A knowledge-assisted multi-classifier fusion method for extremely narrow pulse radar target recognition

Abstract

This invention discloses a method for ultra-narrow pulse radar target recognition using prior knowledge-assisted multi-classifier fusion, belonging to the field of radar target recognition technology. This method introduces prior conceptual information about the target into the automated construction process of a tree-like hierarchical structure, compensating for the knowledge deficiencies of data-based automated construction methods. The difficulty of distinguishing between classification nodes in the constructed tree structure is significantly reduced. Therefore, even in recognition scenarios with numerous categories and insufficient data, this invention's method still exhibits good performance in refined target recognition.

Description

Technical Field

[0001] This invention discloses a method for target recognition in extremely narrow pulse radar that uses prior knowledge to assist in the fusion of multiple classifiers, belonging to the field of radar target recognition technology. Background Technology

[0002] Automatic target identification (ATI) using radar is a technique that determines the type of target by extracting features and classifying the scattered echoes. Extremely narrow pulse echoes reflect the projection of the target's spatial scattering structure onto the radar's line-of-sight and can serve as an information source for radar target identification. Extremely narrow pulse radar refers to a type of radar where the width of a single echo pulse after processing is much smaller than the target size. In extremely narrow pulse radar, the target echo contains multiple extremely narrow pulses, each corresponding to a different scattering point on the target. Therefore, the target's extremely narrow pulse echo can represent the distribution of the target's scattering points along the radar's line-of-sight, and is often referred to as the target's high-resolution range profile (HRRP).

[0003] Researchers often expand the detection dimensions of target information in extremely narrow pulse echo recognition, such as polarization and frequency information, to form a multi-dimensional, multi-quantitative, and multi-meaning feature set. However, faced with complex multi-dimensional features, traditional methods typically perform simple dimensionality reduction and directly classify them using a single classifier, limiting recognition performance. Subsequently, some scholars proposed multi-classifier fusion methods to improve the classification performance of multi-dimensional features. Typical classifier fusion methods are divided into three types: serial, parallel, and hierarchical. Among them, the hierarchical fusion method decomposes a complex multi-target recognition task into a tree-like hierarchical structure. Each base classifier processes the simplified tasks after decomposition, and each base classifier can perform personalized optimization and utilization of the original complex features according to the requirements of the recognition task, achieving better target recognition performance.

[0004] The hierarchical multi-classifier fusion method is also known as a hierarchical classifier. Figure 1 This diagram illustrates a hierarchical tree structure, where gray nodes represent leaf nodes, storing the final output target category; white nodes are divided into "root nodes" and intermediate nodes, each storing its category and a base classifier used to distinguish the categories of its child nodes. How to construct a hierarchical tree structure to organize multiple base classifiers is the core issue in hierarchical classifier design, directly affecting the recognition performance of the hierarchical classifier.

[0005] Currently, in the field of radar target recognition, existing tree-structured hierarchical construction methods include manual construction based on prior concepts and automated construction based on data. Manual construction based on prior concepts refers to constructing a tree structure based on target category concepts recognized by domain experts, with these prior concepts typically related to the requirements of the recognition task. However, since target recognition is based on a feature space, this method ignores the feature distribution between categories, thus limiting its recognition performance. Automated construction based on data calculates the similarity of different categories of data in the feature space and then recursively obtains a tree-structured hierarchical structure using clustering. This organizes categories with high data similarity into the same node base classifier for classification, achieving good recognition results when existing target data is sufficient and complete. However, in real-world recognition scenarios, the ground environment is highly complex, and target categories are diverse, making it difficult to obtain sufficient and complete data for each target category. Automated construction based on data suffers from numerous uncontrollable factors, insufficient performance stability, and poor adaptability to different data types. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a knowledge-assisted multi-classifier fusion method for ultra-narrow pulse radar target recognition. First, the target category concept information determined by the recognition task requirements is represented as prior knowledge, specifically the concept similarity between categories. Second, the aforementioned knowledge is used to assist in the automated construction process of a hierarchical tree structure, compensating for the knowledge deficiencies of data-based automated construction methods. This significantly reduces the difficulty of distinguishing between category nodes in the constructed tree structure. Therefore, even in recognition scenarios with numerous categories and insufficient data, the method of this invention still exhibits good performance in refined target recognition.

[0007] The technical solution of this invention is:

[0008] A knowledge-assisted multi-classifier fusion method for target recognition in extremely narrow pulse radar improves the recognition performance of hierarchical multi-classifier fusion methods by constructing a tree-like hierarchical structure to organize the multi-classifiers by connecting prior conceptual knowledge and data similarity. Specific steps include:

[0009] Step 1: Extract a multi-dimensional feature dataset of the target based on the target's extremely narrow pulse echo data, and divide the extracted multi-dimensional feature dataset into a training set and a test set;

[0010] Step 2: Use the target category concept information as prior knowledge to establish a concept similarity matrix between target categories;

[0011] Step 3: Based on the category similarity matrix, an adaptive tree-like hierarchical structure H is constructed using the spectral clustering algorithm. The category similarity matrix is ​​obtained by fusing the concept similarity matrix and the data similarity matrix between the target categories established in Step 2.

[0012] Step 4: For the classification task of each classification node in the tree-like hierarchical structure H constructed in Step 3, different feature subsets are independently selected based on the training set divided in Step 1 to form a feature set M. The classification node refers to each node in the tree-like hierarchical structure except for the leaf nodes. The classification task of the classification node is to distinguish the categories of each child node connected to the classification node.

[0013] Step 5: Organize multiple base classifiers into a multi-classifier system in a tree-like hierarchical structure H, and input the feature subset extracted in Step 4 into each base classifier, and perform training and parameter optimization in each base classifier.

[0014] Step 6: Input the test set divided in Step 1 into the multi-classifier system organized in Step 5 for target recognition, and complete the knowledge-assisted multi-classifier fusion for ultra-narrow pulse radar target recognition.

[0015] In step 1, multi-dimensional features of the target are manually extracted. These multi-dimensional features include polarization features, high-resolution features, scale features, and frequency features.

[0016] In step 2, the method for establishing the concept similarity matrix between target categories is as follows:

[0017] The first step is to determine the target category concept knowledge based on the requirements of the identification task, and decompose the radar target identification task in the actual complex scenario into three levels. Taking vehicle target identification in complex ground environment as an example: The first level: target identification, mainly to realize the discrimination of target attributes and distinguish vehicle targets from various ground interference objects; The second level: target classification, mainly to realize the determination of target type, such as transport vehicle, engineering vehicle, etc.; The third level: target identification, mainly to realize the determination of the specific model of the target.

[0018] The second step is to represent the category concept knowledge determined in the first step as a concept similarity matrix. The similarity between different categories is determined based on the distance between nodes in the tree structure. The closer two nodes are in the tree structure (i.e., the more common nodes on the path from the root node to both nodes), the greater the concept similarity between the two categories. Conversely, the farther two nodes are in the tree structure (i.e., the fewer common nodes on the path from the root node to both nodes), the smaller the concept similarity. The concept similarity between any two categories is between (0,1). Based on this approach, the formula for calculating the concept similarity between category A1 and category B1 is as follows:

[0019]

[0020] Where path(A1) represents the path from the root node to the node containing category A1, pat(B1) represents the path from the root node to the node containing category B1, path(A1∩B1) represents the common node of the two paths, and length(·) represents the number of nodes in the path.

[0021] Assuming the target contains C categories, a symmetric concept similarity matrix K with dimension C×C can be obtained. c ;

[0022] In step 3, the method for fusing the concept similarity matrix and the data similarity matrix of the target is as follows:

[0023] The first step is to calculate the data similarity matrix between each target category;

[0024] First, based on the feature sample set of each category in the tree structure, effective feature representations are extracted for each category. The feature sample set of the leaf nodes is the training set of the target object, while the feature sample set of the classification node is the union of the feature sample sets of its connected child nodes. An active sampling algorithm is used to design typicality and diversity indicators to evaluate all feature samples in the feature sample set. The most representative sample is then searched as the feature representation of that category. Assuming that the feature sample set X of category A1 contains M samples, it is represented as X = {x} i , i = 1, ..., M}, x i Let x represent the feature vector of a sample in X. i Typicality index R(x) i ) and diversity index D(x i The definition of ) is shown in the following formula:

[0025]

[0026]

[0027] Where, N i For x i Neighboring sample set, δ R Given the Gaussian kernel bandwidth, evaluate all samples in X based on the typicality and diversity indices, and output samples that meet the following conditions:

[0028] argmax(R(x i )+D(x i ))

[0029] The feature vector of this sample is used as the feature representation v(A1) of class A1;

[0030] Subsequently, the similarity relationship of sample data for each category is calculated. The similarity relationship between different categories is determined based on the distance between them in the feature space. The distance between two categories in the feature space is represented by the Euclidean distance between the feature representation vectors of the two categories. The smaller the Euclidean distance between the feature representation vectors of the two categories, the greater the data similarity between the two categories; conversely, the smaller the Euclidean distance between the feature representation vectors of the two categories, the greater the data similarity between the two categories. For category A1 and category B1, the formula for calculating the data similarity between them is as follows:

[0031]

[0032] Where d(·) represents the Euclidean distance, v(A1) is the data knowledge representation vector of category A1, and v(B1) is the data knowledge representation vector of category N1;

[0033] Assuming the target contains C categories, a symmetric data similarity matrix K of dimension C×C can be obtained. d ;

[0034] The second step is to calculate the category similarity matrix between each category.

[0035] First, regarding K... c and K d Normalization is performed to obtain K' c and K' d The fused category similarity matrix K s Modeled as K' c and K' d Suitable linear combination form:

[0036] K s =w1K c '+w2K' d

[0037] Where w1 and w2 are constant weight coefficients, solved through an optimization problem. Specifically, the objective function for optimizing the weight coefficients is defined as follows:

[0038]

[0039] Where I is the identity matrix, the optimization objective function makes the fused category similarity matrix as close as possible to the identity matrix. The goal is to make the similarity between the same category as close as possible to 1 and the similarity between different categories as close as possible to 0. At the same time, since the objective function restricts the sum of weights to 1, the category relationship described by the fused knowledge matrix after weighted summation is consistent with the category relationship described by each matrix before fusion.

[0040] Finally, due to K c and K dBoth are symmetric matrices. The weight coefficients w1 and w2 are solved using the kernel alignment technique in multi-kernel learning. The final weight coefficients can be expressed as:

[0041]

[0042]

[0043] Where alig(·,·) is used to represent the KTA values ​​of two kernel matrices, if for kernel matrix K' c And I, the method for calculating its KTA value is as follows:

[0044]

[0045] <K' c ,I> represents the inner product between two kernel matrices;

[0046] Step 303: Construct a tree-like hierarchical structure H;

[0047] Using the category similarity matrix obtained in step 302, a tree-like hierarchical structure H is adaptively constructed using the spectral clustering algorithm. The basic idea of ​​the spectral clustering algorithm is to construct a similarity graph between different categories based on the similarity matrix, find suitable graph partitions on the similarity graph, and cluster categories with high similarity in the same graph area as much as possible to achieve the clustering purpose.

[0048] The specific method is as follows:

[0049] (1) Using a category similarity matrix and based on the spectral clustering algorithm, the set of all target categories J contained in the parent node f can be obtained. f Clustered into N subsets Jc i {i = 1, 2, 3, ..., N}, and set each subset Jc i Stored in the child node c connected to the parent node i For the set {i = 1, 2, 3, ..., N}, through the first recursive partitioning, all target category sets contained in the "root" node are aggregated into M subsets, with each subset containing Nf1, Nf2, ..., Nf... M ;

[0050] (2) Start the above partitioning process from the “root” node, cluster to obtain multiple subsets and store them in connected child nodes; then recursively execute the above partitioning process on all subsets. After each partition, the number of categories in the resulting subsets is less, until there is only one target category in all partitioned subsets. Store the single target category in the leaf node. Starting from the “root” node, each recursive partitioning process only performs operations on the nodes circled in the red dotted box. After the Nth recursive partitioning process, the generated nodes only store a single target category, and finally complete the construction of the tree hierarchy.

[0051] In step 4, the method for constructing the feature set M is as follows:

[0052] Step 401: For the classification task of each classification node in the tree hierarchy H, calculate the feature relevance and feature redundancy indices based on the training set, and calculate the metrics for relevance and redundancy based on information theory.

[0053] Feature relevance is defined as the mutual information between feature T and the set of categories A to be identified:

[0054] I(T|A)=H(A)-H(T|A)

[0055] Where H(A) is the information entropy of the category set A, and H(T|A) is the conditional entropy of feature T given the category set A;

[0056] Feature redundancy is defined as the mutual information between feature T and feature S:

[0057] I(T|S)=H(T)-H(T|S)

[0058] Where H(T) is the information entropy of feature T, and H(T|S) is the conditional entropy of feature T given feature S;

[0059] Step 402: Based on the combined indicators of feature redundancy and feature relevance, when searching for feature subsets, the fast coherent filtering algorithm is used to score and sort the features, and the desired feature subset is selected. The above feature selection process is executed at each node, and finally the feature set M is formed.

[0060] In step 5, the method for training and parameter optimization is as follows:

[0061] Step 501 organizes multiple base classifiers into a multi-classifier system in a tree hierarchy H. The white nodes in the tree hierarchy store the class of the node and a base classifier used to distinguish the class of the child nodes of the node, while the gray nodes only store the target class of the final output.

[0062] Step 502: Input the feature subset extracted in step 4 into each base classifier, and train and optimize the parameters in each base classifier. Select the decision tree algorithm as the base classifier model. The parameter settings of each base classifier model are consistent. Input the preferred feature subset of each classification node into the base classifier stored in the corresponding node, and train in each base classifier.

[0063] In step 6, the identification method is as follows:

[0064] During the testing phase, when a test sample u is input, the multi-classifier system will find a single path from the root node to the leaf node as the basis for classifying the test sample. Specifically, when the test sample passes through node v, the connected base classifiers stored in that node determine the next node the test sample will pass through. If the next node is a leaf node, then the category corresponding to that leaf node is the classification result of the test sample. If the next node is not a leaf node, the base classifiers stored in that node will continue to search for the next node until the leaf node is finally reached.

[0065] Compared with existing technologies, the method of this invention has the following advantages:

[0066] (1) This invention proposes a knowledge representation method that connects prior concept knowledge and data similarity. The target category concept information determined by the requirements of the identification task is represented as prior knowledge as the concept similarity between each category, and fused with the data similarity constructed in the feature space to form a category similarity matrix, so as to more comprehensively evaluate the similarity relationship between each category.

[0067] (2) This invention proposes a knowledge-assisted multi-classifier fusion method. This method introduces the prior conceptual information of the target into the automated construction process of a tree-like hierarchical structure, making up for the lack of knowledge in data-based automated construction methods. The task differentiation difficulty of each classification node in the constructed tree structure is greatly reduced. Therefore, even in recognition scenarios with a wide variety of categories and insufficient data, the method of this invention still has good performance in refined target recognition. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of a tree-like hierarchical structure;

[0069] Figure 2 This is the overall flowchart of the method;

[0070] Figure 3 It is a tree structure built based on prior conceptual information;

[0071] Figure 4 The process of constructing a tree-like hierarchical structure;

[0072] Figure 5 This is a schematic diagram of the parent-child node relationship in a hierarchical structure.

[0073] Figure 6 The training process for a multi-classifier system. Detailed Implementation

[0074] To enable those skilled in the art to better understand the present application, the technical solutions of the present application embodiments will be clearly and completely described below with reference to the accompanying drawings.

[0075] Combined with appendix Figure 2 This paper will provide a detailed introduction to the knowledge-assisted multi-classifier fusion method for target recognition in ultra-narrow pulse radar proposed in this invention.

[0076] Specifically, it includes the following 6 steps:

[0077] Step 1: Extract a multi-dimensional artificial feature dataset based on the target's extremely narrow pulse echo data, and divide it into a training set and a test set.

[0078] Multi-dimensional features of the target, such as polarization features, high-resolution features, scale features, and frequency features, are manually extracted from the target's extremely narrow pulse echo data to form a multi-dimensional artificial feature dataset, which is then divided into training set and test set.

[0079] Step 2: Based on the target category concept information determined by the requirements of the identification task, establish a concept similarity matrix between each category.

[0080] First, target category concepts are determined based on the requirements of the identification task. The U.S. Air Force Research Laboratory decomposes radar target identification tasks in complex real-world scenarios into three levels, taking vehicle target identification in complex ground environments as an example: Level 1: Target Discirmination, primarily determining target attributes and distinguishing vehicle targets from various ground obstructions; Level 2: Target Classification, primarily determining target types, such as transport vehicles and engineering vehicles; Level 3: Target Identification, primarily determining the specific model of the target. A tree structure diagram based on the above task decomposition is shown below. Figure 3 As shown.

[0081] Secondly, the category concept knowledge is represented as a concept similarity matrix. This invention is based on the nodes in... Figure 3The similarity between different categories is determined by the distance between the nodes in the tree structure shown. The closer two nodes are in the tree structure (i.e., the more common nodes on the path from the root node to both nodes), the greater the conceptual similarity between the two categories. Conversely, the farther apart two nodes are in the tree structure (i.e., the fewer common nodes on the path from the root node to both nodes), the smaller the conceptual similarity. The conceptual similarity between any two categories is between (0,1). Based on this approach, the formula for calculating the conceptual similarity between category A1 and category B1 can be expressed as follows:

[0082]

[0083] Here, path(A1) represents the path from the root node to the leaf node containing type A1, path(A1∩B1) represents the common node of the two paths, and length(·) represents the number of nodes in the path.

[0084] Assuming the target contains C categories, a symmetric concept similarity matrix K with dimension C×C can be obtained. c .

[0085] Step 3: Based on the category similarity matrix, an adaptive tree-like hierarchical structure H is constructed using the spectral clustering algorithm. The category similarity matrix is ​​obtained by fusing the target's concept similarity matrix and data similarity matrix.

[0086] Step 301: Calculate the data similarity matrix between each target category.

[0087] First, based on the feature sample set of each category in the tree structure, effective feature representations are extracted for each category. The feature sample set of the leaf nodes is the training set of the target to be tested, while the feature sample set of the remaining nodes is the union of the feature sample sets of the node's connected child nodes. This invention uses an active sampling algorithm to design typicality and diversity indicators to evaluate all feature samples in the feature sample set, searching for the most representative sample as the feature representation of that category. Assuming that the feature sample set X of category A1 contains M samples, it can be represented as X = {xi, i = 1, ..., M}, x i Let x represent the feature vector of a sample in X. i Typicality index R(x) i ) and diversity index D(x i The definition of ) is shown in the following formula.

[0088]

[0089]

[0090] Where, N i For x iNeighboring sample set, δ R Let X be the Gaussian kernel bandwidth. Evaluate all samples in X based on typicality and diversity indices, and output samples that meet the following conditions:

[0091] argmax(R(x i )+D(x i ))

[0092] The feature vector of this sample is used as the feature representation v(A1) of category A1. Based on the above process, the feature representation of each category is calculated respectively.

[0093] Subsequently, the similarity relationship of sample data for each category is calculated. This invention determines the data similarity between different categories based on the distance between them in the feature space. The distance between two categories in the feature space is represented by the Euclidean distance between the feature representation vectors of the two categories. The smaller the Euclidean distance between the feature representation vectors of the two categories, the greater the data similarity between the two categories; conversely, the smaller the Euclidean distance between the feature representation vectors of the two categories, the greater the data similarity between the two categories. For categories A1 and B1, the formula for calculating the data similarity between them can be expressed as follows:

[0094]

[0095] Where d(·) represents the Euclidean distance, v(A1) is the data knowledge representation vector of category A1, and v(N1) is the data knowledge representation vector of category N1.

[0096] Assuming the target contains C categories, a symmetric data similarity matrix K of dimension C×C can be obtained. d .

[0097] Step 302: Calculate the category similarity matrix between each category.

[0098] First, regarding K... c and K d Normalization is performed to obtain K' c and K' d The fused category similarity matrix K s Modeled as K' c and K' d Suitable linear combination form:

[0099] K s =w1K' c +w2K' d

[0100] Where w1 and w2 are constant weight coefficients. The objective function for optimizing the weight coefficients is defined as follows:

[0101]

[0102] Where I is the identity matrix. The optimization objective function aims to make the fused category similarity matrix as close as possible to the identity matrix, with the goal of making the similarity between the same category as close to 1 as possible and the similarity between different categories as close to 0 as possible. At the same time, since the objective function restricts the sum of weights to 1, the category relationships described by the fused knowledge matrix after weighted summation are consistent with the category relationships described by the individual matrices before fusion.

[0103] Finally, due to K c and K d Both are symmetric matrices, and the weight coefficients w1 and w2 are solved using the kernel alignment technique in multi-kernel learning. The final weight coefficients can be expressed as:

[0104]

[0105]

[0106] Where alig(·,·) is used to represent the KTA (Kernal Target Alignment) values ​​of the two kernel matrices, if for kernel matrix K' c And I, the method for calculating its KTA value is as follows:

[0107]

[0108] <K' c ,I> represents the inner product between two kernel matrices.

[0109] Step 303: Construct a tree-like hierarchical structure H

[0110] Using the category similarity matrix obtained in step 302, a tree-like hierarchical structure H is adaptively constructed using the spectral clustering algorithm. The basic idea of ​​the spectral clustering algorithm is to construct a similarity graph between different categories based on the similarity matrix, find suitable graph partitions on the similarity graph, and cluster categories with high similarity as much as possible in the same graph region, ultimately achieving the clustering goal. Figure 4 A flowchart for constructing a tree hierarchy is provided. In the tree structure, gray nodes represent leaf nodes, and white nodes are divided into "root nodes" and intermediate nodes. Except for the root node, the number in the node indicates the number of categories stored in that node. Leaf nodes store only a single category, so the number in the leaf node is "1".

[0111] for Figure 4 The detailed description of the build process is as follows:

[0112] (1) Using a category similarity matrix and based on the spectral clustering algorithm, the set of all target categories J contained in the parent node f can be obtained. f Clustered into N subsets Jc i {i = 1, 2, 3, ..., N}, and set each subset Jc i Stored in the child node c connected to the parent node i {i=1,2,3,…,N}. A diagram illustrating the hierarchical parent-child node relationship is shown below. Figure 5 As shown. Figure 4 For example, through the first recursive partitioning, all target category sets contained in the "root" node are aggregated into M subsets, with each subset containing Nf1, Nf2, ..., Nf... M .

[0113] (2) Starting from the "root" node, the above partitioning process is executed, resulting in multiple subsets, which are stored in connected child nodes. Then, the above partitioning process is recursively executed on all subsets. After each partition, the number of categories in the resulting subsets decreases until all partitioned subsets contain only one target category. The single target category is then stored in a leaf node. Figure 4 For example, starting from the "root" node, each recursive partitioning process only performs operations on the nodes circled in the red dotted box. After the Nth recursive partitioning process, the generated nodes only store a single target category, thus completing the construction of the tree hierarchy.

[0114] Step 4: For the classification task of each classification node in the tree hierarchy H, independently select different feature subsets based on the training set to form the feature set M. The classification node refers to each node in the tree hierarchy except for the leaf nodes, and its classification task is to distinguish the categories of its connected child nodes.

[0115] Step 401: For the classification task of each classification node in the tree-structured hierarchy H, calculate the feature relevance and feature redundancy indices based on the training set. Calculate the relevance and redundancy metrics based on information theory.

[0116] Feature relevance is defined as the mutual information between feature T and the set of categories A to be identified:

[0117] I(T|A)=H(A)-H(T|A)

[0118] Where H(A) is the information entropy of the category set A, and H(T|A) is the conditional entropy of feature T given the category set A.

[0119] Feature redundancy is defined as the mutual information between feature T and feature S:

[0120] I(T|S)=H(T)-H(T|S)

[0121] Where H(T) is the information entropy of feature T, and H(T|S) is the conditional entropy of feature T given feature S.

[0122] Step 402: Combining the metrics of feature redundancy and feature relevance, the Fast Correlation Based Filter (FCBF) algorithm is used to score and rank the features when searching for a feature subset, and the desired feature subset is selected. The algorithm will not be described in detail here. The above feature selection process is executed at each node, and the feature set M is finally formed.

[0123] Step 5: Organize multiple base classifiers into a multi-classifier system in a tree-like hierarchical structure H, and input the feature subsets extracted in Step 4 into each base classifier for training and parameter optimization. Figure 6 The training process of a multi-classifier system is presented.

[0124] Step 501 organizes multiple base classifiers into a multi-classifier system in a tree-like hierarchical structure H: such as Figure 6 As shown, white nodes in the tree hierarchy store the node category and a base classifier used to distinguish the categories of the node's child nodes, while gray nodes only store the target category of the final output.

[0125] Step 502: Input the feature subset extracted in Step 4 into each base classifier, and train and optimize the parameters in each base classifier. The method of this invention uses a decision tree algorithm as the base classifier model, and the parameter settings of each base classifier model are consistent. The preferred feature subset of each classification node is input into the base classifier stored in the corresponding node, and training is performed in each base classifier.

[0126] Step 6: Multi-classifier system testing.

[0127] During the testing phase, when a test sample u is input, the multi-classifier system finds a single path from the root node to the leaf node, which serves as the basis for classifying the test sample. Specifically, when the test sample passes through node v, the connected base classifiers stored at that node determine the next node the test sample will pass through. If the next node is a leaf node, the category corresponding to that leaf node is the classification result of the test sample; if the next node is not a leaf node, the system continues to use the base classifiers stored at that node to find the next node, until a leaf node is finally reached.

[0128] Example

[0129] This section illustrates the effectiveness of the invention using publicly available simulation datasets. The dataset used consists of extremely narrow pulse echo data of fully polarized targets obtained from simulations conducted by the U.S. Air Force Research Laboratory using nine types of civilian vehicles. The target types include Jeep 93, Jeep 99, Mitsubishi, Toyota Avalon, Toyota Camry, Mazida MPV, Sentra, Toyota Tocoma, and Honda Civic.

[0130] The simulation dataset was set with a center frequency of 9.6 GHz and an azimuth angle ranging from 0 to 360° at 0.0625° intervals. Due to the azimuth symmetry of the target, this experiment selected data from 0 to 180° at 1° intervals and used a subset of the dataset with an elevation angle of 30°, which is closer to the actual application scenario. Furthermore, to better reflect reality, 25 dB of Gaussian white noise was added to the sample data, and 10 Monte Carlo simulations were performed, resulting in 1800 samples for each target. In the experiment, the training and test sets were divided in a 7:3 ratio, resulting in 12600 training samples and 5400 test samples.

[0131] Based on this dataset, 46 artificial features, 38 polarization features, and 8 high-resolution features were extracted to form a multidimensional feature dataset. Finally, the multidimensional dataset was put into the constructed knowledge-assisted hierarchical classification tree for recognition experiments.

[0132] To further evaluate the effectiveness of the method of this invention, a comparative experiment was conducted with three multi-classifier fusion recognition methods that demonstrated excellent performance in the HRRP recognition task. To compare the multi-classifier fusion effect of the method of this invention, the base classifiers of all three multi-classifier fusion methods and the method of this invention were set as decision trees, and the parameter settings of the decision trees were kept consistent. Furthermore, the remaining parameters of these three comparative methods were optimized to achieve the best performance. Table 2 shows the specific parameter settings of the multi-classifier fusion recognition methods used in the experiment. Table 3 shows the recognition results. As can be seen from Table 3, the recognition performance of the method of this invention is superior to the other multi-classifier fusion algorithms.

[0133] Table 2 Parameter settings for the comparison method

[0134]

[0135]

[0136] Table 3 Comparison of recognition results from different methods

[0137]

[0138] 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 method for target recognition using knowledge-assisted multi-classifier fusion in ultra-narrow pulse radar, characterized by the following steps: include: Step 1: Extract a multi-dimensional feature dataset of the target based on the target's extremely narrow pulse echo data, and divide the extracted multi-dimensional feature dataset into a training set and a test set; Step 2: Establish a concept similarity matrix between target categories; Step 3: Construct a tree-like hierarchical structure based on the category similarity matrix. H ; Step 4: Select the tree-like hierarchical structure constructed in Step 3 based on the training set partitioned in Step 1. H The feature subsets of each classification node are selected for the classification task, and all selected feature subsets are combined to form a feature set. M ; Step 5: Arrange multiple base classifiers in a tree-like hierarchical structure. H The system is organized into a multi-classifier system, and the feature subset extracted in step 4 is input into each base classifier. The feature subset is then trained and the parameters are optimized in each base classifier to obtain the trained and optimized multi-classifier system. Step 6: Input the test set divided in Step 1 into the multi-classifier system trained and optimized in Step 5 for target recognition, and complete the knowledge-assisted multi-classifier fusion for ultra-narrow pulse radar target recognition. In step 2, the concept information of the target category is used as prior knowledge to establish a concept similarity matrix between target categories. The method for establishing the concept similarity matrix between target categories is as follows: The first step is to determine the target category concept knowledge based on the requirements of the identification task, and decompose the radar target identification task in the scene into three levels; The second step is to represent the category concept knowledge determined in the first step as a concept similarity matrix. The similarity between different categories is determined based on the distance between nodes in the tree structure. The closer two nodes are in the tree structure (i.e., the more common nodes on the path from the root node to both nodes), the greater the concept similarity between the two categories. Conversely, the farther two nodes are in the tree structure (i.e., the fewer common nodes on the path from the root node to both nodes), the smaller the concept similarity. The concept similarity between any two categories is between (0,1). A 1 and category B 1. The formula for calculating the conceptual similarity between the two is as follows: in, pat ( A 1) Represents the path from the root node to the node containing the category. A The path of node 1, This indicates the path from the root node to the node containing the category. B The path of node 1, This represents the common node of the two paths. Indicates the number of nodes in the path; Obtain a symmetric concept similarity matrix of dimension C×C. C is the number of target categories; In step 3, the category similarity matrix is ​​obtained by fusing the concept similarity matrix and data similarity matrix between the target categories established in step 2. A tree-like hierarchical structure is then constructed based on the category similarity matrix. H The method is as follows: The first step is to calculate the data similarity matrix between each target category; First, based on the feature sample set of each category in the tree structure, effective feature representations are extracted for each category. The feature sample set of the leaf nodes is the training set of the target object, while the feature sample set of the classification nodes is the union of the feature sample sets of the child nodes connected to that classification node. An active sampling algorithm is used to design typicality and diversity indicators to evaluate all feature samples in the feature sample set. The most representative sample is searched as the feature representation of that category. Assuming the category... A 1 Feature sample set X Include M A sample, represented as X ={ x i， i= 1,…, M }, x i express X A feature vector of a sample. x i Typical indicators with diversity indicators The definition is shown in the following formula: in, N i for x i Neighboring sample set δ R Given the Gaussian kernel bandwidth, based on typicality and diversity indicators... X All samples are evaluated, and the output is the sample that meets the following conditions: The feature vector of the sample is used as the category A 1 characteristic representation ; Subsequently, the similarity relationships of sample data for each category are calculated. The similarity between different categories is determined based on their distance in the feature space. The distance between two categories in the feature space is represented by the Euclidean distance between their feature representation vectors. A smaller Euclidean distance indicates greater data similarity between the two categories; conversely, a larger Euclidean distance indicates greater data similarity between the two categories. A 1 and category B 1. The formula for calculating the data similarity between the two is as follows: in, Represents Euclidean distance. For category Data knowledge representation vectors, For category N 1 represents a data knowledge vector; Obtain a symmetric data similarity matrix of dimension C×C. ; The second step is to calculate the category similarity matrix between each category. First of all K c and K d Normalization process is performed to obtain K’ c and K’ d The fused category similarity matrix K s Modeling as K’ c and K’ d Suitable linear combination form: in w 1. w 2 represents constant weight coefficients, which are solved through an optimization problem. Specifically, the objective function for optimizing the weight coefficients is defined as follows: Where I is the identity matrix; Finally, solve for the weight coefficients. w 1. w 2. The final weight coefficients are expressed as follows: in, alig (·,·) is used to represent the KTA values ​​of two kernel matrices. If for kernel matrices... K’ c and I The KTA value is calculated as follows: This represents the inner product between two kernel matrices; The third step is to construct a tree-like hierarchical structure. H ; Using the category similarity matrix obtained in the second step, a tree-like hierarchical structure is adaptively constructed using the spectral clustering algorithm. H ; In the third step, a tree-like hierarchical structure is constructed. H The specific method is as follows: (1) Using a category similarity matrix, the parent nodes are grouped based on the spectral clustering algorithm. f The set of all target categories included J f Clustering N Subset Jc i {i=1,2,3,…, N }, and each subset Jc i Stored in the child node connected to the parent node c i {i=1,2,3,…, N }, through the first recursive partitioning, the set of all target categories contained in the "root" node is aggregated into M There are subsets, and the number of target categories in each subset is . Nf 1. Nf 2、…、 Nf M ; (2) Starting from the "root" node, the above partitioning process is executed, clustering to obtain multiple subsets, which are stored in connected child nodes; then, the above partitioning process is recursively executed on all subsets. After each partition, the number of categories in the resulting subsets is smaller, until there is only one target category in all partitioned subsets. The single target category is stored in the leaf node. Starting from the "root" node, each recursive partitioning process only performs operations on the nodes circled in the red dashed box. N After each recursive partitioning process, the generated nodes store only a single target category, thus completing the construction of the tree-like hierarchical structure.

2. The method for target recognition of ultra-narrow pulse radar using knowledge-assisted multi-classifier fusion according to claim 1, characterized in that: In step 1, multi-dimensional features of the target are manually extracted. These multi-dimensional features include polarization features, high-resolution features, scale features, and frequency features.

3. The method for target recognition of ultra-narrow pulse radar using knowledge-assisted multi-classifier fusion according to claim 2, characterized in that: In step 4, the classification node refers to each node in the tree hierarchy except for the leaf nodes. The classification task of the classification node is to distinguish the categories of the child nodes connected to the classification node.

4. The method for target recognition of ultra-narrow pulse radar using knowledge-assisted multi-classifier fusion according to claim 3, characterized in that: In step 4, a feature set is constructed. M The method is as follows: Step 401: For tree-structured hierarchical structures H The classification task at each node involves calculating feature relevance and feature redundancy indices based on the training set, and then calculating the metrics for relevance and redundancy based on information theory. Feature relevance is defined as feature T and the set of categories to be identified A Mutual information between them: in, H(A) For a set of categories A Information entropy H(T|A) For a given set of categories A Post-feature T Conditional entropy; Feature redundancy is defined as feature T With features S Mutual information between them: in, H(T) Features T Information entropy H(T|S) For a given feature S Post-feature T Conditional entropy; Step 402: Combining the indices of feature redundancy and feature relevance, a fast coherent filtering algorithm is used to score and rank the features when searching for feature subsets, and the desired feature subset is selected. The above feature selection process is executed at each node, and finally a feature set is formed. M .

5. The method for target recognition of ultra-narrow pulse radar using knowledge-assisted multi-classifier fusion according to claim 4, characterized in that: In step 5, the method for training and parameter optimization is as follows: Step 501: Arrange multiple base classifiers in a tree-like hierarchical structure. H Organized into a multi-classifier system, the white nodes in the tree hierarchy store the category of the node, along with a base classifier used to distinguish the categories of the child nodes of that node, while the gray nodes only store the target category of the final output. Step 502: Input the feature subset extracted in step 4 into each base classifier, and train and optimize the parameters in each base classifier. Select the decision tree algorithm as the base classifier model. The parameter settings of each base classifier model are consistent. Input the preferred feature subset of each classification node into the base classifier stored in the corresponding node, and train in each base classifier.

6. The method for target recognition of ultra-narrow pulse radar using knowledge-assisted multi-classifier fusion according to claim 5, characterized in that: In step 6, the identification method is as follows: During the testing phase, when a test sample u is input, the multi-classifier system will find a single path from the root node to the leaf node as the basis for classifying the test sample. Specifically, when the test sample passes through node v, the connected base classifiers stored in that node determine the next node the test sample passes through. If the next node is a leaf node, then the category corresponding to that leaf node is the classification result of the test sample. If the next node is not a leaf node, continue to use the base classifier stored in that node to find the next node until a leaf node is finally reached.

Images