Nerve collapse guided hierarchical prototype learning radar HRRP open set identification system and method
By employing a hierarchical prototype learning method guided by neural collapse, the problems of directional sensitivity and category prototype bias in radar HRRP open set recognition are solved. A hierarchical classification framework is constructed, features are extracted, and confidence scores are generated, achieving accurate recognition and adaptive improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-04-03
AI Technical Summary
Existing open-set recognition methods for radar HRRP based on prototype learning face challenges in addressing the orientation sensitivity of HRRP. This is due to the dispersed features of similar targets and the overlap of dissimilar targets, which increases the difficulty of extracting discriminative features and learning decision boundaries. Furthermore, the category prototypes are easily affected by sample distribution, resulting in a high risk of misclassification.
A hierarchical prototype learning method guided by neural collapse is adopted. Through data preprocessing, HAE module, hierarchical classification network and open set detector module, a hierarchical classification framework based on semantic knowledge is constructed. The prototype estimation is optimized by neural collapse theory. Combined with a composite loss function of cross-entropy loss and center loss, coarse-grained and fine-grained features are extracted, confidence scores are generated and identification is performed through a dual threshold judgment mechanism.
It achieves accurate recognition and improved adaptive performance in different unknown category scenarios, reduces the risk of misclassification, simplifies the difficulty of learning decision boundaries, and improves inter-class discrimination and recognition accuracy.
Smart Images
Figure CN121784690A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar target recognition technology, specifically to a neural collapse-guided hierarchical prototype learning radar HRRP open set recognition system and method. Background Technology
[0002] In radar target recognition technology, the High Resolution Range Profile (HRRP) is a crucial information carrier. It represents the one-dimensional distribution of the target's scattering center along the radar's line of sight, containing geometric information about the target. Compared to two-dimensional synthetic aperture radar (SAR) images and micro-Doppler data, HRRP has lower requirements for the geometric conditions of the detection scene and faster data acquisition speed, meeting the needs of real-time, high-throughput radar automatic target recognition scenarios. In practical applications, the targets to be identified by radar automatic target recognition systems are often non-cooperative targets. During the training phase, it is difficult to establish a recognition template library covering all possible target categories, resulting in the test set often containing targets not present during the training phase—i.e., targets outside the library. However, traditional recognition methods classify all test samples as targets within the library, forcing unknown category samples into a known category, leading to classification errors. Therefore, open-set recognition technology, capable of simultaneously achieving "distinguishing known categories" and "recognizing unknown categories," has become a research focus in the field of HRRP recognition.
[0003] Currently, there is a large body of research on HRRP open set recognition, with various deep learning-based technical solutions, such as full-scale convolutional neural networks, auto-learning, and prototype learning. Among them, prototype learning has become the mainstream technical direction for HRRP open set recognition due to its advantages such as strong generalization ability, intuitive decision logic, and computational efficiency adapted to the real-time requirements of radar. However, existing prototype learning-based HRRP open set recognition methods have two major problems: First, HRRP has extremely strong azimuth sensitivity, which causes the spatial arrangement and intensity characteristics of the scattering center of the same target to change significantly under different azimuth angles. This results in the HRRP features of the same type of target being extremely dispersed in high-dimensional space, making the phenomenon of "dispersed similar samples and overlapping dissimilar samples" in the feature space increasingly prominent, which drastically increases the difficulty of discriminative feature extraction and decision boundary learning. Second, traditional prototype learning methods often use the "feature mean of known class training samples" as the class prototype, but this approach is also affected by the HRRP azimuth sensitivity. This inevitably causes the feature mean prototype to shift towards the azimuth where the sample distribution is dense, resulting in a significant deviation from the true class feature center, thereby increasing the risk of misclassification. Summary of the Invention
[0004] In view of this, the present invention provides a neural collapse-guided hierarchical prototype learning radar HRRP open set recognition system and method, which can achieve accurate differentiation of known categories and effective recognition of unknown categories, while improving the adaptability of target recognition in different unknown category scenarios.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A neural collapse-guided hierarchical prototype learning radar HRRP open set recognition system includes: The data preprocessing module is used to preprocess the raw radar HRRP data to obtain identification input data, and to divide the identification input data into training set and test set during the training phase. The HAE module takes preprocessed HRRP data as input and is trained using a composite loss function that combines cross-entropy loss and center loss to extract and output coarse-grained and fine-grained features. The hierarchical classification network module constructs a tree-like hierarchical structure based on semantic knowledge, and sets a basic classifier at each non-leaf node; the coarse-grained features and fine-grained features output by the HAE module are respectively input to the corresponding hierarchical nodes in the hierarchical classification network module; the hierarchical classification network module is based on the neural collapse theory, and uses the weight vector of the last fully connected layer after the basic classifiers are trained as the category prototype. The open set detector module is used to generate confidence scores by combining the similarity between the features calculated by each node in the hierarchical classification network module and the category prototype based on the input sample, and output the final recognition result through a dual threshold judgment mechanism.
[0006] The data preprocessing module preprocesses the raw HRRP data, specifically including: Perform target region segmentation on the raw HRRP data; Fill the segmented fragments into distance units of a preset length; The inflated HRRP data were then normalized in magnitude. The main peak position of the HRRP data is found by cross-correlation algorithm and the main peak is shifted to a fixed distance cell. Divide the dataset into training and test sets, select some categories as known categories, and the rest as unknown categories.
[0007] The HAE module uses ResNet as its basic backbone network; through a composite loss function, it promotes the clustering of fine-grained categories within the same coarse-grained category and enhances the separation between different coarse-grained categories, thereby reducing misclassification and error propagation of coarse-grained categories. The composite loss function It is composed of a weighted average of the cross-entropy loss and the center loss at each level, as shown in the following formula:
[0008] Where k is the level, and n is the node index within the corresponding level. It is cross-entropy loss. For its weight, It is a central loss. The weights are determined by the cross-entropy loss, which increases the feature space distance between connected child node categories at each classification node, and the center loss, which decreases the feature space distance between connected child node categories at each classification node.
[0009] The cross-entropy loss formula is as follows:
[0010] Where m is the mini-batch size used during training. For the i-th HRRP sample in the mini-batch, This refers to the weight vector corresponding to class j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The weight vector, For the sample The penultimate layer feature vector extracted by the feature encoder, For the sample Corresponding real tags The penultimate layer feature vector, This refers to the bias term for category j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The bias term, N, is the total number of subcategories that the current classification node is responsible for distinguishing; The formula for central loss is as follows:
[0011] in, For the sample The features are the inverse vectors of the distances between the features and the center of each category.
[0012] The hierarchical structure of the hierarchical classification network module includes a root node, coarse-grained category nodes, and fine-grained category nodes; the basic classifier at each non-leaf node consists of two convolutional layers and two fully connected layers.
[0013] The formula for calculating the confidence level generated by the open set detector module is as follows:
[0014] in, This is a scaling factor used to balance the contribution weights of the two indicators in the formula. The L1 norm of the eigenvectors, The score is based on proximity.
[0015] The formula for calculating the proximity score (proxScore) is as follows:
[0016] in, For sample features, The global mean. The L2 norm is the result of subtracting the global mean from the sample features. This is the category prototype weight vector.
[0017] This invention also provides a neural collapse-guided hierarchical prototype learning radar HRRP open set recognition method, comprising the following steps: Step 1: Acquire target radar HRRP data, preprocess it, and divide the data into training and testing sets; Step two: Build and train the recognition model; The recognition model includes a perceptual layer feature encoder (HAE) and a hierarchical classification network. The HAE takes preprocessed HRRP data as input, trains it using a composite loss function that combines cross-entropy loss and center loss, and extracts and outputs coarse-grained and fine-grained features. The hierarchical classification network constructs a tree-like hierarchical structure based on semantic knowledge, and sets a basic classifier at each non-leaf node; based on the neural collapse theory, the weight vector of the last fully connected layer of each basic classifier after training is used as the category prototype. Step 3: Input the test set data into the trained recognition model, calculate the similarity between the sample features and the category prototype, generate confidence by combining the feature norm, and use a dual threshold judgment mechanism to output the fine recognition result of the known category, the known coarse-grained-unknown fine-grained result, or the completely unknown result.
[0018] In step three, the dual threshold judgment mechanism specifically refers to: If the confidence level of a sample at the coarse-grained node is lower than the first threshold, the sample is determined to be of unknown category at both the coarse and fine-grained levels. If the confidence level of a sample at a coarse-grained node is higher than the first threshold, but the confidence level at a fine-grained node is lower than the second threshold, then the known coarse-grained category is output, and the fine-grained category is marked as unknown. If the confidence level of a sample is higher than the corresponding threshold in both the coarse-grained and fine-grained nodes, then the complete known coarse-grained and fine-grained category results will be output.
[0019] In step one, the preprocessing of the original HRRP data specifically includes: target region segmentation, length padding, amplitude normalization, and main peak alignment.
[0020] Beneficial effects: 1. This invention addresses the problem of "scattered similar samples and overlapping dissimilar samples" by constructing a hierarchical classification framework based on semantic knowledge. It decomposes the open set recognition task into two progressive levels: coarse-grained and fine-grained. The coarse-grained level achieves preliminary classification, while the fine-grained level focuses on the feature differences between semantically similar categories, more accurately capturing the discriminative features between categories and thus improving inter-class discrimination. To address the issue of category prototypes being easily affected by sample distribution, this invention utilizes neural collapse theory to optimize prototype estimation, employing the weight vector of the last layer of the neural network to estimate the prototype, thereby reducing prototype estimation errors.
[0021] 2. In the method of this invention, a hierarchical classification open set recognition network framework that integrates semantic knowledge is constructed. By embedding the category hierarchy structure into the loss function and assigning a basic classifier to each non-leaf node according to the tree hierarchy structure to distinguish the categories of its child nodes, the hierarchical decomposition of the target recognition task is realized. At the same time, the feature distribution differences caused by HRRP orientation sensitivity are effectively suppressed, and the difficulty of learning the decision boundary is simplified.
[0022] 3. In the method of this invention, a method for optimizing prototype estimation based on neural collapse theory is proposed. This method is based on a general characteristic of trained neural networks in the feature space: known category features will converge to the weight vector of the last layer of the network, while unknown category features do not have this characteristic. The weight vector of the last layer of the neural network is used as the category prototype, thereby more accurately aligning with the true category center and reducing misjudgments caused by prototype bias. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the HRRP open set recognition method for hierarchical prototype learning radar according to the present invention.
[0024] Figure 2 This is a flowchart of radar data preprocessing in the method of the present invention.
[0025] Figure 3 This is a schematic diagram of the attitude of the vehicle dataset sampled from multiple azimuth angles in the experimental verification of this invention.
[0026] Figure 4 This is a schematic diagram of the HRRP of a truck under multiple azimuth angles in the experimental verification of this invention. Detailed Implementation
[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0028] This invention provides a neural collapse-guided hierarchical prototype learning radar HRRP open set recognition system, including a data preprocessing module, a perceptual layer feature encoder (HAE) module, a hierarchical classification network module, and an open set detector module. The data preprocessing module preprocesses the target data to obtain recognition input data; and during the training phase, it divides the recognition input data into a training set and a test set. The perceptual layer feature encoder (HAE) module and the hierarchical classification network module form a recognition model that outputs confidence scores based on the recognition input data. The recognition model is trained using the HAE module based on the training set. The system outputs two types of features after data processing via HAE: coarse-grained features and fine-grained features, corresponding to different levels in the category hierarchy, preparing for subsequent hierarchical classification and open-set recognition. Based on neural collapse theory, the weight vector of the last fully connected layer after the training of each basic classifier is extracted as the category prototype, replacing the traditional sample mean prototype, making the prototype more closely match the known category features, and preparing for subsequent confidence calculation. The trained recognition model is used to test the test set. The open-set detector module is used to calculate the confidence based on the output results of the recognition model, achieving accurate recognition and unknown category detection under HRRP open set.
[0029] Specifically, the data preprocessing module preprocesses the target data, including target region segmentation, length padding, and amplitude normalization of the original HRRP data. In this embodiment, the specific preprocessing process is as follows: The HRRP (Radar Range Reduction) values of various targets at different azimuths are obtained from the raw target data acquired by radar measurements. High-energy range cells in all raw HRRP data are detected to determine and segment the target's location. The segmented data is then filled into 400 range cells. The amplitude of the filled HRRP data is then normalized. Let x be the original HRRP amplitude value, and min(x) and max(x) be the maximum and minimum amplitude values of the segment, respectively. The amplitude normalization is then expressed as:
[0030] The main peak position of the HRRP data is found by cross-correlation algorithm, and the main peak is shifted to a fixed distance cell to eliminate the influence of target translation sensitivity on feature extraction. The training set and test set are divided, some categories are selected as known categories and the rest are as unknown categories. The HRRP data of known categories are divided into training set and test set in a 7:3 ratio. The data of unknown categories are only used for testing and do not participate in model training.
[0031] In this embodiment, the HAE module is based on ResNet and takes preprocessed HRRP data as input. It is trained by a composite loss function that combines cross-entropy loss and center loss to extract coarse-grained and fine-grained features and promote the clustering of similar features and the separation of dissimilar features. The hierarchical classification network module includes a tree-like hierarchical structure built based on semantic knowledge. A basic classifier is set in each non-leaf node. Combining the neural collapse theory, the weight vector of the last layer of each classifier after training is used as the category prototype.
[0032] The open set detector module calculates the cosine distance between the sample features and the prototype at each classification node, generates confidence scores by combining the L1 norm of the features, and outputs fine-grained recognition results for known categories, coarse-grained results for known categories and fine-grained results for unknown categories, or completely unknown results through a dual threshold judgment mechanism, thereby achieving accurate recognition and detection of unknown categories under the HRRP open set.
[0033] This invention also provides a neural collapse-guided hierarchical prototype learning radar HRRP open set recognition method, which is implemented based on the system of this invention, and the process is as follows: Figure 1 As shown, it includes the following steps: Step 1: Obtain the target data, perform preprocessing, and divide the data into training and testing sets; Step two: Build and train the recognition model; The project constructs a HAE based on ResNet. Based on the training set data, it uses a composite loss function that combines cross-entropy loss and center loss to extract coarse-grained and fine-grained features, and promotes the clustering of similar features and the separation of dissimilar features. Then, based on semantic knowledge, a tree-like hierarchical structure is constructed, a basic classifier is set in each non-leaf node, and combined with the neural collapse theory, the last layer weight vector of each classifier after training is used as the category prototype. Step 3: Input the test set into the trained recognition model. The open set detector calculates the cosine distance between the sample features and the prototype at each classification node, generates confidence by combining the L1 norm of the features, and outputs the known category fine recognition result, the known coarse-grained-unknown fine-grained result, or the completely unknown result through a dual threshold judgment mechanism, thereby achieving accurate recognition and unknown category detection under the HRRP open set.
[0034] The radar data preprocessing specifically involves performing target region segmentation, length padding, and amplitude normalization on the raw HRRP data. The process is as follows: Figure 2 As shown, the specific implementation process of step one is as follows: 101. First, radar measurement is required to acquire the original target data, obtain the HRRP of multiple types of targets in different azimuths, detect the high-energy range cells in all the original HRRP data, and determine and segment the area where the target is located.
[0035] 102. In order to ensure the fixed input size of the neural network, the segmented fragments are filled to 400 distance units.
[0036] The amplitude normalization process is performed on the padded HRRP data. Let x be the original HRRP amplitude value, and min(x) and max(x) be the maximum and minimum amplitude values of the segment, respectively. Then, the amplitude normalization is expressed as:
[0037] The main peak position of the HRRP data is found by cross-correlation algorithm, and the main peak is shifted to a fixed distance cell to eliminate the influence of target translation sensitivity on feature extraction.
[0038] 103. Divide the training set and the test set. Select some categories as known categories and the rest as unknown categories. Divide the HRRP data of the known categories into the training set and the test set in a 7:3 ratio. The data of the unknown categories are only used for testing and do not participate in model training.
[0039] Step two, the construction and feature extraction steps of the HAE specifically include: 201. Using preprocessed HRRP data as input, design a HAE module based on ResNet-18.
[0040] 202. Design a composite loss function This function is composed of a weighted average of the cross-entropy loss and the center loss at each level, as shown in the following formula:
[0041] Where k is the level (k=0 for the root node, k=1 for a coarse-grained node), and n is the node index within the corresponding level. It is cross-entropy loss. For its weight, It is a central loss. Assign weights to them.
[0042] Cross-entropy loss is used to increase the feature space distance between the categories of connected child nodes at each classification node, as shown in the following formula:
[0043] Where m is the mini-batch size used during training, and in this method, m is chosen to be 64. This is the i-th HRRP sample in the mini-batch. This is the weight vector corresponding to category j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The weight vector. For the sample The penultimate layer feature vector extracted by the feature encoder. For the sample Corresponding real tags The second-to-last layer of feature vectors. This is the bias term for category j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The bias term. N is the total number of subcategories that the current classification node is responsible for distinguishing.
[0044] The center loss is used to reduce the feature space distance between the categories of connected child nodes at each classification node, and the formula is as follows:
[0045] in, For the sample The features are the inverse vectors of the distances between the features and the center of each category.
[0046] By using a composite loss function, clustering of fine-grained categories within the same coarse-grained category is promoted, and the separation between different coarse-grained categories is enhanced, thereby reducing misclassification and error propagation of coarse-grained categories.
[0047] 203. Through HAE module training, the data is processed by HAE and outputs two types of features: coarse-grained features and fine-grained features, which correspond to different levels in the category hierarchy structure, thus preparing for subsequent hierarchical classification and open set recognition.
[0048] Step two, the hierarchical classification network generation and prototype estimation steps include: 211. Based on semantic knowledge in the field of radar target recognition, construct a three-layer tree structure containing a root node (layer 0), coarse-grained category nodes (layer 1), and fine-grained category nodes (layer 2).
[0049] 212. Configure a hierarchical classification network. First, set up basic classifiers at the non-leaf nodes of the tree structure. Each classifier consists of two convolutional layers and two fully connected layers. Then, use the coarse-grained features extracted from the samples by HAE. Input coarse-grained classification nodes, fine-grained features Input the corresponding fine-grained classification node, and train the network for each classification node independently to complete the training of the hierarchical classification model.
[0050] 213. Based on the theory of neural collapse, the weight vector of the last fully connected layer after the training of each basic classifier is extracted as the category prototype, replacing the traditional sample mean prototype, so that the prototype is more in line with the known category features, and prepares for subsequent confidence calculation.
[0051] The specific implementation process of step three is as follows: 301. Input the preprocessed test HRRP data into the trained HAE module to extract the coarse-grained and fine-grained features of the test samples. Then, input the features into the trained hierarchical classification network to obtain the preliminary classification results of the test samples at the coarse-grained and fine-grained levels.
[0052] 302. Calculate the similarity between the test sample features and the corresponding hierarchical category prototype. This invention uses cosine distance to quantify similarity; the smaller the distance, the higher the similarity. Then, combine this with the L1 norm of the penultimate layer output feature vector extracted from the test sample via HAE to obtain the final confidence score. The larger the L1 norm, the farther the sample is from the origin, and the higher the confidence score. The confidence score calculation formula is as follows:
[0053] in, This is a scaling factor used to balance the contribution weights of the two indicators in the formula, preventing either indicator from having too large or too small an impact on the final confidence level. Let L1 be the norm of the eigenvector. The score is based on proximity.
[0054] The formula for calculating proximity using cosine distance is as follows:
[0055] in, For sample features, The global mean. The L2 norm (modulus) of the sample features after subtracting the global mean.
[0056] 303. Use a preset threshold to determine the confidence level. If the coarse-grained confidence level is lower than the threshold, both coarse and fine-grained levels are determined to be unknown. If the coarse-grained confidence level is higher than the threshold but the fine-grained level is lower than the threshold, only the coarse-grained known category is output and the fine-grained level is unknown. If both are higher than the threshold, the complete coarse and fine-grained known category results are output, and the identification of known or unknown categories is finally completed.
[0057] To verify the effectiveness of this invention, high-resolution radar range images of vehicles from multiple azimuth angles and of multiple vehicle types were used as samples. A recognition framework including a High-Altitude Area Image (HAE), a hierarchical classification network, and an open-set detector was constructed to achieve open-set recognition of high-resolution radar range images, enabling accurate identification of known categories and effective detection of unknown categories. The dataset and parameter settings are as follows: The dataset was acquired by W-band broadband radar, and the specific parameters are shown in Table 1.
[0058] Table 1 HRRP Measurement Parameters
[0059] The experiment involved 18 vehicle categories, divided into four coarse-grained categories: five industrial and agricultural vehicles, five passenger vehicles, five service vehicles, and three transport vehicles. The known categories consisted of six of these categories, while the unknown categories consisted of the remaining twelve. The specific classification of the experimental samples is shown in Table 2.
[0060] The dataset includes HRRP data reflecting various vehicles under different target poses, such as... Figure 3 As shown, for each vehicle, HRRP data is taken from eight different azimuth angles, each 45° apart. These angles cover a comprehensive range: left front angle (LFO), left front angle (LS), left rear angle (LRO), rear-end collision (TC), right rear angle (RRO), right front angle (RS), right front angle (RFO), and head-on collision (HO). Figure 4 The data shows the HRRP data of trucks in different attitudes. A total of 10,000 samples were collected for each vehicle class at 8 azimuth angles, i.e., 1,250 samples were collected at each azimuth angle. Among them, 7,000 samples of each known class were used as the training set, and the remaining 3,000 samples and all samples of the unknown class were used as the test set.
[0061] Table 2. Description of Experimental Dataset Types
[0062] The proposed HAE module uses ResNet-18 as the backbone network. Each base classifier in the hierarchical classification network consists of two convolutional layers and two fully connected layers, trained using the stochastic gradient descent-momentum (SGD-M) algorithm. The network is initialized with a learning rate of 0.1, and the learning rate is reduced to one-tenth of its initial value every 30 epochs, for a total of 500 epochs. Furthermore, the hyperparameter λ is fixed at 0.1, and the batch size is set to 64.
[0063] Comparative experiment: To evaluate the performance of the method of this invention, the recognition effect of the method under each unknown category group was analyzed, and the performance of the method was evaluated using two indicators: AUROC and OSCR. AUROC refers to the area under the receiver operating characteristic curve, which can evaluate the model's inherent ability to distinguish between "known categories and unknown categories" and is not affected by the choice of decision threshold. OSCR refers to the open set classification rate, which quantifies the overall performance of the model under the dual objectives of "accurately identifying known categories" and "effectively detecting unknown categories." A higher value indicates better overall performance, and it can also evaluate the recognition ability of the method independently of the threshold. The experimental results are shown in Table 3 below.
[0064] Experimental results: Table 3 Performance comparison of various identification methods under different unknown category groups
[0065] As shown in Table 3, the proposed method achieves AUROC and OSCR rates exceeding 80% for industrial and agricultural vehicles, over 81% for passenger vehicles, over 90% for service vehicles, and over 92% for transport vehicles. Furthermore, the proposed method performs optimally across different unknown categories, outperforming other comparative methods and demonstrating its robustness to unknown categories with varying semantic similarity. In summary, the proposed method, through its hierarchical classification structure, can capture subtle yet discriminative differences at a fine-grained level, thereby improving recognition performance in complex scenarios and demonstrating significant advantages. Experimental results comparing it with other traditional methods validate the effectiveness and promising practical application potential of the proposed method.
[0066] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A neural collapse-guided hierarchical prototype learning radar HRRP open set recognition system, characterized in that, include: The data preprocessing module is used to preprocess the raw radar HRRP data to obtain identification input data, and to divide the identification input data into training set and test set during the training phase. The HAE module takes preprocessed HRRP data as input and is trained using a composite loss function that combines cross-entropy loss and center loss to extract and output coarse-grained and fine-grained features. The hierarchical classification network module constructs a tree-like hierarchical structure based on semantic knowledge, and sets a basic classifier at each non-leaf node; the coarse-grained features and fine-grained features output by the HAE module are respectively input to the corresponding hierarchical nodes in the hierarchical classification network module; The hierarchical classification network module is based on the neural collapse theory, and uses the weight vector of the last fully connected layer after the training of each basic classifier as the category prototype. The open set detector module is used to generate confidence scores by combining the similarity between the features calculated by each node in the hierarchical classification network module and the category prototype based on the input sample, and output the final recognition result through a dual threshold judgment mechanism.
2. The system according to claim 1, characterized in that, The data preprocessing module preprocesses the raw HRRP data, specifically including: Perform target region segmentation on the raw HRRP data; Fill the segmented fragments into distance units of a preset length; The inflated HRRP data were then normalized in magnitude. The main peak position of the HRRP data is found by cross-correlation algorithm and the main peak is shifted to a fixed distance cell. Divide the dataset into training and test sets, select some categories as known categories, and the rest as unknown categories.
3. The system according to claim 1, characterized in that, The HAE module uses ResNet as its basic backbone network; through a composite loss function, it promotes the clustering of fine-grained categories within the same coarse-grained category and enhances the separation between different coarse-grained categories, thereby reducing misclassification and error propagation of coarse-grained categories. The composite loss function It is composed of a weighted average of the cross-entropy loss and the center loss at each level, as shown in the following formula: Where k is the level, and n is the node index within the corresponding level. It is cross-entropy loss. For its weight, It is a central loss. The weights are determined by the cross-entropy loss, which increases the feature space distance between connected child node categories at each classification node, and the center loss, which decreases the feature space distance between connected child node categories at each classification node.
4. The system according to claim 3, characterized in that, The formula for cross-entropy loss is as follows: Where m is the mini-batch size used during training. For the i-th HRRP sample in the mini-batch, This refers to the weight vector corresponding to class j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The weight vector, For the sample The penultimate layer feature vector extracted by the feature encoder, For the sample Corresponding real tags The penultimate layer feature vector, This refers to the bias term for category j in the last fully connected layer of the classification node. In the last fully connected layer of the classification node, the corresponding sample Authentic Labels The bias term, N, is the total number of subcategories that the current classification node is responsible for distinguishing; The formula for central loss is as follows: in, For the sample The features are the inverse vectors of the distances between the features and the center of each category.
5. The system according to claim 1, characterized in that, The hierarchical tree structure of the hierarchical classification network module includes a root node, coarse-grained category nodes, and fine-grained category nodes; the basic classifier at each non-leaf node consists of two convolutional layers and two fully connected layers.
6. The system according to any one of claims 1-5, characterized in that, The formula for calculating the confidence level generated by the open set detector module is as follows: in, This is a scaling factor used to balance the contribution weights of the two indicators in the formula. The L1 norm of the eigenvectors, The score is based on proximity.
7. The system according to claim 6, characterized in that, The formula for calculating the proximity score proxScore is as follows: in, For sample features, The global mean. The L2 norm is the result of subtracting the global mean from the sample features. This is the category prototype weight vector.
8. A neural collapse-guided hierarchical prototype learning radar HRRP open set recognition method, characterized in that, Includes the following steps: Step 1: Acquire target radar HRRP data, preprocess it, and divide the data into training and testing sets; Step two: Build and train the recognition model; The recognition model includes a perceptual layer feature encoder (HAE) and a hierarchical classification network. The HAE takes preprocessed HRRP data as input, trains it using a composite loss function that combines cross-entropy loss and center loss, and extracts and outputs coarse-grained and fine-grained features. The hierarchical classification network constructs a tree-like hierarchical structure based on semantic knowledge, and sets a basic classifier at each non-leaf node; based on the neural collapse theory, the weight vector of the last fully connected layer of each basic classifier after training is used as the category prototype. Step 3: Input the test set data into the trained recognition model, calculate the similarity between the sample features and the category prototype, generate confidence by combining the feature norm, and use a dual threshold judgment mechanism to output the fine recognition result of the known category, the known coarse-grained-unknown fine-grained result, or the completely unknown result.
9. The method according to claim 8, characterized in that, In step three, the dual threshold judgment mechanism specifically refers to: If the confidence level of a sample at the coarse-grained node is lower than the first threshold, the sample is determined to be of unknown category at both the coarse and fine-grained levels. If the confidence level of a sample at a coarse-grained node is higher than the first threshold, but the confidence level at a fine-grained node is lower than the second threshold, then the known coarse-grained category is output, and the fine-grained category is marked as unknown. If the confidence level of a sample is higher than the corresponding threshold in both the coarse-grained and fine-grained nodes, then the complete known coarse-grained and fine-grained category results will be output.
10. The method according to claim 8 or 9, characterized in that, Step one involves preprocessing the raw HRRP data, specifically including: target region segmentation, length padding, amplitude normalization, and main peak alignment.