the number of metadata samples and preventing model overfitting. In addition, updating the labels of the training samples can reduce the interference of noise in the training sample labels, improve the model training effect, and enhance the model's accuracy and reliability.

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0059] Figure 1 A flowchart illustrating the training method for a target recognition model based on meta-learning for HRRP noise labels provided by this invention.

[0060] Figure 2 A schematic diagram of generated HRRP data provided by the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of a training device for a target recognition model based on meta-learning HRRP noise labels provided by the present invention. Detailed Implementation

[0062] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0063] Radar target identification refers to determining the type of a target using its radar echo signals. HRRP (High-Resolution Radar Target Recognition) is the vector sum of the projections of target scattering point echoes onto the radar line of sight obtained from broadband signals, containing information about the target's structural dimensions. HRRP-based radar target identification has become a research hotspot in the field. Existing HRRP radar target identification methods are all implemented under the condition that the HRRP data tags are completely correct; however, this condition cannot be met in actual battlefield situations for non-cooperative targets.

[0064] In actual battlefield operations, HRRP data acquisition for non-cooperative targets is often affected by multiple factors. First, electronic jamming and deception play a crucial role in modern warfare. Jamming signals leading to incorrect radar tracking of non-cooperative targets and false radar echoes directly contribute to label noise in HRRP data. Second, the complexities of actual battlefield conditions, especially in environments with dense or obscured targets, mean that inaccurate detection or tracking can cause radar to mistakenly identify a signal from one target as coming from another, or to mix signals from multiple targets, thus generating label noise. Furthermore, from an intelligence gathering perspective, when the enemy employs camouflage and deception strategies to conceal true target information, the radar system will collect incorrect target information, resulting in label noise. Finally, human error in labeling is unavoidable during HRRP data acquisition, also contributing to label noise. In scenarios with label noise, the performance of traditional HRRP target identification methods degrades sharply.

[0065] Figure 1 This is a flowchart illustrating a training method for a target recognition model based on meta-learning using HRRP noise labels provided by the present invention. Figure 1 As shown, the method includes:

[0066] S101. Construct the HRRP-based meta-dataset and the HRRP-based training dataset.

[0067] The metadata set includes multiple first samples, and the original label of each first sample is the correct label.

[0068] The training dataset includes multiple second samples, and the original labels of each second sample contain label noise.

[0069] The number of samples in the metadata set is much smaller than the number of samples in the training dataset.

[0070] S102. The preset diffusion model is trained using the meta dataset, and the trained dataset is added to the meta dataset to obtain a new meta dataset.

[0071] In one possible implementation, the pre-defined diffusion model includes one time-step encoding module, one feature map framework module, four downsampling modules, two intermediate transition blocks, and one image restoration module.

[0072] Specifically, the temporal encoding module includes a first fully connected layer, one GELU activation layer, and a second fully connected layer; the first fully connected layer has 128 input units and 512 output units; the second fully connected layer has 512 input units and 512 output units. The feature map construction module includes one convolutional layer with 1 input channel, 128 output channels, a 1*3 kernel size, and a stride of 1. The downsampling module includes a first GroupNorm layer, a first Swish activation layer, a first convolutional layer, a second Swish activation layer, a second fully connected layer, a third GroupNorm layer, a third Swish activation layer, a third convolutional layer, and a fourth convolutional layer. The parameters of the four downsampling modules are set as follows: the first convolutional layer has 128, 256, 256, and 256 input channels, and 256 output channels for all layers, with a kernel size of 1*3 and a stride of 1; the second fully connected layer has 128 input units and 256 output units for all layers; the third convolutional layer has 128, 256, 256, and 256 input channels, and 256 output channels for all layers, with a kernel size of 1*3 and a stride of 1 for all layers; the fourth convolutional layer has 256 input and 256 output channels for all layers, with a kernel size of 1*3 and a stride of 2 for all layers; the first, second, and third GroupNorm layers each have 32 groups. The intermediate transition module includes a first GroupNorm layer, a first Swish activation layer, a first convolutional layer, a second Swish activation layer, a second fully connected layer, a third GroupNorm layer, a third Swish activation layer, and a third convolutional layer. The parameters of the two intermediate transition modules are set as follows: the first convolutional layer has 256 input channels and 256 output channels, with a kernel size of 1*3 and a stride of 1; the second fully connected layer of both intermediate transition modules has 128 input units and 256 output units; the third convolutional layer of both intermediate transition modules has 256 input channels and 256 output channels, with a kernel size of 1*3 and a stride of 1; the first and second GroupNorm layers each have 32 groups. The upsampling module is implemented as follows: the upsampling module includes a first GroupNorm layer, a first Swish activation layer, a first convolutional layer, a second Swish activation layer, a second fully connected layer, a third GroupNorm layer, a third Swish activation layer, a third convolutional layer, a fourth nearest neighbor interpolation layer, and a fourth convolutional layer.The parameters of the four upsampling modules are set as follows: the first convolutional layer has 256 input channels and output channels of 256, 256, 256, and 128 respectively, with a kernel size of 1*3 and a stride of 1; the second fully connected layer has 128 input units and output units of 256, 256, 256, and 128 respectively; the third convolutional layer has 256 input channels and output channels of 256, 256, 256, and 128 respectively, with a kernel size of 1*3 and a stride of 1; the fourth nearest neighbor interpolation layer has an interpolation factor of 2; the fourth convolutional layer has 256 input channels and output channels of 256, 256, 256, and 128 respectively, with a kernel size of 1*3 and a stride of 2; the first, second, and third GroupNorm layers have 32 groups. The image restoration module includes a first GroupNorm layer, a first Swish activation layer, and a first convolutional layer. Image restoration module parameter settings: The first convolutional layer has 128 input channels, 1 output channel, 1*3 kernel size, and 1 stride; the GroupNorm layer has 32 groups.

[0073] In one possible implementation, a meta-dataset is used to train a pre-defined diffusion model, including a forward diffusion process, a backward generation process, and an HRRP data generation process.

[0074] Specifically, the forward diffusion process includes:

[0075] Meta dataset The data is gradually noise-added, and the noise-addition formula is expressed as:

[0076]

[0077]

[0078]

[0079] in, Indicates the first Initial HRRP data, Indicates the first HRRP moment The noise-adding result, This represents standard Gaussian white noise. express The noise variance at any given time.

[0080] Specifically, the reverse generation process includes:

[0081] Noisy HRRP data corresponding to all HRRP data in the small batch sample set. Time step Noise graph Simultaneously, the input is fed into the constructed deep conditional diffusion network, where , The feature map construction module and time step encoding module are respectively input into the network. The deep conditional diffusion network performs noise estimation on each input noisy HRRP data to obtain the noise map predicted by the network.

[0082] The AdamW optimizer is used to update all parameters to be learned, and the loss function is expressed as follows:

[0083]

[0084] in, This represents the total loss for all input samples in the mini-batch set. This represents the output of the predicted noise map from the deep conditional diffusion network. This represents the i-th HRRP data point in the mini-batch sample set. This represents the time step corresponding to the i-th noisy HRRP data point. This represents the noise map corresponding to the i-th HRRP data point.

[0085] Specifically, the HRRP data generation process includes:

[0086] The formula for generating HRRP data is expressed as follows:

[0087]

[0088]

[0089] in, , , express noise variance at time step This indicates the k-th HRRP data to be generated. Noise graph at time step, This indicates the k-th HRRP data to be generated. Noise graph at time step This represents the output of the predicted noise map from the deep conditional diffusion network during the HRRP data generation stage. Indicates the current time step Generate HRRP data.

[0090] S103. The initial recognition model is pre-trained multiple times using the training dataset to obtain a pre-trained recognition model.

[0091] The recognition model can specifically be a deep convolutional network.

[0092] Because the model is highly robust to label noise in the early stages of training, pre-training can improve the reliability of the model's prediction results.

[0093] S104. Based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample in the training dataset, update the labels of each second sample to generate a new training dataset. Based on the new training dataset, train the pre-trained recognition model until the model converges, thereby obtaining the HRRP-based radar target recognition model.

[0094] Meta-model A and meta-model B have the same architecture, specifically a shallow convolutional neural network.

[0095] In one possible implementation, based on meta-model A and meta-model B, a new meta-dataset, and the original labels of each second sample, the labels of each second sample are updated to generate a new training dataset. Then, based on the new training dataset, the pre-trained recognition model is trained until convergence. This includes: predicting the first predicted label corresponding to each second sample in the training dataset using the pre-trained recognition model; predicting the first predicted label corresponding to each first sample in the new metadata set using the pre-trained recognition model; and updating the parameters of meta-model A and meta-model B based on the original labels and the first predicted labels of each first sample to obtain the first meta-model A. And the first meta-model B; based on the original label and the first predicted label of each second sample, and the first meta-model A, obtain the first weight corresponding to each second sample; based on the original label and the first predicted label of each second sample, and the first meta-model B, obtain the second weight corresponding to each second sample; based on the first weight, the second weight, the original label, and the first predicted label corresponding to each second sample, determine the first soft label corresponding to each second sample; replace the original label of each second sample with the corresponding first soft label to obtain the first new training dataset; train the pre-trained recognition model multiple times based on the new training dataset until the model converges.

[0096] Furthermore, in one possible implementation, in the t-th training round (t is a positive integer greater than or equal to 2), the pre-trained recognition model is trained multiple times based on the new training dataset, including steps A1-A6, as shown below:

[0097] Step A1: The recognition model obtained from the (t-1)th round of training predicts the t-th predicted label corresponding to each first sample in the new metadata set, and updates the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B according to the original label of each first sample and the t-th predicted label, so as to obtain the t-th meta-model A and the t-th meta-model B.

[0098] The parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are generated during the (t-1)th round of training.

[0099] Specifically, in one possible implementation, the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are updated based on the original labels of each first sample and the t-th predicted label to obtain the t-th meta-model A and the t-th meta-model B, including steps a1-a3, as shown below:

[0100] a1. Calculate the difference between the original label and the t-th predicted label of each first sample using the cross-entropy loss function, and construct the meta-model loss gradients corresponding to the (t-1)-th meta-model A and the (t-1)-th meta-model B based on the calculation results. The representations of the (t-1)-th meta-model A and the (t-1)-th meta-model B are as follows:

[0101]

[0102] in, Let represent the meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B. Indicates the index of the first sample. This indicates that the gradient of the parameters of the two meta-models is calculated. Represents the cross-entropy loss function. This represents the t-th predicted label of the first sample. Represents the metadata dataset of the first generation. The label of each sample, Represents the metadata dataset of the first generation. One sample, This represents the parameters of the recognition model obtained in the t-th round of training;

[0103] a2. Based on the preset gradient enhancement model, update the corresponding meta-models A and B of the (t-1)th meta-model to obtain the updated corresponding meta-models A and B. The preset gradient enhancement model is expressed as follows:

[0104]

[0105] in, This represents the meta-model loss gradient corresponding to the updated (t-1)th meta-model A and the (t-1)th meta-model B. Let represent the meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B. Indicates and A vector of the same dimension, where each term is a random variable following a Gaussian distribution. This indicates that the mean is 1 and the variance is... Gaussian distribution, This represents a probability parameter, which can be set to 0.01 for example.

[0106] a3. Based on the updated meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B, update the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B to obtain the t-th meta-model A and the t-th meta-model B, expressed as:

[0107]

[0108] in, Let A represent the parameters of the t-th meta-model A and the t-th meta-model B. Let represent the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B. The learning rates of meta-model A and meta-model B.

[0109] Step A2: Based on the original label and the t-th predicted label corresponding to each first sample, and the t-th meta-model A, obtain the t-th first weight corresponding to each first sample.

[0110] Step A3: Based on the (t-1)th soft label and the tth predicted label corresponding to each first sample, and the tth meta-model B, obtain the tth second weight corresponding to each first sample.

[0111] Step A4: Determine the t-th soft label corresponding to each first sample based on the first weight, second weight, original label, t-th test label, and t-1-th soft label.

[0112] In one possible implementation, the t-th soft label corresponding to each first sample is determined based on the first weight, second weight, original label, first predicted label, and (t-1)-th soft label, and is expressed as follows:

[0113]

[0114] in, This represents the t-th soft tag. Indicates the first weight. Indicates the original label. Indicates the second weight. This represents the (t-1)th soft tag. This represents the t-th test label. The parameters of metamodel A are represented. This represents the parameters of meta-model B.

[0115] Step A5: Replace the original label of each first sample with the corresponding t-th soft label to obtain the t-th new training dataset.

[0116] Step A6: Train the recognition model obtained in the (t-1)th round of training based on the t-th new training dataset to update the model parameters.

[0117] In one possible implementation, the recognition model obtained from the (t-1)th round of training is trained based on the t-th new training dataset to update the model parameters, as shown below:

[0118]

[0119] in, These represent the parameters of the recognition model obtained in the (t-1)th round of training. This represents the parameters of the recognition model obtained in the t-th round of training. This represents the learning rate of the recognition model. This is the index for the second sample. Indicates to Gradient calculation operation The model represents the first The predicted output for each sample, The t-th soft label of the training sample, Indicates the first One input sample.

[0120] Introducing uncertainty through gradient enhancement techniques can improve a model's generalization ability to new tasks and alleviate the overfitting problem of meta-learning optimization frameworks.

[0121] Further optionally, S104 may include: testing the HRRP-based radar target recognition model using a test set.

[0122] The present invention provides a training method for a target recognition model based on meta-learning and HRRP noise labels. This method constructs an HRRP-based meta-dataset and an HRRP-based training dataset. The meta-dataset includes multiple first samples, each with a correct original label. The training dataset includes multiple second samples, each with label noise in its original label. The number of samples in the meta-dataset is much smaller than the number of samples in the training dataset. A pre-defined diffusion model is trained using the meta-dataset, and the trained dataset is then added to the meta-dataset to obtain a new meta-dataset. Finally, the initial recognition model is pre-trained multiple times using the training dataset to obtain a pre-trained recognition model. The model updates the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset. The pre-trained recognition model is then trained on this new training dataset until convergence, resulting in a radar target recognition model based on HRRP. This model utilizes a diffusion model to generate a small-scale clean dataset, which is then used to expand the constructed meta-dataset, increasing the number of metadata samples and preventing model overfitting. Furthermore, updating the labels of the training samples reduces the interference of label noise, improving model training performance and enhancing model accuracy and reliability.

[0123] To further demonstrate the beneficial effects of the present invention, a set of simulation results are also provided, as follows:

[0124] (1) Simulation experimental conditions:

[0125] This invention utilizes HRRP (High-Resolution Rate) measured data from a certain X-band broadband radar for ten categories of civil aircraft. The ten categories are: Airbus A319, Airbus A320, Airbus A321, Airbus A330-200, Airbus A330-300, Airbus A350-941, Boeing 737-700, Boeing 737-800, Boeing 747-89L, and Bombardier CRJ-900. The model and dimensional parameters of these ten aircraft are shown in Table 1.

[0126] Table 1. Information on ten types of aircraft models and their dimensions.

[0127]

[0128] Experimental data setup: Training dataset: 10,000 HRRP data points per class, totaling 100,000; Test dataset: 2,000 HRRP data points per class, totaling 20,000; Small-scale clean-label dataset: 100 data points per class, totaling 1,000.

[0129] This invention compares four classic HRRP target recognition methods: random forest, support vector machine, convolutional neural network, and recurrent neural network. Specific parameter settings are as follows:

[0130] 1) Random Forest (RF): The number of decision trees in RF is set to 200. To ensure the diversity of decision trees, the depth of decision trees is not constrained in the experiment.

[0131] 2) Support Vector Machine: SVM uses grid search to optimize model parameters. Based on the optimization results, the kernel function is set to the radial basis function, the penalty factor C is set to 32, and the kernel function coefficient G is set to 1.

[0132] 3) Convolutional Neural Network: The network structure used in the CNN is ResNet18. During the training phase, the optimization algorithm for the CNN is set to the Adam algorithm, the number of training iterations is set to 200, and the learning rate is set to 0.02.

[0133] 4) Recurrent Neural Network: The input form required by RNN is a subsequence of multiple time steps. Therefore, it is necessary to perform sliding window slicing preprocessing on the HRRP samples to obtain the corresponding HRRP subsequences. Specifically, in the experiment, the sliding window length was set to 32, the step size was 16, and the overlap rate of adjacent windows was 0.5. Finally, each 200-dimensional HRRP sample can obtain a subsequence data with a dimension of 12*32.

[0134] (1) Diffusion model generation experiment: 100 samples were generated for each class. Only the HRRP data generated for the first class is shown here. Figure 2 As shown.

[0135] (2) Symmetrical Noise Experiment: The true labels in the training dataset were uniformly flipped to all other categories. The experiment was conducted at four noise label ratios of 0.2, 0.4, 0.6, and 0.8. The experimental results are shown in Table 2. The table shows the recognition accuracy of different recognition methods under different noise label ratios.

[0136] Table 2 Performance of the present invention and the classic radar HRRP target recognition method in symmetric noise tag scenarios.

[0137]

[0138] (3) Asymmetric noise experiment: The real labels in the training dataset were randomly flipped to another class. The experiment was conducted at four noise label ratios of 0.1, 0.2, 0.3, and 0.4. The table shows the recognition accuracy of different recognition methods under different noise label ratios.

[0139] Table 3 Performance of the classic radar HRRP target recognition method in asymmetric noise tag scenarios

[0140]

[0141] The present invention also provides a target recognition method based on meta-learning HRRP noise labels, comprising: obtaining a first HRRP corresponding to the radar target to be identified; inputting the first HRRP into a radar target recognition model based on HRRP trained according to any of the methods provided in the above method embodiments, so as to identify the type of radar target.

[0142] Figure 3 This is a schematic diagram of the structure of a training device for a target recognition model based on meta-learning and HRRP noise labels provided by the present invention, as shown below. Figure 3 As shown, the device includes:

[0143] Module 31 is used to construct the HRRP-based metadata dataset and the HRRP-based training dataset. The metadata dataset includes multiple first samples, each with a correct original label. The training dataset includes multiple second samples, each with label noise in its original label. The number of samples in the metadata dataset is much smaller than the number of samples in the training dataset.

[0144] Training module 32 is used to train a preset diffusion model using a meta dataset and add the trained dataset to the meta dataset to obtain a new meta dataset.

[0145] Training module 32 is also used to perform multiple rounds of pre-training on the initial recognition model using the training dataset to obtain a pre-trained recognition model.

[0146] The training module 32 is also used to train the pre-trained recognition model based on meta-model A and meta-model B, as well as the new meta-dataset and training dataset, until the model converges, thereby obtaining a radar target recognition model based on HRRP.

[0147] This invention also provides a schematic diagram of the structure of an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0148] Memory, used to store computer programs;

[0149] When a processor executes a program stored in memory, it implements the steps provided in the above method embodiments.

[0150] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0151] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0152] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps provided in the above-described method embodiments.

[0153] For the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple. For specific details and beneficial effects, please refer to the description of the method embodiments.

[0154] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0155] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A training method for a target recognition model based on meta-learning and HRRP noise labels, characterized in that, include: Construct a metadata dataset and a training dataset based on HRRP. The metadata dataset includes multiple first samples, each of which has a correct original label. The training dataset includes multiple second samples, each of which has label noise in its original label. The number of samples in the metadata dataset is much smaller than the number of samples in the training dataset. The preset diffusion model is trained using the meta-dataset, and the trained dataset is added to the meta-dataset to obtain a new meta-dataset; The initial recognition model is pre-trained multiple times using the training dataset to obtain a pre-trained recognition model; Based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample, the labels of each second sample are updated to generate a new training dataset. Based on the new training dataset, the pre-trained recognition model is trained until the model converges, thereby obtaining a radar target recognition model based on HRRP. The step of updating the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset, and training the pre-trained recognition model based on the new training dataset until the model converges, includes: The first predicted label corresponding to each second sample in the training dataset is obtained by predicting using the pre-trained recognition model; The first predicted label corresponding to each first sample in the new metadata set is obtained by predicting the pre-trained recognition model, and the parameters of the meta-model A and the meta-model B are updated according to the original label and the first predicted label of each first sample to obtain the first meta-model A and the first meta-model B. Based on the original label and the first predicted label of each second sample, and the first meta-model A, the first weight corresponding to each second sample is obtained; Based on the original label and the first predicted label of each second sample, and the first meta-model B, the second weight corresponding to each second sample is obtained; Based on the first weight, second weight, original label, and first predicted label corresponding to each second sample, the first soft label corresponding to each second sample is determined; Replace the original label of each second sample with the corresponding first soft label to obtain the first new training dataset; The pre-trained recognition model is trained multiple times based on the new training dataset until the model converges. The step of training the pre-trained recognition model multiple times based on the new training dataset until the model converges includes: In the t-th round of training, t is a positive integer greater than or equal to 2. The recognition model obtained through the (t-1)th round of training predicts the t-th predicted label corresponding to each of the first samples in the new metadata set, and updates the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B according to the original label and the t-th predicted label of each of the first samples, to obtain the t-th meta-model A and the t-th meta-model B. The parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are generated during the (t-1)th round of training. Based on the original label and the t-th predicted label corresponding to each first sample, and the t-th meta-model A, the t-th first weight corresponding to each first sample is obtained; Based on the (t-1)th soft label and the tth predicted label corresponding to each of the first samples, and the tth meta-model B, the tth second weight corresponding to each of the first samples is obtained; Based on the first weight, second weight, original label, t-th test label, and t-1-th soft label corresponding to each first sample, the t-th soft label corresponding to each first sample is determined; Replace the original label of each of the first samples with the corresponding t-th soft label to obtain the t-th new training dataset; The recognition model obtained in the (t-1)th round of training is trained based on the t-th new training dataset to update the model parameters; Wherein, the step of determining the t-th soft label corresponding to each of the first samples based on the first weight, second weight, original label, first predicted label, and t-1-th soft label is expressed as follows: in, This represents the t-th soft tag. Indicates the first weight. Indicates the original label. Indicates the second weight. This represents the (t-1)th soft tag. This represents the t-th test label. The parameters of metamodel A are represented. This represents the parameters of meta-model B.

2. The method according to claim 1, characterized in that, The step of updating the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B based on the original labels and the t-th predicted labels of each of the first samples to obtain the t-th meta-model A and the t-th meta-model B includes: The difference between the original label and the t-th predicted label of each of the first samples is calculated using the cross-entropy loss function. Based on the calculation results, the meta-model loss gradients corresponding to the (t-1)-th meta-model A and the (t-1)-th meta-model B are constructed. The representations of the (t-1)-th meta-model A and the (t-1)-th meta-model B are as follows: in, Let represent the meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B. Indicates the index of the first sample. This indicates that the gradient of the parameters of the two meta-models is calculated. Represents the cross-entropy loss function. This represents the t-th predicted label of the first sample. Represents the metadata dataset. The label of each sample Represents the metadata dataset. One sample, This represents the parameters of the recognition model obtained in the t-th round of training; Based on a preset gradient enhancement model, the corresponding meta-models A and B (the (t-1)th meta-models) are updated to obtain the updated corresponding meta-models A and B (the preset gradient enhancement model is expressed as follows): in, This represents the meta-model loss gradient corresponding to the updated (t-1)th meta-model A and the (t-1)th meta-model B. Let represent the meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B. Indicates and A vector of the same dimension, where each term is a random variable following a Gaussian distribution. This indicates that the mean is 1 and the variance is... Gaussian distribution, Represents probability parameters; Based on the updated meta-model loss gradients corresponding to the (t-1)th meta-model A and the (t-1)th meta-model B, update the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B to obtain the t-th meta-model A and the t-th meta-model B, expressed as: in, Let A represent the parameters of the t-th meta-model A and the t-th meta-model B. Let represent the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B. The learning rates of meta-model A and meta-model B.

3. The method according to claim 2, characterized in that, The step of training the recognition model obtained in the (t-1)th round of training based on the t-th new training dataset to update the model parameters is expressed as follows: in, These represent the parameters of the recognition model obtained in the (t-1)th round of training. These represent the parameters of the recognition model obtained in the t-th round of training. This represents the learning rate of the recognition model. This is the index for the second sample. Indicates to Gradient calculation operation The model represents the first The predicted output for each sample, The t-th soft label of the training sample, Indicates the first One input sample.

4. A target recognition method based on meta-learning using HRRP noise labels, characterized in that, include: The first HRRP corresponding to the radar target to be identified; The first HRRP is input into the HRRP-based radar target recognition model trained according to any of the methods provided in claims 1-3, so as to identify the type of the radar target.

5. A training device for a target recognition model based on meta-learning and HRRP noise labels, characterized in that, include: The construction module is used to construct a metadata dataset and a training dataset based on HRRP. The metadata dataset includes multiple first samples, each of which has a correct original label. The training dataset includes multiple second samples, each of which has label noise in its original label. The number of samples in the metadata dataset is much smaller than the number of samples in the training dataset. The training module is used to train the preset diffusion model using the metadata, and to add the trained dataset to the metadata to obtain a new metadata. The training module is also used to perform multiple rounds of pre-training on the initial recognition model using the training dataset to obtain a pre-trained recognition model. The training module is also used to train the pre-trained recognition model according to meta-model A and meta-model B, as well as the new meta-dataset and the training dataset, until the model converges, thereby obtaining a radar target recognition model based on HRRP. The step of updating the labels of each second sample based on meta-model A and meta-model B, the new meta-dataset, and the original labels of each second sample to generate a new training dataset, and training the pre-trained recognition model based on the new training dataset until the model converges, includes: The first predicted label corresponding to each second sample in the training dataset is obtained by predicting using the pre-trained recognition model; The first predicted label corresponding to each first sample in the new metadata set is obtained by predicting the pre-trained recognition model, and the parameters of the meta-model A and the meta-model B are updated according to the original label and the first predicted label of each first sample to obtain the first meta-model A and the first meta-model B. Based on the original label and the first predicted label of each second sample, and the first meta-model A, the first weight corresponding to each second sample is obtained; Based on the original label and the first predicted label of each second sample, and the first meta-model B, the second weight corresponding to each second sample is obtained; Based on the first weight, second weight, original label, and first predicted label corresponding to each second sample, the first soft label corresponding to each second sample is determined; Replace the original label of each second sample with the corresponding first soft label to obtain the first new training dataset; The pre-trained recognition model is trained multiple times using the new training dataset until the model converges; wherein, the step of training the pre-trained recognition model multiple times using the new training dataset until the model converges includes: In the t-th round of training, t is a positive integer greater than or equal to 2. The recognition model obtained through the (t-1)th round of training predicts the t-th predicted label corresponding to each of the first samples in the new metadata set, and updates the parameters of the (t-1)th meta-model A and the (t-1)th meta-model B according to the original label and the t-th predicted label of each of the first samples, to obtain the t-th meta-model A and the t-th meta-model B. The parameters of the (t-1)th meta-model A and the (t-1)th meta-model B are generated during the (t-1)th round of training. Based on the original label and the t-th predicted label corresponding to each first sample, and the t-th meta-model A, the t-th first weight corresponding to each first sample is obtained; Based on the (t-1)th soft label and the tth predicted label corresponding to each of the first samples, and the tth meta-model B, the tth second weight corresponding to each of the first samples is obtained; Based on the first weight, second weight, original label, t-th test label, and t-1-th soft label corresponding to each first sample, the t-th soft label corresponding to each first sample is determined; Replace the original label of each of the first samples with the corresponding t-th soft label to obtain the t-th new training dataset; The recognition model obtained from the (t-1)th training round is trained based on the t-th new training dataset to update the model parameters; wherein, the t-th soft label corresponding to each first sample is determined based on the first weight, second weight, original label, first predicted label, and (t-1)-th soft label, as expressed as: in, This represents the t-th soft tag. Indicates the first weight. Indicates the original label. Indicates the second weight. This represents the (t-1)th soft tag. This represents the t-th test label. The parameters of metamodel A are represented. This represents the parameters of meta-model B.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-4.

Images