A radar signal pattern inversion method based on a missing relationship network

Abstract

The present application relates to a kind of radar signal pattern inversion methods based on missing relationship network.It includes: sample set construction is carried out to signal data set, constructs missing relationship network according to the missing condition of sample set element, obtains network node sample subset by division, constructs node integrated classifier, calculates each node integrated classifier weight value;Signal sample corresponding to the input signal data is extracted, and the joint integrated classifier is constructed using missing relationship network, node integrated classifier and corresponding weight coefficient are selected, signal sample is input into joint integrated classifier to obtain signal pattern classification recognition result;According to radar signal pattern analysis recognition result, missing element is estimated and predicted using missing relationship network, and signal pattern inversion result is optimized.

Description

Technical Field

[0001] This invention belongs to the field of integrated electronic warfare technology. Background Technology

[0002] With the widespread application of phased array technology in electronic information systems, radar signal patterns have evolved from fixed patterns to patterns generated by selecting corresponding waveform combinations from a waveform library based on the combat mission. Radar signal pattern analysis has become a crucial aspect of electronic signal reconnaissance. Due to the complexity of the electromagnetic environment on the actual battlefield, real-world signal data often exhibits missing elements in some signal sequences. Current signal pattern analysis methods primarily target datasets with complete signal sequence elements, and can be categorized into traditional methods such as Support Vector Machines (SVM) and intelligent methods such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Residual Neural Networks (ResNet). However, none of these methods can perform signal pattern classification, prediction, or other inversion analysis on signal data with missing sequence elements. Summary of the Invention

[0003] This invention addresses the problems existing in the prior art by providing a radar signal pattern inversion method based on missing relation networks, thus solving the problem that the prior art cannot classify and invert signal patterns with missing data.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] Step 1: Analyze and process the training signal data to construct a sample set, construct a missing relation network, divide the network node sample subsets, construct a node ensemble classifier, and calculate the weights of each node ensemble classifier;

[0006] Step 2: Extract the corresponding signal samples from the input signal data, select the node ensemble classifier using the missing relation network and the corresponding weight coefficients to construct a joint ensemble classifier, and input the signal samples into the joint ensemble classifier to obtain the radar signal pattern classification and recognition results;

[0007] Step 3: Based on the radar signal pattern recognition results, select all corresponding inverse missing child nodes using the missing relationship network, and estimate and predict the missing elements by calculating the nearest neighbor weighted value.

[0008] Based on the above technical solution, the present invention can be further improved as follows.

[0009] Furthermore, step 1, constructing the missing relationship network, includes dividing sample subsets into network nodes based on the missing elements of the sample elements, constructing positive and negative missing relationships between nodes based on the inclusion relationship of missing elements, and adding the samples of all negative missing node subsets after removing redundant elements to the corresponding node sample subsets.

[0010] Furthermore, the node ensemble classifier construction in step 1 includes first generating a set of strongly discriminative elements and a set of weakly discriminative elements using pre-classification and post-classification methods, then constructing the base classifier of the ensemble classifier by selecting corresponding elements from the set of strongly discriminative elements and the set of weakly discriminative elements, and finally integrating the classification results through the simple majority voting results of each classifier.

[0011] Furthermore, the step 1 node ensemble classifier weight calculation includes calculating the corresponding ensemble classifier weight based on the corrected conditional entropy value of the missing element set.

[0012] Furthermore, step 2, the joint ensemble classifier identification, includes selecting the corresponding child nodes and the node ensemble classifiers and their weights corresponding to the corresponding positive missing nodes based on the element missing status of the input signal using the missing relation network, and constructing a joint ensemble classifier by inputting the signal sample into the joint ensemble classifier to obtain the final signal pattern classification and identification result.

[0013] The beneficial effects of this invention are: on the one hand, by constructing a missing relation network to obtain the missing relations corresponding to each subset, the sample subsets are expanded; on the other hand, by using ensemble learning methods to flexibly construct weighted classifiers to improve classification accuracy, the signal pattern inversion analysis of the missing signal data of sample elements is finally realized. Attached Figure Description

[0014] Figure 1 Flowchart of a radar signal pattern inversion method based on missing relational networks.

[0015] Figure 2 Flowchart for constructing a signal sequence sample set.

[0016] Figure 3 Flowchart for constructing a missing relation network. Detailed Implementation

[0017] The implementation flowchart of the present invention is as follows: Figure 1 As shown, the steps of the preferred embodiment are as follows:

[0018] 1) Signal pattern classification training

[0019] ①Construction of signal sequence sample set

[0020] Perform the following steps on all radar pulse descriptor (PDW) sequences under each signal pattern in the radar signal data: Figure 2 The process shown is used to construct a signal sample set. First, the PDW sequence is spliced ​​according to the signal time stamp. Then, the PDW sequence is cleaned, filtered, and statistically analyzed to obtain the signal sequence sample length. Finally, correlation analysis is performed on the signal sequence with missing elements to determine the location of the missing elements.

[0021] Assuming that the length of each signal sequence sample is N, then a certain signal sample can be represented as A. i ={a i,1 ,...,a i,N}, where the element is represented as a i,j =PDW j Missing elements are represented as

[0022] ② Construction of missing relationship networks

[0023] The process of constructing a missing relation network is as follows: Figure 3 As shown, the sample set Λ is first divided into missing network nodes {Λ} based on the missing elements of the samples under each signal pattern. sub1 ,...,Λ subk At this point, each node contains only signal samples with the same missing element. Then, the inverse and forward missing relationships of the network nodes are defined, with node Λ... subi The set of missing elements is defined as v i Its inverse missing node missing element set is v i A subset of, satisfying Its positive missing node missing element set is v i The parent set, satisfying Finally, iterate through all nodes and assign node Λ. subi After removing the corresponding elements from all samples within the inverse missing node, the expanded node is obtained by adding them to the node.

[0024] ③ Construction of Node Ensemble Classifier

[0025] For nodes To construct the ensemble classifier, pre-classification methods such as the Kolmogorov-Smirnov Test (KStest) and F-test are first used, along with post-classification methods such as GRLVQI relevance analysis, Wilk's Lambda ratio analysis, and Multi-Discriminant Analysis Loading Fusion (MDA Loadings Fusion, MLF) to select elements with better discriminative power, forming a strong discriminative element set. The remaining elements form a weak discriminative element set. The base classifiers of the ensemble classifier are constructed by selecting corresponding elements from the strong and weak discriminative element sets. Then, the sample set is divided into a training set and a test set. A portion of the training set is used for error prediction, and the remainder is used to construct the base classifiers. The classification results are integrated using a simple majority vote among the base classifiers. The error prediction samples are input into the classifier to obtain the classification error. If the error exceeds a set threshold, the number of base classifiers is increased. This process is iterated until the error falls below the set threshold, at which point the node ensemble classifier training is complete.

[0026] ④ Calculation of weights for the node ensemble classifier

[0027] Based on the missing element set v i The corrected conditional entropy value is used to calculate the corresponding classifier weights. For nodes with an empty set of missing elements, the information entropy of each element is calculated. For nodes with a non-empty set of missing elements, the corrected conditional entropy is calculated. The corrected conditional entropy is the sum of the conditional entropy and the information entropy, avoiding estimation errors caused by limited samples. Its calculation formula is:

[0028] H CCE (v i )=H(v i )+H(v i |v j )

[0029] The formula for calculating weights using modified conditional entropy is:

[0030]

[0031] Here c i This is an estimate of classification accuracy. This represents the number of node samples.

[0032] 2) Signal pattern classification and recognition

[0033] ① Extract signal sequence samples

[0034] The input signal is first spliced ​​according to the time stamp, then cleaned, filtered, and subjected to correlation analysis to determine the location of missing elements, and finally the signal sequence sample is obtained.

[0035] ② Joint ensemble classifier recognition

[0036] Based on the missing elements of the input signal sequence sample, the corresponding child node and the node ensemble classifier corresponding to the positive missing node are selected in the missing relation network for identification. The identification result is multiplied by the corresponding weight coefficient to obtain the final signal pattern classification and identification result.

[0037] 3) Missing element estimation and prediction

[0038] Based on the radar signal pattern analysis and identification results, the inverse missing node samples of the signal pattern are selected according to the missing relationship network, and the weighted mean of the nearest neighbor samples is calculated for the missing elements as the output of the estimated prediction result.

Claims

1. A radar signal pattern inversion method based on missing relation networks, characterized in that: Step 1: Analyze and process the training signal data to construct a sample set, construct a missing relation network, divide the network node sample subsets, construct a node ensemble classifier, and calculate the weights of each node ensemble classifier; Step 2: Extract the corresponding signal samples from the input signal data, select the node ensemble classifier using the missing relation network and the corresponding weight coefficients to construct a joint ensemble classifier, and input the signal samples into the joint ensemble classifier to obtain the radar signal pattern classification and recognition results; Step 3: Based on the radar signal pattern recognition results, select all corresponding inverse missing child nodes using the missing relationship network, and estimate and predict the missing elements by calculating the nearest neighbor weighted value; The signal pattern classification training in step 1 is as follows: Step 11: The PDW sequence is spliced ​​according to the time stamp of the signal data. Then, the PDW sequence is cleaned, filtered, and statistically analyzed to obtain the length of the signal sequence sample. Finally, correlation analysis is performed on the signal sequence with missing elements to determine the location of the missing elements. Step 12: Divide the sample subsets into network nodes according to the missing elements of the sample elements. Construct the positive and negative missing relationships between nodes according to the missing element inclusion relationship. Use the missing relationship network to remove the redundant elements from all negative missing node subset samples and add them to the corresponding node sample subset. Step 13: Generate a set of strongly discriminative elements and a set of weakly discriminative elements using pre-classification and post-classification methods. The base classifier of the ensemble classifier is constructed by selecting corresponding elements from the set of strongly discriminative elements and the set of weakly discriminative elements. Finally, the classification results are integrated by simple majority voting results of each classifier. Step 14: Calculate the corresponding ensemble classifier weights based on the corrected conditional entropy values ​​of the missing element set; The missing relation network construction process is as follows: First, the sample set Λ is divided into missing network nodes {Λ} according to the missing elements of the samples under each signal pattern. sub1 ,...,Λ subk }, at this point, each node contains only signal samples with the same missing element; then, the inverse and forward missing relationships of the network nodes are defined, node Λ subi The set of missing elements is defined as v i Its inverse missing node missing element set is v i A subset of v j ⊂v i Its positive missing node missing element set is v i The parent set that satisfies v j ⊃v i Finally, iterate through all nodes and select node Λ. subi After removing the corresponding elements from all samples within the inverse missing node, the expanded node Λ is obtained by adding the samples into that node. * subi .

2. A radar signal pattern inversion method based on a missing relation network according to claim 1, characterized in that: The signal joint ensemble classifier identification method in step 2 is as follows: Step 21: The input signal is spliced ​​according to the time stamp, then cleaned, filtered, and correlation analysis is performed to determine the location of missing elements, and finally the signal sequence sample is obtained; Step 22: Based on the missing elements of the input signal, use the missing relation network to select the corresponding child nodes and the node ensemble classifiers and their weights corresponding to the corresponding positive missing nodes to construct a joint ensemble classifier. Input the signal sample into the joint ensemble classifier to obtain the final signal pattern classification and recognition result.

Images